Artificial Intelligence Reveals Subtle Mouse Seizure Behaviors That Predict Fatal Outcomes

Abstract

Behavior Predicts Seizure Outcomes

Shen Y, Thomas J, Chen X, Zelidon J, Najeeb M, Hahn A, Zhang P, Sathyanesan A, Gu B. Behavior Decoding Delineates Seizure Microfeatures and Associated Sudden Death Risks in Mouse Models of Epilepsy. Annals of Neurology. 2025. doi:10.1002/ana.78032. Online ahead of print. PMID: 40944492

Objective: Behavior and motor manifestations are distinctive yet often overlooked features of epileptic seizures. Seizures can result in transient disruptions in motor control, often organized into specific behavioral sequences that can inform seizure types, onset zones, and outcomes. However, refined analysis of behaviors in epilepsy remains challenging. Current manual video inspection approaches are subjective, time-consuming, and often overlook the intricate behavioral dynamics and action kinematics. This study investigates whether artificial intelligence (AI)-aided tools can unravel complex behavior repertoire that can delineate seizure outcomes in a data-driven manner. Methods: We utilized two AI-aided tools, DeepLabCut (DLC), and Behavioral Segmentation of Open Field in DLC, to decode underexplored behavior and action domains of induced seizures in a population of 32 inbred mouse strains that mimic human genetic diversity and a mouse model of Angelman syndrome. Results: Our automated behavior classification tool identified 63 interpretable behavior groups (BGs). Analysis of these BGs demonstrates significant differential behavior usage and complexity that can delineate distinct seizure states, unravel intrinsic seizure progression over time, and inform mouse sex, strain backgrounds, and specific pathogenic mutations. We also identified seizure behavior transition dynamics and action/subaction kinematics, like hindlimb motions, that can determine the risks of sudden unexpected death in epilepsy (SUDEP). Interpretation: These behavior microfeatures can facilitate preclinical mechanistic studies and antiseizure medication screening at scales. These findings also underscore the translational potential of video-based seizure behavior decoding in both inpatient and outpatient settings, including analyzing videos captured by home surveillance devices and ubiquitous smartphones.

Commentary

An ongoing challenge in the fight against sudden unexpected death in epilepsy (SUDEP) is the inability to accurately predict who is most at risk and when that risk is highest. Although SUDEP is the leading cause of epilepsy-related mortality, reliable, time-sensitive predictors remain lacking, limiting the development of effective interventional strategies.^1,2 While several SUDEP risk factors have been identified, most notably those related to seizures themselves, such as the frequency of generalized tonic-clonic seizures, these measures provide only a coarse estimate of vulnerability. ³ They do not identify which individuals will ultimately succumb to SUDEP or which specific seizures carry the greatest danger. As a result, a fundamental and clinically urgent question remains unanswered: what makes a seizure fatal? To date, most efforts to address this question have focused on neurocardiac and neurorespiratory mechanisms and their underlying anatomical pathways. ⁴

New work led by Bin Gu offers a compelling new perspective, suggesting that critical clues may lie in seizure semiology. ⁵ Semiology refers to the observable signs and symptoms of a seizure. A central emphasis of semiologic analysis is the characterization of seizure-related motor behaviors, which have long been used to inform seizure localization and guide clinical decision-making. ⁶ In this study, Shen and colleagues introduce a novel artificial intelligence (AI)-aided framework to decode complex, fine-scale seizure behaviors in mice, revealing subtle behavioral signatures that distinguish fatal from nonfatal seizures. These findings suggest that how a seizure unfolds behaviorally, at a resolution beyond conventional human scoring, may carry previously unrecognized predictive power for seizure outcome.

Using two open-source AI-based video behavior analysis tools, DeepLabCut (DLC) and Behavioral Segmentation of Open Field in DLC, Shen and colleagues identified 63 distinct behavior groups (BGs) from mouse video recordings (eg, pushing backward, leaping, and balancing). These BGs were classified with high accuracy using pose estimation derived from 28 tracked body parts.

These BGs then served as the foundation for quantifying a rich set of behavioral microfeatures, capturing not only how often specific behaviors occurred but also how they evolved over time. Analyses included BG usage patterns, transition dynamics between behavioral states, measures of behavioral diversity and complexity (Shannon entropy), and detailed kinematic features reflecting the duration and movement structure of individual BG bouts.

