PDualNet: a deep learning framework for joint prediction of Parkinson's disease progression subtype and MDS-UPDRS scores

Abstract

Parkinson’s disease is one of the most common and complex neurodegenerative diseases, characterized by remarkable motor and cognitive decline. As it is a highly heterogeneous disorder, i.e., the specific symptoms, their severity, and their progression rate manifest significant interpersonal variability, multiple progression subtypes can be defined. The identification and prediction of these subtypes is crucial for understanding the disease’s state and future trajectory, advancing prognostic accuracy and personalized treatment planning. At the same time, the ability to predict future MDS-UPDRS scores, provides an objective assessment of symptoms, supporting clinicians in tracking disease progression and evaluating treatment efficacy. To address both critical objectives, we introduce PDualNet, a novel dual-task framework that jointly models and predicts the disease progression and severity based on longitudinal clinical patient data. Our approach involves two key components: (i) an unsupervised module that maps the single-visit data of each patient, onto a “Single-Visit Embedding (SiVE) space”, and (ii) a supervised part, that utilizes the pre-trained SiVE embeddings to learn a compact representation of the longitudinal data of each patient, representing the “Disease State Embeddings (DiSE)”. These embeddings drive two parallel decoders: one predicting the progression subtype, and the other forecasting the future MDS-UPDRS I–III scores. After analysing patient visit data from up to six years after baseline, each consisting of 89 clinical features, we trained and evaluated PDualNet on 579 participants from the Parkinson’s Progression Markers Initiative. The resulting model, demonstrated remarkable performance on both classification and regression tasks, while additional validation on 490 participants from the Parkinson’s Disease Biomarkers Program cohort, confirmed its robust performance and strong generalization capabilities.