Small-sample ERP-assisted screening for depression and audio multimodal fusion based on the MODMA dataset

Event-related potentials (ERPs) may provide objective neurophysiological markers for depression-assisted screening. We benchmarked structured ERP features and strictly paired ERP-audio fusion in small-sample, subject-level settings.

Two experiments on MODMA used leakage-controlled leave-one-subject-out (LOSO) validation, with all preprocessing, tuning and fusion restricted to training folds. Experiment I included 52 ERP subjects retained after quality control (QC) (23 with major depressive disorder (MDD) and 29 healthy controls (HC)), each represented by a 60-feature vector, to compare machine learning (ML) and deep learning (DL) models. This paper additionally evaluated convolutional neural network (CNN)-based architectures (EEGNet and ShallowConvNet) on a multichannel ERP time-series representation (28 channels × 129 samples) constructed from subject-level averaged cue-locked and dot-locked ERPs. Experiment II used 37 strictly paired ERP-audio subjects (16 MDD, 21 HC) with 178 audio features to evaluate unimodal and decision-level fusion under balanced and screening objectives.

In Experiment I, SVM-RBF achieved area under the ROC curve (AUC) 0.859 vs 0.588 for TinyMLP ($\Delta$AUC = +0.271). On ERP time-series inputs, EEGNet achieved AUC = 0.640 and ShallowConvNet achieved AUC = 0.540, both inferior to the engineered-feature SVM baseline. In Experiment II, the best unimodal baseline was audio SVM-RBF (AUC = 0.753). Fusion achieved AUC to 0.762 and increased specificity (Spec) to 0.810 under the balanced objective, and increased recall (Rec) to 0.813 with a false-negative rate (FNR) of 0.188 under the screening objective.

This study provides a leakage-controlled, strictly paired, subject-level evaluation framework for depression-assisted screening and shows that ERP-audio fusion offers objective-dependent complementary value.