Using logs from the Lingual AB online English platform, the study examined sequences of student responses to vocabulary, grammar, reading and listening activities. The researchers reported recognition errors below 8% for each of the three categories, with 40 of 720 Lexical Retrieval Deficit cases, 60 of 840 Syntactic Parsing Deficit cases and 55 of 840 Discourse Coherence Deficit cases misclassified. The model’s F1-score across five-fold cross-validation was 0.91 ± 0.015, with a 95% confidence interval from 0.895 to 0.925.

The study involved 200 selected participants aged 12 to 15 whose platform records were collected from September 2023 through January 2024. The raw dataset contained 185,000 interaction records with 12 fields each, including timestamps, response time, correctness, help clicks, prompt views, answer revisions, retry counts and session identifiers. After cleaning and preprocessing, the dataset contained 188 users, approximately 172,000 logs and 9,380 samples. Each sample consisted of 20 consecutive interactions from one student.

Ground-truth labels were not generated automatically. Three senior English as a Second Language instructors, each with more than eight years of teaching experience, independently reviewed the final 20 interaction logs for each user against a rubric covering error frequency, response latency and help-seeking behavior. Lexical Retrieval Deficit required a vocabulary error rate above 40% and average response latency above 15 seconds; Syntactic Parsing Deficit required a grammatical error rate above 35% and clause-segmentation accuracy below 60%; and Discourse Coherence Deficit required a cross-sentence referential error rate above 30%. Fleiss’ Kappa was 0.83, and cases with disagreement, approximately 8%, were resolved through moderated consensus. The protocol received institutional review-board approval, and written informed consent was obtained from participants and their legal guardians.

The knowledge map contained 847 entities and 3,412 relationships, representing knowledge points, topic items and student nodes. Its relations included knowledge-point prerequisites, topic tests, student mastery and links between broader and narrower knowledge points. Entities and relations were represented with TransE embeddings, while two stacked LSTM layers with 128-dimensional hidden states processed normalized behavioral sequences. The fusion mechanism used cross-attention and a gate derived from each LSTM state to modulate the relevant knowledge embedding, rather than merely concatenating the two information sources.

Comparisons favored the combined model. Support Vector Machine and Random Forest baselines achieved accuracies of 0.78 and 0.82, respectively; GRU and single LSTM models reached 0.85 and 0.87; and the knowledge-graph-enhanced MLP reached 0.88. A Transformer baseline achieved 0.90, compared with 0.92 for the full KG-LSTM. Paired tests of five-fold F1 scores against the Transformer reported t(4) = 3.82, p = 0.019 and Cohen’s d = 1.52. Ablation results were lower when major components were removed: accuracy fell to 0.87 without the knowledge map, 0.86 without the LSTM and 0.84 without the fusion layer, while removal of attention produced an accuracy of 0.89.

The model also generated diagnostic paths linking behavioral evidence to specific knowledge nodes. In a worked example, the system recorded four weeks of vocabulary accuracy below 60%, an average grammar-filling response time of 18.6 seconds and seven to nine long sentences encountered in each reading passage. The knowledge map connected these recorded observations into a path from low vocabulary mastery through slower syntactic parsing to increased reading-processing load. After edge weights were combined, the syntactic-parsing path scored 0.84, compared with 0.76 for the lexical-retrieval alternative, so the system identified grammar difficulty as the main obstacle and reading comprehension as a secondary risk.

These findings apply to behavioral indicators observed in an online learning environment, not to clinically diagnosed specific learning disorders such as dyslexia. The authors describe persistent errors, prolonged response times and unusual help-seeking behavior as early-warning signals and quantitative references, not substitutes for formal clinical assessment. The practical application scenarios in online and classroom settings were presented as potential uses rather than interventions validated in the study.

Several limitations constrain the results. The model was tested on data from one platform and one age group, and the study did not empirically validate classroom deployment or teaching interventions. Platform-specific logging standards, privacy protection, and inference speed at high concurrency remain deployment challenges; the paper reports that latency could exceed 300 milliseconds with more than 500 concurrent users and that direct transfer to other schools or age groups could reduce accuracy by 3% to 6%. The framework also assumes that the knowledge graph and relevant entity attributes are complete. Future work is proposed to address cross-platform validation, incomplete graphs, richer relation types, larger samples and alternative sequence-model architectures.

Subject of Research: A KG-LSTM model for classifying behavioral English-learning difficulty patterns from online platform logs

Article Title: Intelligent diagnosis method of English learning disabilities based on knowledge mapand long-term and short-term memory network

Article References: Lin, C. (2026). Intelligent diagnosis method of English learning disabilities based on knowledge mapand long-term and short-term memory network. Discover Artificial Intelligence, 6(1), Article 1281. https://doi.org/10.1007/s44163-026-02095-1

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02095-1

Keywords: English learning difficulty, knowledge graph, LSTM, behavioral data, online learning, diagnostic model, junior high school, interpretable AI, cross-validation, Intelligent, diagnosis, method

Fundus photography captures surface-level vascular and optic-disc information, whereas OCT provides high-resolution cross-sectional detail of retinal layers but does not include the same surface-level cues. OHDM-FPN was designed to combine these complementary representations for eye-disease detection, including early-stage disease detection. For fundus images, the framework used EfficientNetV2-Small as its encoder.

OCT images were processed by GhostNet with dilated residual blocks. Each modality passed through a feature pyramid network using atrous spatial pyramid pooling to develop multi-scale representations before the two streams were brought together. The fusion stage used a cross-modality multi-scale fusion block to align and combine modality-specific features.

This block incorporated channel-spatial attention and deformable convolutions. A lightweight classification head then predicted the disease category, while hyperparameter tuning with the Salmon Salar optimization algorithm was used to address class imbalance and improve generalization. Experiments were conducted on a combined fundus–OCT benchmark dataset assembled from EyePACS and Duke OCT.

Against the study’s comparison models, the reported OHDM-FPN results were described as averaging 6.3% better in accuracy and 5.8% better in AUC than single-modality models. The researchers also reported a 9.1% improvement in early-stage disease detection rates compared with OCT-only baselines. Attention heat maps were used to examine the regions influencing the model’s classifications.

The reported maps showed that OHDM-FPN focused on clinically relevant retinal regions, providing an interpretability aid for ophthalmologists. This visualization was presented as a way to inspect where the network placed emphasis when making predictions. The findings are computational results from benchmark datasets rather than results from a prospective clinical trial.

Consequently, the reported performance does not establish how OHDM-FPN would perform in routine clinical use. Within the study’s stated scope, the results indicate that fusing fundus and OCT features improved classification performance relative to the reported single-modality baselines. OHDM-FPN combined modality-specific feature extraction, multi-scale processing, attention, deformable convolutions and optimized classification in one framework for early eye-disease detection.

Subject of Research: Evaluation of a deep-learning model for classifying eye diseases from combined fundus and OCT images

Article Title: Optimized multi-scale dual-modality feature pyramid network for early eye disease classification using fundus and OCT imaging

Article References: Kiruthiga, R., & Kumar, A. S. (2026). Optimized multi-scale dual-modality feature pyramid network for early eye disease classification using fundus and OCT imaging. Complex & Intelligent Systems. https://doi.org/10.1007/s40747-026-02507-w

Image Credits: AI Generated

DOI: 10.1007/s40747-026-02507-w

Keywords: multimodal AI, eye disease detection, fundus photography, optical coherence tomography, deep learning, feature fusion, diabetic retinopathy, macular edema, glaucoma, interpretable AI, Optimized, multi-scale

The new tool, described in Nature Biomedical Engineering, is called SPACER — short for a deep learning framework designed to decode the spatial codes of cell recruitment directly from spatial transcriptomics data. Unlike conventional AI models that act as inscrutable black boxes, SPACER was built from the ground up to be interpretable. It does not simply predict where cells will end up; it explains which molecular features and spatial relationships drive those predictions, offering researchers a transparent window into the logic of tissue organization.

