MIL-O-PD: a two-stage multiple instance learning and heuristic optimization framework for unpaired multimodal Parkinson's diagnosis

Parkinson's disease (PD) is a common neuro-degenerative disorder. Recent studies have used non-invasive biomarkers such as electroencephalography (EEG) and speech signals for PD diagnosis. However, majority of them optimize models at the instance level, whereas clinical diagnosis is a subject-level decision making process. In this work, we propose MIL-O-PD, a novel two-stage subject-level multimodal framework that formulates PD diagnosis. The stage-1 implements a Multiple Instance Learning approach with modality specific encoders and an attention-based aggregation mechanism. Stage-2 focuses incorporating a Gray Wolf Optimization-based post-hoc feature optimization module and final subject-level classifier. It supports representation-level multimodal fusion under strict subject-independent conditions. The proposed framework achieves up to 70.5% subject-level accuracy and 69.7% F1-score, while outperforming alternate pooling mechanisms. Further, the analysis of attention weights highlights modality-specific behavior, supporting the interpretability of the model. Overall, this study presents a principled framework for subject-level analysis of PD designed for unpaired multimodal data addressing the importance of aligning learning objectives with real-world diagnostic processes.