In the survey’s definition, a world model is a learned predictive model that maps histories of observations and actions to a latent or explicit representation sufficient to forecast future states, observations, rewards, values or other decision-relevant quantities. A common architecture encodes observations and past actions into a compact latent state, propagates that state through a learned transition model under candidate actions and applies task-specific output heads to predict observations, rewards or values. The latent state can act as a learned belief-like summary of history, filtering noise and retaining information needed for planning. Not every model reconstructs the environment: a planning-sufficient model may predict only rewards, policies and values during imagined action sequences. MuZero is used in the survey as a conceptual example of this principle because its action-conditioned latent rollouts support search, but its game-domain performance is not treated as evidence that it is suitable for physical deployment.

The distinction between world models and adjacent components is important for interpreting results. A value model estimates expected return, while a policy model estimates action preferences; either may guide behavior without containing a predictive transition structure. Perception systems, representation learners and online state estimators are likewise classified as enabling components when they support prediction but do not themselves forecast how the world will evolve under actions. The review therefore focuses on models that either predict explicit dynamics or contain an action-conditioned latent rollout structure suitable for decision-making. This boundary also explains why some prominent generative systems are discussed cautiously. Action-conditioned video models and learned interactive simulators may qualify when they respond to interventions, while generative video models without meaningful action conditioning are treated only as adjacent technologies. Photorealistic output by itself does not demonstrate physical consistency or planning utility.

The six-dimensional taxonomy reveals why no architecture is universally superior. State abstractions may consist of explicit physical variables, predicted pixels, compact latent variables, probability distributions over hidden states or factorized object representations. Temporal dynamics can be deterministic or stochastic, recurrent or Markovian, discrete- or continuous-time, or temporally abstract. Uncertainty may be represented through ensembles, probabilistic outputs, latent stochastic variables, Bayesian approximations or implicit mechanisms. Structural priors may encode local linearity, conservation laws, objects, relations or sparse interaction graphs. Inputs may consist of state measurements, a single perceptual channel or asynchronous combinations of vision, touch and proprioception. Finally, a model can be coupled to model predictive control, trajectory optimization, imagination-based policy learning, tree search or no decision process at all. These dimensions are compositional rather than mutually exclusive, so one system can combine a compact latent state, stochastic recurrent dynamics, ensemble-based epistemic uncertainty, object structure, multimodal sensing and model predictive control.

Each combination carries a different compromise. Deterministic latent models can generate fast, memory-efficient imagined trajectories, which is valuable when an optimizer must evaluate hundreds or thousands of candidates within a robot’s control cycle. Their single predicted path also simplifies gradient-based trajectory optimization. The limitation is equally important: if several futures are plausible, a deterministic model can collapse them into an implausible average. In contact-rich manipulation, a small pose change can alter whether an object slips, sticks or impacts another object, so a mean trajectory may violate physical behavior. Stochastic and probabilistic models preserve distributions over futures and can distinguish inherent variability from uncertainty caused by limited data or distribution shift. Ensembles, particles and latent random variables make that uncertainty available to planners, but each sample increases latency and long probabilistic rollouts may accumulate variance. Continuous-time and hybrid formulations can better represent irregular sampling, delays and abrupt contact transitions, yet they also introduce the difficulty of integrating numerical dynamics, identifying modes and propagating uncertainty across discontinuities.

Multimodal world models try to overcome the blind spots of individual sensors. Vision provides global geometry and semantic context but can fail under occlusion or changing illumination; proprioception reports an agent’s internal configuration but does not directly resolve external contact; tactile and force sensing can anchor contact physics but is spatially local and temporally intermittent. Early fusion concatenates modality-specific features, late fusion combines separately encoded representations, and probabilistic fusion treats each sensor as a partial posterior over the hidden state. Such combinations can improve effective observability and reduce ambiguity about contact, but they must handle sensor dropout, corrupted measurements, conflicting evidence, asynchronous sampling and delays. Visuo-tactile transformers and multimodal state estimators are therefore classified as useful enabling components rather than standalone world models unless their representations are coupled to a predictive transition or rollout mechanism. The design target is not merely a state that can be decoded across modalities, but one that remains sufficiently informative about the consequences of candidate actions.

Structured and object-centric models represent another route to generalization. Instead of propagating one monolithic latent vector, they assign representations to entities and model pairwise or relational interactions through message passing. Their explicit locality, object factorization and permutation symmetry can improve scaling to scenes with different numbers of entities and may support systematic transfer to new layouts, which unstructured representations can find difficult. Graph-based formulations are particularly compatible with sparse physical interactions, while object slots can preserve identity through time. These benefits come with stronger assumptions: object discovery can become unstable under visual ambiguity, assumed relations may not match the true contact structure, and perception, graph inference and dynamics must be learned together. Temporal abstractions offer a different way to control long-horizon error by predicting transitions between landmarks or subgoals instead of every fine-grained control step. Such abstraction can shorten planning horizons, but its adequacy depends on whether the selected milestones retain the information needed for safe decisions.

Control integration is where theoretical model design meets the practical danger of model error. In model predictive control, a learned model generates candidate futures, an optimizer selects an action sequence, and the physical system executes only the first action before the process repeats. Receding-horizon feedback can partially absorb disturbances and model mismatch, while optimization provides a natural place to impose actuator, collision and state constraints. It can also magnify error. An optimizer may “cheat” by steering toward regions where the learned model is inaccurate but predicts an attractive outcome, a form of planner exploitation. The same risk occurs when policies are trained largely on imagined rollouts that gradually diverge from physical reality. The review highlights short rollouts, frequent replanning, uncertainty-aware objectives, terminal-value bootstrapping and hybrid model-free learning as mechanisms for reducing this problem. It also emphasizes that methods should be evaluated under matched planners and compute budgets, because improvements attributed to a model may actually come from a different decision layer.

Real-time constraints make the trade-offs measurable. For a planning horizon H, K candidate trajectories and S uncertainty realizations, rollout computation grows approximately with the product of those quantities, with S equal to one for deterministic prediction. Multimodal perception adds further delay through separate encoders, temporal alignment and fusion. The review reports that specialized neural model-predictive-control systems have operated onboard quadrotors at 50 hertz, but it argues that such results require detailed context. Evaluations should report model inference time, complete planning time, target hardware and achieved control frequency rather than presenting prediction quality alone as evidence of deployability. In model-based reinforcement learning, imagination can improve data efficiency by supplying training experiences beyond those collected from hardware. Restricting those experiences to short horizons limits bias, while real interaction or model-free learning stabilizes policy improvement. A task-oriented model may also deliberately sacrifice pixel reconstruction if its latent state retains what is needed for planning and value prediction.

The survey recommends evaluation that links prediction to behavior. One-step error and pixel-level reconstruction can overemphasize task-irrelevant backgrounds and visual details while underweighting sparse, consequential events such as impacts, slips or changes in object state. Recommended reporting includes one-step and multi-step open-loop errors at several horizons, closed-loop rollout consistency, likelihood or proper scoring rules for stochastic models, and calibration measures such as negative log-likelihood, Brier score or expected calibration error. These should be paired with task success, return, constraint violations, robustness and sample efficiency under fixed data budgets. Comparisons should include representative deterministic, stochastic and model-free baselines, report means and variance across multiple random seeds, and disclose computational cost and failure rates. Strong deployment claims should additionally use hardware-in-the-loop or physical-system evaluation. Simulation results remain useful, but the review warns that benchmark improvements on simplified environments may not establish meaningful gains in deployment.

Transfer from simulation to hardware remains a decisive obstacle because real machines introduce imperfect friction estimates, unmodeled compliance, actuator delay, sensor distortion and external disturbances. Domain randomization can vary appearance, while dynamics randomization can vary mass, friction coefficients, joint damping and delays; together these strategies can reduce dependence on simulator-specific behavior. The review says they rarely suffice alone, noting that calibration, safety layers and hybrid control architectures remain important. Transfer claims should specify whether evaluation is zero-shot, few-shot adaptation or continual online adaptation, document randomized parameters and ranges, and report stress tests involving sensor noise, latency, mismatch and disturbances. The remaining challenges include long-horizon consistency, physically valid contact prediction, reliable adaptation to changing payloads and wear, data-efficient learning, multi-agent coordination and safety under rare events. World models are therefore emerging as a promising predictive core for physical AI, but the survey’s central conclusion is conditional: their usefulness depends not on producing a visually convincing future, but on remaining decision-relevant, uncertainty-aware and physically credible after their predictions enter a closed control loop.

