PLOS Computational Biology

Evidence accumulation is a fundamental aspect of human decision-making. However, how the precise temporal structure of evidence shapes the accumulation process has not been systematically studied. As a result, current understanding of evidence accumulation remains largely limited to its time-averaged behavior. We tested human subjects in a visual estimation task in which they inferred the angular position of an unknown source from a noisy stimulus sequence. Introducing systematic temporal perturbations, i.e., breaks of different durations and at different positions in the otherwise regular evidence sequence, revealed that subjects actively compensated for the memory loss endured during the break by dynamically enhancing evidence integration and memory maintenance immediately after the break. We derived a new time-continuous Bayesian updating model that is dynamically constrained by optimal performance-effort trade-offs. With two free parameters determining the overall resource-efficiencies of encoding and memory maintenance, the model accurately predicts the rich dependencies of subjects accumulation behavior on the evidence schedule, including subjects individual tendencies to emphasize either early (primacy) or late (recency) samples in the evidence sequence. Our results suggest that evidence accumulation is a non-stationary, dynamically controlled process that optimally balances the information gained from incoming evidence against the cognitive effort required to acquire and maintain it. The proposed model is general and should apply broadly across many task domains.

Connectome-constrained neural network models aim to link anatomical connectivity with functional computation by training networks whose architectures reflect biological circuits. Because such models are increasingly used to infer neural mechanisms, it is important to assess their robustness to variations in training conditions and model selection criteria. Here we retrain ensembles of connectome-constrained models under nominally identical conditions and compare their correspondence to experimentally measured response properties in the Drosophila motion pathway. While task performance remains similar across models, the identification of biologically plausible circuit solutions is unstable across retraining runs. In particular, model clusters selected by lowest validation task error do not reliably correspond to experimentally observed neural tuning, and small variations in performance metrics can reorder cluster rankings. These results indicate that, in this framework, similar task performance does not reliably identify biologically plausible circuit solutions. Task error alone is therefore insufficient for mechanistic identification, and additional model-selection criteria are needed.

Hidden (semi-) Markov Models (HsMMs) are increasingly being used to segment neurophysiological signals into sequences of latent cognitive processes. The idea: different processes will leave distinct traces in trial-level recordings of (multivariate) neuro-physiological signals. Markov models, equipped with an emission model of these traces and a latent process model describing the progression through the different latent processes involved in a task, can then be used to infer the most likely process for any time-point and trial. However, the currently used HsMMs remain limited in two important ways. First, they cannot account for subject-level heterogeneity in the latent and emission process. Instead, a single group-level model is assumed to explain the entire data. Second, they cannot account for the potentially non-linear effects of experimental covariates on the latent and emission process. To address these problems, we present a modeling framework in which the HsMM parameters of the emission and latent process are replaced with mixed additive models, including smooth functions of experimental covariates and random effects. We derive all necessary quantities for empirical Bayes and fully Bayesian inference for all parameters and provide a Python implementation of all estimation algorithms. To demonstrate the advantages offered by this framework, we apply such a multi-level model to an existing lexical decision dataset. We show that, even in such a simple task, not all subjects rely on the same processes equally and that at least two semi-Markov states, previously believed to reflect distinct processes, might actually relate to the same cognitive process.

Machine learning models have seen a lot of success in medical image segmentation domain. However, one of the challenges that they face are confounders or shortcuts: spurious correlations or biases in the training data that affect the resulting models. One example of such confounders for surgical machine learning is the setup of surgical equipment, including tools and lighting. Using the task of identification of safe and dangerous zones of dissection in laparoscopic cholecystectomy images and videos as a use-case, we inspect two equipment-induced biases: the presence of surgical tools in the field of view and the position of lighting. We propose methods for evaluating the severity of these biases and augmentation-based methods for mitigating them. We show that our tool bias mitigations improve the models' consistency under tool movements by 9 percentage points in the most inconsistent cases, and by 4 percentage points on average. Our lighting bias mitigations help reduce fraction of true dangerous zone pixels that may be predicted as safe under light changes from 5% to 1.5%, without compromising segmentation quality.