Spatial transcriptomics, the technology that underpins SPACER, allows scientists to measure gene activity across thousands of positions within an intact tissue section, preserving the physical context that single-cell sequencing loses when cells are dissociated. This produces extraordinarily rich datasets — but also extraordinarily difficult ones. A single tissue slice can contain millions of measurements, and the relationships between neighboring cells form a web of interactions too complex for traditional statistical tools to untangle. Machine learning seemed like an obvious solution, but standard neural networks tend to hide their reasoning inside millions of weighted connections, leaving biologists with accurate predictions and no understanding of the underlying biology.

SPACER sidesteps this problem through a design philosophy rooted in interpretability. Rather than treating the tissue as an undifferentiated blob of data, the model explicitly represents the physical distances between cells and the gene expression profiles of the cell types involved. It then learns which combinations of molecular signals and spatial arrangements are most predictive of a given cell type appearing in a given location. Because these features are biologically meaningful — a chemokine gradient, a cell–cell contact, a stromal barrier — researchers can read the model’s internal logic directly, rather than reverse-engineering it after the fact.

To test the system, the researchers turned to cancer, where cellular recruitment is a matter of life and death. Whether T cells successfully infiltrate a tumor is one of the strongest predictors of response to immunotherapy, yet many tumors remain stubbornly resistant to immune invasion. When the team applied SPACER to spatial transcriptomics data from multiple cancer types, two previously underappreciated factors emerged. The first was the local antigen presentation machinery — the molecular apparatus that cells use to display fragments of proteins on their surfaces, flagging themselves for immune inspection. Tumors with robust antigen presentation in specific microanatomical niches were far more successful at recruiting T cells than those where this machinery was absent or suppressed.

The second discovery was more surprising: tumor mucins, a family of heavily glycosylated proteins that coat the surface of many cancer cells, appeared to play a decisive role in modulating T cell recruitment. Mucins have long been associated with physical barriers and immune evasion, but SPACER’s analysis suggested they actively shape the spatial distribution of immune cells within the tumor microenvironment. This finding reframes mucins not merely as passive structural components but as active participants in the cellular choreography of the tumor, potentially offering a new target for therapeutic intervention.

The researchers did not stop at cancer. They also deployed SPACER on a mouse model of severe checkpoint inhibitor-induced myocarditis, a rare but potentially fatal inflammatory condition of the heart that can occur as a side effect of cancer immunotherapy. In this model, the AI analysis pointed to CD4-positive T cells — a subset of immune cells typically associated with helper and regulatory functions — as the key effectors driving the disease. This conclusion challenges the assumption that cytotoxic CD8-positive T cells are the primary culprits in immunotherapy-related heart inflammation, and it suggests that therapeutic strategies aimed at CD4-positive cells might be more effective in preventing or treating this dangerous complication.

Beyond its specific biological findings, SPACER represents a broader shift in how computational biology approaches the problem of tissue organization. The field has generated vast repositories of spatially resolved molecular data, but the bottleneck has shifted from data collection to data interpretation. Tools like SPACER aim to close that gap by embedding biological reasoning directly into the architecture of the model, so that the output is not just a prediction but a hypothesis — a testable statement about how cells find their way to particular destinations within living tissue.

The implications extend well beyond the two systems studied here. If the rules of cellular recruitment can be decoded in tumors and inflamed hearts, the same framework could be applied to neurodegenerative disease, fibrosis, organ transplantation, and developmental biology — anywhere that the spatial arrangement of cells determines function. And because SPACER is interpretable, its conclusions can be directly interrogated and validated by experimental biologists, creating a feedback loop between computational prediction and laboratory confirmation that accelerates discovery.

For clinicians, the near-term promise is more concrete. Understanding why some tumors recruit T cells and others do not could help identify patients likely to respond to checkpoint inhibitors before treatment begins, sparing others from unnecessary toxicity. Uncovering the role of CD4-positive T cells in immunotherapy-related myocarditis could lead to earlier detection and targeted prevention strategies for a complication that currently forces many oncologists to abandon otherwise life-saving therapies. In both cases, the path forward runs through the same insight: that the spatial code of cellular recruitment, once decoded, becomes a map — and a map, unlike a black box, can be read, questioned, and ultimately rewritten.

Subject of Research: An interpretable AI model that decodes the spatial rules of cell recruitment from tissue spatial transcriptomics data

Article Title: An interpretable AI model decodes the rules of cell recruitment to specific destinations

Article References: An interpretable AI model decodes the rules of cell recruitment to specific destinations. (2026). Nature Biomedical Engineering. https://doi.org/10.1038/s41551-026-01803-7 (Author names are not included in the publisher’s citation export.)

Image Credits: AI Generated

DOI: 10.1038/s41551-026-01803-7

Keywords: SPACER, spatial transcriptomics, cell recruitment, interpretable AI, tumor microenvironment, T cells, antigen presentation, tumor mucins, checkpoint inhibitor myocarditis, CD4-positive T cells, cancer immunotherapy, computational biology

The central problem the team set out to solve is twofold. First, resistance in the tuberculosis bacillus is not determined by single mutations acting alone; variants interact epistatically, meaning the effect of one mutation can depend on the presence of others elsewhere in the genome. Second, whole-genome sequencing pipelines vary in quality, and a variant call that looks convincing in one isolate may be technically noisy in another. Conventional machine learning classifiers typically consume a fixed vector of variant features and ignore both the set-like nature of genomic variation and the quality metrics attached to each call in the variant call format, or VCF.

VAMP-Net answers these challenges with a dual-pathway design. Path-1 employs a Set Attention Transformer to model the genome’s variants as a permutation-invariant set. Because a bacterial genome carries a variable number of mutations and there is no biologically meaningful ordering among them, the network must produce the same output regardless of the order in which variants are presented. Set attention layers, originally introduced in the Set Transformer framework, achieve this by exchanging information among set elements through attention blocks, allowing the model to capture epistatic dependencies between variant pairs without imposing an artificial sequence structure.

Path-2 takes a different slice of the evidence: the quality metrics embedded in the VCF, such as depth, mapping quality and the fraction of supporting reads. A one-dimensional convolutional neural network processes these metrics to produce an adaptive confidence score for each variant. The two pathways are then fused so that the final classification reflects not only which variants a genome carries but how reliable the calls are. The authors characterize the result as a learned integrated audit: the model does not simply trust raw sequencing depth, which can be inflated by repetitive regions or contamination, but learns to prioritize the fraction of supporting reads and relative confidence over raw depth as a way of mitigating technical noise.

The training data came from a public compendium associating more than 12,000 tuberculosis isolates with quantitative resistance phenotypes, and the complete dataset is openly available through the European Bioinformatics Institute. The code is also public, hosted on GitHub, which matters for the auditable character the authors emphasize. In clinical genomics, a model whose decision process cannot be inspected is difficult to deploy, because clinicians and laboratories need to know whether a resistance call rests on a canonical resistance mutation or on a sequencing artifact.

Interpretability was tested in two layers. The first layer uses Integrated Gradients, a feature attribution method that assigns each input feature an importance score by integrating gradients along a path from a baseline input to the actual input. Applied to VAMP-Net, the attribution analysis successfully recovered the canonical resistance targets: rpoB for the rifamycins, embB for ethambutol and katG for isoniazid. These genes are the well-established determinants that the World Health Organization’s technical guide on next-generation sequencing for resistance detection is built around, so recovering them is a strong sanity check that the model learned genuine biology rather than shortcuts.

The second layer of interpretability is more scientifically provocative. The attribution analysis also discovered high-impact novel loci that are not part of the canonical resistance catalogue. Functional enrichment analysis confirmed that these novel variants are not randomly scattered across the genome but constitute coherent metabolic modules, with a statistical significance of p equal to 0.00239, centered on cell-wall remodeling. This finding is biologically plausible: the tuberculosis cell wall is rich in mycolic acids and related lipids, and prior literature has documented that rifampicin-resistant rpoB mutants upregulate the phthiocerol dimycocerosate biosynthetic pathway, and that cross-resistance between clofazimine and bedaquiline arises through upregulation of the mmpL5 efflux transporter. The enriched modules suggest that resistance in the bacillus is embedded in compensatory metabolic rewiring rather than in isolated target mutations.