Subject of Research: Technical review of learned world models, uncertainty representation and closed-loop control for physical AI

Article Title: A survey of world models for physical AI with uncertainty representation and control

Article References: Kirchner, S., Purschke, N., & Knoll, A. (2026). A survey of world models for physical AI with uncertainty representation and control. Discover Artificial Intelligence, 6(1), Article 1037. https://doi.org/10.1007/s44163-026-02122-1

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02122-1

Keywords: world models, physical AI, robotics, uncertainty quantification, model predictive control, model-based reinforcement learning, embodied AI, sim-to-real transfer, multimodal sensing, safe control, survey, world

Although the models performed similarly in separating sensitized from non-sensitized participants, the Bayesian network supplied an auditable structure and the ability to update probability estimates in either direction. Its output was also well calibrated: the out-of-fold Brier score was 0.171, the calibration slope was 0.9998, with a 95% confidence interval of 0.9563 to 1.0434, and the intercept was −0.0001, with a 95% interval of −0.0522 to 0.0520. The study defines serological sensitization as an allergen-specific IgE concentration of at least 0.35 kUA/L for one or more allergens in a 19-allergen panel. This is a laboratory outcome and is not equivalent to clinically manifest allergy, which also depends on relevant exposure and compatible symptoms. The network therefore estimates the probability of seropositivity rather than the probability that a person will experience an allergic reaction.

The analytic cohort had a median age of 23 years, an interquartile range of 12 to 48 years, and included 51.2% women. Overall, 47.1% of participants met the laboratory definition of sensitization. Sensitized participants were younger on average than non-sensitized participants, with mean ages of 28.8 and 33.1 years, respectively, and men were more common among those classified as sensitized: 50.8% compared with 43.7% among non-sensitized participants. Physician-diagnosed asthma was reported by 19.5% of sensitized participants, compared with 9.3% of non-sensitized participants. The association was much weaker for self-reported allergic symptoms: among the 2,251 participants with a valid symptom response, symptoms were present in 66.7% of sensitized and 63.7% of non-sensitized participants. These cross-sectional differences describe associations in the cohort and do not establish a causal relationship between sensitization and either clinical outcome.

For probabilistic stratification, the researchers divided posterior sensitization probabilities into three prespecified groups: low risk below 0.40, intermediate risk from 0.40 to 0.70 and high risk above 0.70. In the cohort, the observed sensitization rates were 13.8%, 46.1% and 81.6% in those respective groups. The model’s three strata corresponded exactly to its three total-IgE tertiles: every conditional probability in the low-IgE group fell below 0.40, every value in the middle group fell between 0.40 and 0.70, and every value in the high group exceeded 0.70. Eosinophil category changed probabilities within these groups but did not move any participant across a stratum boundary. The nearly equal group sizes—2,767, 2,759 and 2,759—reflect the construction of equal-width IgE tertiles rather than an independent property of the observed outcomes.

The contribution of eosinophils was therefore modest but measurable in continuous prediction. A Bayesian network using total IgE alone had an AUC of 0.7985, compared with 0.8081 when eosinophil category was added, a difference of 0.0096. However, the categorical net reclassification improvement was zero because no participant changed among the three probability strata. This distinction matters because a marker can shift a posterior probability and improve ranking without changing a person’s assigned stratum. The underlying sensitization table nevertheless showed numerical variation within IgE groups. For example, the probability was 0.129 in the low-IgE, reference-eosinophil cell, rising to 0.211 with mild eosinophilia. Middle-IgE probabilities ranged from 0.443 to 0.569, while high-IgE probabilities ranged from 0.788 to 0.868. These values describe variation in estimated serological sensitization probability within the model, without establishing clinical relevance for any particular group.

The network’s graph made these dependencies inspectable. Age and sex were root nodes; total IgE was conditioned on age and sex; eosinophil percentage was conditioned on age and total IgE; and sensitization was conditioned on both immunological markers. Asthma and allergic symptoms were placed downstream of sensitization, allowing the model to represent serological sensitization as distinct from clinical outcomes. The model used discretized variables: age was divided into five bands, total IgE into log-tertiles of no more than 26.0 kU/L, 26.0 to 97.7 kU/L and more than 97.7 kU/L, and eosinophil percentage into reference values below 4%, mild values of 4% to 8% and marked values above 8%. Bayesian-information-criterion evaluation did not support any direct edge from race or ethnicity to downstream variables, so that variable was removed from the predictive architecture and retained only as a descriptive cohort variable.

The retained network contained 70 free conditional-probability parameters and could be queried through exact variable-elimination inference. Its maximum in-degree was two, which kept the treewidth low and made inference computationally tractable. The conditional probability table for sensitization encoded nine combinations formed from three IgE tertiles and three eosinophil categories. The network was refitted from scratch within each of five stratified cross-validation folds, including re-estimation of the sensitization table, so the out-of-fold scores reflected generalization rather than in-sample fitting. Confidence intervals were derived from 2,000 stratified bootstrap resamples. Because the Bayesian network’s continuous posterior probability takes only nine values—one for each sensitization-table cell—the plotted calibration curve was a staircase, making the Brier score and calibration intercept and slope more informative summaries than the curve alone.

One distinguishing feature was bidirectional inference. The model could use a confirmed asthma diagnosis as evidence to update its estimate of sensitization, rather than using sensitization only to predict asthma. In the asthma subsample of 8,275 participants, the prior probability of sensitization was 0.472. Entering asthma as present raised the posterior to 0.651, a shift of 0.179, while asthma classified as absent lowered it to 0.442. Incorporating the reverse-direction evidence increased AUC from 0.808 to 0.814, a gain of 0.006. The study characterizes this as a structural capability that would require a separately trained reverse model in a conventional unidirectional classifier, rather than as evidence of superior predictive performance. The allergic-symptoms node could not undergo an equivalent empirical evaluation because the relevant questionnaire was unavailable for 72.8% of the analytic cohort.

Uncertainty also varied across parts of the network. Sparse evidence combinations produced wider bootstrap intervals than combinations represented by large numbers of participants. The low-IgE, marked-eosinophilia cell had a probability of 0.132 but was estimated from only 36 people, making its parameter less precise than, for example, the high-IgE, mild-eosinophilia cell at 0.868, which was based on 724 participants. The investigators distinguished irreducible biological heterogeneity, represented by genuinely mixed outcomes within some marker groups, from estimation uncertainty that could potentially be reduced with more data. The largest calibration deviation occurred in the lowest probability bin, where the mean predicted probability was about 0.10 and the observed proportion was about 0.33; this bin included the sparsely populated low-IgE, marked-eosinophil cell. At a chosen threshold of 0.520, the model’s confusion matrix corresponded to 74.4% accuracy, 77.0% precision, 65.2% sensitivity and 82.6% specificity.

The findings position the Bayesian network as a methodological demonstration rather than a clinical decision tool. Its principal strengths are transparent conditional probability tables, native uncertainty propagation and flexible evidence updates, while the study does not establish that using it would improve patient outcomes. Important limitations include the cross-sectional design, reliance on one historical NHANES cycle, the modest effect of eosinophils on stratification, and the inability of a binary laboratory outcome to model the full chain from allergen-specific sensitization through relevant exposure to compatible symptoms. The model also cannot distinguish allergic asthma from other asthma endotypes because the survey question records physician-diagnosed asthma without that clinical resolution. Independent external validation, prospective assessment against clinical endpoints and data linking sensitization to symptoms are needed before the framework could move beyond research use.

Subject of Research: Cross-sectional evaluation of a Bayesian network for probabilistic stratification of laboratory-defined IgE sensitization in NHANES 2005–2006 data.

Article Title: Bayesian network-based risk stratification of allergic sensitization using NHANES population data

Article References: Shayakhmetova, A., Tasbolatuly, N., Sultangaziyeva, A., Shurenov, M., Bissarinova, A., & Kurishbay, N. (2026). Bayesian network-based risk stratification of allergic sensitization using NHANES population data. Frontiers in Artificial Intelligence, 9, Article 1920823. https://doi.org/10.3389/frai.2026.1920823

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1920823

Keywords: Bayesian networks, IgE sensitization, NHANES, risk stratification, model calibration, eosinophils, machine learning, uncertainty quantification, allergic sensitization, Bayesian, network-based, risk

The selection criterion was the downstream efficiency of power-envelope calculation rather than residual volume alone. Across three random-seed runs, the task-aware procedure consistently chose a joint box for the five primitive-state residuals, whereas the residual-volume control selected a sample-covariance ellipsoid. Power envelopes were constructed from synchronized sets of primitive current and bus-voltage forecasts, allowing the framework to tailor joint uncertainty geometry to the final monitoring task.

At the 90% nominal level, orbit-balanced joint coverage across the primitive electrical states was 0.892, with a 95% complete-orbit bootstrap confidence interval of 0.883–0.899. At the 95% nominal level, coverage was 0.947, with a 95% confidence interval of 0.943–0.952. Simultaneous coverage of both power channels across all three forecast horizons was 0.913 at the 90% nominal level and 0.957 at the 95% nominal level. These were empirical results from the chronologically subsequent July 2023 telemetry, not guarantees of future performance.

