The study, published in Frontiers in Artificial Intelligence, placed 55 participants — 29 men and 26 women aged 18 to 61 — in front of a virtual audience of four characters arranged in a semicircle, delivered through a wireless Meta Quest Pro headset. Each participant gave six short oral presentations on familiar topics such as a personal introduction or a description of their hometown. The experimental design systematically varied two audience characteristics: the gendered appearance of the virtual characters, which could be entirely male, entirely female, or mixed, and the audience’s social attitude, expressed through nonverbal cues such as posture, gaze direction and facial expressions that conveyed either agreement, attentiveness and happiness, or disagreement, distraction and unhappiness.

What made the experiment technically distinctive was its multimodal data capture. The Meta XR Interaction SDK recorded head and hand positions, gaze direction and facial muscle activations derived from FACS blendshapes directly through the headset, used without controllers. An external camera in front of the speaker recorded each session for later expert evaluation. From these raw signals, the researchers extracted 33 interpretable multimodal nonverval behavioral features organized into four categories: gaze features such as gaze diversity and entropy, body movement features including horizontal displacement amplitude, facial expression features tracking ten specific blendshape frequencies, and gesture and posture features covering palm orientations, hand configurations and body openness.

The evaluation framework combined subjective and objective measures. Participants completed the Self-Assessment Manikin to gauge emotional valence and the Personal Report of Confidence as a Speaker to measure anxiety after each presentation, alongside a 7-point self-reported performance rating. An experienced public speaking expert then evaluated all 330 recorded presentations through a dedicated Unity-based replay application, rating overall performance and 20 behavioral items across five dimensions: verbal behavior, gestures, posture, body movements and gaze direction. The order of presentations was randomized to prevent direct comparisons between sessions from the same participant.

Statistical analysis using Aligned Rank Transform ANOVA revealed a striking pattern. The virtual audience’s social attitude significantly influenced participants’ emotional experience and self-reported performance: speakers reported lower anxiety and more positive emotions when facing a positive audience than a negative one, and rated their own performance more favorably after supportive presentations. However, no significant effect of audience gendered appearance was detected on any measure — anxiety, emotional state, self-reported performance or expert-rated performance. More surprisingly, speaker gender showed no significant influence on emotional experience or overall public speaking performance either, contrasting with previous literature reporting higher public speaking anxiety among female speakers.

The expert did, however, identify gender differences in specific nonverbal behaviors. Male speakers displayed more open gestures and appeared more comfortable in terms of body protection behaviors, while female speakers were rated as more static during their presentations and needing greater improvement in gesture use. Notably, the expert found no gender differences in verbal behavior. But these observations could reflect either genuine behavioral variations or subjective perceptions on the part of the evaluator — a distinction the researchers addressed with machine learning.

To determine whether nonverval behaviors objectively differed by gender, the team trained a Multilayer Perceptron classifier to predict speaker gender from the automatically extracted features. Each of the 330 presentations was divided into three equal-duration segments, producing 990 observations split with rigorous stratification so that subjects in the training set never appeared in the test set. Hyperparameter optimization used random search over 10 configurations, with Recursive Feature Elimination combined with 5-fold cross-validation to identify the optimal feature subset. The final model achieved a weighted F1-score of 0.81 and an area under the ROC curve of 0.86 on the independent test set, with the small gap between training and test performance suggesting good generalization and limited overfitting.

SHapley Additive exPlanations (SHAP) analysis then revealed which behavioral features drove the predictions. Horizontal displacement amplitude — the maximum distance a speaker moved along the x-axis — emerged as the most important feature for gender classification. Female speakers tended to exhibit smaller horizontal displacement amplitudes, used neutral palm orientations less frequently while displaying downward-oriented palms more often, and directed their gaze less uniformly toward the audience. Male speakers generally displayed larger displacement amplitudes, more frequent neutral palm orientations, and more balanced gaze behavior distributed evenly across audience members. Female speakers also showed greater variability in facial expressions, particularly through more frequent Upper Lid Raiser movements, an action unit commonly associated with surprise, fear or anger.

The convergence between expert evaluation and automatic analysis is significant: both approaches identified distinct nonverval behavioral patterns between male and female speakers, even though neither found differences in overall performance. The automatic analysis went further, detecting gaze and facial expression differences that the expert could not systematically assess — partly because the VR headset partially obscured speakers’ faces, making facial expression evaluation impractical, and partly because the multi-window replay interface limited accurate tracking of gaze dynamics.

These findings carry important implications for automated public speaking training systems. If comparable levels of performance can be achieved through distinct behavioral profiles, a universal evaluation model may be insufficient to accurately assess speakers or provide effective guidance. Accounting for individual behavioral variability may represent a critical consideration in developing personalized feedback systems. The researchers also caution that while personalized feedback may better reflect individual communication patterns, it could inadvertently reinforce existing gender stereotypes — a trade-off future work must investigate. The study’s limitations include reliance on a single expert evaluator, exclusive focus on nonverval features in the automatic analysis, and an experimental setting that may not have elicited sufficiently high stress levels for gender-related differences in emotional experience to become detectable. Future research could extend the computational approach to verbal and vocal features, design virtual audiences whose gendered appearance is complemented by gender-congruent nonverval behaviors, and examine whether these behavioral differences extend to real-life public speaking situations.

Subject of Research: Neuroscience / AI

Article Title: How gender shapes public speaking in virtual reality: a multimodal analysis

Article References: Ristorcelli, M., Casanova, R., & Ochs, M. (2026). How gender shapes public speaking in virtual reality: a multimodal analysis. Frontiers in Artificial Intelligence, 9, Article 1954571. https://doi.org/10.3389/frai.2026.1954571

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1954571

Keywords: virtual reality, public speaking, gender differences, machine learning, nonverbal behavior, multimodal analysis, SHAP explainability, virtual audience, gender, shapes, public, speaking

The study, published in the International Journal of Machine Learning and Cybernetics, was led by Anish Kumar Vishwakarma and Rakesh Ranjan, who set out to fix a structural weakness in how deep learning has been applied to agricultural image classification. Most existing pest recognition systems are built as multi-stage pipelines: one model extracts features from an image, and a separate classifier makes the final decision. This division of labor sounds tidy, but it carries real costs. Each stage adds computational overhead, and because the feature extractor is not trained jointly with the classifier, the learned representations can miss the subtle visual cues that separate one pest species from another. For insects that differ only in the pattern of wing veins or the shade of a marking on the thorax, that loss of discriminative detail is exactly where mistakes happen.

Vishwakarma and Ranjan’s answer is an end-to-end architecture built on a modified MaxViT backbone, a vision transformer design first introduced at the European Conference on Computer Vision in 2022. MaxViT is unusual among transformers because it combines two kinds of attention in a single network: window attention, which captures fine-grained local relationships within small patches of the image, and grid attention, which lets distant regions of the image exchange information. This multi-axis attention is what allows the model to notice both the texture of an insect’s body and its overall shape and position in the frame, two kinds of evidence that classical convolutional networks often handle separately.

The key innovation, however, lies in what the authors call a Multi-Stage Feature Fusion Head. Deep networks process images hierarchically: early layers detect simple edges and textures, middle layers assemble those into parts and patterns, and deep layers encode abstract semantic concepts. In many architectures, the low-level texture information is effectively discarded by the time the network reaches its final decision. The fusion head explicitly reintegrates these complementary levels, combining low-level texture features with high-level semantic features so that the final classification draws on the full spectrum of visual evidence. For visually similar pest categories, this fusion proved decisive, since fine surface texture often carries the discriminating signal that abstract features alone cannot supply.

Training such a system on a specialized agricultural dataset presents its own challenge: there is simply no jute pest corpus large enough to train a transformer from scratch. The researchers therefore employed a two-phase transfer learning strategy, starting from representations pretrained on ImageNet, the massive general-purpose image database that has become the standard starting point for computer vision. In the first phase, the pretrained backbone is adapted to the general characteristics of the target domain, stabilizing optimization before the model is asked to master the finer distinctions among pest species. In the second phase, the network fine-tunes on the jute pest data itself, allowing the ImageNet-learned features to reshape around the specific visual grammar of insects on leaves and stems. This staged approach avoids the instability that can occur when a large pretrained model is abruptly exposed to a narrow new domain.

