Functional near-infrared spectroscopy-based computer-aided diagnosis of major depressive disorder using explainable artificial intelligence: Comparison with conventional machine learning

Abstract

Background and objective Only a limited number of explainable artificial intelligence (XAI) models have been developed for the functional near-infrared spectroscopy (fNIRS)-based computer-aided diagnosis (CAD) of major depressive disorder (MDD). In this study, we implemented an XAI model based on a convolutional neural network (CNN) specifically designed to highlight the differences in interhemispheric asymmetry between patients with MDD and healthy controls (HCs). Methods The proposed XAI model was applied to the fNIRS data acquired from 48 patients with MDD and 68 HCs during a verbal fluency task. Layer-wise relevance propagation (LRP), a well-known XAI technique that identifies the contribution of individual input data to the model prediction, was employed. Results The performance of the proposed model was validated using ten-fold cross-validation, which reported an average accuracy of 81.17 %, a sensitivity of 79.5 %, and a specificity of 82.38 %. The LRP provided additional insights into the model's decision-making process, revealing that channels within both the right and left dorsolateral prefrontal cortices (DLPFC) were critical for classification. Notably, the relevance score distribution in the right DLPFC closely aligned with the Fisher score distribution observed in conventional machine learning models using handcrafted features. Furthermore, the analysis of symmetrical channel pairs with high relevance scores revealed distinct interhemispheric asymmetry in the waveforms of patients with MDD. Conclusion In summary, we developed an XAI model tailored for fNIRS-based CAD of MDD that not only demonstrated high diagnostic performance but also effectively visualized the decision-making processes of the deep learning model.