The study, published in BioMedical Engineering OnLine, was conducted by a team led by Yeongjun Cho of the Biomedical Engineering Research Center at the Asan Institute for Life Sciences, together with colleagues from Asan Medical Center and the University of Ulsan College of Medicine in Seoul. Rather than attempting a full clinical trial, the authors framed their work explicitly as an internal methodological evaluation: a controlled, sample-level test of whether a particular training strategy could improve the way a neural network represents neonatal movement data. That framing matters, because it sets honest boundaries around what the results do and do not demonstrate, a distinction that is often lost in the breathless coverage of medical artificial intelligence.

The data came from eleven preterm neonates assessed at three to five months of corrected age, the age the babies would have reached had they been born at term. From these recordings, the researchers constructed 1,957 multimodal samples. Each sample combined two complementary views of the same movement episode. The first was a motion history image, or MHI, a computer-vision representation that collapses a short video sequence into a single image in which recent motion appears brighter and older motion fades, effectively freezing the spatial pattern of a baby’s movement into a form a vision model can digest. The second was a stream of inertial measurement unit, or IMU, signals from a wearable sensor, capturing the raw accelerations and rotations of the baby’s limbs that a camera might miss, especially when the infant is partially occluded or lying in an awkward position.

Every segment was labeled as fidgety movement or non-fidgety movement, and the labels, together with metadata about each segment, were organized in a unified manifest so that training, validation, and testing could be conducted in a controlled and reproducible way. The architecture at the heart of the framework was built on ImageBind, a multimodal embedding model designed to pull different sensory modalities into a shared representational space. In this shared space, a video-derived MHI feature and an IMU signal from the same movement episode should land close together, while samples from different movement classes should drift apart. Crucially, both training and final test-stage inference used only the MHI and IMU inputs, with the FM and nFM labels reserved for supervised loss optimization rather than being fed to the model as features.

The genuinely novel ingredient was the loss function, the mathematical objective that tells the network what it means to learn well. The team compared two established metric-learning losses and found each had complementary weaknesses. Triplet loss, a workhorse of metric learning, enforces local distance constraints: it pulls an anchor sample closer to samples of the same class than to samples of a different class by a defined margin. That local discipline is effective but myopic; it says nothing about the global geometry of the entire embedding space, and networks trained with it can develop distorted structures in which classes are separated in some regions and entangled in others. Tuned constellation loss, by contrast, shapes the global arrangement of embeddings but can leave local neighborhoods poorly constrained, allowing confusingly similar samples to sit near one another.

To resolve this tension, the researchers designed a combined loss that jointly enforces local distance constraints and global embedding structure. The intuition is architectural in the broadest sense: rather than choosing between a microscope and a telescope, the combined objective applies both simultaneously, demanding that individual samples obey local margins while the overall constellation of classes maintains a well-separated global layout. When the combined-loss model was evaluated on the predefined internal test partition, it outperformed both single-loss baselines, achieving an accuracy of 0.838, an F1-score of 0.828, an area under the receiver operating characteristic curve of 0.909, and an area under the precision-recall curve of 0.894. In a screening context, where the balance between catching true cases and avoiding false alarms is everything, the AUROC and AUPRC figures are particularly meaningful, since they summarize performance across all possible decision thresholds rather than at a single arbitrary cutoff.

Perhaps the most clinically relevant finding was the effect on class-wise balance. Screening tools fail in two distinct ways: they miss affected infants, or they flood clinics with false positives that erode trust and waste resources. Single-loss training in this dataset tended to sacrifice one class for the other, a failure mode that the combined loss appeared to mitigate. By shaping both the local and global geometry of the embedding space, the combined objective produced representations in which fidgety and non-fidgety samples were more evenly and reliably separated, suggesting that loss-function design, often treated as an engineering afterthought, can be a first-order determinant of clinical model behavior.

The authors are careful, and rightly so, about the limits of what this pilot demonstrates. The evaluation was sample-level and internal: the 1,957 test samples were drawn from the same eleven infants whose data trained the model, even if individual samples were held out. That means the reported metrics measure how well the framework learns to represent movement patterns it has, in some sense, already seen variants of, not how well it generalizes to a completely new baby. Neonatal movement data is notoriously subject-dependent; each infant has a unique movement signature shaped by gestational age, postnatal course, and temperament, and models can exploit these idiosyncrasies as shortcuts. The team explicitly states that their findings motivate further validation in larger subject-held-out cohorts, where the test infant is entirely unseen during training. Until such studies are done, the 0.909 AUROC should be read as a proof of concept rather than a clinical performance figure.