The evaluation was carried out on the publicly available Jute Pest Dataset hosted in the UCI Machine Learning Repository, which covers seventeen jute pest species. The results were striking across every standard metric: overall classification accuracy of 98.43 percent, precision of 98.28 percent, recall of 98.28 percent, and an F1-score of 98.26 percent. The tight agreement among these numbers matters, because it indicates the model performs consistently across all seventeen classes rather than excelling on a few common pests while failing on rare ones. To guard against the possibility that a single favorable train-test split had flattered the results, the authors also ran five-fold stratified cross-validation, and the model’s performance held steady across different data partitions, confirming genuine generalization rather than memorization.

The comparison against established baselines was equally one-sided. The proposed framework consistently outperformed VGG16, ResNet50, DenseNet201, the original Vision Transformer, the Swin Transformer, and ConvNeXt, a lineup that spans a decade of computer vision architecture design. It also surpassed the previously best-performing published method for this task, a feature fusion framework paired with a mountain gazelle optimizer that had reported 96.78 percent accuracy. The new system’s improvement to 98.43 percent represents a meaningful gain in a domain where the remaining errors often involve the most visually confusable species pairs, precisely the cases where misidentification leads to the wrong pesticide or the wrong intervention.

Accuracy alone, however, is not enough for a tool that farmers and agronomists are expected to trust. A model that achieves high scores by latching onto background clutter, such as the color of the soil or the lighting conditions under which photos were taken, will fail unpredictably in the field. To probe what the network was actually looking at, the researchers applied SHAP, or SHapley Additive exPlanations, a technique from game theory that assigns each input region a contribution score for the model’s decision. The analysis showed that the model concentrates its attention on biologically meaningful pest characteristics rather than irrelevant background regions. That transparency is a quiet but important result: it suggests the network has learned something resembling an entomologist’s eye, and it gives practitioners a way to verify that predictions rest on the insect itself.

The broader context is the economics of jute and the promise of precision agriculture. Pest infestations can significantly reduce both crop yield and fiber quality, and jute remains a cornerstone crop for Bangladesh and eastern India, where it sustains farming livelihoods and supplies the global market for natural fiber. Automated, image-based pest identification could allow farmers to detect infestations early, target interventions only where needed, and reduce the blanket pesticide spraying that carries both financial and environmental costs. The authors note that their framework achieves a practical balance between classification accuracy, computational efficiency, and interpretability, the three qualities that determine whether a laboratory model can survive deployment in an intelligent agricultural decision support system.

There are, of course, the usual caveats that separate a benchmark result from a field-ready product. The model was trained and tested on a curated public dataset, and real-world deployment will confront images taken with varied cameras, under harsh sunlight, of partially obscured insects on wind-blown plants. The researchers themselves frame the work as a foundation for practical jute pest monitoring rather than a finished deployment. Still, the trajectory is clear: a single end-to-end transformer, trained with a disciplined two-phase strategy and equipped with a fusion mechanism that preserves both texture and meaning, now classifies seventeen pest species more reliably than any method before it. For a crop that has been cultivated for millennia and identified by eye for just as long, the golden fiber may soon have a digital sentinel watching over it.

Subject of Research: Deep learning-based automated recognition of jute pest species for sustainable agriculture

Article Title: A hierarchical feature fusion transformer for robust Jute Pest recognition in sustainable agriculture

Article References: Vishwakarma, A. K., & Ranjan, R. (2026). A hierarchical feature fusion transformer for robust Jute Pest recognition in sustainable agriculture. International Journal of Machine Learning and Cybernetics, 17(10), Article 494. https://doi.org/10.1007/s13042-026-03330-y

Image Credits: AI Generated

DOI: 10.1007/s13042-026-03330-y

Keywords: jute pests, vision transformer, MaxViT, feature fusion, deep learning, precision agriculture, transfer learning, SHAP explainability, image classification, crop protection, computer vision, sustainable farming

The technical core of the framework is a domain-adapted multilingual transformer encoder, a neural network architecture that converts text into dense numerical representations capturing semantic meaning. Transformer models, which underpin modern language technologies, process words in relation to one another across an entire sentence, allowing them to disambiguate terms that shift meaning with context. In agriculture this matters enormously: a phrase describing leaf spots, wilting, or discoloration must be mapped to the correct pathological category, and the same symptom vocabulary may differ across languages and regions. AgroNLP-SimAssist shares one semantic representation across three prediction heads, each tailored to a distinct task, so that the contextual understanding built from a farmer’s query remains consistent whether the system is classifying a disease, ranking advisory recommendations, or extracting named entities such as crop names, pests, and locations from the text.

Training and validation drew exclusively on publicly available data sources, a deliberate choice to counter the reproducibility problems the authors see in the field. The pipeline incorporates multilingual agricultural reports from the Food and Agriculture Organization, crop disease descriptions from the PlantVillage knowledge base, and market price narratives from Agmarknet, India’s agricultural marketing information network. At the heart of the evaluation effort is a dataset of 3,200 multilingual farmer queries spanning English, Hindi, and Kannada. The researchers split this corpus in the conventional manner, assigning 70 percent of queries to training, 15 percent to validation, and 15 percent to testing, which allows the model to learn from the bulk of the data while retaining held-out examples for honest performance measurement.

The reported results are striking. On disease classification, the system achieved 96.1 percent accuracy, outperforming existing classifiers against which it was compared. Advisory relevance, a measure of whether the recommendations the system generates actually match the agronomic situation described in the query, reached 94.3 percent. Named entity recognition, the task of pulling structured information such as crop and disease names out of free-form text, scored 92.7 percent on the F1 measure, a metric that balances precision and recall by taking their harmonic mean. For a system operating across three languages and three distinct tasks simultaneously, these figures suggest that the shared-representation approach does not force a trade-off between breadth and depth, a concern that often dogs multitask machine learning systems.

Perhaps the most distinctive contribution, and the one the authors emphasize as filling a genuine gap, is the advisory simulation engine. Most agricultural NLP papers report accuracy on static test sets and stop there, leaving a chasm between offline benchmark performance and what happens when thousands of real users hit the system at once during a disease outbreak or a market crisis. AgroNLP-SimAssist confronts this directly by simulating deployment at scale. In the reported experiments, the system scaled to 10,000 concurrent users while maintaining an inference latency of 0.35 seconds per query, fast enough for interactive conversational use even under heavy load. Latency and concurrency testing of this kind is standard practice in commercial software engineering but has been rare in agricultural AI research, and its inclusion here signals a push toward systems that are genuinely deployable rather than merely publishable.

Interpretability receives equally careful treatment. The framework integrates SHAP values, a technique from explainable AI that quantifies how much each input feature contributes to a particular model output. For every diagnosis or recommendation, the system can show which words or phrases in the farmer’s query pushed the model toward its conclusion. The authors frame this as enabling expert validation: an agronomist reviewing the system’s reasoning can check whether the model is attending to genuinely diagnostic language, such as symptom descriptions, or latching onto spurious patterns. This matters because black-box AI in high-stakes domains like food production faces justified skepticism, and advisory systems that cannot explain themselves are unlikely to earn the trust of extension workers or the farmers they serve.

The significance of the multilingual design is hard to overstate in the Indian context, where hundreds of millions of farmers operate in languages other than English and where extension services are chronically stretched. Kannada, spoken primarily in Karnataka, and Hindi, the most widely spoken language in northern India, are both covered alongside English, meaning the system can serve populations that are typically the last to benefit from digital agriculture tools. By grounding the entire pipeline in open data from FAO, PlantVillage, and Agmarknet, the researchers have also made it possible for other teams to reproduce their results, extend the model to additional languages, and audit its behavior, all of which are difficult or impossible when systems are trained on proprietary datasets locked inside companies or labs.

The three tasks the system unifies map closely onto the actual information needs of a working farmer. Disease diagnosis answers the urgent question of what is wrong with a crop. Advisory generation translates that diagnosis into actionable guidance on treatment and management. Market intelligence, drawn from price narratives and trend estimation, helps the farmer decide when and where to sell, closing the loop from production to economics. Traditional intent-oriented frameworks, as the authors note, might recognize that a farmer is asking about a disease but stop short of delivering the diagnosis itself, let alone connecting it to price trends. By handling diagnosis, advisory ranking, and entity extraction within one architecture, AgroNLP-SimAssist avoids the fragmentation that has limited earlier agricultural NLP deployments.