Interpreting serological data remains challenging, particularly in low prevalence or cross reactive contexts, where antibody responses often show substantial overlap between exposed and unexposed individuals and may depart from normal distributional assumptions. Conventional cutoff based approaches often yield inconsistent or biased estimates of seroprevalence. Here, we present a decisional framework based on finite mixture models (FMMs) that enhances the robustness and interpretability of serological analyses. Beyond simply applying mixture models, our framework integrates multiple methodological innovations : (i) systematic comparison of Gaussian and skew normal mixture models to accommodate asymmetric antibody distributions; (ii) rigorous model selection using the Cramer von Mises test (p > 0.01) combined with a parsimonious score (APS) to prioritize models with well separated clusters; and (iii) hierarchical clustering of posterior probabilities to collapse latent components into biologically meaningful seronegative and seropositive groups. Applied to chikungunya virus (CHIKV) data from Bangladesh, the framework produced prevalence estimates consistent with ROC based methods while probabilistically identifying borderline cases. Validation on SARS CoV 2 and dengue datasets further demonstrated its generalizability: for SARS CoV 2, the approach identified up to five latent clusters with high sensitivity (up to 100%) and specificity (up to 100%), enabling discrimination by disease severity. For dengue, it revealed interpretable subgrouping consistent with background exposure and subclinical infection, despite limited confirmed cases. By integrating distributional flexibility, robust goodness of fit testing, and biologically guided cluster consolidation, this decisional FMM framework provides a reproducible and scalable method for serological interpretation across pathogens and epidemiological settings, addressing key limitations of threshold based classification.

PurposeColonic motility disorders, including diarrhea-predominant irritable bowel syndrome and slow-transit constipation, impose a major clinical burden. Although high-resolution colonic manometry reveals characteristic spatiotemporal motor patterns, such as high-amplitude propagating contractions and cyclic motor pattern in healthy individuals, these patterns are often altered or absent in disease. Understanding how these patterns arise from underlying pacemaker, neural, and mechanical mechanisms is essential for improving treatment strategies. MethodsWe developed a biophysical whole-colon model that integrates an Interstitial Cells of Cajal-inspired oscillator network, enteric nervous system reflexes, a pressure-gated modulation element motivated by rectosigmoid brake behavior, and a nonlinear tube law describing colon wall mechanics. The model simulates spatiotemporal pressure patterns along the colon and allows systematic variation of physiological parameters associated with pacemaker activity, neural reflex control, and distal gating. ResultsA small set of parameters reproduces three illustrative motility patterns corresponding to healthy motility, diarrhea-predominant irritable bowel syndrome, and slow-transit constipation. The simulated pressure maps recapitulate key features observed in high-resolution manometry, including propagation direction, regional patterning of contractions, and case-specific changes in amplitude and coordination. Sensitivity analysis suggests that proximal excitation strength and waveform morphology strongly influence global motility metrics. ConclusionOur study presents a simple, biophysical framework for reproducing clinically observed colonic motor patterns and exploring their disruption in disease. More broadly, the model may help interpret clinical manometry in mechanistic terms and support hypothesis-driven in silico studies of colonic motility disorders.

Patterns in the vegetation across arid and semiarid regions may be explained as a form of self-organization driven by water scarcity, and are often modeled through reaction-diffusion dynamics. Recent work has shown that similar mathematical models generate patterns on networks. However, these studies have focused on idealized topologies with no reference to natural pattern-forming systems. Our study aims at bridging these two fields: we employ a physical reaction-diffusion vegetation model, and gradually modify the topology of the diffusion network by adding random shortcuts over a 2-dimensional grid, interpolating between a regular lattice and a random network. We found that network topology strongly shapes both the resulting vegetation patterns and the precipitation range that supports them. Three behavioral regimes emerge. On a regular lattice, high-regularity patterns develop reflecting local diffusion processes. On a random network, the system is dominated by global pressure towards homogenization yielding either a uniform state or a single patch. In the intermediate shortcut density range, as the network topology resembles a small world network, the interaction between the two scales of diffusion generates two kinds of disordered patterns: low-regularity patterns with a well-defined characteristic wavelength, and irregular patterns characterized by a broad patch size distribution. These disordered patterns resemble real-world observations and, in our model, they show different responses to changing precipitation. Although we focused on dryland vegetation, we suggest that network-mediated diffusion could lead to similar mechanisms in a wide variety of pattern-forming systems. HighlightsO_LIWe study vegetation pattern formation over different diffusion network topologies. C_LIO_LITwo kinds of stable disordered patterns states develop over small world topologies. C_LIO_LILow-regularity patterns with a well-defined characteristic wavelength. C_LIO_LIIrregular patterns characterized by a broad patch size distribution. C_LIO_LIThese different kinds of disordered states show different relations to precipitation. C_LI

