Explainable artificial intelligence based on feature optimization for age at onset prediction of spinocerebellar ataxia type 3

Abstract

Existing treatments can only delay the progression of spinocerebellar ataxia type 3/Machado-Joseph disease (SCA3/MJD) after onset, so the prediction of the age at onset (AAO) can facilitate early intervention and follow-up to improve treatment efficacy. The objective of this study was to develop an explainable artificial intelligence (XAI) based on feature optimization to provide an interpretable and more accurate AAO prediction. A total of 1,008 affected SCA3/MJD subjects from mainland China were analyzed. The expanded cytosine-adenine-guanine (CAG) trinucleotide repeats of 10 polyQ-related genes were genotyped and included in related models as potential AAO modifiers. The performance of 4 feature optimization methods and 10 machine learning (ML) algorithms were compared, followed by building the XAI based on the SHapley Additive exPlanations (SHAP). The model constructed with an artificial neural network (ANN) and feature optimization of Crossing-Correlation-StepSVM performed best and achieved a coefficient of determination (R2) of 0.653 and mean absolute error (MAE), root mean square error (RMSE), and median absolute error (MedianAE) of 4.544, 6.090, and 3.236 years, respectively. The XAI explained the predicted results, which suggests that the factors affecting the AAO were complex and associated with gene interactions. An XAI based on feature optimization can improve the accuracy of AAO prediction and provide interpretable and personalized prediction.