Compared with residual-volume selection, the task-aware method reduced the normalized power interval score by 34.3% at the 90% nominal level and by 30.3% at the 95% level. The paired absolute reductions were 1.522, with a 95% confidence interval of 1.505–1.540, and 1.552, with a 95% confidence interval of 1.534–1.570, respectively. Because the selected region was a non-negative box, its zero algebraic gap and full containment of the resulting interval image follow by construction; the empirical analyses quantify its coverage and interval sharpness.

The findings demonstrate task-aware geometry selection for BUPT-1 multi-horizon power monitoring. The evidence is limited to the studied satellite, telemetry period, three forecast horizons, five primitive electrical states and evaluation protocol, so the results do not establish performance for other spacecraft or operating conditions.

Subject of Research: Task-aware conformal uncertainty quantification for multi-horizon spacecraft power monitoring

Article Title: Task-aware conformal power envelopes for multi-horizon spacecraft monitoring

Article References: Lu, N., Cui, L., & Zhi, X. (2026). Task-aware conformal power envelopes for multi-horizon spacecraft monitoring. PLOS One, 21(10), e0360910. https://doi.org/10.1371/journal.pone.0360910

Image Credits: AI Generated

DOI: 10.1371/journal.pone.0360910

Keywords: spacecraft monitoring, conformal prediction, power envelopes, uncertainty quantification, iTransformer, electrical forecasting, BUPT-1, telemetry, multivariate residuals, Task-aware, conformal, power

The core challenge the researchers tackled is one that has long frustrated the community: climate projections are only as trustworthy as the models behind them, and the skill of those projections depends on whether models can reproduce the long-term, low-frequency variability of the system. Checking that ability against real observations is essential, but doing so in a statistically rigorous way is fiendishly difficult when the model simulation carries systematic errors and the observations are sparse along satellite ground tracks. The new method, developed by Jean-Michel Brankart, Damien Héron, Lisa Weiss, Thierry Penduff and Pierre Brasseur, estimates the low-frequency component of ocean variability directly from native observations, using statistics from a prior long-term ensemble simulation as the starting point.

At the heart of the approach lies a Monte Carlo Markov Chain (MCMC) sampler, a statistical engine that explores a probability distribution by constructing a stochastic process whose attractor is the distribution one wants to sample. The team used a modified variant of the classic Metropolis/Hastings algorithm, one that includes localization of the prior covariance. Localization is a trick borrowed from ensemble data assimilation: because the prior ensemble contains only 30 members, its estimated correlations over long distances are unreliable, so the method drops those small long-range correlations using a Gaussian function of distance and time difference, with a correlation length scale of roughly 100 kilometres and a time scale of about six months. What remains is a local, trustworthy correlation structure that the sampler can exploit.

The estimation problem itself is formulated in a Bayesian framework, where the target is the posterior probability distribution obtained by conditioning the model-based prior distribution on the observations. Crucially, the researchers did not assume the model was perfect. They augmented the state vector with additional free parameters: a time-constant bias field for sea surface height, temperature and salinity; a multiplicative factor to adjust the prior ensemble variance; and another factor to adjust the observation error variance, which is dominated by representativity error from unresolved time scales. Including these auxiliary parameters makes the prior distribution non-Gaussian, which is precisely why linear estimation methods such as standard Kalman filters cannot be applied directly and why the MCMC sampler is needed.

The computational innovation that makes the whole thing feasible at scale is a clever way of generating random perturbations. In a naive implementation, sampling a large-dimension proposal distribution would require computing a linear combination of basis vectors, a cost that grows prohibitively with the size of the state vector. Instead, the researchers exploit the structure of the localized covariance: they generate a moderate-size sample of vectors whose covariance, when raised to the sixth Schur-power, matches the localized correlation matrix. By taking six different vectors from that sample and computing their Schur product, together with one vector from the prior ensemble, they obtain a pseudo-random perturbation direction with exactly the right localized covariance, at the cost of just six element-wise products. With 30 vectors in the localizing sample, the combinatorial effect yields more than ten billion distinct perturbation directions, enough to span all relevant degrees of freedom of the system.

This algorithmic trick is not just elegant; it is fast. The researchers show that the method is well suited to parallel computing with MPI and GPUs, and efficient enough to solve problems with about 10^8 variables and 10^7 observations. In their application, they achieved a speedup of roughly a factor of 100 compared to running on CPUs alone. The overall cost is linear in the size of the state vector and the observation vector, and the authors note that the main practical limitation was the memory required to store the prior ensemble on the GPUs. A classic implementation of the Ensemble Kalman Filter on the same problem would have been substantially more expensive, given the density of observations and the cubic scaling of local updates with the number of observations used.

To demonstrate the method, the team reconstructed the low-frequency variability of Mediterranean sea surface height using along-track altimetric observations from 12 satellite missions between 1993 and 2018, obtained from the Copernicus Marine Environment Monitoring Service. The prior statistics came from a 30-member ensemble simulation of the Mediterranean Sea based on the NEMOMED12 configuration of the NEMO ocean model, run at 1/12-degree resolution over 75 vertical levels and driven by realistic atmospheric forcing between 1979 and 2020. The ensemble dispersion was generated by tiny perturbations of the initial conditions, so that the spread reflects only the intrinsic variability of the nonlinear system. A low-pass Lanczos filter extracted the interannual component, discarding periods shorter than about 2.5 years.

The resulting posterior ensemble estimate was assessed against independent observations using cross-validation, with the altimetric data split into two subsets so that observations not used in the estimation could serve as a check. The rank histogram of the independent observations in the updated ensemble came out essentially flat, indicating that the ensemble solution is globally unbiased and only slightly overdispersive. Self-consistency checks comparing the full-observation experiment with the two partial-observation experiments suggested that the posterior uncertainty might be slightly underestimated, possibly because the statistical assumptions, such as uncorrelated observation errors, give too much confidence in the observations. The posterior ensemble mean also became remarkably similar to the Aviso mapped product, despite the two being obtained through very different processing routes, which the authors interpret as evidence that both approaches made consistent use of the same along-track data.

Beyond the reconstruction itself, the method produced consistent ensemble estimates of the model bias and of the observation error variance. The bias estimate revealed that the model simulation has a too weak cyclonic circulation in the Balearic Sea and the South-Western Algerian basin, and too weak Iera-Petra and Pelops gyres in the Eastern basin, features already noted in earlier work. The observation error standard deviation, dominated by representativity error from unresolved scales, was found to be larger in regions of intense mesoscale activity, such as where eddies detach from the Algerian current. However, the adaptive estimate of the missing prior ensemble variance proved difficult to control by the observations; the various Markov chains spread from their background value without converging towards a common pattern, so the researchers computed the missing variance as a diagnostic instead, revealing that the model largely underestimates its own uncertainty along coasts and in shallow basins.

The broader implications of this work reach well beyond the Mediterranean. The authors argue that producing reliable ensemble reanalyses, forecasts and projections requires that the main sources of modelling uncertainty be adequately represented, and their diagnostic framework provides a concrete way to identify where variance is missing in the model world. They suggest that moving towards a stochastic formulation of the ocean model, as has been proposed for NEMO and recently developed in the CROCO model, would not only improve the solution of modelling and estimation problems but also help understand which observations are most needed to fill the information gaps. With the code and data openly available, the method stands as a significant step forward in combining numerical ensembles with satellite data, and it is likely to influence future improvements in data assimilation for climate and ocean science.

Subject of Research: Ensemble estimation of Mediterranean low-frequency sea level variability using an MCMC sampler combining observed and simulated data

Article Title: Efficient ensemble estimation using an MCMC sampler: a reconstruction of the Mediterranean low-frequency variability combining observed and simulated sea level

Article References: Brankart, J.-M., Héron, D., Weiss, L., Penduff, T., & Brasseur, P. (2026). Efficient ensemble estimation using an MCMC sampler: a reconstruction of the Mediterranean low-frequency variability combining observed and simulated sea level. Ocean Science, 22(5), 3207-3224. https://doi.org/10.5194/os-22-3207-2026

Image Credits: AI Generated

DOI: 10.5194/os-22-3207-2026

Keywords: MCMC sampler, Mediterranean Sea, sea level, satellite altimetry, ensemble simulation, Bayesian inference, data assimilation, low-frequency variability, ocean modelling, uncertainty quantification, GPUs, NEMO ocean model

The motivation is straightforwardly clinical. Molecular genetic testing remains the gold standard for diagnosing conditions such as sickle-cell disease, beta-thalassemia, BRCA1-linked hereditary cancer syndromes, and cystic fibrosis, but it is expensive, logistically demanding, and unavailable in many low-resource settings. Computational screening tools that flag high-risk patients from routinely collected measurements—complete blood count indices, inflammatory markers, targeted gene-expression assays—could triage patients toward confirmatory testing earlier and more cheaply. Deep neural networks have been proposed widely for this purpose because they can learn complex, non-linear relationships among biomarkers without hand-engineered diagnostic rules. Yet most such tools produce a single softmax point estimate that is often overconfident, particularly for atypical inputs or cases lying in regions of genuine class overlap. In a diagnostic setting, an overconfident wrong answer is arguably more dangerous than a low-confidence one, because it removes the clinician’s incentive to order the confirmatory test that would catch the error.

