A Bayesian-optimized design for an interpretable convolutional neural network to decode and analyze the P300 response in autism

Objective.P300 can be analyzed in autism spectrum disorder (ASD) to derive biomarkers and can be decoded in brain-computer interfaces to reinforce ASD impaired skills. Convolutional neural networks (CNNs) have been proposed for P300 decoding, outperforming traditional algorithms but they (a) do not investigate optimal designs in different training conditions; (b) lack in interpretability. To overcome these limitations, an interpretable CNN (ICNN), that we recently proposed for motor decoding, has been modified and adopted here, with its optimal design searched via Bayesian optimization.Approach.The ICNN provides a straightforward interpretation of spectral and spatial features learned to decode P300. The Bayesian-optimized (BO) ICNN design was investigated separately for different training strategies (within-subject, within-session, and cross-subject) and BO models were used for the subsequent analyses. Specifically, transfer learning (TL) potentialities were investigated by assessing how pretrained cross-subject BO models performed on a new subject vs. random-initialized models. Furthermore, within-subject BO-derived models were combined with an explanation technique (ICNN + ET) to analyze P300 spectral and spatial features.Main results.The ICNN resulted comparable or even outperformed existing CNNs, at the same time being lighter. BO ICNN designs differed depending on the training strategy, needing more capacity as the training set variability increased. Furthermore, TL provided higher performance than networks trained from scratch. The ICNN + ET analysis suggested the frequency range [2, 5.8] Hz as the most relevant, and spatial features showed a right-hemispheric parietal asymmetry. The ICNN + ET-derived features, but not ERP-derived features, resulted significantly and highly correlated to autism diagnostic observation schedule clinical scores.Significance.This study substantiates the idea that a CNN can be designed both accurate and interpretable for P300 decoding, with an optimized design depending on the training condition. The novel ICNN-based analysis tool was able to better capture ASD neural signatures than traditional event-related potential analysis, possibly paving the way for identifying novel biomarkers.