Authors :
Presenting Author: Patrick Kwan, Prof. – Monash University
Haris Hakeem, Dr. – haris.hakeem@monash.edu; Wei Feng, Mr. – Monash University; Richard Shek-kwan Chang, Dr. – Monash University; Duong Nhu, Dr. – Monash University; Martin Brodie, Prof. – Epilepsy Unit, Division of Cardiovascular and Medical Sciences, Western Infirmary; Xinshi Wang, Dr. – The First Affiliated Hospital of Wenzhou Medical University; Huiqin Xu, Dr. – The First Affiliated Hospital of Wenzhou Medical University; Mohsen Farazdaghi, Dr. – Shiraz Epilepsy Research Center, Shiraz University of Medical Sciences; Ali A. Asadi-Pooya, Dr. – Shiraz Epilepsy Research Center, Shiraz University of Medical Sciences; Zhibin Chen, Dr. – Monash University; Zongyuan Ge, Dr. – Monash University
Rationale:
Patients with drug-resistant epilepsy (DRE), defined as failure of two appropriately chosen antiseizure medications (ASMs) to achieve seizure freedom, are recommended to undergo evaluation for non-drug therapies. Selection of the first two ASM regimens is therefore of critical importance for the individual patients. Our team has recently developed a deep learning model to predict treatment success with the first ASM in adults with newly diagnosed epilepsy. Here, we develop another model to predict the response to the second ASM regimen.Methods:
A total of 467 patients (232 males/235 females; median age at initiation of second ASM regimen: 34 years interquartile range: 25-47 years) treated with a second ASM regimen were recruited from Glasgow (UK) (n=353), Wenzhou (China) (n=85), and Shiraz (Iran) (n=29). Patients received either substitution monotherapy or duotherapy. Multilayer perceptron algorithm was adopted. The following variables were input into the model: the same 16 clinical factors (legend of figure 1) included in our prior published model for predicting response to the first ASM; names of the first and second ASMs; duration, maximal daily dose, reason of failure (inefficacy/intolerance), and outcome (seizure free or not) at 12 months after initiation of the first ASM. In the first experiment, the model was trained on Glasgow cohort and validated on the other two cohorts. In the second experiment, the 3 cohorts were pooled for model training and 5-fold cross-validation. Model performance was assessed using standard metrics. Shapley additive explanations (SHAP) analysis was used to assess the relative contribution of input variables to the model. Results:
In the first experiment, the area under the receiver operating characteristic curve (AUC) was 0.69 (95% Confidence Interval [CI]: 0.679-0.701) and F1 score was 0.52 (95%CI: 0.503-0.537) (table 1). In the second experiment, the AUC and F1 were 0.64 (95%CI: 0.621-0.659) and 0.56 (95%CI: 0.541-0.579) respectively for Wenzhou cohort and 0.77 (95%CI: 0.748-0.792) and 0.36 (95%CI: 0.341-0.379) for Shiraz cohort. SHAP analysis showed that reason of failure of the first ASM, outcome at 12 months after initiation of the first ASM, and presence of psychiatric comorbidities were top contributors to the model in both experiments (Figure 1).
Conclusions:
Our study demonstrated the feasibility of using deep learning to predict the outcome of the second ASM regimen on individual level. With improved performance, this model may assist the selection of subsequent ASM regimens after failure of the first ASM, thereby reducing the proportion of patients with DRE.
Funding:
The study is supported by the National Health and Medical Research Council of Australia, Ideas Grant GNT2010382.