# Development and validation of interpretable multimodal clinical-radiomics models for predicting epileptogenic foci and surgical outcomes in tuberous sclerosis complex: A multicenter study

## TL;DR

A noninvasive AI model using clinical and imaging data accurately predicts epileptogenic foci and surgical outcomes in tuberous sclerosis complex patients.

## Contribution

Development of a validated, interpretable multimodal clinical-radiomics model with superior predictive performance for epi foci and surgical outcomes in TSC.

## Key findings

- The ensemble super-learner model achieved AUCs of 0.92 (prospective) and 0.91/0.87 (retrospective) for epi foci prediction.
- The model predicted 1-, 3-, and >5-year surgical outcomes with AUCs of 0.93, 0.91, and 0.92, respectively.
- A web-based tool improved junior clinicians' accuracy from 0.61 to 0.78, matching senior clinicians' accuracy of 0.80.

## Abstract

Precise localization and resection of epileptogenic (epi) foci from multiple cortical foci determine surgical outcomes in the tuberous sclerosis complex (TSC). Although the use of intracranial electroencephalography (EEG) for detecting epileptic discharges remains the gold standard for identifying epi foci, its invasiveness and cost limit clinical application. We aimed to develop and validate a noninvasive, clinically applicable predictive model for epi foci identification and surgical outcome assessment in patients with TSC. This multicenter study focused on three retrospective cohorts and one prospective cohort from three comprehensive epilepsy centers from June 2013 to October 2024. Comprehensive clinical and imaging data (CT, MRI, and 18F-FDG PET) of cortical foci were collected. Nineteen individual machine learning (ML) models and three ensemble ML models (voting, averaging and super-learner [SL]) were developed on the basis of the clinical and radiomics features of cortical foci. Model performance was evaluated by using the area under the curve (AUC), accuracy, precision, specificity, and sensitivity values, along with the F1 score, with additional validation being conducted via decision curve analysis (DCA) and calibration curves. Follow-up data were collected at 1, 3, and >5 years to validate the ability of the ML models to predict long-term postoperative outcomes. Non-epi foci were clustered by using the k-means algorithm to investigate the mechanisms underlying postoperative epileptogenic transformation. A web-based tool was developed to provide a user-friendly interface for clinical application. A total of 665 cortical foci (epi foci, n = 161; non-epi foci, n = 504) were included in this study. The model integrating multimodal clinical-radiomics features performed better than the individual models based only on single-modal clinical or radiomics features did. The ensemble SL model using clinical-radiomics features demonstrated the best stability and superior predictive performance compared to those of individual models and an additional two ensemble models in prospective (AUC: 0.92) and two retrospective cohorts (AUCs: 0.91 and 0.87); moreover, it outperformed previously reported prediction models. In addition, the SL model effectively predicted 1-, 3- and >5-year surgical outcomes (AUCs: 0.93, 0.91, and 0.92, respectively). K-means revealed two clusters of non-epi foci, including those foci with epileptogenic potential and those without, which were potentially confirmed by the follow-up data. The web-based tool significantly increased the accuracy of junior clinicians (from 0.61 to 0.78), which matched the accuracy of senior clinicians (0.80). The multimodal clinical-radiomics model represents a noninvasive tool for predicting epi foci, guiding preoperative evaluation, addressing diagnostic discrepancies and enabling personalized treatment strategies in patients with TSC. The clinical application of artificial intelligence (AI)-driven clinical-radiomics models provides a useful tool and auxiliary reference for clinicians in preoperative epileptogenic foci prediction.

In this multicenter study, we developed and validated noninvasive predictive models for epi foci localization and surgical outcomes in patients with TSC by integrating multimodal clinical-radiomics features from three retrospective cohorts and one prospective cohort from June 2013 to October 2024. Nineteen individual and 3 ensemble machine learning (ML) models were constructed, with the ensemble super-learner (SL) model demonstrating superior performance in prospective (AUC: 0.92) and two retrospective cohorts (AUCs: 0.91 and 0.87); additionally, it outperformed other reported prediction models. The clinical-radiomics SL model predicted the 1-, 3- and >5-year surgical outcomes (AUCs: 0.93, 0.91, and 0.92, respectively) of cortical foci. Furthermore, a web-based tool of predictive models was developed, and it significantly increased the accuracy of junior clinicians (from 0.61 to 0.78), which matched that of senior clinicians (0.80). The clinical-radiomics SL model represents a promising tool to predict epi foci, address diagnostic discrepancies among clinicians of varying levels of expertise, and enable personalized treatment strategies while minimizing the invasiveness and financial strain of TSC patients. Due to its robust performance, this model demonstrates considerable translational potential in both TSC patients and all epilepsy patients with multiple epi foci. The clinical application of this model provides an AI-driven tool and auxiliary reference for clinicians in preoperative epileptogenic foci prediction.