The paper’s appendix consolidates this mechanistic picture with systematic epistatic interaction tables. For each drug, the authors list the top variant pairs extracted by the attention mechanism, mapping every genomic coordinate against the H37Rv reference annotation and, for canonical loci, against the literature underlying the WHO catalogue. The tables highlight pairs involving efflux transporters such as mmpL5 and mmpL4, the fluoroquinolone target gyrA at a non-QRDR background position, the ethambutol-operon arabinosyltransferase embC, and the PhoPR two-component regulator phoR, which controls complex-lipid biosynthesis and stress response. An independent, model-agnostic statistical epistasis screen largely disjoint from the primary resistance-determinant loci corroborated the strongest pairs, ranking them by interaction significance and reporting odds ratios for carrying both variants versus neither.

Systematic ablation of the quality-aware pathway demonstrated that removing it degrades performance, confirming that the confidence-scoring branch contributes information beyond the variant set alone. The authors frame this as establishing a new paradigm for robust, auditable and clinically actionable resistance prediction, one that bridges genomic pathogenicity with technical reliability. In practice, this means a laboratory could inspect not only which loci drove a resistance call but also how the model weighted the sequencing evidence, making the classification traceable from mutation to metric to decision.

The work positions VAMP-Net as a tool for both diagnostic classification and mechanistic discovery in clinical genomics. Tuberculosis remains one of the world’s leading infectious killers, and drug resistance complicates treatment decisions that must often be made quickly. Whole-genome sequencing is increasingly used to guide therapy, but machine learning classifiers in this domain have been criticized for limited interpretability and for ignoring sequencing quality. By combining permutation-invariant set attention over variants with an explicit quality audit, and by validating its learned features against both canonical targets and independent statistical screens, the Algiers team has produced a system whose predictions come with a readable trail of evidence. Whether the discovered cell-wall remodeling modules translate into new biological understanding of how the bacillus tolerates antibiotics will require experimental follow-up, but the framework demonstrates that a single network can be simultaneously accurate, transparent and technically self-critical.

Subject of Research: Machine learning

Article Title: VAMP-Net: a variant-quality-aware, interpretable permutation-invariant multi-path network for auditable MTB drug-resistance

Article References: Boutorh, A., Baghdadi, K. H., & Daoud, A. (2026). VAMP-Net: a variant-quality-aware, interpretable permutation-invariant multi-path network for auditable MTB drug-resistance. Applied Intelligence, 56(16), Article 508. https://doi.org/10.1007/s10489-026-07496-4

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07496-4

Keywords: tuberculosis, drug resistance, deep learning, genomics, interpretable AI, VAMP-Net, machine learning, antimicrobial resistance, variant-quality-aware, interpretable, permutation-invariant, multi-path

Now a team of researchers in Morocco has introduced a deep learning framework that attacks this problem from a direction that traditional tools cannot reach. The framework, called GasGAT, was developed by Youssef Said and colleagues at the Laboratory of Information Processing, University Hassan II in Casablanca, together with a collaborator at the Hassania School of Public Works. Their work, published in Frontiers in Blockchain, models smart contracts not as lines of code to be scanned for known anti-patterns, but as semantic graphs in which functions, state variables, and control-flow constructs become nodes, and their interactions become edges. A Graph Attention Network, or GAT, then learns to navigate this graph and identify the structural causes of excessive gas consumption.

The innovation matters because existing gas optimization tools are largely confined to static heuristics and local pattern matching. Tools such as GasOptiScan and GasChecker excel at detecting pre-defined anti-patterns, like costly storage operations syntactically enclosed within loops. But they operate at a syntactic level and cannot easily model complex state dependencies or detect emergent inefficiencies that arise from the non-local interaction of multiple, seemingly optimized functions across a contract’s call graph. Machine learning approaches that treat contracts as flat feature vectors, such as opcode counts or function sizes, are similarly constrained by the expressiveness of manually chosen features. Neither family of methods can reason about long-range execution paths that span multiple functions.

GasGAT’s graph representation is designed precisely to expose these hidden dependencies. The researchers built a dedicated parser that transforms the Solidity Abstract Syntax Tree of each contract into a directed semantic graph. Nodes correspond to key semantic entities: public, internal, and private functions; control-flow constructs such as loops and conditionals; persistent state variables; and gas-intensive operations like the SSTORE instruction that writes to storage. Edges capture function call dependencies, control-flow relations, read and write accesses to state variables, and inter-procedural data and execution dependencies. Each node is described by a 16-dimensional feature vector encoding its type, function visibility, payable, view, and pure flags, normalized loop nesting depth, parameter count, source position, constructor status, and out-degree.

A concrete example illustrates why this representation is so powerful. Consider a batch-processing function that loops over an array of identifiers and calls an internal helper function, which in turn writes to a storage map. The expensive SSTORE operation is syntactically distant from the loop, buried inside the helper. The graph conversion creates a multi-hop path connecting the loop to the eventual storage write, and the attention mechanism can propagate along this entire execution chain. The model learns to attribute gas costs not to isolated syntax, but to the holistic interaction of iteration, calling context, and state mutation. Purely local, intra-procedural analyses would miss this pattern entirely, because no single function appears wasteful in isolation.

To train and evaluate the framework, the team retrieved source code for more than 100,000 unique verified smart contracts via the Etherscan API, filtering to retain contracts written in Solidity version 0.8.0 or later with at least 50 confirmed transactions. The final corpus comprised 40,000 contracts from the Ethereum mainnet, spanning decentralized finance protocols, NFT marketplaces, DAO governance contracts, and general-purpose utilities. Each contract was labeled using Slither static analysis to compute an aggregate inefficiency score based on the weighted presence of known gas-related anti-patterns. Contracts in the top quartile were labeled gas-inefficient, those in the bottom quartile gas-efficient, and the ambiguous middle 50 percent were deliberately excluded to establish a clear benchmark. A random subset of 500 contracts was independently reviewed by two smart contract security experts, yielding high inter-rater agreement with a Cohen’s Kappa of 0.89.

The GasGAT architecture consists of three stacked graph attention layers with an initial linear projection from 16-dimensional input features to a 128-dimensional hidden space. The first two layers use four attention heads with concatenation, while the final layer uses a single head. Residual connections and LayerNorm are applied after each layer, and the graph-level representation is obtained by concatenating global mean and max pooling over node embeddings, feeding a 256-dimensional vector to a three-layer MLP classifier. The model trains with the Adam optimizer, dropout of 0.3, and weighted cross-entropy loss to correct an 84-to-16 percent class imbalance. Remarkably, the entire framework contains only 94,530 trainable parameters and runs on a CPU with 16 GB of RAM, completing five-fold cross-validation in roughly three hours.

Under five-fold stratified cross-validation on 27,879 labeled graphs, GasGAT achieved a mean accuracy of 94.92 percent, with a standard deviation of 1.04 percent, and a macro F1-score of 88.69 percent, with a standard deviation of 2.70 percent. Precision stabilized at 96.94 percent, with a standard deviation of only 0.43 percent across all five folds, confirming the robustness of the classification capability. A McNemar statistical test yielded a statistic of 633.12 with p far below 0.05, confirming that the predictions are statistically significant and not attributable to chance. Training loss converged typically between epochs 5 and 20, with early stopping triggered before epoch 40 in all folds, indicating efficient optimization and strong generalization to unseen contracts.

The most striking result, however, concerns what happens when correlated node features are removed. The researchers rebuilt all graphs using a revised feature vector that completely eliminates two features directly tied to the labeling heuristics: a per-node gas weight estimate and membership in gas-intensive operations. When these features were removed, a graph convolutional network baseline dropped by 16.15 percentage points, from 99.05 to 82.90 percent F1, revealing that its near-perfect performance was primarily driven by memorizing the feature-label correlation. GraphSAGE and GIN dropped by 2.59 and 2.29 points respectively. GasGAT alone showed zero performance change, from 88.69 to 88.69 percent, providing direct empirical evidence that its attention mechanism learns genuine structural graph dependencies rather than exploiting label-feature correlations inherent in the heuristic labeling process.