The team, led by V. S. Keerthana and Tarak Hussain, trained a feed-forward multilayer perceptron with three hidden layers of 128, 64, and 32 units, each stabilized by batch normalization and regularized by dropout with a rate of 0.3. The network classified 1,000 patient records into five genetic-disease-associated groups using twelve clinical and molecular biomarkers, including hemoglobin concentration, fetal hemoglobin fraction, serum ferritin, sweat chloride, the proportion of sickled red blood cells, BRCA1 expression, and TP53 mutation status. Class distribution was nearly balanced, with 196 to 205 samples per category, avoiding the confound of extreme class imbalance. Performance was assessed with stratified five-fold cross-validation against logistic regression, random forest, support vector machine, and gradient boosting classifiers, all running on the same standardized features.

The accuracy results were humbling for the deep learning camp. The neural network achieved 94.30 percent cross-validated accuracy, plus or minus 1.36 percent—numerically lower than all four classical baselines, which ranged from 94.90 to 95.50 percent. An exploratory McNemar’s test on paired predictions suggested the gaps against random forest (95.50 percent, p = 0.0018) and support vector machine (95.30 percent, p = 0.0063) crossed the threshold of statistical significance, while logistic regression and gradient boosting were statistically indistinguishable from the network. The authors are careful to note that their pooled-fold significance test violates the independence assumption McNemar’s test requires, since cross-validation folds share overlapping training sets, so these p-values should be read as suggestive rather than confirmatory. Still, the asymmetry in paired disagreement counts was striking: across all five folds, the network corrected only one test sample that random forest got wrong, while random forest corrected thirteen samples the network misclassified. The finding aligns with a growing body of evidence that tree-based and kernel methods remain highly competitive with deep learning on structured tabular data of this size.

Where the deep learning approach pulled ahead was in uncertainty. The technique at the center of this advantage, Monte Carlo dropout, exploits the fact that dropout—normally a training-time regularizer that randomly silences a fraction of neurons—can be kept active at inference time. Instead of one deterministic forward pass, the network runs up to 200 stochastic passes per input, each with a different random pattern of dropped units, producing a collection of softmax vectors. The mean of these vectors serves as the final predicted probability distribution, while the Shannon entropy of that mean quantifies total predictive uncertainty. Crucially, this total decomposes into two components with different clinical meanings: aleatoric uncertainty, the average of the per-pass entropies, reflects ambiguity inherent to the data itself, such as borderline biomarker presentations; epistemic uncertainty, computed as the mutual information between the mean prediction and the per-pass entropies, reflects model doubt that could in principle be reduced with more or more representative training data. In this dataset, aleatoric uncertainty dominated, accounting for 69.9 percent of total predictive entropy—suggesting most low-confidence predictions reflect genuine biological overlap between disease classes rather than a fixable data deficiency.

The uncertainty estimates proved convergent and usable. Accuracy and mean predictive entropy stabilized by 50 to 100 stochastic passes, with the bootstrap standard deviation of the entropy estimate falling nearly 5.5-fold from 0.0123 nats at 10 passes to 0.0022 nats at 200. The estimates also tracked prediction quality: when held-out test predictions were sorted into low, medium, and high uncertainty terciles by predictive entropy, the error rate roughly doubled from 6.0 percent in the lowest tercile to 12.0 percent in the highest, and 60 percent of all misclassifications landed in the high-uncertainty group despite that group containing only a third of samples. The conditional probability of high uncertainty given an incorrect prediction, 0.600, was nearly twice the 33 percent base rate expected if uncertainty carried no information about correctness. Incorrect predictions, as a population, were shifted toward substantially higher entropy than correct ones.

The most immediately practical result came from selective prediction, sometimes called abstention analysis. By ranking held-out predictions in increasing order of predictive entropy and discarding the most doubtful ones, the effective accuracy of retained predictions rose sharply: removing the most uncertain 20 to 30 percent of predictions pushed retained-sample accuracy above 95 percent, up from the full-coverage figure of 93.33 percent. The area under the risk-coverage curve, 0.0662, summarizes this gain in a single statistic. The authors frame this as a candidate decision-support mechanism—automatically accepting confident predictions while referring uncertain ones for physician review or confirmatory genetic testing—rather than an autonomous diagnostic tool. The computational cost of the approach is negligible in absolute terms: the network has 12,165 trainable parameters, trains in about half a second on a standard CPU, and Monte Carlo inference with 100 passes on the entire 150-sample test set took under 50 milliseconds, roughly 121 times the deterministic cost but still trivial for a screening application that is not latency-critical.

Calibration told a more nuanced story. On the held-out test partition, the Monte Carlo dropout model achieved 93.33 percent accuracy with macro F1 of 0.933, macro ROC-AUC of 0.971, and Matthews correlation coefficient of 0.917. But its expected calibration error of 0.076 was numerically the worst among the six model and inference configurations evaluated, exceeding all four classical baselines, which ranged from 0.024 for the support vector machine to 0.066 for gradient boosting. Yet Monte Carlo averaging substantially reduced the negative log-likelihood relative to deterministic inference, 0.370 versus 0.445—a metric that penalizes overconfident wrong predictions far more heavily than bin-wise calibration measures do. The authors interpret this as evidence that no single calibration metric tells the whole story: the benefit of Monte Carlo dropout is concentrated in limiting the damage of confident mistakes rather than in aligning confidence with accuracy across probability bins. An ablation study across seven dropout configurations revealed a clear monotonic trade-off, with lower dropout rates yielding better-calibrated but less richly quantified uncertainty, and higher rates producing poorer calibration but larger, more clearly separated uncertainty estimates.

Robustness testing added further confidence, with one instructive surprise. Accuracy was consistent across five random initialization seeds at 93.07 percent plus or minus 0.33 percent, and held up even when training data was cut to a quarter of available samples. When zero-mean Gaussian noise was injected into test features at inference time, accuracy degraded sharply beyond a noise standard deviation of 0.5 standardized units, falling from 93.33 percent to 69.33 percent at 1.0 and 59.33 percent at 2.0—but mean predictive entropy rose monotonically and substantially across the same range, from 0.522 to 0.840 nats. This co-movement of increasing uncertainty with degrading accuracy is exactly the property a trustworthy uncertainty estimate should have under distribution shift: the model becomes appropriately less confident on corrupted inputs rather than remaining artificially assured. The surprise came from the training-data experiment, where mean predictive entropy was lowest, not highest, when the model trained on the smallest subset—contrary to the naive expectation that less data should produce more epistemic uncertainty. The authors present this pattern in full and flag it as an open question for multi-seed replication rather than burying it.

Every result in the study rests on internal computational validation of a single publicly available dataset of 1,000 records whose real-world clinical versus synthetic provenance could not be independently confirmed, and no external or prospective clinical testing was performed. The authors also emphasize that the composite five-class formulation—spanning conditions that would never share a differential-diagnosis pathway at the bedside—is a deliberately heterogeneous methodological benchmark for stress-testing the uncertainty pipeline, not a proposed simultaneous diagnostic instrument; a real deployment would use separate, condition-specific models triggered by the marker subset clinically indicated for a given presentation. Permutation importance analysis showed the classification signal was governed mainly by direct molecular and hematological markers, with sickled red blood cell percentage and BRCA1 expression jointly carrying most of the discriminative ability, while age, gender, and family history contributed essentially nothing once those direct markers were available. The broader lesson for the field is twofold: do not default to deep learning for structured genetic disease data without empirical validation, since classical algorithms matched or beat it here under standard, non-tuned hyperparameters; but recognize that the deep learning approach’s real advantage was not accuracy at all, rather the ability to know what it did not know—a capacity the authors argue is worth more than a single percentage point of accuracy in any screening workflow where a confident error can delay the right test.