The authors applied this framework to video recordings of seizures induced by the inhaled chemoconvulsant flurothyl in 31 inbred strains from the Collaborative Cross, ⁷ the widely used C57BL/6J strain, and a mouse model of Angelman Syndrome. Flurothyl reliably produces a stereotyped seizure sequence composed of an initial myoclonic seizure followed by a generalized clonic seizure, allowing behavioral microfeatures to be compared across the preictal baseline period and two distinct seizure states. Importantly, flurothyl also causes seizure-related death in a subset of animals, enabling direct comparison of behavioral differences preceding fatal versus nonfatal seizures.

One of the most compelling insights of the study is the identification of fine-grained behavioral signatures that predict seizure-related mortality, pointing to novel behavioral biomarkers relevant to SUDEP. The authors identified 13 BG transitions that showed different probabilities in fatal and nonfatal seizures, suggesting that seizures with divergent outcomes engage distinct neural circuits.

Moving beyond individual transitions, the authors used machine learning to uncover hidden associations among BGs. This analysis identified probabilistic “itemsets” that capture common behavioral sequences distinguishing fatal from nonfatal seizures, revealing behavioral pathways that funnel seizures into high-risk trajectories. These micro-behavioral patterns were then leveraged to build a predictive model of seizure outcome. When applied to video recordings of generalized seizures, the model correctly classified fatal seizures 74% of the time, substantially outperforming human observers, whose accuracy was only 57%. This performance gap highlights the limitations of coarse visual assessment and the added value of fine-scale behavioral decoding.

Finally, kinematic analysis revealed an unexpected but predictive feature of fatal seizures: knee joint dynamics. Fatal generalized seizures were associated with increased knee joint angles and faster flexion-extension movements, reflected in elevated instantaneous angular velocity. These subtle joint-level features were sufficient to discriminate fatal from nonfatal outcomes, underscoring the potential of micro-behavioral and kinematic signatures as objective biomarkers of seizure-related mortality.

Beyond mortality prediction, the study yielded additional insights. These included seizure type- and sex-dependent behavioral differences, effects of genetic background on seizure behaviors, and the identification of strain-specific phenotypes not detected by traditional Racine scoring, including a previously unrecognized seizure phenotype in Angelman syndrome mice.

This study has important implications for future research, alongside some notable limitations. In preclinical rodent models, seizure severity is typically assessed using variants of the Racine scale, a manually applied scoring system that is somewhat subjective and offers limited resolution, with ordinal scores usually ranging from 1 to 5 or 7. ⁸ In contrast, the AI-assisted approach described here enables automated, objective, and highly granular behavioral analysis and, importantly, outperforms Racine-style scoring in discriminating seizure behaviors across genotypes and mouse strain backgrounds. These findings were achieved using two-dimensional video with modest resolution (480p) and frame rates (30 fps), underscoring the robustness of the behavioral signatures and suggesting that even greater sensitivity may be possible with higher resolution, higher frame rate, or three-dimensional recordings. Moreover, the method can be implemented with basic off-the-shelf cameras and open-source software, without extensive coding expertise, making it broadly accessible. Coupling this framework with electrophysiological recordings could further link behavioral microfeatures to neural circuit dynamics and seizure outcomes.

Several limitations should be considered. Although behavioral microfeatures discriminated between fatal and nonfatal seizures, all seizures analyzed were experimentally induced. Whether similar signatures generalize to fatal spontaneous seizures, which more closely align with the clinical definition of SUDEP, remains unknown. In addition, mice were recorded in large beakers, which may have limited naturalistic behavior.

Beyond the current application, this methodology could be extended to other animal models and, in principle, to clinical settings that use video monitoring, such as epilepsy monitoring units. ⁹ With further development, it may also be adapted for home surveillance or smartphone-based recordings to detect seizures, including potentially fatal events during sleep. Translating this approach into a SUDEP prevention strategy for humans, however, presents practical challenges. Although body part tracking is feasible in infants, ¹⁰ who often sleep uncovered, it becomes substantially more difficult in older individuals due to occlusion by bedding, representing a significant obstacle to real-time monitoring. Addressing these challenges, potentially through integration with complementary monitoring strategies such as multimodal wearable devices, could help unlock the full translational potential of video-based behavioral biomarkers. In the meantime, this approach is already helping to decode the intricate behaviors that distinguish fatal from nonfatal seizures in preclinical models, providing an essential foundation for uncovering the mechanisms and pathways that lead to seizure-related death. By advancing behavioral quantification, this work shifts the landscape of SUDEP research, transforming video from a passive record into an active source of neurobiological insight.