Limitations and open questions remain, as they do with any early-stage system. The evaluation dataset of 3,200 queries, while multilingual, is modest by the standards of large-scale language modeling, and performance on languages, crops, or disease presentations not represented in the training data is an obvious area for future work. The published version is an early-release, peer-reviewed accepted manuscript that the journal notes is subject to further edits before the final version of record, so some details may be refined. The authors also declare no competing interests and note that open access funding was provided by the Manipal Academy of Higher Education. Still, the combination of strong benchmark numbers, transparent open-data provenance, built-in explainability, and, crucially, demonstrated scalability under simulated real-world load makes this one of the more complete visions of what an intelligent agricultural advisory system can look like. If precision agriculture is to reach the smallholders who need it most, systems like this one, tested not just for accuracy but for the messy realities of deployment, point toward a plausible path.

Subject of Research: A multilingual transformer-based decision support system for crop disease diagnosis, agricultural advisory generation, and market intelligence

Article Title: AgroNLP-SimAssist: a multilingual transformer-based decision support simulator for crop disease diagnosis, advisory generation, and market intelligence

Article References: Thirugnanam, P., Mathivanan, S. K., M R, S. K., Ramaswamy, S., T S, S., & S.K.B, S. (2026). AgroNLP-SimAssist: a multilingual transformer-based decision support simulator for crop disease diagnosis, advisory generation, and market intelligence. BMC Plant Biology. https://doi.org/10.1186/s12870-026-10040-8

Image Credits: AI Generated

DOI: 10.1186/s12870-026-10040-8

Keywords: agricultural NLP, transformers, precision agriculture, crop disease diagnosis, multilingual AI, decision support systems, SHAP explainability, PlantVillage, Agmarknet, smart farming, simulation engine, open datasets

The research, led by Chinyere P. Okechukwu of Lovely Professional University together with colleagues at Nnamdi Azikiwe University and Ogbonnaya Onu Polytechnic, assembled 43 years of quality-controlled monthly rainfall records, spanning 1981 to 2023, from 38 stations across the country. Rather than treating Nigeria as a single statistical unit, the team first asked whether the data themselves reveal natural rainfall regions. Using ensemble clustering on a 15-dimensional feature space capturing seasonal, magnitude and variability characteristics, they identified four homogeneous rainfall zones, with a Silhouette coefficient of 0.375 and a Davies–Bouldin index of 0.696 quantifying the coherence of the resulting groups. The zones include a Coastal Humid region, a Transitional Sub-Humid belt, a Northern Arid/Sahelian zone and a smaller cluster of anomalous or outlier stations that behave unlike their neighbours.

With the regional map in hand, the authors turned to trend detection. Mann–Kendall tests, a standard non-parametric tool for identifying monotonic change in noisy hydrological series, revealed statistically significant wetting trends in two zones that might surprise observers accustomed to narratives of Sahelian desiccation. The Transitional Sub-Humid zone showed a trend statistic of Z = 3.516 with a p-value of 0.0004 and a slope of 0.607 millimetres per year, while the Northern Arid/Sahelian zone recorded Z = 5.065, p < 0.0001, and a slope of 0.529 millimetres per year. Crucially, the increases were concentrated in the June-to-September monsoon season, the window when most of the region’s crops depend on rain and when saturated soils are most vulnerable to flooding.

Trends in the mean, however, tell only part of the story. Extreme rainfall events, not averages, destroy infrastructure and displace communities. To characterise those extremes, the team fitted Generalized Extreme Value (GEV) models, the workhorse of hydrological risk analysis, and allowed the distribution’s parameters to vary through time rather than assuming a stationary climate. Model comparison using the Bayesian Information Criterion showed that non-stationary formulations were consistently preferred, with ΔBIC values ranging from 7.52 to 189.07 across zones, evidence that the statistical behaviour of rainfall extremes has genuinely shifted over the four-decade record. Interestingly, when it came to estimating return levels, a stationary LogNormal distribution provided the best overall fit, a reminder that model choice in extreme-value analysis remains a careful balancing act between flexibility and parsimony.

The return-level estimates carry direct engineering consequences. The 100-year rainfall, the depth expected to be exceeded on average once per century, ranged from 382.6 millimetres in the Northern Arid zone to 643.8 millimetres in the Coastal Humid zone. Because these figures are rising against a background of non-stationarity, drainage systems, culverts and flood defences designed decades ago using historical stationarity assumptions are increasingly likely to be undersized. The authors argue that their zone-specific return levels offer a more defensible basis for climate-resilient infrastructure design than national averages or legacy design manuals.

The predictive phase of the framework compared a battery of machine-learning models across the four zones, and the results underline a central lesson of modern applied machine learning: no single algorithm wins everywhere. Random Forest achieved the strongest performance in the Outlier, Transitional and Northern Arid zones, with Nash–Sutcliffe Efficiency (NSE) values ranging from 0.654 to 0.894, while Support Vector Regression took the lead in the Coastal Humid zone with an NSE of 0.735. To test whether the models had genuinely learned transferable rainfall physics rather than station-specific quirks, the team employed Leave-One-Cluster-Out cross-validation, training on three zones and predicting the held-out one, which yielded NSE values from 0.220 to 0.766, a more sobering but honest picture of spatial generalisation.

Deep learning added a further twist. Across repeated runs, standard Long Short-Term Memory (LSTM) networks consistently outperformed their Bidirectional counterparts in the Coastal Humid, Transitional and Northern Arid zones, posting mean NSE values of 0.625, 0.778 and 0.870 respectively, against 0.465, 0.368 and 0.353 for the Bidirectional LSTM. The two architectures performed comparably only in the Outlier/Anomalous zone. Because bidirectional models read the input sequence forwards and backwards, they are often assumed to extract richer temporal context; here, that extra capacity conferred no consistent advantage, suggesting that for monthly rainfall prediction the relevant information is largely causal, carried by antecedent conditions rather than future context.

What elevates the study beyond a leaderboard of algorithms is its commitment to explainability. Using SHapley Additive exPlanations (SHAP), a game-theoretic method that attributes each prediction to individual input features, the authors found that Lag-1 rainfall, the previous month’s total, was the dominant predictor in every zone, with mean absolute SHAP values between 51.0 and 92.8 millimetres. More strikingly, trend-related features carried meaningful attribution weight, revealing that the models had absorbed the long-term intensification signal embedded in the record, particularly in the Sahelian north. In other words, the networks were not merely memorising seasonality; they were detecting the slow drift of a changing climate.

Equally notable is what the models did not find. Indices of the El Niño–Southern Oscillation (ENSO), the Atlantic Multidecadal Oscillation (AMO) and the Intertropical Convergence Zone (ITCZ) proved non-significant across all cluster stations, indicating that, at the monthly scale and within these zones, local antecedent rainfall and embedded trends dominate predictability over large-scale teleconnections. For forecasters in West Africa, that is an actionable result: it suggests that skilful monthly outlooks can be built primarily from local observation networks, even where global climate-index-based forecasting underperforms.

The framework’s broader significance lies in its integration. Regionalisation, trend analysis, non-stationary extreme-value modelling, deep learning prediction and SHAP-based attribution are usually pursued as separate exercises; here they form a single pipeline that connects data-driven rainfall zoning to physically interpretable, zone-specific insight. The authors position the work as a foundation for climate-resilient infrastructure design, flood risk management and adaptive water resource planning in Nigeria and in other monsoon-driven regions facing similar gradients. As rainfall extremes continue to drift away from the stationary past, tools that are simultaneously accurate and transparent may prove as important as the rain gauges themselves.

Subject of Research: Explainable machine learning for non-stationary rainfall modelling and extreme value analysis in Nigeria

Article Title: A non-stationary and explainable machine learning framework for adaptive rainfall modelling

Article References: Okechukwu, C. P., Anabike, I. C., Ubaoji, K. I., Ugorji, C. C., Paul, R. U., & Obulezi, O. J. (2026). A non-stationary and explainable machine learning framework for adaptive rainfall modelling. Theoretical and Applied Climatology, 157(10), Article 617. https://doi.org/10.1007/s00704-026-06583-3

Image Credits: AI Generated

DOI: 10.1007/s00704-026-06583-3

Keywords: rainfall modelling, machine learning, Nigeria, extreme value analysis, GEV distribution, LSTM, SHAP explainability, Mann-Kendall trend test, monsoon, climate non-stationarity, flood risk, Sahel

The research, led by Mengyuan Jing, Haoxiang Lu, and Qing Liu with colleagues at Lanzhou University Second Hospital, the Guangdong Cardiovascular Institute, and partner institutions, set out to build an explainable machine learning model that could predict functional improvement after ablation in patients with atrial fibrillation combined with heart failure. Rather than relying on a single measurement, the team fused three complementary streams of information: the geometry of the left atrium and pulmonary veins as seen on cardiac CT angiography, quantitative radiomics features extracted from the left atrial wall, and a handful of routine clinical variables. The result was a combined model, dubbed COMB, that achieved an area under the receiver operating characteristic curve of 0.866 in the training set, 0.803 in the validation set, and 0.845 in an independent testing set drawn from other hospitals.