The interpretability of the attention mechanism transforms predictive performance into actionable developer insights. In a real case study from the test set, a contract misclassified by an XGBoost baseline but correctly identified as gas-inefficient by GasGAT contained a public batch function that iterated over an array and invoked a registration function, which delegated state modification to an internal helper. Because the expensive SSTORE operation was not syntactically enclosed within the loop, no local static-analysis feature was activated. GasGAT assigned its highest attention weights to the edges forming the execution path between the loop and the deep state-modifying operation, with scores of 0.31, 0.47, and 0.52, and one edge carrying an attention score of 0.9965. A hub node exhibited seven outgoing edges with attention weights above 0.96, a pattern entirely absent from Slither’s detector catalogue. The highlighted attention path translates directly into refactoring guidance: developers can extract state-modifying logic to make gas costs visible at the call site, apply batch storage updates instead of repeated individual writes, or restructure the loop to minimize cross-function delegation.

The researchers also conducted supplementary experiments on the full dataset, reintegrating the previously excluded middle 50 percent of contracts. In a three-class classification setting distinguishing gas-efficient, moderately optimized, and gas-inefficient contracts, GasGAT achieved 92 to 94 percent accuracy compared to 88 to 90 percent for a local baseline. In a regression formulation predicting a continuous inefficiency score, GasGAT reached an R-squared of 0.84 to 0.88 versus 0.72 to 0.78 for the baseline. A preliminary evaluation on 2,000 contracts written in Solidity versions 0.5 through 0.7, without retraining, exhibited stable qualitative behavior with limited performance degradation, suggesting that the learned representations capture higher-level execution structures rather than version-specific syntactic artifacts.

The authors are transparent about the limitations of their approach. The polarized dataset construction simplifies the classification task and may underrepresent moderately optimized contracts encountered in real-world development. The ground-truth labels are derived from static-analysis heuristic scores rather than large-scale dynamic execution profiling, because publicly available datasets validated through systematic runtime execution traces remain unavailable at the scale of 40,000 contracts. The dataset was restricted to Solidity 0.8.0 and later, so direct generalization to legacy contracts cannot be assumed without further validation. Training is computationally intensive due to graph attention operations, though inference requires only a single forward pass whose cost scales linearly with the number of nodes and edges, compatible with interactive developer assistance scenarios.

What distinguishes GasGAT from prior work is not marginal accuracy improvement but a paradigm shift from localized pattern matching to holistic structural analysis. To the best of the authors’ knowledge, it represents one of the first applications of graph attention networks specifically to the gas optimization problem, with an explicit focus on inter-procedural structural reasoning and developer-oriented interpretability. The framework identifies three categories of inter-procedural gas inefficiency that are fundamentally invisible to Slither’s detectors: delegation chains where expensive storage operations are triggered indirectly through multi-level call chains, hub functions that delegate execution to multiple state-modifying helpers through high-attention edges, and cross-function state accumulation where repeated mutations emerge only when considering the full inter-procedural execution path.

The implications extend beyond Ethereum. As blockchain applications proliferate across finance, supply chains, insurance, gaming, and e-voting systems, the economic viability of decentralized applications depends critically on controlling execution costs. GasGAT demonstrates that data-driven program analysis can complement manual rule engineering, offering a scalable alternative that learns structural abstractions not explicitly encoded in static analysis features. The attention mechanism functions not merely as a predictor but as a diagnostic tool, guiding developers toward concrete refactoring opportunities that enhance both performance and security. For future work, the team plans to extend the framework toward multi-class and regression-based formulations for finer-grained identification of gas anti-patterns, and to generalize it to legacy Solidity versions and other EVM-compatible blockchains, strengthening its applicability to real-world smart contract ecosystems.

Subject of Research: Computer Science

Article Title: GasGAT: a graph attention network framework for smart contract gas optimization

Article References: Said, Y., Khaddar, A. M., Hassine, L., Eddaoui, A., & Chafiq, T. (2026). GasGAT: a graph attention network framework for smart contract gas optimization. Frontiers in Blockchain, 9, Article 1838214. https://doi.org/10.3389/fbloc.2026.1838214

Image Credits: AI Generated

DOI: 10.3389/fbloc.2026.1838214

Keywords: Ethereum, smart contracts, gas optimization, graph attention network, machine learning, blockchain, interpretable AI, GasGAT, graph, attention, network, framework

The clinical backdrop is stark. Aboriginal mothers experience significant perinatal mental health disparities shaped by social, historical and cultural determinants, from intergenerational trauma to systemic barriers in mainstream health services. Routine risk screening is supposed to guide follow-up care, but conventional approaches often miss cultural considerations entirely. Standard screening questionnaires and risk scores tend to catalogue deficits and symptoms while overlooking the protective factors that matter in Aboriginal communities: connection to family, cultural identity, pride, and community belonging. The Baby Coming You Ready program was designed to flip that framing, using a strengths-based, culturally adapted digital assessment that lets mothers tell their own story rather than simply answer deficit-oriented checklists.

Yet even a well-designed culturally safe tool can be undermined by how its results are used. According to the research team, some clinicians continue to rely primarily on conventional risk scores, which can neglect the broader contextual circumstances influencing Aboriginal mothers, including cultural strengths and protective factors. That reliance limits the program’s full impact. The question the team set out to explore was whether machine learning could help: could an algorithm, trained on BCYR data, predict perinatal psychological distress in a way that is both accurate and transparent enough to support culturally safe clinical conversations rather than replace them?

The study analyzed deidentified pilot data from 293 Aboriginal mothers recruited across six sites in Western Australia. After excluding five records with incomplete outcome data, the analytic dataset included 288 participants. The researchers faced a familiar obstacle in clinical machine learning: the dataset was modest in size, which constrains how well complex models can learn. Their solution was generative AI data augmentation. Using a technique known as a Tabular Variational Autoencoder, or TVAE, the team generated synthetic training examples that preserved the statistical structure of the real data while expanding the effective training set. Crucially, synthetic data was used only to augment training; model performance was always assessed on held-out real test data, ensuring that the reported results reflect genuine generalization rather than artifacts of artificial samples.

For the prediction engine itself, the researchers chose an Explainable Boosting Machine, a member of a newer family of models called generalized additive models with structured interactions. Unlike deep neural networks, whose internal reasoning is notoriously opaque, an Explainable Boosting Machine is intrinsically interpretable: each feature’s contribution to a prediction is expressed as a transparent, inspectable function, and pairwise interactions between features remain visible to the user. This matters enormously in a clinical and cultural context where a black-box score that simply labels a mother as high risk could reinforce exactly the deficit-focused, paternalistic practices the BCYR program was built to counter. An interpretable model, by contrast, can show which factors drove an estimate, inviting scrutiny and dialogue.

The results were striking. The best internal performance came from the TVAE-augmented Explainable Boosting Machine at an augmentation ratio of 1.0, meaning the synthetic data roughly doubled the training set. That configuration achieved an accuracy of 0.867, an F1 score of 0.817, and an area under the receiver operating characteristic curve, or AUC, of 0.934. In practical terms, an AUC above 0.9 indicates the model ranked mothers by their estimated risk of perinatal psychological distress with high discrimination across repeated internal train-test splits. The F1 score, which balances precision and recall, suggests the model maintained a reasonable trade-off between flagging mothers who genuinely need support and avoiding false alarms that could burden families and services.

Just as important as the numbers is what the model revealed when asked to explain itself. The explanations identified self-blame, loneliness, family pride, day-to-day management, and sleep as influential contributors to model-estimated risk, alongside related contextual factors. The pattern is telling: alongside recognized distress markers such as loneliness, disturbed sleep and self-blame, the model gave weight to protective and contextual factors like family pride and the capacity to manage daily life. That mixture reflects the strengths-based philosophy of the underlying assessment and stands in sharp contrast to conventional risk scores, which typically reduce a mother’s situation to a narrow symptom count. The researchers are explicit that these explanations are intended as prompts for culturally safe discussion and clinical reflection, not as verdicts delivered by a machine.