Synaptic plasticity is metabolically expensive, yet animals continuously update their internal models without exhausting energy reserves. However, when artificial neural networks are trained, the network parameters are typically updated on every sample that is presented, even if the sample was classified correctly. Inspired by the human negativity bias and error-related negativity, we propose memorized mistake-gated learning--a biologically plausible plasticity rule where synaptic updates are strictly gated by current and past classification errors. This reduces the number of updates the network needs to make by 50% [~] 80%. Mistake gating is particularly well suited in two cases: 1) For incremental learning where new knowledge is acquired on a background of pre-existing knowledge, 2) For online learning scenarios when data needs to be stored for later replay, as mistake-gating reduces storage buffer requirements. The algorithm can be implemented in a few lines of code, adds no hyper-parameters, and comes at negligible computational overhead. Learning on mistakes is an energy efficient and biologically relevant modification to commonly used learning rules that is well suited for continual learning.

Knowing the mortality rates associated with infection by a pathogen is essential for effective preparedness and response. Here, harnessing the flexibility of a Bayesian approach, we produce an estimate of the Infection Fatality Ratio (IFR) for A(H5N1) conditional on explicit assumptions, and quantify the uncertainty thereof. We also apply the method to first-wave COVID-19 data up to March 2020, demonstrating the estimates that could be obtained were the model available then. Our analysis uses World Development Indicators (WDI) from the World Bank, the A(H5N1) WHO confirmed cases and deaths tracker by country (2003-2024), and COVID-19 cases and deaths data from John Hopkins University (January and February 2020). Since infectious disease dynamics are typically influenced by local socio-economic factors rather than political borders, individual countries are placed within clusters of countries sharing similar WDIs relevant to respiratory viral diseases, with clusters derived by performing Hierarchical Clustering. To estimate the IFR, we fit a Negative Binomial Bayesian Hierarchical Model for A(H5N1) and COVID-19 separately. We explicitly modelled key unobserved parameters with informative priors from expert opinion and literature. By modelling underreporting, our analysis suggests lower fatality (15.3%) compared to WHO's Case Fatality Ratio estimate (54%) on lab-confirmed cases. However, credible intervals are wide ([0.5%, 64.2%] 95% CrI). Therefore, good preparedness for a potential A(H5N1) pandemic implies adopting scenario planning under our central estimate, as well as for IFRs as high as 70%. Our approach also returns a COVID-19 IFR estimate of 2.8% with [2.5%, 3.1%] 95% CrI which is consistent with literature.

ObjectivePathological beta-band oscillations (13 to 30 Hz) in the subthalamic nucleus (STN) are a hallmark of Parkinsons disease and a primary target for deep brain stimulation therapy, yet the specific pattern of synaptic reorganization that drives their emergence remains incompletely understood. We developed a GPU-accelerated computational framework to systematically investigate combinations of synaptic changes across basal ganglia pathways that produce Parkinsonian beta oscillations while satisfying literature-based electrophysiology constraints. ApproachWe implemented a biophysically detailed spiking network model of the STN, external globus pallidus (GPe), and internal globus pallidus (GPi) in JAX (a high-performance numerical computing Python library), achieving a 490-fold speedup over conventional CPU-based simulation. Using the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) we optimized 10 network parameters across two stages: first establishing a healthy baseline matching primate electrophysiology data, then searching within biologically motivated bounds for synaptic modifications that reproduce Parkinsonian firing rates and beta power. Fixed in-degree connectivity ensured optimized parameters produced scale-invariant dynamics from 450 to 45000 neurons. All simulations ran on a single cloud GPU instance at 84 cents per hour. Main ResultsThe optimizer converged on a coordinated pattern of synaptic reorganization dominated by asymmetric changes within the STN-GPe reciprocal loop: STN to GPe excitation increased 2.21-fold while GPe to STN inhibition collapsed to 0.11-fold of its healthy value. STN to GPi and GPe to GPi pathways changed minimally (1.06-fold and 1.45-fold respectively). This configuration transformed asynchronous firing (beta: 0.4 percent of spectral power) into synchronized bursting with prominent beta oscillations (49.4 percent), with firing rate changes matching experimental observations. Network dynamics were invariant across a 100-fold range of network sizes (firing rate deviation less than 2.4 Hz; all metrics p less than 0.001 across 10 random seeds at 45000 neurons). We implemented a simplified deep brain stimulation model for validation purposes, which achieved complete beta suppression (49.4 percent to 0.0 percent) and restored GPi output to healthy levels. SignificanceThese results suggest that pathological beta oscillations emerge from a specific pattern of synaptic reorganization, namely the reduction of GPe inhibitory feedback to STN. The GPU-accelerated optimization framework, running on commodity cloud infrastructure, demonstrates an accessible platform for parameter exploration in neural circuit models and a foundation for generating synthetic training data for adaptive deep brain stimulation algorithms.