Those numbers deserve unpacking, because the area under the curve, or AUC, is the workhorse metric for judging how well a model separates patients who will improve from those who will not. A value of 0.5 would mean the model performs no better than a coin flip, while 1.0 would indicate perfect discrimination. Scores consistently above 0.80 across three separate patient cohorts, including an external test set the model had never encountered during development, represent solid, clinically meaningful performance. Equally important is how the team arrived at that performance. Patients from the primary institution were randomly divided into training and validation sets in a seven-to-three ratio, while two other organizations contributed entirely separate testing cohorts, a design that guards against the model simply memorizing the quirks of one hospital’s scanner or patient population.

The technical heart of the study lies in what the researchers chose to measure. Cardiac CT angiography, already a routine part of pre-ablation planning because it maps the pulmonary veins before catheter insertion, provides exquisitely detailed three-dimensional images of the left atrium, the chamber where fibrillation typically originates. From these images the team extracted morphological features describing the shape of the left atrium and the pulmonary veins, capturing subtle geometric signatures such as chamber distortion and remodeling that a human reader might overlook. In parallel, they applied radiomics, a computational approach that converts medical images into hundreds of quantitative descriptors of texture, intensity, and spatial pattern, to the left atrial wall itself, the thin muscular sleeve where ablation lesions are created and where fibrotic change often determines whether the procedure succeeds.

From this high-dimensional feature space, the researchers distilled two focused models. Two shape features and three left atrial wall radiomics features survived rigorous screening and were used to construct what they called the Shape model and the Wall model, respectively. Each model produced a continuous score, the Shape_score and the Wall_score, which quantified how strongly a given patient’s cardiac anatomy resembled that of patients who went on to improve. These scores were then integrated with the clinical variables that proved most informative, namely gender, hyperlipidemia, blood urea level, and the type of atrial fibrillation, to yield the final combined model. The parsimony is striking: out of the vast number of features that could have been included, the final predictor rests on just a handful of imaging and clinical inputs, all obtainable from examinations and blood tests that are already standard practice.

What elevates the study beyond a typical prediction exercise is its commitment to explainability. Black-box algorithms have long been a stumbling block for clinical adoption, because physicians are understandably reluctant to act on a probability they cannot interrogate. The researchers addressed this by applying SHAP, or Shapley additive explanations, a technique borrowed from cooperative game theory that assigns each input feature a precise contribution to every individual prediction. In effect, SHAP reveals which factors pushed a particular patient’s predicted probability up or down, and by aggregating these contributions across the cohort, the team identified the features with the greatest impact on outcomes in both the Shape and Wall models. This transparency allows a cardiologist to see, for example, whether an unfavorable prediction stems from pronounced atrial remodeling, abnormal wall texture suggesting fibrosis, or a clinical factor such as persistent rather than paroxysmal fibrillation.

The clinical stakes of this kind of tool are considerable. Patients with atrial fibrillation and concomitant heart failure represent a particularly vulnerable group, and ablation in these individuals carries procedural risk, substantial cost, and a demanding recovery. If a model can reliably flag patients unlikely to experience functional improvement, clinicians could counsel them more honestly, weigh alternative management strategies such as optimized medical therapy, or reserve ablation for those with the greatest expected benefit. Conversely, identifying patients with a high predicted probability of improvement could support earlier referral and shared decision-making grounded in quantitative evidence rather than clinical intuition alone. Because the model’s inputs come from CT angiography and routine laboratory work, it could in principle be deployed without any additional testing beyond what pre-ablation workups already include.

The multicenter architecture of the study strengthens its claims in another important way. Models trained and tested within a single institution often flatter themselves, absorbing center-specific artifacts in scanner protocol, image reconstruction, and patient mix. By including 240 patients in the training set, 101 in the validation set, and 75 in external testing sets from two other organizations, the team demonstrated that the model’s discrimination held up across different hospitals and imaging environments. The ethics committees of all three participating institutions approved the retrospective study, and the requirement for individual informed consent was waived because of its retrospective design, with the work conducted in accordance with the Declaration of Helsinki.

The authors themselves are careful about the limits of what they have shown. The study was retrospective, meaning it looked backward at patients who had already undergone ablation rather than prospectively assigning the model to guide care in real time. The team explicitly states that further prospective evaluation is required before the model can be implemented clinically, a caveat that applies to nearly every prediction algorithm now emerging in cardiovascular medicine. Prospective validation would involve using the COMB model to generate predictions before ablation and then tracking whether those predictions match observed outcomes, ideally across diverse populations and healthcare systems. Questions about calibration, the agreement between predicted probabilities and actual event rates, and about how the model behaves in subgroups underrepresented in the data, would also need answers.

Even with those caveats, the study offers a compelling glimpse of where cardiac imaging and artificial intelligence are converging. The same CT scan obtained to map a patient’s pulmonary veins becomes, through shape analysis and radiomics, a quantitative portrait of atrial health, and machine learning converts that portrait, together with a few clinical facts, into an individualized forecast of recovery. The SHAP framework keeps the forecast auditable, showing clinicians exactly which anatomical and biological signals drove the conclusion. For the millions of patients whose fibrillation and heart failure travel together, and for the physicians deciding whether to offer them ablation, a tool that turns pre-procedural images into evidence-based expectations could reshape the conversation, replacing uncertainty with a number that both doctor and patient can understand, question, and trust.

Subject of Research: Explainable machine learning on cardiac CT angiography to predict functional improvement after atrial fibrillation ablation in patients with heart failure

Article Title: Explainable machine learning model based on cardiac CT angiography for predicting functional improvement after atrial fibrillation ablation: a multicenter study

Article References: Jing, M., Lu, H., Liu, Q., Jing, Y., Lei, F., Yang, X., Chen, G., Xi, H., Xin, W., Zhu, H., Sun, Q., Zhang, Y., Ren, J., Ren, W., Liu, Z., Wang, G., & Zhou, J. (2026). Explainable machine learning model based on cardiac CT angiography for predicting functional improvement after atrial fibrillation ablation: a multicenter study. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02822-1

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02822-1

Keywords: atrial fibrillation, heart failure, catheter ablation, cardiac CT angiography, machine learning, radiomics, left atrium, SHAP explainability, predictive modeling, multicenter study, BMC Medical Imaging, cardiac imaging

The study, published in the journal Smart Agricultural Technology, addresses a persistent gap in the literature on honey yield prediction. Previous efforts have ranged from simple regression equations linking climate to harvests in the United Kingdom two decades ago, to fuzzy inference systems, radial basis function interpolation, and modern algorithms such as random forests and gradient boosting applied in Australia, Spain, Italy, and Turkey. Yet the authors found that relatively few such studies are indexed in major databases like Scopus and Web of Science, and many of them keep their datasets closed, making it impossible for other researchers to verify results, compare models fairly, or build on the findings. The new work is explicitly designed to break that pattern.

At the heart of the study is a dataset of 49 records of average honey yield per hive, drawn from 15 apiaries in southwest Australia between 2011 and 2018. The main nectar source in that region is the marri tree, Corymbia calophylla, which flowers around February. Because reliable yield records are scarce and often confidential, the researchers deliberately kept the predictor set simple: minimum and maximum temperatures and rainfall, downloaded from weather stations of the Australian Government’s Bureau of Meteorology located near each apiary. From these they engineered 45 climatic features, including monthly mean temperatures, monthly rainfall, counts of days above 40 degrees Celsius and below 25 degrees Celsius, each indexed by how many months before flowering the measurement was taken.

The team formulated the problem as a supervised multiclass classification task, mirroring the methodology of the seminal 2020 study on marri honey harvests so that results could be compared directly. Yields below 20 kilograms per hive were labeled a poor year, between 20 and 40 kilograms a moderate year, and above 40 kilograms a good year. Exploratory analysis showed that the climatic variables overlap heavily across the three classes, which rules out simple threshold rules and statistical models, and instead points toward machine learning methods capable of capturing complex, nonlinear relationships.