The vision emerging from this work is a collaborative one. In the imagined clinical workflow, an interpretable model surfaces the factors shaping a mother’s estimated risk, and a clinician, ideally working within a culturally safe framework and in partnership with Aboriginal health workers, uses those factors to open a conversation: about loneliness, about sleep, about family strengths and daily pressures. The AI becomes a mirror and a discussion aid rather than an authority. This is what the authors mean by interpretable generative AI supporting culturally contextualised prediction: the technology’s transparency is not a technical nicety but the very feature that makes it compatible with culturally responsive care, where mothers’ own narratives and community context must remain central.

The team is careful about the limits of the study, and appropriately so. This was an exploratory model-development exercise using pilot data from a single program, evaluated with repeated internal train-test splits rather than on an entirely new population. The authors state that broader community-governed external validation and prospective clinical workflow evaluation are required before any clinical implementation. Community governance is a key phrase: for Aboriginal health research, legitimacy depends on Aboriginal communities themselves controlling how data is used and how tools are deployed, a principle consistent with Indigenous data sovereignty movements worldwide. External validation would test whether the model’s performance and its explanations hold in different settings, while prospective evaluation would examine how the tool behaves in real clinics, with real time pressures and real therapeutic relationships.

Even with those caveats, the study offers a template that reaches far beyond perinatal mental health. It demonstrates that generative AI can be used to stretch small, culturally specific datasets without compromising honest evaluation, and that intrinsically interpretable models can deliver competitive accuracy while remaining auditable by clinicians and communities. As health systems worldwide grapple with deploying AI in sensitive domains, the example set by Aboriginal mothers, researchers and clinicians in Western Australia suggests a different path from the black-box norm: one where the world’s oldest living cultures help define what trustworthy, culturally safe artificial intelligence should look like, and where the measure of success is not just a high AUC but a better, more respectful conversation between a mother and those caring for her.

Subject of Research: Interpretable generative AI for culturally safe perinatal mental health prediction among Aboriginal mothers

Article Title: The world’s oldest living culture leads the way with interpretable generative AI: Aboriginal women shaping the future of perinatal mental health services

Article References: Wang, G., Bennamoun, H., Kwok, W. H., Raza, R., Sohel, F., Kelly, B., Ratajczak, T., Reid, C., Marriott, R., Walker, R., Kotz, J., & Reibel, T. (2026). The world’s oldest living culture leads the way with interpretable generative AI: Aboriginal women shaping the future of perinatal mental health services. PLOS Digital Health, 5(10), e0001709. https://doi.org/10.1371/journal.pdig.0001709

Image Credits: AI Generated

DOI: 10.1371/journal.pdig.0001709

Keywords: Aboriginal health, perinatal mental health, interpretable AI, generative AI, Explainable Boosting Machine, data augmentation, TVAE, Baby Coming You Ready, cultural safety, machine learning, PLOS Digital Health, Western Australia

The clinical stakes are considerable. Brain tumors produce symptoms that vary wildly depending on their type, grade, and location, ranging from headaches and seizures to motor impairment, sensory loss, and personality changes. Despite decades of progress in treatment, the five-year survival rate for malignant central nervous system tumors in the United States remains just 35.7 percent, a figure that underscores how much depends on early and accurate characterization. Radiologists reading MRI scans must first determine whether anything is wrong and then distinguish between tumor types that can look deceptively similar on T1-weighted images, particularly gliomas and meningiomas, which share overlapping morphological features that challenge even experienced eyes.

The researchers, led by Yasir Hussein Shakir with colleagues across institutions in Malaysia, Iraq, and Oman, identified four recurring weaknesses in the existing literature. Most published models handle either binary screening or multiclass classification, but not both, forcing hospitals to maintain and deploy two separate pipelines. Few models offer any visual explanation for their predictions, a roadblock to clinical trust. Accurate models rarely ship with usable software that a radiologist could actually integrate into a workflow. And nearly all prior studies validate on a single dataset, leaving generalizability an open question. The new system, called Modified DenseNet121, was designed to address all four gaps simultaneously.

Architecturally, the model starts from DenseNet121, a network whose defining feature is dense connectivity: every layer concatenates its feature maps with those of all preceding layers, enabling extensive feature reuse. The team chose this backbone because it carries fewer parameters than comparable ResNet designs and is less prone to vanishing or exploding gradients when fine-tuned on relatively small medical image collections. They stripped the ImageNet-pretrained network’s top classification layers, applied global average pooling to obtain a 1024-dimensional feature vector, and then attached a custom regularization stack: a dropout layer with a rate of 0.50, a 128-unit dense layer with ReLU activation for feature compression, and batch normalization to stabilize activations across training batches.

From that shared representation, two output branches diverge. The first is a single-unit sigmoid layer that predicts the probability that a scan is abnormal. The second is a three-unit softmax layer that assigns probabilities to glioma, meningioma, and pituitary tumors. A clever labeling strategy keeps the two tasks from interfering: every image contributes to the binary loss, but only abnormal images contribute to the multiclass loss, enforced through a binary mask that zeroes out the multiclass loss for healthy scans. This masking prevents the model from being misled by meaningless tumor labels on healthy subjects while allowing the shared backbone to learn robust representations for both tasks. The combined loss was optimized with the Adamax optimizer at a learning rate of 0.0001, a batch size of 32, and early stopping after ten epochs without validation improvement.

The benchmarking was unusually thorough. The team trained and evaluated the model against six competing architectures, including the unmodified DenseNet121 baseline, ConvNeXtSmall, ConvNeXtTiny, EfficientNetB0, NASNetMobile, and Xception, all pretrained on ImageNet and fine-tuned under identical preprocessing and evaluation conditions. Two large public MRI datasets served as the testbeds: a Kaggle collection of 7,023 T1-weighted contrast-enhanced scans and a Mendeley collection of 12,064 scans, both balanced across glioma, meningioma, pituitary, and normal classes. On the first dataset, Modified DenseNet121 achieved 99.85 percent binary accuracy and 99.00 percent multiclass accuracy, the best multiclass result overall, edging out the baseline by 0.28 percent. On the second, it reached 99.62 percent binary and 99.00 percent multiclass accuracy, statistically comparable to the baseline while retaining the advantage of a single dual-output framework.

Perhaps the most striking result came from a zero-shot evaluation on a third, entirely unseen dataset of 3,264 images. The model, trained on the merged 19,087 images from the first two datasets, achieved 99.75 percent binary accuracy and 99.20 percent multiclass accuracy without any additional fine-tuning, with bootstrap resampling over 1,000 iterations yielding narrow 95 percent confidence intervals. The authors are careful, however, to interpret this conservatively: because no dedicated image-level overlap analysis was performed between the public datasets, the result demonstrates external transfer performance rather than definitive proof of complete dataset independence. An ablation study further showed that while individual regularization components did not uniformly help, the full configuration delivered the best overall trade-off, with dropout removal causing the largest multiclass accuracy drop.

Interpretability comes through Gradient-weighted Class Activation Mapping, or Grad-CAM, a technique that computes the gradients of the predicted class score with respect to the feature maps of the network’s final convolutional block. Those gradients are globally average-pooled to produce importance weights for each feature map, which are then combined into a coarse localization map, resized to the original image resolution, and overlaid on the MRI with a jet colormap at 40 percent transparency. When a scan is classified as normal, no heatmap is generated; when abnormal, the heatmap highlights the tumor-relevant regions that drove the decision. In the published examples, the activations land where radiologists would expect: on an extra-axial hemispheric mass for a meningioma case, on a large intra-axial frontal-parietal lesion for a glioma, and on the sellar region for a pituitary tumor, while healthy scans show cold, inactive maps.

The entire system is wrapped in a PyQt5 desktop application that functions as a research prototype rather than a certified clinical device. A clinician uploads a scan and receives both predictions at once, along with softmax confidence scores and the Grad-CAM overlay. Measured on a laptop with an AMD Ryzen 7 5800H processor and an NVIDIA RTX 3070 GPU, typical inference latency ranged from 80 to 130 milliseconds, with a median of 127.99 milliseconds and end-to-end latency of roughly 217 milliseconds including image loading, preprocessing, and heatmap rendering, though occasional outliers reached several seconds. Crucially, the interface includes agree and disagree buttons that log physician feedback, along with filenames and probability scores, into a CSV file, creating a curated dataset for future active learning and model refinement.

