Segment-wise split leakage in hierarchical multimodal prediction: a case study on DAIC-WOZ

In interview-based prediction, inputs are often constructed at the utterance or segment level, whereas labels are assigned at the participant level.In hierarchical multimodal prediction, this mismatch causes the same participant-level label to be repeatedly assigned across multiple text and audio segments.When these segments are randomly divided between training and evaluation sets, segment-wise split leakage can artificially inflate predictive performance.Rather than learning clinically meaningful patterns, the model may instead memorize participant-specific identity cues that appear across the split.This study focuses on the DAIC-WOZ dataset and examines how predictive performance changes as the leakage strength r increases from a participant-wise split.We evaluate a multimodal framework consisting of frozen text and audio encoders, sequence modeling, and late fusion using mean absolute error (MAE), root mean squared error (RMSE), and Pearson correlation.Performance is summarized on the training, validation, and test sets at the early-stopped epoch.The results provide a practical case study demonstrating why participant-wise split should be regarded as a fundamental principle in segmented multimodal prediction.