Subject of Research: Machine learning

Article Title: Deep learning-based genetic disease prediction with Monte Carlo (MC) dropout uncertainty estimation: a comparative study with classical machine learning baselines

Article References: Keerthana, V. S., & Hussain, T. (2026). Deep learning-based genetic disease prediction with Monte Carlo (MC) dropout uncertainty estimation: a comparative study with classical machine learning baselines. Frontiers in Genetics, 17, Article 1969295. https://doi.org/10.3389/fgene.2026.1969295

Image Credits: AI Generated

DOI: 10.3389/fgene.2026.1969295

Keywords: deep learning, Monte Carlo dropout, uncertainty quantification, genetic disease prediction, machine learning benchmarking, model calibration, clinical decision support, precision medicine, Deep, learning-based, genetic, disease

The package, developed by JoonHo Lee and Alison Hooper and released under an MIT license, grew out of a practical evaluation problem in early childhood education. Analysts studying pre-kindergarten programs often want to know what proportion of enrolled children have families receiving nutrition assistance. Sometimes they hold individual student records; sometimes they have only published counts, often with different denominators across sites and sometimes with values deliberately suppressed by the publisher. Before sitemix, preparing proportions and their standard errors from these heterogeneous inputs required ad hoc calculations scattered across scripts, with assumptions rarely stated explicitly. The package replaces that improvisation with a common set of functions that return a standardized table of estimates, standard errors, and provenance information for every site and year.

The central technical insight of sitemix is that a standard error is meaningless until you specify what the proportion is intended to describe. Administrative records may cover an entire cohort, a selected subset, or a probability sample, and each interpretation demands a different variance calculation. The package’s default binomial model treats the observed count as arising from independent Bernoulli trials with a common underlying probability, so the variance of the observed proportion is the familiar theta times one minus theta divided by n. Under this model, uncertainty concerns repeated outcomes at a fixed sample size. Alternatively, analysts may want the proportion in a fixed finite population of N individuals, in which case uncertainty comes from which n individuals were sampled, and a finite-population correction adjusts the sampling variance. Crucially, observing the entire population gives zero design-based sampling variance, even when the population consists of a single person. The package forces users to choose between these targets rather than silently applying one formula to all data.

Once the target is defined, sitemix offers a menu of established calculations with their assumptions stated openly. For proportions of exactly zero or one, where the plug-in variance collapses to zero, the package supplies a Wilson-derived standard error surrogate by default, with options for the Agresti-Coull adjustment or a retained zero variance. For reporting on transformed scales, the default is the arcsine square-root transformation, whose first-order working standard error is simply one divided by two n, an approximation that depends only on the denominator. An optional Anscombe adjustment refines this transformation for small samples. The package also implements a bias-corrected binomial variance that replaces the denominator n with n minus one for interior cells. None of these boundary surrogates or working approximations guarantee confidence-interval coverage, and the documentation says so explicitly, a refreshing departure from software that presents approximations as exact.

Perhaps the most technically sophisticated part of the package concerns covariance between overlapping indicators. Suppose an analyst wants to know not just the share of children eligible for free and reduced-price meals and the share enrolled in SNAP, but how these indicators vary together. Separate totals for two assistance programs do not reveal how many children participate in both; only joint records or co-occurrence counts can estimate that covariance. When complete joint information is available, sitemix computes the plug-in covariance from the observed joint proportion minus the product of the marginal proportions, divided by n. For a single categorical variable partitioned into mutually exclusive categories, complete counts determine a multinomial covariance matrix whose rows sum to zero, preserving the constraint that category proportions total one. The package validates these matrices against the properties required by their construction rule, recording scale, indicator order, and sampling design metadata alongside every covariance carrier.

The package’s treatment of published aggregate counts reveals what such summaries actually retain. When numerator-denominator pairs match the record-based inputs, the aggregate route reproduces the record-based estimates and standard errors exactly, a property the authors demonstrate across 250 simulated site-year cells. But separate marginal counts for several indicators generally do not determine covariance between them. When only margins are available, sitemix returns a diagonal covariance matrix explicitly labeled as working independence, recording whether the margins describe the same, different, or unknown sampled units. Equal denominators alone do not establish that the units are the same. For complete, unsuppressed margins from the same units, the package computes Fréchet identification intervals, the mathematically sharp bounds on the joint proportion given the observed margins: the joint share must lie between the maximum of zero and the sum of the two proportions minus one, and the minimum of the two marginal proportions. These are identification intervals, not confidence intervals, and the simulated example shows how wide the joint-proportion range can remain without co-occurrence counts, falling on an endpoint at 13 of 50 sites.

Suppression handling is another distinctive feature. Publishers routinely withhold small or sensitive counts, and software that treats withheld values as observed zeros would silently corrupt the analysis. sitemix keeps suppressed values distinct from observed zeros throughout, requires users to explicitly acknowledge sensitivity values, stores them in separate columns, and excludes them from ordinary covariance and Fréchet calculations. A dedicated suppression report audits the data before estimation begins. The diagnostic function sm_diagnose checks objects, rows, and covariance matrices, returning results on a severity scale from error through warning, note, and ok, so analysts can review missing values, small denominators, and covariance compatibility before feeding the output into another model.

The validation effort behind the package is unusually thorough for statistical software. The authors checked all 263 scalar and finite-population configurations against reference values computed from closed-form identities, score-quadratic roots, and finite-population sample enumeration, with the largest absolute discrepancy at 8.88 times ten to the minus sixteenth, essentially machine precision. All 118 categorical and Fréchet assertions passed, covering simplex constraints, positive semidefiniteness, sparse categories, boundary surrogates, and census behavior. The full test suite passed 6,918 expectations across 53 files with no failures, and test-based line coverage reached 91.37 percent. The authors also examined the accuracy of the arcsine working variances by computing exact variances across a grid of sample sizes and probabilities, finding that the unadjusted working variance could be anywhere from 0.405 to 10.23 times the exact variance across interior grid points, a sobering reminder that transformed-scale approximations demand scrutiny at small samples.

To illustrate the workflow, the authors provide prek_sim, a simulated panel of 7,845 student-year rows from 50 sites over 2021 to 2025, containing four overlapping binary indicators: eligibility for free and reduced-price meals, enrollment in SNAP, participation in WIC, and receipt of TANF. Site-year cells contain between 6 and 134 children. The examples walk through estimating site-level proportions, computing covariance from jointly observed indicators, reproducing record-based estimates from matching counts, and examining what marginal counts alone cannot recover. A final empirical Bayes illustration shows how the estimate-standard error pairs can feed a normal-normal shrinkage analysis on the arcsine scale, where sites with small denominators move noticeably in rank after borrowing strength across sites, though the authors caution that rank uncertainty is not quantified and the normal approximation requires assessment for sparse cells.

The impact of sitemix lies in making uncertainty preparation repeatable and transparent. Each estimate retains its denominator, reporting scale, suppression status, and method information, so analysts can review how a table was produced when updating it or sharing it with colleagues. The count interfaces extend the workflow to settings where individual records are unavailable, while the Fréchet machinery helps analysts decide which additional counts they actually need. The package does not pool proportions across sites, fit hierarchical models, or handle complex survey designs; its only cross-site model is an optional experimental variance smoother. Instead, it focuses on preparing estimates and uncertainty summaries for subsequent analysis, leaving the modeling choices where they belong: with the analyst, justified separately. For a field drowning in published proportions of unknown provenance, that discipline may prove quietly transformative.

Subject of Research: Statistics; Software; Data Science

Article Title: sitemix: An R package for site- and group-level proportions and uncertainty from individual records and aggregate counts

Article References: Lee, J., & Hooper, A. (2026). sitemix: An R package for site- and group-level proportions and uncertainty from individual records and aggregate counts. SoftwareX, 36, Article 103131. https://doi.org/10.1016/j.softx.2026.103131

Image Credits: AI Generated

DOI: 10.1016/j.softx.2026.103131

Keywords: R package, statistics, uncertainty quantification, proportions, empirical Bayes, open-source software, education research, covariance, sitemix, package, site-, group-level

The framework, described in the journal Applied Intelligence, is designed for the reality of industrial equipment, where new fault modes keep emerging while early-stage fault samples remain scarce. When a new fault class appears, there may be only a few real vibration recordings available, far too few to train a deep network from scratch. T-UIL responds by first employing a diffusion model to generate candidate augmented samples for the newly emerging fault classes. Diffusion models, the same family of generative architectures behind recent image-generation breakthroughs, learn to synthesize data by gradually denoising random signals into realistic samples. Here, they manufacture synthetic vibration data that mimics the rare new faults, alleviating the lack of supervised information that would otherwise cripple learning.

But generated data is a double-edged sword. If the diffusion model produces samples that are inconsistent with the actual incremental task, training on them can mislead the classifier rather than help it. This is where the framework’s most distinctive component enters. The researchers designed an uncertainty-guided Trustworthy Data Engine, or TDE, to rigorously screen the candidate samples before they are allowed anywhere near the training pipeline. The engine quantifies the predictive uncertainty of the previous model when it looks at each generated sample. Samples that leave the old model confident and consistent with the current task are kept; samples that trigger high uncertainty are discarded. In effect, the previous model acts as a skeptical referee, judging whether each synthetic example genuinely belongs to the new fault class as the system understands it.