The authors are candid about the limits of what they have built. The evaluation relied exclusively on retrospective, publicly available curated data, not on real clinical workflows, and no formal usability testing with radiologists has been conducted. The Grad-CAM visualizations were judged subjectively rather than quantitatively against radiologist-annotated tumor masks, and WHO tumor grading was not addressed. The team outlines a roadmap for the future that includes prospective multicenter validation with de-identified patient data, extension to tumor grading and pixel-level segmentation, quantitative attribution validation, cryptographic and perceptual hashing to formally verify image-level independence between datasets, and eventual web-based deployment toward regulatory readiness. For now, the work stands as a rigorously benchmarked, reproducible, and interpretable foundation, one that brings AI-based brain tumor diagnosis a tangible step closer to the radiologist’s desk.

Subject of Research: Dual-task deep learning with Grad-CAM interpretability for brain tumor classification from MRI scans

Article Title: A modified dual-task DenseNet121 with Grad-CAM interpretability for brain tumor classification from MRI

Article References: Shakir, Y. H., Hasan, A. A., Al-Yasiri, A. Y., Al Mandhari, E. A. A., Abdulnabi, M., Al-Hadabi, A. S. Y., Alghatrifi, I., & Al Mamari, W. H. S. (2026). A modified dual-task DenseNet121 with Grad-CAM interpretability for brain tumor classification from MRI. Discover Artificial Intelligence, 6(1), Article 1502. https://doi.org/10.1007/s44163-026-02467-7

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02467-7

Keywords: brain tumor classification, deep learning, DenseNet121, Grad-CAM, MRI, dual-task learning, transfer learning, interpretable AI, PyQt5, neuro-oncology, computer-aided diagnosis, convolutional neural networks

The clinical motivation is stark. The 2021 World Health Organization classification describes more than 100 distinct tumor entities of the central nervous system, and even with intensive multimodal treatment the five-year survival rate for high-grade gliomas remains below 10 percent. Interpreting MRI scans is time-consuming, and inter-observer variability among radiologists can reach 30 percent in challenging cases, while specialist expertise is unevenly distributed across the globe. Deep learning has repeatedly posted impressive numbers on standardized brain tumor benchmarks, yet three persistent gaps have blocked clinical translation: models are computationally heavy, they are rarely validated on data from different scanners and protocols, and their softmax outputs are notoriously overconfident on ambiguous or out-of-distribution inputs.

The new architecture attacks all three weaknesses at once. Its backbone is EfficientNet-B3, pretrained with the Noisy Student semi-supervised procedure on ImageNet-1K and the unlabeled JFT-300M corpus, which contributes roughly 10.28 million of the model’s 10.97 million total parameters. That figure is 7.8 times smaller than the 86 million parameters of ViT-B/16 and about 92 percent smaller than VGG-16’s 134 million, making the system plausible for deployment in resource-constrained clinical settings where giant models are impractical.

The architectural centerpiece is the Adaptive Multi-Scale Module, or AMSM, inserted after each of three hierarchical feature stages of the backbone. Tumors vary enormously in spatial extent: an infiltrative glioma spreads across a wide region, while a small pituitary adenoma may occupy only a few pixels in the sellar region. AMSM handles this diversity with three parallel depthwise-separable dilated convolutional branches at dilation rates 1, 2, and 4, each enlarging the receptive field without losing resolution, plus an identity branch that preserves the unmodified features. A small two-layer network reads globally pooled descriptors of all four branches and emits softmax gating weights that sum to one, so the fused output is a learned convex combination of multi-scale evidence rather than a fixed aggregation. Because depthwise-separable factorization requires only C times the sum of the kernel size squared and the channel count, instead of the channel count squared times kernel size squared of a standard convolution, the module costs roughly 6.6 times fewer parameters than equivalent conventional branches: 692,490 parameters instead of 4,546,320.

The second innovation concerns trust. Instead of a plain softmax head, the network ends in an Evidential Deep Learning classifier that recasts classification as estimation of a Dirichlet distribution. Logits pass through a softplus function to produce non-negative evidence, Dirichlet parameters are set to evidence plus one, and the uncertainty mass, or vacuity, is defined as the number of classes divided by the Dirichlet strength. Large accumulated evidence yields low uncertainty; thin evidence yields a flag. Training combines an evidential cross-entropy built on the digamma function, a KL divergence term that pulls non-evidence entries toward a uniform Dirichlet prior and is annealed over the first ten epochs, and a focal correction with class weights to handle imbalance. The result is deterministic, single-pass uncertainty quantification at an inference cost of 16.8 milliseconds per image, roughly 20 times faster than MC-Dropout with 20 stochastic forward passes, which took 340.5 milliseconds.

Performance on the primary benchmark, BRISC 2025, a 6,000-image expert-annotated T1-weighted dataset derived from BraTS 2021 with official splits of 5,000 training and 1,000 test images, was 99.20 percent accuracy and 99.30 percent macro-F1, with an AUC-ROC of 0.9995 and only 0.95 gigaflops of compute. Five-fold cross-validation returned a mean accuracy of 99.06 percent with a standard deviation of 0.17 percent, and bootstrap confidence intervals on the test set spanned 98.6 to 99.7 percent. Of the eight errors among 1,000 test images, seven carried uncertainty scores in the 90th percentile of the test distribution, and the model achieved perfect recall on the No Tumor class, meaning no normal scan was falsely flagged as cancerous. Misclassified samples carried 2.6 times higher mean uncertainty than correctly classified ones, and the misclassification detection AUROC reached 0.9461.

Perhaps the most scientifically interesting result is a negative one. During development the team built a Dual Attention Module combining efficient channel attention and spatial attention and tested four architectural variants. In isolation the attention module improved in-domain calibration and achieved the highest uncertainty separation ratio at 2.94 times, but stacking it after AMSM degraded both in-domain accuracy and, more damagingly, cross-dataset generalization on the external PMRAM-BD cohort, a brain MRI collection from Bangladeshi clinical centers acquired under a different protocol. The authors concluded that channel-spatial reweighting after scale-gated fusion encourages dataset-specific feature emphasis at the expense of the domain-invariant spatial representations that AMSM captures, and they excluded the module from the deployed model. It is a rare public example of a deep learning paper arguing against adding more attention.

Cross-dataset evaluation is where the work departs from most of the literature. Without any fine-tuning, the model trained only on BRISC 2025 scored 97.41 percent on PMRAM-BD, a generalization gap of 1.79 percentage points that is smaller than those of ResNet-50, VGG-16, and the EfficientNet-B3 backbone alone, though not smaller than those of ViT-B/16, Swin-Tiny, or DenseNet-121, which achieved the highest absolute external accuracy of 98.07 percent in a single run. The authors are careful here: they frame their model as a strong external performer rather than the best external generalizer, and multi-seed testing showed DenseNet-121’s external advantage was fragile, dropping to 94.88 percent when averaged over three random seeds while the proposed model stayed above 96.35 percent. VGG-16 fared worst, losing 5.94 percentage points, evidence that sheer parameter count does not buy cross-domain robustness. Notably, mean predictive uncertainty rose significantly on the external cohort, with a rank-sum test p-value of 0.000064, showing the model signals reduced confidence exactly when it faces an unfamiliar imaging distribution.

The paper is equally candid about limitations. McNemar’s paired tests against six baselines yielded no statistically significant differences after Holm-Bonferroni correction, which the authors attribute to the near-ceiling performance: at 98.7 to 99.2 percent accuracy on a balanced 1,000-sample test set, only 6 to 11 prediction pairs differ between any two models, so the test cannot resolve gaps below roughly half a percentage point. The evidential model’s expected calibration error of 0.0263 is higher than the softmax baselines, a known property of Dirichlet-parameterized models under standard confidence binning. Classification is performed on 2D slices without volume information, both datasets lack patient-level identifiers so all metrics are image-level, and inference at 19.2 milliseconds per image is slower than ResNet-50 because of an inefficient CUDA implementation of depthwise convolution rather than higher computational load. Interpretability comes from UA-GradCAM heatmaps and a novel Multi-Scale Attention Decomposition that logs the gating weights per class: glioma predictions preferentially activate the rate-4 dilated branch, consistent with infiltrative growth, while pituitary predictions favor the rate-1 branch, consistent with small focal lesions, a pattern the authors compare to a radiologist reporting whether local enhancement or regional mass effect supports a diagnosis.