A central innovation of the paper is its layered approach to interpretability. The researchers applied three complementary techniques to understand which variables drive the predictions. First, they computed feature importance in a random forest model based on the mean decrease in impurity, which revealed that rainfall variables dominate: rainfall five, one, and eight months before flowering emerged as the three most influential features, followed by minimum temperatures three and eleven months before flowering. Second, they quantified how frequently each variable appears at different depths of the decision trees inside the random forest, on the premise that the most relevant features sit closer to the root. This analysis largely corroborated the importance rankings, with rainfall eight and five months before flowering again standing out, while the count of extremely hot days above 40 degrees Celsius had almost no influence.

Third, the team used SHAP, or Shapley Additive Explanations, a method rooted in cooperative game theory that distributes the prediction among the input features according to their average marginal contribution. The SHAP analysis added class-specific nuance. For poor harvests, the minimum temperature eleven months before flowering was among the strongest explanatory variables, with low values pushing predictions away from that class. For moderate harvests, rainfall seven months before flowering carried the greatest weight, with high values contributing positively to that prediction. For good harvests, high values of rainfall eight and five months before flowering were the clearest signals. Taken together, the interpretability analyses consistently point to rainfall in the months leading up to flowering as the most reliable indicator of a strong honey season.

On the modeling side, the researchers benchmarked nine widely used algorithms: logistic regression, k-nearest neighbors, support vector machines, decision trees, multilayer perceptrons, random forests, linear discriminant analysis, gradient boosting, and Naive Bayes. They evaluated each under a battery of experimental conditions, including with and without feature selection, with four normalization schemes (no normalization, division by the maximum, min-max scaling, and z-score standardization), and under two cross-validation strategies, leave-one-out and stratified 10-fold. The effects were algorithm-dependent. Distance-based and margin-based methods such as k-nearest neighbors, support vector machines, logistic regression, and multilayer perceptrons benefited from min-max and z-score scaling, while tree-based models like random forests and decision trees were indifferent to normalization, and Naive Bayes sometimes performed worse when the data were rescaled.

Feature selection based on the interpretability results, using the nine most important variables, improved several models, most dramatically linear discriminant analysis, which jumped from an accuracy of 0.65 to 0.80 under both validation schemes. A k-nearest neighbors model with k set to three also reached roughly 0.80 accuracy. To squeeze out further gains on the small dataset, the team applied SMOTE, a synthetic minority over-sampling technique that balances the classes during training, along with bagging and a Voting Classifier that combines the probability outputs of multiple base models. The best configuration paired k-nearest neighbors with a support vector machine under soft voting, achieving an accuracy of 0.82 and a Matthews correlation coefficient above 0.71. That figure represents a substantial improvement over the 0.67 accuracy reported in the foundational Australian study, achieved here without any satellite or remote sensing data at all.

The authors are careful about what these numbers mean. With only 49 samples and no independent external test set, the reported performance reflects cross-validation estimates rather than proven generalization to new regions or production systems. Hyperparameter searches proved unstable on such a small dataset, so the team deliberately used default settings to keep the comparison controlled and reproducible. They frame the contribution not as a single performance record but as a replicable workflow: the raw climatic data, the yield database, the preprocessing scripts, the feature selection procedure, and the model evaluations are all openly available on GitHub, so that any research group can rerun the pipeline, adapt it to local flowering calendars, and retrain the models on regional data.

The practical implications reach beyond the laboratory. Because the model classifies seasons as low, moderate, or high yield potential using weather records that are essentially free, it could power a simple web or mobile tool in which a beekeeper selects an apiary location, the system links it to the nearest weather station, and the season’s outlook appears as an early alert. Such a signal could guide decisions on feeding schedules, which follow Farrar’s rule that a hive’s honey production capacity grows exponentially with bee population, as well as the preparation of honey supers, transhumance planning, labor allocation, and harvest logistics. Cooperatives, technical advisors, and public institutions could use territory-scale forecasts to plan assistance programs and climate adaptation strategies. The authors stress that the tool is meant to support, not replace, the beekeeper’s expert judgment, and that validation on new seasons and regions, along with the eventual addition of satellite vegetation data and hive sensor information, remains the necessary next step before the approach can be deployed at scale.

Subject of Research: Machine learning classification of honey yield per hive using climatic variables

Article Title: Machine learning models combined with feature importance methods for honey yield classification: A replicable approach

Article References: Ahumada-García, R., Zabala-Blanco, D., Monzón, V. H., Sánchez, I., da Silva, N. F. F., Rosa, T. C., Ferreira, A. I. S., López-Cortés, X., Flores-Calero, M., & Vasquez-Iglesias, P. (2026). Machine learning models combined with feature importance methods for honey yield classification: A replicable approach. Smart Agricultural Technology, 15, Article 102575. https://doi.org/10.1016/j.atech.2026.102575

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102575

Keywords: machine learning, honey yield prediction, beekeeping, random forest, SHAP explainability, feature importance, climate variables, rainfall, SMOTE, cross-validation, precision apiculture, reproducible research

The research, conducted by Furquan Ahmad of the National Institute of Technology Patna in India and Warit Wipulanusat of Thammasat University in Thailand, addresses a long-standing challenge in geotechnical engineering. Predicting the factor of safety against liquefaction requires untangling complex nonlinear interactions among soil properties, seismic loading conditions, and in-situ measurements. Traditional simplified procedures, developed over decades from case histories of past earthquakes, rely on empirical charts and hand calculations that can miss subtle parameter interactions. Machine learning offers a way forward, but conventional neural networks are often criticized as opaque black boxes, and their training can be slow and unstable. The new framework tackles both problems at once, pairing a fast-learning network architecture with a transparent layer of explainability analysis.

At the heart of the framework is the Extreme Learning Machine, or ELM, a type of single-hidden-layer feedforward neural network introduced in the mid-2000s. Unlike conventional neural networks that iteratively adjust all their internal weights through backpropagation, an ELM assigns its hidden-layer weights randomly and then solves for the output weights analytically in a single step. This makes training dramatically faster while often achieving excellent generalization. However, the random initialization of hidden-layer parameters can leave performance on the table. The researchers solved this by coupling the ELM with four metaheuristic optimization algorithms, each inspired by different natural processes: the Adaptive Opposition Slime Mould Algorithm, Harris Hawks Optimization, Particle Swarm Optimization, and the Reptile Search Algorithm. These algorithms search the space of possible network parameters to find configurations that minimize prediction error, effectively tuning the network far beyond what random initialization alone could achieve.

Each optimizer brings a distinct search philosophy. Particle Swarm Optimization, one of the oldest and most widely used metaheuristics, mimics the social behavior of bird flocks, with candidate solutions sharing information about promising regions of the search space. Harris Hawks Optimization models the cooperative hunting strategies of hawks, dynamically switching between exploration and exploitation phases to chase down optimal solutions. The Reptile Search Algorithm draws on the hunting and social behaviors of crocodiles, while the Adaptive Opposition Slime Mould Algorithm simulates the oscillating foraging patterns of slime moulds, enhanced with an opposition-based learning mechanism that evaluates candidate solutions and their mirrored counterparts to accelerate convergence. By benchmarking all four against the same ELM backbone and the same dataset, the study provides a fair head-to-head comparison of how each optimizer shapes predictive performance.

Before any model training began, the researchers applied a disciplined feature screening process grounded in information theory. Using a database of 309 case histories based on cone penetration test, or CPT, measurements, they employed Mutual Information Analysis and Shannon Entropy Analysis to quantify how much information each candidate input variable carries about the liquefaction outcome. Mutual information measures the reduction in uncertainty about one variable given knowledge of another, while Shannon entropy, a concept borrowed from communication theory, quantifies the inherent uncertainty or information content of a variable. This pre-processing step ensures that the models focus on genuinely informative features rather than redundant or noisy ones, a critical consideration when working with real-world geotechnical data that often contains correlated measurements.

The performance results are striking. The best model, ELM-AOSMA, achieved coefficients of determination of 0.9912 on the training dataset and 0.9900 on the testing dataset, with root mean square errors of just 0.0146 and 0.0160 respectively. In practical terms, an R-squared value approaching 0.99 means the model explains nearly all of the variance in the liquefaction factor of safety across the case database, leaving only a tiny fraction unexplained. The models were validated through 5-fold cross-validation, a robust technique in which the data is partitioned into five subsets and the model is repeatedly trained and tested on different combinations to guard against overfitting. Beyond the headline metrics, the evaluation battery included the Nash-Sutcliffe coefficient, weighted mean absolute percentage error, and the A20-index, which measures the proportion of predictions falling within twenty percent of observed values.