The logic behind this screening is rooted in a growing body of work on uncertainty-aware deep learning. Researchers in the field have argued that predictive uncertainty is a promising tool for trustworthy fault diagnosis, particularly in safety-critical applications where a confidently wrong prediction can be worse than an abstention. By measuring how uncertain the previous model is about each candidate, TDE substantially improves the consistency between the generated data and the current incremental task. The screened high-confidence augmented samples then pass through to the next stage, where they are jointly used with the real few-shot samples and replayed historical samples from earlier fault classes. Replay, a classic continual-learning strategy, gives the model fresh exposure to old knowledge so that learning new faults does not erase the old ones.

On top of this data mixture, T-UIL incorporates two optimization techniques for unified incremental learning. Supervised contrastive learning pulls together the feature representations of samples from the same fault class while pushing apart representations of different classes, sharpening the decision boundaries even when data is scarce. Knowledge distillation, meanwhile, encourages the current model’s outputs to stay aligned with those of the previous model on old classes, preserving established knowledge while the network absorbs the new fault modes. The combination means that a single training session simultaneously optimizes for new-class accuracy, old-class retention, and representation quality, rather than stitching together separate fixes after the fact.

The team evaluated the framework extensively on two public bearing datasets: the Paderborn University dataset and the Southeast University dataset. Bearing fault diagnosis is a standard proving ground for industrial AI because vibration signals encode rich information about mechanical damage, from inner-race defects to outer-race wear, and benchmark datasets allow fair comparison across methods. The experimental results demonstrate that, compared with several benchmark incremental learning algorithms, the proposed method achieves better classification accuracy across different sample-limited incremental tasks. In other words, whether the new fault class arrives with five examples or twenty, T-UIL consistently outperformed conventional approaches that either ignore the data-scarcity problem or trust their generated data blindly.

Perhaps more important for real-world deployment, the framework also improves the calibration performance and reliability of model predictions. Calibration refers to the alignment between a model’s stated confidence and its actual accuracy: a well-calibrated model that says it is ninety percent sure should be right about ninety percent of the time. Overconfident models are dangerous in industrial settings because operators rely on stated confidence to decide whether to shut down equipment. By screening out uncertain synthetic samples and training on a disciplined mixture of data, T-UIL produces predictions whose confidence is more honest, a property the authors highlight as a key benefit alongside raw accuracy.

The significance of this work extends beyond bearings. Intelligent equipment in practical industrial scenarios faces the continuous emergence of new fault modes, and the authors frame their contribution as a general answer to the incremental few-shot fault diagnosis problem. The field has seen a wave of related efforts, from few-shot generative adversarial networks for data-imbalanced diagnosis to prototype recalibration frameworks that avoid reserving exemplar samples, and from Bayesian deep learning frameworks for uncertainty quantification to DDPM-based augmentation methods for imbalanced rolling bearing data. T-UIL distinguishes itself by unifying these threads: generative augmentation, uncertainty-guided screening, replay, contrastive learning, and distillation all operate inside one coherent pipeline rather than as bolted-on patches.

There are honest limits to what the current evidence shows. The published evaluations rest on two public bearing datasets, and the authors note that the data used in the study are available from the corresponding author upon reasonable request, with all experiments conducted on publicly available benchmarks. Whether the uncertainty-guided screening engine transfers to other machinery, other sensor modalities, or noisier industrial environments remains a question for future work. The framework also depends on the quality of the underlying diffusion model: if the generator cannot synthesize plausible samples for a genuinely novel fault physics, the screening engine has little trustworthy material to work with. Still, the core insight stands on its own. In an era when generative AI can manufacture unlimited synthetic data, the hardest problem is no longer making more data but deciding which of it deserves to be believed.

For engineers watching the field, the message is viral in the best sense: the machines that learn fastest may be the ones that learn to doubt. T-UIL shows that a model’s own uncertainty, usually treated as a nuisance to be minimized, can be repurposed as a quality-control instrument for the very data used to teach it. As factories fill with sensors and intelligent equipment, and as fault modes evolve faster than maintenance teams can collect examples, frameworks that combine generative imagination with skeptical screening may define the next generation of trustworthy industrial AI. The work was supported by the Wenzhou Special Project for the Open-Competition Global-Talent Recruitment, and the authors declare no conflicts of interest regarding the publication of the manuscript.

Subject of Research: Artificial Intelligence

Article Title: T-UIL: an incremental few-shot fault diagnosis framework based on uncertainty-guided trustworthy sample screening

Article References: Yu, Q., Hu, L., Lang, C., & Tian, J. (2026). T-UIL: an incremental few-shot fault diagnosis framework based on uncertainty-guided trustworthy sample screening. Applied Intelligence, 56(16), Article 507. https://doi.org/10.1007/s10489-026-07517-2

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07517-2

Keywords: fault diagnosis, incremental learning, few-shot learning, diffusion model, uncertainty quantification, trustworthy AI, bearing datasets, knowledge distillation, T-UIL, incremental, few-shot, fault

The framework, developed by Sadaf Tabatabaee and Mohammad T. Khasawneh, was trained to distinguish glioma, meningioma, pituitary tumor, and no-tumor cases from contrast-enhanced T1-weighted MRI images. The corrected experimental cohort contained 7,023 images, split into 5,712 training images and 1,311 held-out test images, with data augmentation applied only to the training subset to reduce the risk of information leaking between splits. On the untouched test set, SYNAPSE-X achieved 98.41 percent accuracy, 98.2 percent macro-F1, and 98.4 percent weighted-F1 — figures comparable to the best convolutional and transformer-based pipelines previously reported on similar public four-class brain MRI datasets.

What sets the study apart is what happened next. The researchers point out that a model can be highly accurate on average yet dangerously overconfident in its errors, a phenomenon documented since Guo and colleagues showed in 2017 that modern neural networks are frequently miscalibrated. Calibration means that when a model says it is 90 percent sure, it should be right roughly 90 percent of the time on similar predictions. To measure this, the team used three complementary metrics: expected calibration error, which quantifies the gap between stated confidence and observed accuracy across confidence bins; the Brier score, a classical measure of how close predicted probabilities are to actual outcomes; and negative log-likelihood, which harshly penalizes confident mistakes.

Before any correction, SYNAPSE-X showed moderate calibration error despite its stellar accuracy. The team then applied temperature scaling, a simple post hoc technique that rescales the model’s raw output logits with a single learned parameter before the softmax function converts them into probabilities. Crucially, the temperature parameter was fitted only on a validation subset, never on the test data or the external cohort, preventing calibration leakage. The result was striking: expected calibration error fell from 0.041 to 0.018, a relative reduction of roughly 56 percent, while the Brier score dropped from 0.036 to 0.021 and negative log-likelihood fell from 0.092 to 0.061. Because temperature scaling only reshapes probability values and never changes which class the model picks, classification accuracy remained exactly the same — but the confidence numbers became genuinely meaningful.

With calibrated confidence in hand, the researchers implemented the framework’s most distinctive feature: selective referral. Instead of forcing every scan through the same automated decision pathway, predictions whose calibrated confidence fell below a threshold of 0.85 were flagged for expert review, while high-confidence predictions were retained for decision-support reporting. At this operating point, 93.3 percent of internal test cases were retained and 6.7 percent were referred to human specialists. Among the retained cases, accuracy rose to 99.12 percent, pushing retained-case risk below one percent. Sensitivity analyses across thresholds from 0.70 to 0.95 traced out the expected risk-coverage trade-off: more conservative thresholds reduced risk among retained cases but increased the referral burden on clinicians.

The error profile analysis added a clinically reassuring pattern. Referred and misclassified cases were not scattered randomly across the dataset; they clustered in categories where visual overlap is expected, particularly between glioma and meningioma, followed by meningioma and pituitary tumor. Meningioma showed the highest referral tendency, consistent with the fact that meningiomas can share enhancement patterns and boundary characteristics with other tumor types depending on location and imaging plane. No-tumor cases, by contrast, carried the lowest referral burden and the strongest high-confidence behavior. This suggests the model’s uncertainty is concentrated exactly where a radiologist would expect diagnostic difficulty — a property that makes confidence-based triage genuinely informative rather than arbitrary.