Even with those caveats, the study’s design choices point toward a practical future for the technology. The reliance on ordinary video plus a lightweight wearable sensor means the data acquisition could, in principle, happen in a clinic or even at home, without the expensive and stressful infrastructure of a dedicated motion-capture laboratory. The unified manifest approach, in which labels and segment metadata are organized alongside the raw modalities, reflects a growing recognition in biomedical machine learning that rigorous data governance is as important as model architecture. And the finding that combining losses improves robustness is a transferable lesson for the wider field of multimodal health AI, where researchers frequently must fuse signals as different as pixels and accelerations.

The work, approved by the Institutional Review Board of Asan Medical Center and registered with the Clinical Research Information Service, was funded by Korea’s Ministry of Trade, Industry and Energy, the National Research Foundation of Korea, Asan Medical Center, and the Korean Academy of Rehabilitation Medicine. Its significance lies less in any single number than in the demonstration that fidgety movement, a subtle qualitative phenomenon long accessible only to trained human eyes, can be encoded in a shared multimodal representation learned from cameras and inertial sensors. If subsequent subject-held-out studies replicate and extend these results, the quiet wriggling of a preterm infant, recorded for a few minutes on an ordinary camera and a small wearable, could become one of the earliest and most accessible windows into neurological risk, giving clinicians and families time, the resource that matters most in early intervention.

Subject of Research: Multimodal machine learning for automated classification of fidgety movements in preterm neonates as an early screen for cerebral palsy risk

Article Title: An internal methodological evaluation of a multimodal VIDEO-IMU metric-learning framework for fidgety movement classification in preterm neonates: a controlled sample-level study

Article References: An internal methodological evaluation of a multimodal VIDEO-IMU metric-learning framework for fidgety movement classification in preterm neonates: a controlled sample-level study. (n.d.). https://doi.org/10.1186/s12938-026-01638-y

Image Credits: AI Generated

DOI: 10.1186/s12938-026-01638-y

Keywords: fidgety movements, preterm neonates, cerebral palsy, machine learning, metric learning, multimodal learning, inertial measurement unit, motion history image, loss function, neonatology, biomedical engineering, early screening

The new model, described in the journal Medical & Biological Engineering & Computing, is called AMGFM, short for Adversarial–Metric Graph Fusion Model. It was developed by Jing Zhu, Aohan Zhang, Xiaowei Li, Yizhou Li, and Bin Hu of the Gansu Provincial Key Laboratory of Wearable Computing at Lanzhou University, with Hu also affiliated with the School of Medical Technology at the Beijing Institute of Technology. Their work addresses a problem that has dogged the field of physiological psychiatry for years: the signals that betray depression come from many different parts of the body at once, and each of those signals speaks a different statistical language.

Electroencephalography, or EEG, records the tiny electrical voltages produced by firing neurons through electrodes on the scalp. Eye-tracking devices capture where a person looks, how quickly their gaze shifts, and how long they fixate on particular images. Pupil measurements track the moment-to-moment dilation of the eye, which reflects activity of the autonomic nervous system. Each of these modalities carries information relevant to depression — altered neural oscillations, atypical attentional biases toward negative stimuli, and abnormal pupillary responses have all been documented in patients — but combining them computationally is far from trivial. The signals differ in sampling rate, dimensionality, noise characteristics, and statistical distribution, so a naive fusion often produces a model that leans heavily on one modality while ignoring the others.

AMGFM attacks this misalignment problem with two complementary learning strategies. The first is adversarial learning, a technique borrowed from generative adversarial networks in which two neural networks compete against each other. In this case, the system learns feature transformations that pull the distributions of the different modalities into a shared space, reducing the global discrepancies between, say, EEG features and eye-movement features. The second strategy is metric learning, which shapes that shared space so that examples of the same class — depressed or non-depressed — cluster tightly together while examples from different classes are pushed far apart. This combination enforces what the authors call intra-class compactness and inter-class separability, effectively sculpting a latent space in which a simple classifier can draw a clean boundary between patients and healthy controls.

On top of this aligned representation sits the model’s most distinctive component: a hierarchical graph fusion module. Rather than simply concatenating all the features together, the module explicitly models interactions at different levels — the information carried by each modality alone, the complementary information carried by each pair of modalities, and the joint information carried by all three together. A graph neural network then adaptively weighs these unimodal, bimodal, and trimodal interactions, deciding dynamically how much each level of interaction should contribute to the final decision. This design gives the model a degree of interpretability that is rare in deep learning systems, because researchers can inspect which interaction pathways the network relied on for a given classification.

The team evaluated AMGFM on two datasets, EFVP and AADP, which contain physiological and behavioral recordings from human participants collected under ethical approval from the Second Hospital of Lanzhou University and Gansu Provincial Hospital respectively, with informed consent from all participants. Across both datasets, the model consistently outperformed unimodal baselines that used only EEG, only eye movement, or only pupil data, as well as representative multimodal fusion approaches from the recent literature. The headline figure — up to 90.33 percent accuracy — places the system among the strongest reported results for physiological depression recognition, a task where typical accuracies have often hovered well below that threshold.

