An Explainable Multimodal Neural Network Architecture for Predicting Epilepsy Comorbidities Based on Administrative...

Area of Science:

• Neurology
• Artificial Intelligence
• Health Informatics

Background:

• Epilepsy is a complex neurological disorder with frequent seizures.
• Epilepsy patients experience significant physical and psychological comorbidities.
• Individualized comorbidity risk prediction is crucial, considering factors like medication.

Purpose of the Study:

• To develop a machine learning model for predicting individualized, time-dependent comorbidity risks in epilepsy patients.
• To assess the performance and interpretability of the proposed model against existing methods.
• To validate the model's generalization and stability on independent datasets.

Main Methods:

• Utilized inpatient and outpatient administrative health claims data from approximately 19,500 U.S. epilepsy patients.
• Developed a multimodal neural network architecture named Deep personalized LOngitudinal convolutional RIsk model (DeepLORI).
• Employed a game theoretic approach for feature relevance identification and model interpretability.

Main Results:

• DeepLORI demonstrated superior predictive performance compared to several existing methods.
• The model's predictions were interpretable at the individual patient level.
• Model predictions were explainable, aligning with existing epilepsy disease knowledge.
• Validation on ~97,000 patients confirmed good generalization and stable performance over time.

Conclusions:

• DeepLORI effectively predicts individualized comorbidity risks in epilepsy patients.
• The model offers interpretable and explainable insights, enhancing clinical utility.
• This approach advances personalized medicine for epilepsy management by improving risk prediction.