Perhaps the most demanding test came from outside the development data entirely. The trained model was applied, without any retraining, fine-tuning, architecture modification, external recalibration, or threshold re-selection, to an independent external brain MRI cohort of 3,264 images. Under this direct-transfer regime, SYNAPSE-X achieved 97.86 percent accuracy, 97.7 percent macro-F1, and 97.9 percent weighted-F1, with an external expected calibration error of just 0.026. At the same 0.85 referral threshold, the external referral rate was 8.4 percent and retained-case accuracy remained high at 98.74 percent. The modest performance drop relative to internal testing is exactly what dataset shift would predict, and the fact that calibration and referral behavior survived the transfer intact strengthens the case that the framework’s reliability signals are robust rather than artifacts of a single dataset.

The authors are careful to position the work honestly. Both the primary and external cohorts are public image-level datasets lacking complete patient-level identifiers, scanner metadata, and institutional acquisition information, which means patient-level independence could not be verified. The study is therefore framed as research-stage reliability assessment rather than clinical validation, and the external testing is described as direct-transfer evaluation rather than proof of deployment readiness. This cautious framing aligns with reporting guidelines such as CLAIM and TRIPOD+AI, and with a broader literature showing that radiology AI models often lose performance across hospital sites, sometimes because they learn site-specific acquisition signatures instead of genuine disease features. Distributional shift, prior work has shown, can degrade the reliability of confidence values themselves, not just accuracy — which is precisely why calibration-aware external testing matters.

Technically, the underlying classifier is a compact Vision Transformer in the tradition introduced by Dosovitskiy and colleagues, in which each 224-by-224 pixel MRI image is divided into non-overlapping 16-by-16 patches, producing 196 patch embeddings. Positional embeddings are added, and the sequence passes through transformer encoder blocks built on multi-head self-attention, residual connections, layer normalization, and feed-forward layers before a four-unit softmax head produces class probabilities. This architecture allows the model to capture not only local texture but also lesion shape and spatial context, which matters when tumor discrimination depends on relationships across distant image regions. Training used categorical cross-entropy loss, AdamW optimization with weight decay, dropout, and training-only augmentation, with checkpoint selection based on validation loss and macro-F1.

The broader significance of SYNAPSE-X lies in its conceptual shift. Conventional classification studies report one aggregate accuracy number, implicitly treating every prediction as equally dependable. By jointly evaluating accuracy, probability calibration, confidence-stratified correctness, risk-coverage behavior, and external transfer stability, the framework converts brain tumor MRI classification into a reliability-aware selective-referral problem in which the AI acts as a triage-aware assistant rather than an infallible oracle. The authors suggest that future work should move to patient-level, multi-institutional datasets with scanner diversity, prospective evaluation, and radiologist-in-the-loop studies to determine whether calibrated selective referral can support safe clinical workflows. For now, the study offers a compelling demonstration that the next generation of medical AI may be judged not by how often it is right, but by how honestly it communicates the moments when it is not sure.

Subject of Research: Calibration-aware Vision Transformer classification of brain tumor MRI with uncertainty-guided selective referral

Article Title: SYNAPSE-X: a calibration-aware vision transformer framework for selective referral in multi-class brain tumor MRI classification

Article References: Tabatabaee, S., & Khasawneh, M. T. (2026). SYNAPSE-X: a calibration-aware vision transformer framework for selective referral in multi-class brain tumor MRI classification. Neural Computing and Applications, 38(20), Article 813. https://doi.org/10.1007/s00521-026-12608-9

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12608-9

Keywords: brain tumor, MRI, Vision Transformer, deep learning, calibration, temperature scaling, selective referral, uncertainty quantification, medical imaging, risk-coverage analysis, external validation, decision support

Therapeutic peptides, short chains of amino acids that can punch holes in bacterial membranes or disrupt essential microbial processes, have long been touted as alternatives to conventional antibiotics. They are structurally diverse, chemically tunable and often engage multiple microbial targets at once, which makes it harder for pathogens to develop resistance through a single mutation. Yet the authors are blunt about a widespread misconception: peptides do not inherently prevent resistance. Bacteria evolve proteases that degrade them, efflux pumps that expel them, and membrane modifications that reduce susceptibility, meaning any discovery strategy that ignores these adaptive mechanisms is designing for a battlefield that has already shifted.

The core critique centres on fragmentation. Current AI-driven workflows typically consist of isolated tools: a classifier that labels sequences as antimicrobial or not, a generative model that proposes novel sequences, and an explainability layer bolted on afterwards. Each component may perform well in benchmarks, but they rarely exchange information coherently. A model that achieves high accuracy on a static test set provides little translational value if it cannot tell researchers why a particular sequence is predicted to be active, how confident it is, or what happens when a pathogen evolves a new resistance mechanism mid-campaign. The result is a discovery process that is technically impressive but operationally brittle.

Interpretability is a particular sore point. Deep learning architectures dominate modern peptide prediction, but they function largely as black boxes. Post hoc explainability methods such as SHAP and LIME can attribute importance to individual features, yet these attributions are sensitive to model architecture, training conditions and input perturbations, and they rarely reveal stable, global patterns of antimicrobial activity. Compounding the problem, many pipelines rely on embedding-based representations derived from protein language models, which improve predictive performance but introduce an abstraction layer that decouples model outputs from interpretable biochemical features. Even when attribution maps are produced, translating them into actionable design decisions becomes difficult because the underlying feature space lacks direct biological meaning.

Compounding the interpretability deficit is the problem of static pipelines. Models are typically trained on in vitro datasets and deployed as fixed predictors, with limited incorporation of experimental feedback for iterative refinement. This is particularly problematic in peptide discovery, where discrepancies between in vitro and in vivo performance are common. A peptide predicted to be highly active may fail under physiological conditions because of degradation, serum binding, immune interactions or altered pharmacokinetics. The authors note that this decoupling is not only methodological but also organisational: computational and experimental groups often operate under different constraints, timescales and throughput capacities, with AI systems generating millions of candidates while only a small fraction can realistically be synthesised and tested.

To address these limitations, the researchers propose three interdependent pillars. The first is explainable and uncertainty-aware predictive modelling, which combines local attribution methods with global interpretability strategies and integrates uncertainty quantification techniques such as deep ensembles, Bayesian neural networks and conformal prediction. The second is conditional and rule-aware generative peptide design, which embeds biological constraints directly into the generative process rather than applying them as post hoc filters. The third is orchestrated modular systems governed by design–test–learn principles, in which predictive, generative and interpretability-oriented modules interact iteratively, allowing experimental observations to propagate systematically across the entire discovery process.

Uncertainty awareness deserves particular attention. Antimicrobial peptide datasets are often sparse and biased, especially for resistant phenotypes, making point predictions unreliable under distributional shift. Conformal prediction offers distribution-free confidence guarantees that are particularly valuable in low-data settings, while ensemble-based approaches can identify regions of elevated uncertainty and support candidate ranking when the model encounters sequences unlike anything in its training data. Integrated within active learning frameworks, these estimates can prioritise sequences expected to maximise experimental information gain, transforming candidate selection from a simple ranking exercise into a principled decision-making problem under limited experimental budgets.

To demonstrate these concepts in practice, the team presented a worked example integrating interpretable prediction, uncertainty-aware prioritisation, rule extraction and adaptive model updating. They curated approximately 86,000 peptides from the Peptipedia v2.0 database, removed non-canonical sequences, excluded peptides outside the range of 5 to 60 amino acids, and reduced redundancy through MMseqs2 clustering at 80 percent sequence identity, yielding roughly 47,000 non-redundant peptides. For each sequence, a panel of 41 descriptors was calculated, encompassing amino acid composition, physicochemical properties, sequence complexity, entropy-related measures and structural propensity estimates. An ensemble of ten predictive models, trained under distance-aware data partitioning to better approximate prospective screening scenarios, achieved an overall Matthews correlation coefficient of 0.65 on a held-out evaluation set of approximately 14,000 peptides.

Global SHAP analysis applied to this ensemble revealed that descriptors associated with amino acid composition, charge distribution, physicochemical balance and sequence complexity emerged among the most influential variables, consistent with established mechanistic characteristics of membrane-active antimicrobial peptides. These explanatory signals were then converted into explicit decision rules using surrogate modelling approaches, generating twelve human-readable criteria that can support candidate selection and optimisation. Prediction uncertainty was estimated through ensemble variability and incorporated directly into the prioritisation process, distinguishing between sequences suitable for immediate experimental validation and candidates that may require additional evidence before prioritisation. Application of this multi-criteria strategy to an external screening pool of approximately 16,000 candidate peptides resulted in the prioritisation of 606 sequences for further evaluation.