Understanding how genetic variation in pathogens influences clinical phenotypes observed in infected hosts is a fundamental challenge in evolutionary genomics and public health. Phenotypic traits such as infection severity are often non-randomly distributed within the pathogens phylogeny, suggesting the existence of evolutionary determinants but also violating the independence assumption underlying classical genome-wide association studies and potentially leading to inflated false positive rates. We present MutaPhy, a phylogeny-based method aimed at detecting correlations between a binary host phenotype and the corresponding pathogen genome by directly utilizing the hierarchical structure of phylogenetic trees. MutaPhy encompasses three different scales: (i) a subtree scale, on which relevant clades over-representing the phenotype of interest are detected using permutation-based tests; (ii) a tree scale, which agglomerates local signals into a global association statistics; and (iii) a site scale, whereby candidate mutational events on branches leading to significant clades are examined using ancestral sequence reconstruction. We evaluate the statistical behavior and detection performance of MutaPhy using simulations under diverse evolutionary scenarios. We also compare this tool to several existing phylogenetic association methods. As illustrative applications, we apply MutaPhy to dengue virus and hepatitis C virus datasets associated to clinical phenotypes in human hosts. Our results highlight the ability of the proposed approach to detect viral lineages associated to over-represented phenotypes while revealing limited evidence for robust mutation-level associations in these particular datasets. Altogether, MutaPhy provides a framework for guiding genotype-phenotype association analyses by leveraging phylogenetic structure, thereby reducing false positive findings and improving the interpretability of association signals.

Emerging multi-omic profiling has made it feasible to subtype disease using multiple molecular layers. However, inconsistent preprocessing, heterogeneous implementations, variable evaluation, and limited reproducibility often constrain method selection. Here, we systematically benchmark 22 publicly available unsupervised approaches for bulk data on the TCGA-BRCA cohort across five modalities (RNA-seq, miRNA, DNA methylation, copy numbers, single nucleotide polymorphisms) and validate findings in two independent datasets, enabling a multi-layered comparison of performance, heterogeneous data support and interpretability. Most approaches fuse multi-omic data to produce a two-cluster solution largely aligned with ER status, with higher-resolution approaches further refining these into four coherent subclasses (angiogenic luminal, oxidative-phosphorylation/HER2-low luminal, immune-inflamed basal-like, and hyper-proliferative basal-like). Our benchmarking results indicate that methods based on similarity networks can efficiently produce stable, reliable partitions. Matrix factorisation and Bayesian factorisation algorithms produce rich latent representations, allowing quantification of feature and modality contributions, albeit at higher computational cost. Consensus clustering can be used on a case-by-case basis and refine partitions into more robust and generalisable findings. We aggregate our insights into a decision workflow that aligns with study goals, data characteristics, and computational resources, enabling optimal analytic strategies. This comprehensive assessment provides a practical roadmap for investigators seeking to extract reproducible, biologically meaningful subtypes from complex multi-omic datasets. We higlight the different technical and practical benefits and trade-offs that shape the selection and development of multi-omic approaches applied in precision oncology.

Functional spinal cord repair in zebrafish is governed by regeneration-favorable biochemical and mechanical cues within the lesion microenvironment. Alterations in extracellular matrix composition and stiffness are closely associated with axon regeneration. However, experimentally dissecting the interplay between mechanical signals and axonal regrowth in vivo remains technically challenging. Here, we present an agent-based modeling framework to simulate stiffness-mediated axonal growth trajectories across the lesion. We use this model to explore potential mechanisms underlying the characteristic growth patterns observed during zebrafish spinal cord regeneration. Computational predictions were qualitatively compared with confocal imaging data obtained from larval zebrafish. These phenomenological comparisons revealed a close agreement between simulated and experimentally observed axon growth, indicating that experimentally observed patterns could be governed by transient changes in the stiffness profile of the spinal cord and lesion microenvironment. Hence, our computational framework provides an in silico platform for investigating the role of mechanical cues in axon regeneration in the injured spinal cord.

