Developing a multivariable deep learning model to predict psychiatric illness in patients with epilepsy

Comprehensive Summary

Epilepsy care extends beyond seizure control and suppression. Psychiatric comorbidities, most notably depression, anxiety, and psychosis, affect nearly one-third of patients, which could complicate medication adherence and seizure outcomes. To address this, researchers in this study conducted a retrospective analysis of 2,258 epilepsy patients treated between 2013-2023 to determine whether clinical and demographic variables could predict psychiatric risk. By leveraging 32 distinct clinical and demographic variables ranging from EEG findings and seizure duration, to birth history and specific antiseizure medication (ASM) profiles, they developed deep learning neural network models using the keras and neuralnet frameworks. The results showed that psychiatric comorbidities were found in 27.6% of the cohort, and the models achieved a training accuracy of up to 97.16%. Further, SHAP-based interpretability analyses revealed that while a younger age of onset and longer seizure duration increased risk, specific ASMs like valproate and lamotrigine were associated with lower predicted psychiatric risk within this dataset.

Outcomes and Implications

This study marks a significant shift toward a more proactive and integrated approach to neuropsychiatric care. By using data already present in Electronic Health Records, the keras and neuralnet models could serve as an early warning during limited neurology visits (in terms of time), and help flag high-risk individuals for early intervention before psychiatric symptoms reach a point of crisis. This allows for a more individualized approach to risk profiling, potentially influencing ASM selection based on a patient’s psychiatric risk profile. While the model shows high sensitivity and accuracy, researchers emphasize that clinical implementation still requires external validation across multiple different populations to ensure the generalizability of results. Future studies should consider the inclusion of genetic markers and longitudinal treatment histories to transition this tool into a decision support system that improves long-term patient outcomes and quality of life.