The authors position the system for retrospective clinical investigation rather than immediate deployment, noting that prospective validation across multiple sites with radiologist consensus as ground truth remains necessary. Even so, the combination of near-state-of-the-art accuracy, a 10.97 million parameter footprint, deterministic single-pass uncertainty, and rigorous multi-cohort evaluation offers a template for what clinically credible medical AI should look like. The code, notebooks, and experimental results are publicly available on GitHub, and the open datasets, BRISC 2025 on Kaggle under CC BY 4.0 and PMRAM-BD on Mendeley Data, mean any laboratory can reproduce the numbers. In a field often criticized for leaderboard chasing and overconfident claims, a model that knows what it does not know, and a paper that says so, may be the most viral result of all.

Subject of Research: Uncertainty-aware deep learning classification of brain tumors from MRI scans

Article Title: A hybrid scale attention network for uncertainty aware brain tumor MRI classification

Article References: Josh, T. J., Mim, M. M. R., Joy, M. A. R., Assaduzzaman, M., Moon, M. M. H., Noori, S. R. H., Fahad, N., & Liew, T. H. (2026). A hybrid scale attention network for uncertainty aware brain tumor MRI classification. Discover Artificial Intelligence, 6(1), Article 1493. https://doi.org/10.1007/s44163-026-02386-7

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02386-7

Keywords: brain tumor, MRI, deep learning, uncertainty quantification, evidential deep learning, EfficientNet, multi-scale fusion, attention mechanisms, cross-dataset validation, medical imaging AI, lightweight neural network, interpretable AI

The work, carried out by Simone Amirato, Fabrizio Angiulli, and Fabio Fassetti of the University of Calabria in Italy, addresses what the authors call the Anomaly Explanation problem. Their idea is deceptively simple: an explanation of an anomaly is a compact subset of features, a subspace, in which the anomalous examples deviate most sharply from the normal population. If a fraud-detection system flags a transaction, the explanation might boil down to three or four attributes, such as transaction amount, merchant category, and time of day, that together make the record look nothing like ordinary behavior. The difficulty is that finding the best such subset is a combinatorial nightmare. With d features, the number of possible subsets grows as the power set of the feature set, a search space so vast that exhaustive enumeration is hopeless even for moderately sized problems.

RLAX, short for Reinforcement Learning-based Anomaly eXplanation, recasts this search as a Markov Decision Process. Each state in the process is a candidate subspace, and each action corresponds to adding a single feature to the current selection. The agent starts from the empty set and progressively builds up an explanation, one feature at a time. What guides it is a reward signal derived from a quantity the authors call the Subspace Margin. This margin measures the difference between the anomaly scores of known outlier examples and known inlier examples within the current subspace, normalized by the square root of the subspace size. The anomaly score itself is a classic distance-based measure: the average distance from a point to its K nearest neighbors, which serves as a proxy for local density. When the margin grows after adding a feature, the agent earns a positive reward; when the margin shrinks, the reward is negative.

To learn a policy over this enormous action space, the team turned to Proximal Policy Optimization, or PPO, a policy-gradient algorithm that has become a standard tool in the reinforcement learning repertoire. The choice is deliberate. Value-based methods such as Deep Q-Networks struggle when the number of actions is very large, and in RLAX the number of actions equals the number of features, which can easily reach into the hundreds or thousands in real applications. Policy gradient methods, by contrast, handle large action spaces gracefully. The authors used the Stable-Baselines3 implementation of PPO with a custom actor-critic policy that masks inadmissible actions, keeping the default architecture of two 64-unit tanh layers for both the policy and value networks. Once training is complete, the learned policy is distilled into a ranking of features, either through the empirical mean of discounted rewards each action collected during training or directly from the policy’s action probabilities, and candidate subspaces are read off from the top of that ranking.

A key design element is the Explanation Gain, a measure that trades off separation accuracy against explanation size. Formally, the gain of a subspace is its separation quality minus a penalty proportional to the fraction of features it uses, with a tunable coefficient alpha that defaults to 0.5. This penalty is what pushes the agent toward compact explanations. The experimental analysis shows why it matters: the subspace that maximizes raw separation, measured by the Area Under the ROC Curve, may require more than 30 percent of the available features, whereas the best-gain solution often involves only a small fraction of them while preserving comparable discriminative performance. In other words, the framework resists the temptation to throw every feature at the problem and instead delivers explanations an analyst can actually read.

The journal version of the work extends an earlier conference paper with a Subspace Refinement Strategy that makes the exploration genuinely non-monotone. In addition to add actions, the agent can now remove features it previously selected or issue a stop action to terminate the episode early. This flexibility matters for two reasons. First, remove actions let the agent backtrack out of locally suboptimal regions of the search space, revising earlier selections that looked promising in isolation but hurt in combination. Second, the stop action prevents the agent from wandering through suboptimal territory merely to exhaust its step budget. The authors impose a budget on add operations to prevent loops, but leave remove operations unconstrained since they are inherently bounded by the number of prior additions. Episodes can end through the stop action, budget exhaustion, or a patience criterion that triggers after several consecutive negative rewards.

The empirical evaluation spans benchmark datasets from the ODDS repository, including arrhythmia with 274 features, mnist with 100 features, and several others, alongside two substantially larger real-world datasets: Internet Advertisements with 1,558 features and Swarm Behaviour with 2,400 features. Against competitors including V-GAN, SOD, HiCS, a genetic subspace-selection method, and random feature selection, RLAX achieved the highest Explanation Gain on every benchmark dataset, indicating the most favorable balance between discriminative power and compactness. Statistical significance against random selection was confirmed with one-sided paired t-tests adjusted using Holm’s correction. The refinement strategy also proved its worth in training dynamics: on representative datasets it produced markedly shorter episodes, saving roughly 30 percent of execution time while achieving improved AUC at the same feature subset sizes. On synthetic scalability tests, RLAX’s runtime grew only moderately, by factors of roughly 1.5 to 1.9 as the number of instances increased twentyfold, a growth rate substantially lower than that of the genetic and SOD baselines.

Perhaps the most striking evidence for the framework’s practical value comes from a qualitative case study on the cardio dataset, which contains cardiotocography measurements of fetal heart rate with well-established clinical semantics. For a representative anomaly, RLAX selected four top features: the number of prolonged decelerations per second, the histogram variance of the fetal heart rate, the percentage of time with abnormal long-term variability, and the number of fetal movements per second. Each of these maps directly onto recognized indicators of fetal distress in the obstetric literature. Prolonged decelerations are documented as strong markers of potential fetal hypoxia and acidosis, abnormal long-term variability is one of the four fundamental characteristics used to interpret a cardiotocography trace, reduced fetal movements are a classic clinical warning sign, and heart rate variance captures deviations in oscillation amplitude in either direction. An algorithm with no medical knowledge, in other words, recovered a clinically coherent explanation.

The authors are careful about the limits of their approach. Because RLAX explores the search space through single-feature modifications, strongly non-additive interactions between features, where an informative combination only pays off after several individually unrewarding additions, can be harder to discover, though the remove actions partially mitigate this. They also note that the method assumes access to a small set of labeled anomalies and inlier examples, which in practice may come from a detector’s output or from a domain expert. Future work, they write, will explore domain-informed reward shaping, alternative separation measures, and settings with heterogeneous features and weaker supervision. The source code has been released publicly, and all datasets used in the study are openly available.

What makes RLAX more than a technical curiosity is the broader shift it represents in how machine learning systems might communicate their judgments. As automated decisions increasingly touch medicine, security, and environmental monitoring, the demand is growing not just for accurate detectors but for systems that can articulate the basis of a flag in terms a human can verify. By demonstrating that a reinforcement learning agent can navigate a combinatorially explosive search space and return explanations that align with clinical knowledge, the Calabria team offers a concrete template for that kind of accountability. The anomalies that matter most are often the ones we cannot explain away, and a tool that points precisely at the handful of features responsible turns a cryptic alarm into the beginning of an investigation rather than the end of one.