Despite the evolution of epidemiological analysis and modeling tools, difficulties still remain, especially in developing countries, regarding the availability and use of these tools. Often expensive, requiring high technical expertise, demanding constant connectivity of several or sometimes even significant resources, these tools, although efficient, present a major gap with the operational realities of health districts. It is in this context that we introduce Episia, an open-source Python library designed and conceived to provide a framework to facilitate epidemiological analysis and modeling. It integrates a suite of compartmental epidemic models (SIR, SEIR, SEIRD) with a sensitivity analysis using the Monte Carlo method, a complete biostatistics suite validated against the OpenEpi reference standard, as well as a native DHIS2 client for automated data ingestion. Developed in Burkina Faso, it is optimized and aims not only to address these health challenges encountered in Africa but also remains a versatile tool for global health informatics.

Invasive insect species pose serious threats to agriculture and ecosystems, with their spread increasingly accelerated by global trade and climate change. To support prevention and mitigation efforts, it is essential to map the regions where these pests can survive and thrive. Here, we apply MaxEnt, a leading species distribution modeling framework, to estimate current (2020) and future (2040-2060) suitable habitats for five major invasive insects across the contiguous United States: brown marmorated stink bug, corn earworm, spongy moth, root weevil, and spotted lanternfly. To account for an uncertain climatic future, these projections are generated under four shared socioeconomic pathways, which reflect a range of plausible climate change scenarios. Beyond forecasting distributions, we examine several key modeling decisions, especially those often overlooked in practice. In particular, we find that background sampling strategies play a critical role in model calibration and that a hybrid sampling approach with a moderate buffer bias provides better predictive accuracy. We also show that permutation importance scores, commonly used to rank environmental variables, are highly sensitive to small changes in the background data and should be interpreted with caution. Finally, to bridge the gap between ecological modeling and applied machine learning, we provide a self-contained, math-focused background to MaxEnt aimed at practitioners outside of traditional ecological fields. Overall, this work delivers reproducible modeling workflows and critical insights into building robust, transparent, and ecologically meaningful MaxEnt models for climate-informed species distribution analysis.

Cefiderocol uptake in Enterobacterales depends partly on TonB-dependent catecholate transporters, including CirA, yet the functional interpretation of CirA missense variation remains limited by an absence of large experimental phenotype datasets. Here we describe a structure-informed Siamese graph neural network (GNN) framework designed to prioritise CirA missense variants that are likely to impair transporter function and thereby contribute to reduced cefiderocol susceptibility. Because large experimental datasets of CirA missense phenotypes are not available, we trained the model on a synthetic mutant set generated from structurally motivated rules applied to the CirA reference structure (AlphaFold model, UniProt P17315). Each residue was represented using protein language model embeddings, backbone geometry, and amino-acid identity, and paired wild-type and mutant graphs were compared through a shared encoder. On synthetic held-out benchmarks, the model achieved strong classification performance on a position-held-out split (macro-F1 = 0.989 against synthetic labels). Applied to a collection of Escherichia coli CirA protein sequences, the framework prioritised a subset of variants as high-confidence non-benign candidates and assigned many others to review or abstain categories, reflecting predictive uncertainty outside the synthetic training distribution. A post-hoc severity-ranking scheme triages disruptive candidates for follow-up. This framework demonstrates that structure-informed synthetic data generation paired with Siamese GNN inference can bridge the gap between sequence-level genomic surveillance and mechanistic functional prediction of outer-membrane transporter variants.

A central goal of neuroscience is to understand how neural circuit architecture supports learning. While recent work has clarified the computational role of depth in sensory cortical hierarchies, it remains unclear why predominantly feedforward, non-convolutional circuits such as the cerebellum and olfactory system also contain multiple processing layers. Theoretical work in deep learning has shown that two-hidden-layer networks can achieve classification capacity that scales quadratically with the number of intermediate neurons, but these results rely on nonlocal synaptic optimization and are therefore difficult to reconcile with biological learning rules. Here, we show analytically and numerically that a two-hidden-layer network with feedforward gating can achieve quadratic capacity using local three-factor Hebbian learning when intermediate activity is sparse. This architecture supports efficient one-shot learning and, in settings where backpropagation requires many repeated weight updates, offers an advantage in learning speed. Beyond random perceptron tasks, the model also performs well on structured cerebellum-related tasks, including reinforcement-learning-based motor control. Mapping the model onto cerebellar microcircuitry further suggests functional roles for dendritic compartmentalization, branch-specific inhibition, and disinhibitory interneuron pathways. Together, these results extend the Marr-Albus-Ito framework by showing how the presence of multiple intermediate layers in cerebellum-like circuits can support fast, local, and high-capacity learning.