Perhaps the most surprising finding came from the team’s systematic analysis of how different kinds of emotional stimuli affect recognition performance. Participants in these paradigms are typically shown images or clips designed to provoke emotional responses, on the assumption that depressed individuals react differently to negative or emotionally charged content. But the Lanzhou analysis found that neutral stimuli — material with no particular emotional valence — actually provided stronger discriminative cues than emotional stimuli. This echoes earlier clinical observations that depressed patients show distinctive patterns of processing even neutral faces, and it suggests that screening protocols might not need elaborate emotional provocation to work well. A simpler, calmer stimulus set could make future screening sessions shorter and less stressful for participants.

The second analytical insight concerns which modality should serve as the anchor for aligning the others. The experiments indicated that EEG works best as the target modality — the reference signal into which eye movement and pupil features are aligned. That makes physiological sense: EEG offers a rich, high-dimensional, and relatively direct window into neural activity, while eye and pupil signals are noisier proxies of underlying cognitive and autonomic states. For engineers designing future multimodal systems, the practical lesson is that the choice of alignment target is not arbitrary, and treating EEG as the backbone of a multi-signal depression screener is a principled starting point rather than a guess.

The significance of this work extends beyond the specific accuracy numbers. Depression affects hundreds of millions of people worldwide and is recognized by public health authorities as a leading cause of disability, with prevalence studies among groups such as Chinese university students underscoring how widespread the condition is among young populations. Earlier detection means earlier intervention, and objective, sensor-based screening could eventually bring detection into settings where psychiatrists are scarce — primary care clinics, university health centers, or even wearable devices. The Lanzhou group has a long track record in this area, with prior work on EEG-based cognitive interfaces, mutual-information fusion of EEG and pupil signals, transformer networks for EEG and eye-tracking fusion, and graph neural network approaches to depression detection. AMGFM represents a synthesis of those threads into a single, more principled architecture.

Important caveats remain. The EFVP and AADP datasets are not publicly available because they contain sensitive human physiological data, and the model’s source code is likewise not yet in a public repository, though the authors say both may be shared with qualified researchers under appropriate ethical and data-use agreements. Independent validation on external datasets, across different populations and recording hardware, will be essential before any clinical deployment. Depression is also a heterogeneous condition, and a classifier trained on one cohort may not generalize to another. Still, the convergence of adversarial alignment, metric learning, and hierarchical graph fusion into a model that hits 90 percent accuracy marks a genuine step forward. If the results hold up in broader testing, the quiet electrical whispers of the brain — read alongside the flicker of an eye — may one day become a routine part of how medicine listens for depression.

Subject of Research: Multimodal machine learning for depression identification using EEG, eye movement, and pupil signals

Article Title: AMGFM: A multimodal fusion model for depression identification based on adversarial metric learning

Article References: AMGFM: A multimodal fusion model for depression identification based on adversarial metric learning. (n.d.). https://doi.org/10.1007/s11517-026-03676-z

Image Credits: AI Generated

DOI: 10.1007/s11517-026-03676-z

Keywords: depression, EEG, eye movement, pupil dynamics, multimodal fusion, adversarial learning, metric learning, graph neural network, machine learning, mental health screening, physiological signals, AMGFM

The blind spot, which the authors call direction-blindness, arises from a property most people never think about when they compute a distance. Standard Euclidean distance is symmetric: it depends only on the absolute magnitude of the difference between two feature values, not on whether one value sits above or below the other. That means a feature whose informative signal for a label lies entirely on one side of its average value — say, a gene whose under-expression flags a disease, or a word whose absence marks a document topic — is, from the perspective of pairwise distances, indistinguishable from a feature whose signal is symmetrically distributed around the mean. The classifier sees proximity, but not direction, and can collapse two semantically very different patterns into a single similarity score.

Adaptive metric-learning techniques such as LMNN and NCA can address this problem, but they do so inside the classifier, downstream of any feature selection step. The Sassari team, Filippo Casu, Andrea Lagorio and Giuseppe A. Trunfio, took a different route: expose directional evidence at the representation level, before the classifier ever runs. Their approach is an embedded feature-selection method, meaning it learns which features matter while simultaneously fitting a predictive model, rather than scoring features one at a time with a filter statistic. Embedded methods are attractive because they account for interactions among features and labels, and because they return a ranking that can be plugged into any downstream algorithm.

