Smartphone-Based Multimodal Digital Biomarker Integration for Parkinson’s Disease Screening and Diagnostic Support

Abstract

Timely identification of Parkinson’s disease (PD) is often delayed because of clinical heterogeneity and limited awareness of early symptoms. Digital biomarkers obtained via smartphones offer scalable screening potential. However, unimodal assessments may lack sufficient sensitivity or specificity given the multidimensional nature of PD. The aim of this study was to develop and validate a smartphone-based, multimodal digital biomarker framework for PD screening and diagnostic support. The study progressed through two phases: an initial version [n = 368; 233 PD, 135 healthy controls (HC)] was used for data-driven task refinement, and a final version (n = 296; 204 PD, 92 HC) containing optimized motor (Touch, Swipe, Balloon, Spiral, Wave), visual, and speech tasks and a refined questionnaire task was evaluated. Feature selection and speech subtask selection were performed exclusively within the training set using stratified cross-validation. Random Forest and XGBoost classifiers were trained using (1) single-task features, (2) all-task multimodal features, and (3) selected task subsets. The primary outcome was area under the receiver operating characteristic curve (AUROC) on the independent test set. In the final version, single-task models demonstrated heterogeneous performance (Random Forest AUROC range 0.5689–0.8397), with the questionnaire (0.8397) and Touch task (0.7789) performing best individually. The all-task multimodal model achieved AUROC 0.8620. A reduced multimodal subset combining Touch, Spiral, and questionnaire features yielded the highest discriminative performance (AUROC 0.9053). Additional feature- and task-level analyses showed significant multivariate group differences (Hotelling’s T2p < 0.001) and stronger inter-feature association in PD compared with HC, providing interpretability. A smartphone-only multimodal digital biomarker framework achieved high discrimination between PD and controls. Multimodal integration outperformed unimodal approaches, supporting the potential utility of scalable, smartphone-based tools for PD screening and diagnostic support. External validation in broader populations is warranted. Parkinson’s disease is often difficult to identify early because symptoms vary widely between individuals and access to clinical evaluation can be limited. Smartphones may provide a practical and accessible way to support screening through the use of digital biomarkers, which are measurable patterns collected through digital devices that reflect health conditions. In this study, we developed and tested a smartphone-based system for Parkinson’s disease screening and diagnostic support. The system combined several types of information, including hand movement tasks, drawing tasks, speech recordings, visual tasks, and symptom questionnaires. We analyzed data from 664 participants, including people with Parkinson’s disease and healthy controls. Artificial intelligence methods were used to determine how accurately the smartphone tasks could distinguish people with Parkinson’s disease from healthy individuals. Some single tasks performed reasonably well, particularly the questionnaire and touch-based finger movement task. However, combining multiple tasks produced better results than using any single task alone. The best-performing model combined a touch task, a spiral drawing task, and questionnaire responses, showing high accuracy in identifying Parkinson’s disease. These findings suggest that smartphone-based multimodal digital biomarkers may provide a scalable and accessible tool for Parkinson’s disease screening and diagnostic support. Further validation in broader populations is still needed.