Protein synthesis in cell-free protein synthesis systems often exhibits non-intuitive input-output relationships. In the PURE system, a reconstituted cell-free system, protein production peaked at low elongation factor Tu (EF-Tu) concentrations and decreased at higher concentrations, resulting in a characteristic bell-shaped profile. Here, we investigated the origin of this behavior using a detailed mechanistic model of translation in the PURE system, designated as ePURE, which describes reaction dynamics of hundreds of molecular species and reactions. Our computational analysis suggested that excess EF-Tu sequesters the initiator tRNA (tRNAfMet) into non-productive EF-Tu{middle dot}GTP{middle dot}Met-tRNAfMet complexes, thereby depleting the pool of initiator tRNA available for translation initiation. This suppression arises from competition for a limited molecular resource rather than from direct inhibition. Based on this mechanism, we predicted that increasing the concentrations of tRNAfMet and methionyl-tRNA formyl-transferase would eliminate the bell-shaped dependence, and experimentally confirmed this prediction. Under these modified conditions, the bell-shaped response disappeared and protein production was enhanced. These findings demonstrate how mechanistic computational models can reveal hidden constraints underlying non-intuitive input-output relationships in complex biochemical networks and guide the rational optimization of cell-free protein synthesis systems.

Stratified epithelial tissues such as the skin epidermis maintain barrier integrity during development and homeostasis through the coordinated action of cell proliferation, differentiation, delamination, and tissue-scale mechanical forces. During development, the orientation of cell division within the basal layer plays a pivotal role in tissue stratification; however, the mechanical principles linking the orientation of the division plane to these processes across developmental stages remain poorly understood. Here, we expand a recently developed three-dimensional vertex model for stratified epithelia, composed of the basement membrane, basal, and suprabasal layers, to study the mechanical and structural impact of cell divisions with a wider range of orientations. The model integrates developmental stage via specific changes in heterotypic interfacial tensions (arising from actomyosin cortical contractility and adhesion molecules at the basal-suprabasal interface) and tissue stiffness that have been quantified previously in experiments. By systematically varying background mechanical parameters, we investigate how heterotypic tension, division orientation, and tissue fluidity collectively influence the outcome of cell division. Our goal is to uncover the strategies that the embryo may employ to generate stratified phenotypes at different developmental stages, recognizing that these strategies might evolve over time. Although our focus is on the embryonic developmental stages of the epidermis, this framework may also be extended to investigate transformed cells, such as in cancer, to explore how altered division orientation contributes to precancerous or transformed phenotypes.

BackgroundManual video-based evaluation of surgical skills can be time-consuming and delays trainee feedback. Artificial intelligence (AI) offers opportunities to automate aspects of assessment while maintaining clinician oversight. We developed an interpretable spatiotemporal model that classifies surgical expertise directly from endoscopic video in standardized training tasks and generates saliency-based "highlights reels" showing the most influential frames. MethodsAn RGB pipeline combining InceptionV3 for spatial feature extraction and a gated recurrent unit (GRU) for temporal modeling was trained on the JIGSAWS dataset. The model outputs novice, intermediate, or expert labels. A rolling-window, low-latency evaluation at 30 fps with a stride of 10 frames was used. A motion-augmented variant fused RGB with optical-flow features. Spatial and temporal saliency maps highlighted key decision-making regions. ResultsThe RGB model achieved 95% accuracy (F1: 92% expert, 86% intermediate, 99% novice). Performance was strongest for novice and expert trials, while intermediate trials showed the lowest recall, consistent with greater ambiguity around the intermediate skill level. Saliency maps consistently emphasized tool-tissue interactions and peaked during technically demanding phases. The optical-flow variant underperformed, approximately 38% accuracy, which may reflect sensitivity to global camera motion and other non-informative motion patterns. ConclusionsThis interpretable AI pipeline accurately classifies surgical skill while producing intuitive visual highlights. Future work will refine highlight thresholds and validate on laparoscopic inguinal hernia repair for realworld deployment.