The core idea of PSDFS is a mathematical trick the authors call a signed-deviation lift. Every original feature is split into two nonnegative channels: one measuring how far a value sits above the training-fold mean, and one measuring how far it sits below. This two-sided hinge-style expansion, which echoes classical constructions in multivariate adaptive regression splines and in convex nonnegative matrix factorization, has a powerful consequence. In a conventional nonnegative reconstruction model, the contribution of a feature is a monotonically increasing function of its value, so low feature values can never act as explicit evidence for a label. The lift breaks that monotonicity barrier while keeping every weight nonnegative — a constraint that preserves the stability of multiplicative updates and keeps feature contributions easy to interpret. The researchers prove formally that the class of functions representable after lifting strictly contains the class representable without it.

Splitting each feature into two channels, however, creates a new hazard. If a sparsity penalty acts independently on the lifted components, the optimizer may select only one half of a pair — an XOR-like artifact the authors quantify with a diagnostic called the SplitRatio. Under an unpaired penalty, this ratio often approaches one, meaning the reported selection fragments the identity of the original feature and produces rankings that are sensitive to an arbitrary decomposition. PSDFS solves this with a paired group-sparsity penalty that treats the above- and below-center channels of each feature as a single selection unit, ranking the original features directly by the combined norm of their two channels. The result is a single, coherent ranking in the original feature space, with no split artifacts by construction.

The optimization machinery is deliberately conservative. For fixed training-fold statistics, the objective — a Frobenius reconstruction loss plus the paired group penalty — is convex in the weight matrix, and the multiplicative updates descend monotonically with a sublinear O(1/T) bound on average suboptimality, properties the authors establish in a series of propositions with full proofs. Empirically, most of the objective decrease happens within roughly ten iterations, and a budget of forty updates proved safe across every benchmark tested. Additional propositions show that the optimum carries genuine directional semantics: an active above-center weight requires positive correlation between the above-center deviation and the label residual, and vice versa, so the relative magnitudes of the two channels reveal which side of the mean carries the signal.

The experimental evaluation is unusually thorough. The team compared PSDFS against seven recent embedded feature-selection methods — LRMFS, LSMFS, LRDG, GRRO, RFSFS, SRFS and SCNMF — on sixteen benchmark datasets drawn from the Mulan repository and the Yahoo topic collection, spanning text, image and bioinformatics domains with anywhere from 194 to nearly 14,000 instances and up to 374 labels. All methods ran on identical stratified folds with no per-dataset hyperparameter tuning, a discipline that guards against information leakage. Evaluated with ML-kNN as the downstream classifier, PSDFS achieved the best average rank on all seven reported metrics, including Micro-F1, Macro-F1, Hamming Loss, precision-recall areas and label-ranking measures. After Holm–Bonferroni correction, its improvements were statistically significant against every baseline on Micro-F1 and Macro-F1, and against all but one baseline — GRRO — on Hamming Loss.

Targeted ablations confirm that each design element earns its place. Removing the lift collapses performance back toward the monotonicity-limited baseline, with the lifted variant winning on twelve to fourteen of sixteen datasets across the core metrics. Replacing the nonnegative constraint with an unconstrained solver degrades results on nearly every metric, suggesting that nonnegativity regularizes the solution and prevents the two channels from canceling each other algebraically. A robustness check with a second fixed-metric classifier, binary relevance with an RBF-kernel logistic regression, preserved the ranking of methods, reducing the risk that the advantage is an artifact of ML-kNN’s neighborhood bias. The method is also fast: at an average of 1.37 seconds per fold, it outruns most of its competitors, only trailing the two baselines that solve simpler linear systems.

Perhaps the most practically valuable analysis concerns low-resource regimes. When the training split is subsampled to as little as five percent, PSDFS degrades gracefully and is never overtaken by any baseline, though its separation from the strongest reconstruction-based competitors narrows to a statistical tie. The authors identify the mechanism precisely: the lift estimates twice as many channel weights from disjoint subsets of the data, so its contribution is the first casualty of scarce labels. They offer concrete deployment guidance — use the method as-is when the sample size exceeds roughly twice the lifted dimensionality, and treat fine-grained directional annotations with caution below that threshold. The authors also note two honest limitations: the model remains additive and does not capture higher-order feature interactions, and it has not been validated inside end-to-end deep visual pipelines. A reference implementation is publicly available on GitHub, and given how widely fixed-metric distance classifiers are deployed, the idea of making feature selection direction-aware seems likely to travel far beyond the sixteen benchmarks where it was born.

Subject of Research: Direction-aware feature selection for distance-based multi-label classification

Article Title: Direction-aware multi-label feature selection via paired signed-deviation lifting

Article References: Direction-aware multi-label feature selection via paired signed-deviation lifting. (n.d.). https://doi.org/10.1007/s10618-026-01265-0

Image Credits: AI Generated

DOI: 10.1007/s10618-026-01265-0

Keywords: multi-label learning, feature selection, machine learning, distance-based classifiers, nonnegative reconstruction, group sparsity, ML-kNN, signed-deviation lift, metric learning, data mining, benchmark evaluation, interpretable AI