The study also went beyond standard statistical scores with a suite of diagnostic analyses. Regression Error Characteristic curves were used to visualize the trade-off between prediction tolerance and the fraction of samples accurately predicted, while rank analysis and curve fitting analysis compared the relative standing of the four hybrid models. Convergence analysis tracked how quickly each optimizer approached its best solution, and computational efficiency was measured directly. ELM-AOSMA converged within approximately 400 iterations while requiring the lowest computational time of the four models, at 220.68 seconds. For engineers who may need to run liquefaction assessments across many sites or within probabilistic reliability frameworks that demand thousands of repeated evaluations, this combination of accuracy and speed is a meaningful advantage over slower, more computationally demanding alternatives.

Perhaps the most consequential contribution is the framework’s commitment to explainability. Using SHapley Additive exPlanations, a technique from cooperative game theory that allocates credit for each prediction to individual input features, and Partial Dependence with Individual Conditional Expectation analysis, which traces how predictions change as each variable varies, the researchers identified the dominant drivers of liquefaction behavior. Four parameters emerged as the most influential: the effective vertical stress, denoted sigma prime v; the normalized and cleaned cone penetration tip resistance, qc1Ncs; the total vertical stress, sigma v; and the magnitude scaling factor, MSF, which adjusts cyclic stress ratios for earthquakes of different magnitudes. This hierarchy aligns with established geotechnical understanding, since liquefaction resistance fundamentally depends on how tightly soil particles are confined and how densely they are packed, but the explainable AI analysis quantifies these relationships in a way empirical charts cannot.

The implications for practice extend to performance-based geotechnical design and reliability-oriented assessment. Modern seismic design codes increasingly demand not just a binary safe-or-unsafe verdict but a quantified factor of safety that can feed into probabilistic risk calculations. A model that is both highly accurate and computationally lightweight can serve as a surrogate within Monte Carlo simulations or first-order reliability methods, where the simplified empirical procedures would be evaluated tens of thousands of times. The interpretability layer also matters for professional acceptance: engineers are understandably reluctant to trust black-box predictions for safety-critical infrastructure, and SHAP-based attribution gives them a window into the reasoning behind each assessment, enabling sanity checks against engineering judgment.

The work arrives amid a broader wave of machine learning adoption in geotechnical engineering, with researchers applying hybrid models to slope stability, sheet pile walls, embankment performance in seismic zones, and liquefaction potential mapping in regions from the Indo-Gangetic plain to Kolkata. What distinguishes this study is its end-to-end rigor: information-theoretic feature selection, systematic optimizer benchmarking, cross-validated performance evaluation, and explainable AI diagnostics all integrated within a single framework. As earthquake-prone cities continue to grow and climate-driven changes alter groundwater conditions that govern soil saturation, tools that can rapidly and transparently assess liquefaction risk will become increasingly valuable. The ELM-AOSMA framework demonstrates that the next generation of geotechnical prediction tools need not force a choice between accuracy, speed, and interpretability; with careful design, they can deliver all three at once.

Subject of Research: Hybrid machine learning prediction of seismic soil liquefaction using cone penetration test data

Article Title: An information-theoretic and explainable hybrid extreme learning machine framework for seismic soil liquefaction assessment

Article References: Ahmad, F., & Wipulanusat, W. (2026). An information-theoretic and explainable hybrid extreme learning machine framework for seismic soil liquefaction assessment. Earth Science Informatics, 19(11), Article 195. https://doi.org/10.1007/s12145-026-02248-x

Image Credits: AI Generated

DOI: 10.1007/s12145-026-02248-x

Keywords: soil liquefaction, earthquake engineering, extreme learning machine, metaheuristic optimization, SHAP explainability, cone penetration test, mutual information, Shannon entropy, geotechnical engineering, machine learning, factor of safety, seismic hazard

The research team, led by Miad Islam of Saint Leo University together with collaborators from institutions in Bangladesh, the United Kingdom, and the United States, built a framework that combines three complementary neural network architectures into a single sepsis prediction engine. The first component is a one-dimensional convolutional neural network, or 1D-CNN, which excels at scanning sequences of clinical measurements and picking out local patterns, much as a radiologist might scan an image for telling features. Layered on top of that is a bidirectional long short-term memory network, or BiLSTM, a recurrent architecture that reads the patient’s clinical timeline in both directions, capturing how early vital-sign changes foreshadow later laboratory abnormalities and how recent developments reframe earlier ambiguity. The final piece is a Transformer attention block, the same family of architecture that powers modern language models, which allows the network to weigh the relative importance of different clinical variables and time points dynamically rather than treating every input as equally significant. Together, these components form a model capable of learning the subtle, multi-scale signatures that precede sepsis onset.

The privacy mechanism at the heart of the study is federated learning, a training paradigm in which the data never moves. Instead of shipping records to a central server, each participating hospital trains the model locally on its own patients. Only the resulting model parameters, essentially long lists of numerical weights, are transmitted to a coordinating server, which averages them using an algorithm known as Federated Averaging, or FedAvg. The updated global model is then sent back to each site for another round of local training. Over many such rounds, the shared model gradually absorbs the statistical diversity of all participating institutions without any single record, diagnosis, or lab value ever being exchanged. For the purposes of this study, the researchers simulated this arrangement by partitioning a MIMIC-IV-style ICU dataset across three virtual hospital clients, allowing them to evaluate how the federated approach behaves under realistic distributed conditions while working with data that posed no privacy risk.

The results are striking for a model trained under such constraints. The federated hybrid framework achieved an accuracy of 93.6 percent and an area under the receiver operating characteristic curve, or AUROC, of 0.959, a standard measure of a classifier’s ability to distinguish septic from non-septic patients across all possible thresholds. Precision came in at 0.871, meaning that when the model raised an alarm it was usually justified, and specificity reached 98.24 percent, indicating it rarely flagged healthy patients falsely. The F1-score, which balances precision against recall, was 0.759, a figure that reflects the inherent difficulty of detecting a condition that is rare relative to the ICU population. For comparison, the researchers also trained a centralized baseline model on all the data pooled together, and it reached an AUROC of 0.997. That gap is real, but the federated model’s performance remains firmly in the range clinicians would consider useful, and it was achieved without the data centralization that privacy law makes impractical.

Perhaps even more consequential for real-world deployment is the framework’s communication efficiency. In federated systems, the cost of repeatedly shipping model weights between hospitals and the coordinating server can become prohibitive, particularly for institutions with limited bandwidth. Across all federated training rounds in this study, the total communication cost was just 36.09 megabytes, a figure small enough to travel over ordinary hospital networks in seconds. That efficiency matters because it suggests the approach could scale to many more participating sites without the coordination overhead becoming a bottleneck. In distributed healthcare environments, where connectivity is often uneven and IT infrastructure varies widely between a major academic medical center and a community hospital, keeping the communication footprint light is not a luxury but a prerequisite.

A model that clinicians cannot understand is a model they will not trust, so the researchers turned to SHAP, or SHapley Additive exPlanations, a technique borrowed from cooperative game theory that assigns each input variable a quantified contribution to every individual prediction. The SHAP analysis identified the Sequential Organ Failure Assessment score, a composite measure of dysfunction across six organ systems, as the most influential predictor of sepsis risk, followed by lactate level, white blood cell count, creatinine, and procalcitonin. Each of these is a familiar marker to any intensivist: lactate rises when tissues are starved of oxygen, white blood cells surge or crash during systemic infection, creatinine signals kidney injury, and procalcitonin is a well-established biomarker of bacterial sepsis. The fact that the model’s attention converged on variables that already carry clinical weight is reassuring, because it suggests the network learned genuine physiology rather than exploiting statistical artifacts in the data.

The authors are candid about the limits of their work, and their honesty is itself instructive. The study used a MIMIC-IV-style dataset rather than live records from multiple hospitals, and the federated setup was a simulation rather than a deployment across real institutions with genuinely heterogeneous patient populations, equipment, and documentation practices. The researchers also note that they did not formally define an adversary or threat model. They did not assess whether a malicious server or a colluding client could infer patient-level information from the exchanged model updates, a known vulnerability in federated systems where gradient information can sometimes leak details about training data. Techniques such as differential privacy, secure aggregation, and homomorphic encryption exist to close these gaps, and the authors flag rigorous validation on real multi-institutional ICU records as a necessary next step before any claim of scalable, secure deployment can be made.