Subject of Research: A reinforcement learning framework for explaining anomalies through compact feature subspaces

Article Title: RLAX: A Reinforcement Learning-Based Framework for Explaining Anomalous Data

Article References: Amirato, S., Angiulli, F., & Fassetti, F. (2026). RLAX: A Reinforcement Learning-Based Framework for Explaining Anomalous Data. Machine Learning, 115(10), Article 246. https://doi.org/10.1007/s10994-026-07182-z

Image Credits: AI Generated

DOI: 10.1007/s10994-026-07182-z

Keywords: reinforcement learning, anomaly explanation, explainable AI, feature selection, PPO, subspace margin, anomaly detection, machine learning, interpretable AI, outlier detection, cardiotocography, high-dimensional data

The tool, called SPACER, was developed by a large multi-institutional team led by researchers at UT Southwestern Medical Center, Baylor College of Medicine and UT MD Anderson Cancer Center. Its central innovation is a commitment to interpretability at a level that most machine learning systems abandon. Rather than burying its reasoning inside millions of opaque parameters, SPACER is built as a multi-instance learning framework that outputs gene-level recruitment rules directly from spatial transcriptomics data. In practical terms, a researcher can point the model at a tissue slide and receive back a ranked list of genes that either attract or repel a given cell type, with the logic connecting each gene to each cellular encounter laid bare for inspection and experimental testing.

The technical challenge the team confronted is worth appreciating. Cell recruitment in tissue is governed by an interplay of chemoattractants, adhesion molecules, extracellular matrix barriers, antigen presentation and countless other cues, all acting across microscopic distances. Standard computational approaches to spatial data tend to describe correlations between cell types and their surroundings, or to infer ligand–receptor communication from bulk expression patterns, but they rarely recover the causal grammar of recruitment. SPACER reframes the problem as one of learning from many small spatial instances—local patches of tissue in which a target cell either engaged with or was excluded from a particular neighborhood—and extracting the gene programs that consistently distinguish successful recruitment from exclusion. The result is a model whose internal logic maps onto biological mechanisms that experimenters can actually interrogate.

To establish credibility, the researchers benchmarked SPACER against existing methods across an unusually broad foundation of evidence: 37 spatially resolved transcriptomics datasets spanning multiple human tumor types as well as mouse heart tissue. Across a range of tasks designed to test whether a method can recover biologically meaningful recruitment signals, SPACER delivered improvements of fourfold to twenty-one-fold over prior approaches. That margin is not a marginal tuning gain; it reflects a qualitative difference in what the model can extract. The team also validated the framework against complementary data types, including single-cell RNA sequencing paired with T cell receptor sequencing, immunopeptidomics datasets cataloguing the peptide fragments displayed on cell surfaces, and newly generated Visium HD maps of mouse heart tissue.

The first major biological finding concerns mucins, the large, heavily glycosylated proteins best known for forming the protective mucus layer of airways and gut. SPACER’s analysis indicated that mucins expressed by tumor cells are associated with physical barriers that prevent T cells from engaging with the cancer cells around them. This is a mechanistically plausible picture: mucins can dramatically alter the biophysical properties of a cell’s surface and its surrounding matrix, and prior work has implicated them in shielding tumor cells from immune surveillance. What SPACER adds is a systematic, data-driven link between tumor mucin expression and the spatial exclusion of T cells across cancer types—a link that had been suspected but never decoded at this scale from spatial data.

The clinical implications of that finding are immediate. When the team examined cohorts of patients treated with immune checkpoint inhibitors—the blockbuster immunotherapy drugs that unleash T cells against tumors—mucin-associated recruitment barriers predicted resistance to treatment. The same signal also correlated with patient prognosis in pan-cancer analyses of The Cancer Genome Atlas, including in mesothelioma and kidney renal clear cell carcinoma. In other words, a computational readout of why T cells fail to reach tumor cells, derived purely from spatial gene expression maps, carries information that could in principle stratify patients before they embark on a course of therapy. That is precisely the kind of translational bridge that spatial biology has promised but rarely delivered.

The second headline discovery concerns what actually draws T cells into productive contact with tumor cells. Across cancer types, SPACER identified local antigen presentation as the dominant correlate of T cell engagement. This finding resonates with a core principle of immunology—T cells recognize their targets through peptide fragments displayed on major histocompatibility molecules—but seeing it emerge as the strongest spatial signal, outcompeting chemokine gradients and other candidate attractants, gives the principle new quantitative weight. The team reinforced the point with immunopeptidomics data, confirming that the peptides presented by tumor cells align with the T cell specificities inferred from receptor sequencing. Recruitment, in this view, is not a generic homing process but a antigen-informed negotiation between the immune cell and its immediate environment.

The third application took the team out of oncology and into a dangerous side effect of cancer therapy. Immune checkpoint inhibitors occasionally trigger myocarditis, a potentially fatal inflammation of the heart muscle, and which immune cells drive the damage has been debated. Applying SPACER to spatial maps of mouse heart tissue from a model of checkpoint inhibitor–induced myocarditis, together with single-cell RNA and T cell receptor sequencing, the researchers found that CD4-positive T cells—helper T cells more often associated with coordination than killing—are the dominant functional effectors in the severe form of the disease. Heart-infiltrating T cell clonotypes, expanded and largely distinct from those circulating in the blood, pointed to a local antigen-driven response within the myocardium, sharpening a picture previously suggested by work on α-myosin–reactive T cells.

Beyond these specific discoveries, the study sketches a broader vision for how interpretable machine learning could reshape tissue biology. The authors and their collaborators have made SPACER freely available through GitHub, and the underlying datasets—including human Visium and Visium HD data, gastric cancer CosMx maps, Xenium data paired with T cell receptor sequencing, and the mouse heart resources—are deposited in public repositories such as NIH GEO and Zenodo. That openness matters. A discovery engine for recruitment biology only earns trust if its rules can be independently tested, and the interpretability that defines SPACER is designed exactly for that cycle of computational prediction and experimental verification. The framework also proved computationally efficient, an unglamorous but essential quality for a tool intended to be applied across the rapidly growing volume of spatial datasets now being generated worldwide.

What makes the work resonate beyond its immediate findings is the way it reframes the relationship between artificial intelligence and biology. Foundation models for spatial omics are proliferating, and they are impressive at prediction, but their black-box character limits their scientific usefulness when the goal is mechanism. SPACER demonstrates the opposite philosophy: build interpretability in from the ground up, and the model becomes a hypothesis generator that hands experimenters concrete, testable molecular rules. The mucin barrier hypothesis, the centrality of local antigen presentation, and the CD4-driven pathology of immunotherapy myocarditis are all claims that can now be pursued in the laboratory—and, if they hold up, translated into biomarkers and therapeutic strategies. In decoding the spatial codes of cell recruitment, the study suggests that the most valuable AI in biomedicine may be the kind that shows its work.

Subject of Research: Interpretable machine learning for decoding cell recruitment rules from spatially resolved transcriptomics data

Article Title: Interpretable AI model reveals spatial codes of cell recruitment in tissues

Article References: Yao, J., Shim, J. S., Wong, K., Zhu, J., Liu, Y., Kang, S. W., McCord, K., Burt, B. M., Choi, J. M., Lee, C., Lubo Julio, I. C., Solis Soto, L. M., Hernandez, S., Colbert Maresso, K., Liu, Y., Dai, Y., Chu, T., Lu, W., Khan, K., … Wang, T. (2026). Interpretable AI model reveals spatial codes of cell recruitment in tissues. Nature Biomedical Engineering. https://doi.org/10.1038/s41551-026-01804-6

Image Credits: AI Generated

DOI: 10.1038/s41551-026-01804-6

Keywords: spatial transcriptomics, interpretable AI, SPACER, T cell recruitment, tumor microenvironment, mucins, immune checkpoint inhibitors, antigen presentation, myocarditis, CD4 T cells, multi-instance learning, cancer immunotherapy