The authors are careful to note the limitations of this demonstrative application. Extensive hyperparameter optimisation, large-scale benchmarking and exhaustive comparative evaluation were intentionally omitted, and the implementation does not include generative modelling components, explicit modelling of resistance evolution, active learning-based experimental design or feedback derived from wet-laboratory validation. Nevertheless, the example illustrates how complementary sources of information can be transformed into actionable decision support throughout the discovery process. Looking forward, the researchers suggest that advances in active learning, foundation models, autonomous experimentation and scientific agents may further accelerate the development of closed-loop discovery ecosystems. Ultimately, they argue, the impact of artificial intelligence on antimicrobial peptide development will depend not only on predictive performance but on the ability to continuously integrate prediction, experimentation, interpretation and learning within a unified and evolving discovery framework, one capable of adapting as fast as the pathogens it is designed to defeat.

Subject of Research: Toward explainable and adaptive artificial intelligence systems for antimicrobial peptide discovery under resistance pressure

Article Title: Toward explainable and adaptive artificial intelligence systems for antimicrobial peptide discovery under resistance pressure

Article References: Islas-Ávila, A. L., García-Vinuesa, J., Soto-García, N., Soto-García, M., Gutiérrez-Cárdenas, B., Inostroza-Munoz, A., Davari, M. D., Murgas, L., Uribe-Paredes, R., Martínez-Rebollar, A., Sepulveda-Yañez, J., & Medina-Ortiz, D. (2026). Toward explainable and adaptive artificial intelligence systems for antimicrobial peptide discovery under resistance pressure. Frontiers in Drug Discovery, 6, Article 1820604. https://doi.org/10.3389/fddsv.2026.1820604

Image Credits: AI Generated

DOI: 10.3389/fddsv.2026.1820604

Keywords: antimicrobial resistance, antimicrobial peptides, explainable AI, uncertainty quantification, generative design, design–test–learn, machine learning, drug discovery, SHAP, conformal prediction, Toward, explainable

The framework was developed by Faizan Ullah and colleagues at Forschungszentrum Jülich, RWTH Aachen University, Arab Open University, and Yeungnam University. Its purpose is deliberately different from the usual goals of medical imaging AI. Rather than deciding whether a tumor is present or delineating its boundaries for clinical use, NeuroFAIR Curator acts as a data steward. It enriches metadata, screens technical annotation consistency, characterizes image quality, flags statistically unusual masks, prioritizes samples for human review, and scores the dataset against the FAIR principles of Findability, Accessibility, Interoperability, and Reusability. The authors are explicit that the system does not determine the clinical correctness of tumor annotations; it surfaces computational signals that experts can then investigate.

The test case was BRISC 2025, a public dataset of 6,000 two-dimensional contrast-enhanced T1-weighted MRI images distributed as JPEG files with binary PNG segmentation masks. The images span four categories—glioma, meningioma, pituitary tumor, and no tumor—across axial, coronal, and sagittal anatomical planes. NeuroFAIR Curator generated 6,000 sample-level curation records linked with 15,586 external manifest entries, producing 21,586 unified records, and processed 4,793 image-mask pairs. The resulting manifest serves simultaneously as a dataset description and as a curation log, carrying inferred sequence descriptors, anatomical planes, image statistics, mask statistics, consistency checks, quality flags, and model-derived signals for every sample.

Two neural networks form the computational core of the label and mask screening. A ResCNNClassifier, trained for 120 epochs with class weighting, label smoothing, MixUp augmentation, and an exponential moving average of weights, achieved 98.30 percent accuracy, 98.53 percent macro F1, and 99.79 percent macro AUC on the held-out test set. Per-class F1 ranged from 0.9736 for meningioma to a perfect 1.0000 for the no-tumor class. The classifier’s confidence and uncertainty signals, together with any disagreement between the recorded label and the model prediction, are stored as curation signals rather than diagnostic outputs. A UNet segmenter, trained for 90 epochs with a combined binary cross-entropy, Dice, and focal loss, achieved a Dice score of 82.30 percent, an IoU of 72.69 percent, precision of 83.64 percent, recall of 84.61 percent, a Hausdorff distance of 7.84 pixels, and an HD95 of 4.11 pixels at a validation-selected threshold of 0.95. Boundary quality reached 0.9964.

Explainability was built into the workflow. Grad-CAM saliency maps were generated from the final convolutional layer for the 20 most uncertain test images, and their fidelity was quantified using deletion area under the curve, or DAUC, which measures how much class confidence drops when the most salient pixels are progressively deleted. The mean DAUC was 0.2996, with per-class means of 0.3145 for glioma and 0.2933 for meningioma. Lower values indicate that the importance ordering is faithful to what the network actually relies on. The authors stress that these maps quantify faithfulness of the explanations but do not constitute clinical localization validation.

One of the most intriguing findings concerns mask morphology. The framework stratified supplied mask-area ratios by tumor class and anatomical plane and applied class- and plane-specific 1.5 times interquartile-range limits. This exploratory screening flagged 181 of 4,793 masks, or 3.78 percent, as statistically atypical. The authors are careful to note that these flags are not confirmed annotation errors; they may reflect tumor size, slice position, morphology, annotation convention, or genuine annotation problems, and without clinical-reader assessment these explanations cannot be distinguished. The flags are retained as review candidates rather than grounds for automatic exclusion, a stance consistent with prior research showing that inter-observer variability in tumor delineation can be substantial.

Technical integrity checks told a reassuring story about the dataset’s file-level organization. No missing tumor masks, empty tumor masks, dimension mismatches, filename inconsistencies, unknown planes, or corrupted samples were detected. Image-quality screening, by contrast, revealed a small but real tail of problematic images: 51 low-contrast cases, 60 blur outliers, 208 intensity outliers, and 180 anomalies detected by an Isolation Forest. Across the corpus, 456 samples carried at least one flag and 184 carried multiple flags simultaneously, while 5,544 samples had zero flags.

The framework’s most distinctive contribution is the Curation Uncertainty Index, or CUI, a transparent seven-component score that fuses classification uncertainty, segmentation uncertainty, mask abnormality, metadata incompleteness, image-quality anomaly, image-mask mismatch, and metadata-validation error. The baseline weights were prespecified as a hierarchical prior rather than fitted to test data, with model uncertainty receiving 0.36, statistical and structural mask evidence 0.16, reuse-quality evidence 0.28, and sparse technical validation signals 0.20. A one-at-a-time sensitivity analysis perturbed each weight by up to 50 percent and showed that the ranking remained stable under most perturbations, with Spearman correlations above 0.99. The largest ranking change occurred when segmentation uncertainty was reduced by half, highlighting which signals drive the prioritization.

Active curation prioritization demonstrated the practical value of the CUI. At a review budget of just 1 percent of the dataset, the CUI ranking captured 49 of the 456 uniquely flagged samples, a 10.75 percent capture rate that represents a 99 percent reduction in review workload compared with random inspection. At budgets of 5 and 10 percent, capture rose to 15.13 and 17.11 percent respectively. The CUI assigned 5,001 samples, or 83.35 percent, to low curation priority and 999 samples, or 16.65 percent, to moderate priority, giving curators a concrete, defensible ordering for their limited expert time.

The FAIR readiness assessment, scored on an 18-sub-criterion rubric, showed the largest gains where the dataset was weakest. Findability improved from 87.50 to 100.00 and Interoperability from 36.00 to 91.98, reflecting the addition of structured descriptors, harmonized metadata, quality indicators, and provenance fields. Accessibility and Reusability remained unchanged at 95.00 and 95.98 under the declared rubric, a pattern the authors attribute to the specific instrument and dataset rather than a universal result. Taken together, NeuroFAIR Curator converts a widely reused public brain tumor MRI corpus into a metadata-enriched, technically checked, quality-screened, review-prioritized, and FAIR-aligned neuroinformatics resource. The authors note that future work should evaluate the framework with blinded expert review, multicenter DICOM or NIfTI datasets, three-dimensional annotations, and external calibration across different scanners and acquisition protocols, since the operating threshold and CUI weights are specific to the BRISC 2025 validation distribution rather than universal parameters.

Subject of Research: AI-assisted metadata enrichment, quality screening, curation prioritization, and FAIR readiness assessment of public brain tumor MRI data

Article Title: NeuroFAIR curator: an AI-assisted framework for metadata enrichment, quality screening, curation prioritization, and FAIR readiness of brain tumor MRI data

Article References: Ullah, F., Abbas, Z., Abrar, M., Amin, F., & Shah, N. J. (2026). NeuroFAIR curator: an AI-assisted framework for metadata enrichment, quality screening, curation prioritization, and FAIR readiness of brain tumor MRI data. Frontiers in Neuroinformatics, 20, Article 1907796. https://doi.org/10.3389/fninf.2026.1907796

Image Credits: AI Generated

DOI: 10.3389/fninf.2026.1907796

Keywords: brain tumor MRI, FAIR data, metadata enrichment, data curation, neuroinformatics, annotation integrity, uncertainty quantification, active learning, image quality control, BRISC 2025, Curation Uncertainty Index, deep learning