Those caveats notwithstanding, the study lands at a moment when the tension between data hunger and data privacy has become the defining challenge of clinical artificial intelligence. The most accurate models are typically those trained on the largest and most diverse datasets, yet the most diverse datasets are precisely the ones that privacy law keeps fragmented across institutions. Federated learning offers a mathematically principled compromise, and pairing it with architectures that capture both local temporal patterns and long-range dependencies, as this hybrid CNN-BiLSTM-Transformer design does, may prove to be a template for predictive medicine well beyond sepsis. Acute kidney injury, respiratory failure, and cardiac deterioration are all conditions where early, distributed, privacy-preserving prediction could change outcomes.

What makes this work compelling as a piece of the larger puzzle is its demonstration that privacy and performance need not be framed as a zero-sum trade. A model trained without ever seeing another hospital’s records came within a few percentage points of the centralized ideal, and it did so with a communication budget measured in tens of megabytes and with explanations that map onto established clinical intuition. The road from a three-client simulation to a network of hospitals sharing a living model is long, and it will require threat modeling, regulatory negotiation, and prospective clinical validation. But the direction of travel is clear. If the next generation of ICU decision support can be trained collectively while respecting the sovereignty of every patient record, the hours that matter most in sepsis may finally be spent on treatment rather than detection.

Subject of Research: Privacy-preserving federated deep learning for sepsis prediction in intensive care units

Article Title: A hybrid deep learning framework for privacy-preserving sepsis prediction in distributed ICU environments using federated learning simulation

Article References: Islam, M., Mohiuddin, T., Rahman, M. A., Sharfuddin, M., Islam, M. S., Sunny, S. R., & Lokesh, E. (2026). A hybrid deep learning framework for privacy-preserving sepsis prediction in distributed ICU environments using federated learning simulation. Discover Social Science and Health. https://doi.org/10.1007/s44155-026-00489-1

Image Credits: AI Generated

DOI: 10.1007/s44155-026-00489-1

Keywords: federated learning, sepsis prediction, deep learning, ICU, electronic health records, privacy-preserving AI, Transformer, BiLSTM, SHAP explainability, clinical decision support, health informatics, FedAvg

The scale of the underlying problem is enormous. Nearly sixty million people worldwide live with dementia, a syndrome most commonly caused by Alzheimer’s disease and marked by progressive deterioration of memory, speech and judgment, and that figure is expected to more than double by 2050. Because irreversible brain damage typically occurs long before a clinical diagnosis is made, the most effective defence against the disease is early detection paired with risk-reducing lifestyle interventions. Yet conventional diagnostic approaches, which combine psychometric testing, imaging and lifestyle data, are time- and resource-intensive, often requiring multiple specialist referrals and producing delays that were further exacerbated by the COVID-19 pandemic. Simplified toolkits such as the CAIDE and BDSI risk scores were designed to help, but they operate over fixed, pre-specified horizons, target narrow age groups, and in some analyses add little discriminative value beyond knowing a patient’s age alone.

Previous machine learning models trained on large clinical datasets promised improvements but were typically locked into a single fixed prediction window, such as twenty-nine months or one year. The research team behind the new study hypothesised that this rigidity is itself a core limitation: defining one prediction horizon discards valuable data and can introduce sampling bias when right-censored data, meaning patients whose follow-up ended before the horizon, is handled poorly. Their solution was elegant. They introduced a feature called HORIZON, which encodes the number of months between a patient’s baseline visit and each subsequent visit, and set it to zero for immediate diagnosis. Rather than building separate models for each time frame, the algorithm treats the prediction window as just another input, learning how the relationship between risk factors and dementia changes, or does not change, as the horizon stretches out. The approach is conceptually related to a statistical technique known as survival stacking, which recasts time-to-event analysis as a classification problem, but here the researchers exploited the naturally rich multi-visit structure of their dataset instead of generating synthetic samples.

The data came from the Uniform Data Set of the National Alzheimer’s Coordinating Center, a collaboration spanning more than 42 American research centres, covering 47,400 patients between June 2005 and February 2023. After rigorous preprocessing, including the removal of variables with more than seventy percent missing values and the elimination of highly correlated features, the team arrived at 134,117 visit records from 47,017 unique patients. Critically, variable selection was clinically informed from the outset: a consultant psychiatrist working in an NHS memory assessment service reviewed candidate variables against two criteria, routine availability in secondary care and minimal reliance on subjective expert judgment, so that the final inputs could plausibly be collected by junior doctors or medical assistants. Recursive feature elimination, embedded within a nested cross-validation pipeline, then narrowed the field to the twenty most informative variables.

Six machine learning algorithms competed in the study, and the K-Nearest Neighbour classifier emerged as the strongest. Its logic is intuitively clinical: to classify a new patient, the model compares their profile against the hundred most similar patients with known dementia status and weighs each equally, effectively mimicking a clinician reasoning from a wealth of prior comparable cases. For same-visit diagnosis, the model achieved a sensitivity of 0.951 and a specificity of 0.810, with an area under the receiver operating characteristic curve of 0.959. Across future horizons from 24 to 120 months, sensitivity remained high, ranging between 0.808 at two years and 0.833 at ten years, while specificity declined from 0.856 to 0.706 as the horizon lengthened, an expected trade-off in risk prediction. Class imbalance was addressed with SMOTE-NC oversampling applied within each horizon group, and all preprocessing was confined to training folds to prevent information leakage.

The flexible-horizon model was also benchmarked against a Cox proportional hazards survival model, the conventional statistical approach for time-to-event outcomes. The comparison revealed complementary strengths: the KNN model achieved a higher geometric mean of sensitivity and specificity at four of the five evaluated horizons and substantially greater specificity from 48 months onwards, whereas the Cox model delivered greater sensitivity at longer horizons. Importantly, the machine learning approach offers practical advantages beyond raw performance, supporting point-of-assessment diagnosis, something a survival model cannot do, and fitting into standard supervised learning pipelines without requiring assumptions such as proportional hazards or non-informative censoring.

Transparency was built into the analysis through Shapley additive explanations, a technique that assigns each feature a contribution value for every individual prediction. Globally, the two most influential features were functional independence and the prediction horizon itself, followed by cognitive measures including the animal naming test, the Mini-Mental State Examination and the geriatric depression scale. The explanations aligned with established clinical understanding: lower cognitive test scores, loss of functional independence, poorer recall of significant dates and difficulty assembling tax records all pushed the model’s risk estimate upward. Local explanations, presented as waterfall plots, showed how individual features shifted the risk score for specific patients, the kind of case-by-case interpretability that clinicians increasingly demand before trusting an algorithm. Notably, age ranked lower than expected, likely because its effect is partially absorbed by the horizon feature.

Fairness received equally serious attention. The team evaluated performance across subgroups defined by sex, race and age, computing equal opportunity, predictive equality and a novel G-mean equality ratio. Balanced discrimination proved broadly similar across most groups, but the trade-offs differed: males showed higher sensitivity but lower specificity than females, Asian and Other race subgroups showed high sensitivity with lower specificity, and the small 18-44 age group showed lower sensitivity but very high specificity. The authors are careful to frame these results as an initial internal fairness assessment rather than proof of equity, noting that some subgroup estimates rested on small sample counts, and they caution that false negatives and false positives carry different clinical consequences, from delayed assessment to unnecessary anxiety and referral burden.

The researchers have deployed the model as an interactive web application built with Streamlit, intended strictly for demonstration, clinician feedback and future validation rather than autonomous diagnosis. The app runs password-protected on a small cloud instance, retains no user data after each session, and has not yet undergone formal cybersecurity assessment, clinical safety review or regulatory evaluation. They estimate that collecting the model’s twenty inputs, many of which are brief: the MMSE takes roughly five to ten minutes, the short-form geriatric depression scale five to seven, and the animal naming task just sixty seconds, could be completed in around thirty minutes within a suitably prepared secondary-care workflow, compared with a conventional memory clinic appointment of an hour and a half or more. Because the feature set includes modifiable risk factors such as blood pressure, body mass index and depression measures, clinicians could in principle use local explanations to simulate how changing specific factors might alter a patient’s projected risk.

The team is candid that the model is not clinically ready. All training and validation data originate from the American NACC cohort, which may not fully represent demographics, referral pathways or assessment procedures in UK secondary care or other health systems. The raw KNN outputs are not fully calibrated as absolute probabilities, particularly at shorter horizons, and should be interpreted as model-estimated risk scores. The model also treats all-cause dementia as a single label, collapsing distinctions between Alzheimer’s, vascular and other subtypes. The authors’ staged pathway forward is methodical: retrospective evaluation on historical data from a real NHS memory clinic, silent prospective testing in which outputs are generated but not acted upon, clinician usability assessment, workflow impact analysis, and finally governance, cybersecurity and regulatory review. If those hurdles are cleared, a flexible-horizon, explainable and fairness-tested algorithm could become a genuine decision-support companion, one that complements rather than replaces clinical judgment, concentrates specialist time on the hardest cases, and helps bring dementia detection closer to the moment when intervention still matters most.

Subject of Research: A flexible-horizon machine learning model for dementia early detection and risk prediction in secondary care using responsible AI principles

Article Title: A Flexible-Horizon Clinical Decision Support Model for Dementia Early Detection and Risk Prediction in Secondary Care: A Responsible AI Approach

Article References: Ez-zizi, A., Seelam, K., Leggett, L., & Malik, B. R. (2026). A Flexible-Horizon Clinical Decision Support Model for Dementia Early Detection and Risk Prediction in Secondary Care: A Responsible AI Approach. Journal of Medical Systems, 50(1), Article 132. https://doi.org/10.1007/s10916-026-02458-2

Image Credits: AI Generated

DOI: 10.1007/s10916-026-02458-2

Keywords: dementia, machine learning, responsible AI, clinical decision support, risk prediction, NACC Uniform Data Set, K-nearest neighbours, SHAP explainability, algorithmic fairness, secondary care, survival analysis, early detection

A new study published in Annals of Gastroenterological Surgery asks whether machine learning could tip that balance. Researchers from the University Hospital of Split in Croatia analyzed the records of 1,547 adults operated on for suspected appendicitis between January 2020 and June 2024, ultimately assembling a rigorously curated dataset of 623 patients. Of these, 66 had a histologically normal appendix, 248 had uncomplicated appendicitis, and 309 had complicated appendicitis confirmed by pathologists. Crucially, every case was verified by histopathology, eliminating the diagnostic guesswork that plagues many prior studies in which patients were classified by clinical or radiological impression alone.

The team built and compared four machine learning models: logistic regression, random forest, balanced random forest, and XGBoost, an extreme gradient boosting algorithm. All models relied exclusively on routinely available clinical and laboratory data—age, sex, symptom duration, fever, nausea, vomiting, rebound tenderness, pain migration, white blood cell count, neutrophil and lymphocyte percentages, platelet indices, C-reactive protein, and serum sodium. No imaging data were included. Missing values, which ranged from zero to 5.3 percent across variables, were imputed with a bagged trees algorithm carefully configured to prevent data leakage between training and test sets.

The methodological centerpiece of the study was a nested cross-validation design, with five inner and outer folds repeated ten times, yielding fifty independent test evaluations. Hyperparameters were tuned exclusively on inner validation data, and performance was then measured on untouched outer test folds, keeping performance estimates unbiased. On top of that, the researchers applied bootstrap resampling with 2,000 iterations to confirm that results were stable and not artifacts of a particular split of the data.

The most distinctive innovation, however, was threshold optimization. Rather than accepting a default probability cutoff of 0.5, the team deliberately shifted decision thresholds to meet six predefined minimum sensitivity targets, ranging from 0.95 to 0.995. Within each outer fold, thresholds were derived from pooled inner validation predictions, fixed, and then applied unchanged to the test set. The logic is clinical, not statistical: in appendicitis, missing a true case is far more dangerous than operating on a healthy appendix, so a decision support tool must be calibrated to catch virtually every case of appendicitis while identifying the small minority of patients who might safely avoid immediate surgery.

The results were revealing. For detecting acute appendicitis, logistic regression outperformed the more complex algorithms, achieving an area under the ROC curve of 0.765. At the strictest sensitivity target of 0.995, the model detected 99.8 percent of appendicitis cases, with specificity of just 0.038. Translated to a clinical scale, using the observed case distribution, this operating point would miss approximately two appendicitis cases per 1,000 patients while correctly flagging about four patients without appendicitis who could be spared surgery. At the more lenient sensitivity target of 0.95, roughly 41 appendicitis cases per 1,000 would be missed, but approximately 24 patients per 1,000 without appendicitis would avoid an unnecessary operation. The numbers lay bare the inherent trade-off between diagnostic safety and operative selectivity.

For distinguishing complicated from uncomplicated disease, the random forest model performed best, with an AUC of 0.785 and a maximum achievable sensitivity of 0.997 paired with specificity of 0.057. This capability carries real clinical weight: complicated appendicitis generally demands urgent surgery and carries higher morbidity and mortality, whereas some centers now attempt conservative, non-operative management of uncomplicated cases. A model that reliably identifies complicated disease could support more urgent operative decisions, temper enthusiasm for non-operative strategies, or prompt additional imaging and closer monitoring when the predicted risk is ambiguous.

To address the notorious black-box problem that undermines clinical trust in machine learning, the researchers applied SHAP analysis, a framework from explainable artificial intelligence that quantifies each feature’s contribution to individual predictions. For appendicitis detection, neutrophil percentage, the neutrophil-to-lymphocyte ratio, the platelet-to-lymphocyte ratio, sex, and lymphocyte percentage emerged as the most influential predictors. For complication prediction, C-reactive protein, age, symptom duration, and the neutrophil-to-lymphocyte ratio dominated. These rankings align closely with the pathophysiology of appendicitis, in which bacterial invasion of the appendiceal mucosa triggers a systemic inflammatory cascade, and they mirror the feature importances reported by other research groups using different algorithms and populations.

The authors are careful to frame what these models are and are not. The cohort consisted exclusively of adults already selected for surgery by a surgeon’s judgment, so the models should not be treated as stand-alone diagnostic tools for undifferentiated abdominal pain in the emergency department, nor as replacements for ultrasound or CT where those are routinely available. Instead, the researchers envision them as adjunctive decision-support and reassessment tools: a low-risk prediction should trigger repeat examination, laboratory reassessment, or additional imaging rather than discharge, particularly in settings with limited imaging access or after equivocal imaging findings. Notably, the superiority of the simple linear model over tree-based methods in this adult cohort contrasts with findings in pediatric populations, suggesting that inflammatory markers relate to appendicitis in a more linear fashion in adults and underscoring the need for population-specific models.

Limitations remain. The study was retrospective and single-center, and its performance may reflect local diagnostic pathways, surgeon thresholds, and imaging practices that do not transfer to other institutions. The dataset is also surgically enriched, meaning there are no true negative patients who avoided surgery, so specificity must be interpreted conditionally within the operated population. Still, the combination of histopathological verification, rigorous nested cross-validation, bootstrap stability testing, deliberate sensitivity-first threshold calibration, and transparent SHAP-based interpretability marks this work as a methodologically serious step toward machine learning that could meaningfully reduce negative appendectomies. The authors call for prospective, multicenter external validation before clinical deployment, along with future exploration of multimodal models integrating imaging and of tools to help select patients with uncomplicated appendicitis for non-operative management—a direction that could ultimately reshape how one of surgery’s most routine emergencies is decided.

Subject of Research: Machine learning with sensitivity-focused threshold optimization to reduce negative appendectomies in adults with suspected acute appendicitis

Article Title: Can Machine Learning Reduce Unnecessary Surgeries? A Retrospective Analysis Using Threshold Optimization to Prevent Negative Appendectomies in Adults

Article References: Males, I., Kumric, M., Boban, Z., Vrdoljak, J., Pecenkovic, D., Ivanda, M., Grahovac, M., Pogorelic, Z., & Bozic, J. (2026). Can Machine Learning Reduce Unnecessary Surgeries? A Retrospective Analysis Using Threshold Optimization to Prevent Negative Appendectomies in Adults. Annals of Gastroenterological Surgery, 10(5), 1631-1645. https://doi.org/10.1002/ags3.70225

Image Credits: AI Generated

DOI: 10.1002/ags3.70225

Keywords: machine learning, appendicitis, negative appendectomy, logistic regression, random forest, XGBoost, threshold optimization, SHAP explainability, clinical decision support, C-reactive protein, nested cross-validation, emergency surgery