Parkinson's disease diagnosis using machine learning and multimodal biomarkers: A systematic review.
TL;DR
Multimodal fusion architectures, particularly cross-attention transformers integrating motor, voice, and gait data, consistently outperformed unimodal approaches by 4%-8% in the area under the receiver operating characteristic curve (AUC), with the largest gains in early stage detection and differential diagnosis tasks.
Abstract
Parkinson's disease (PD) is the second most common neurodegenerative disorder and the fastest-growing neurological condition globally, with an estimated 10.5 million people living with PD in 2026 and projections exceeding 14.2 million by 2040 according to the World Health Organization. Despite this staggering burden, diagnosis continues to depend on clinical motor signs that appear only after 60%-80% of dopaminergic neurons in the substantia nigra have been irreversibly lost, and specialist misdiagnosis rates reach 25% at early disease stages even in dedicated movement disorder centers. This systematic review synthesizes evidence from 239 studies drawn from 5847 screened records (2016-2026), critically evaluating unimodal diagnostic approaches encompassing dopamine transporter single-photon emission computed tomography (DAT-SPECT), quantitative magnetic resonance imaging (MRI), alpha-synuclein (α-syn) seed amplification assay (SAA), neurofilament light chain (NfL), genetic risk markers, voice and speech biomarkers, gait and movement metrics, electroencephalography (EEG), and handwriting analysis, alongside their integration into multimodal machine learning (ML) and deep learning (DL) frameworks with explainable artificial intelligence (XAI) components. Across 87 quantitative studies, the α-syn SAA achieved a sensitivity of 84%-92% and specificity of 93%-98% in prodromal cohorts. DAT-SPECT achieved a pooled sensitivity and specificity of approximately 90% but could not differentiate PD from atypical parkinsonian syndromes (APS). Multimodal fusion architectures, particularly cross-attention transformers integrating motor, voice, and gait data, consistently outperformed unimodal approaches by 4%-8% in the area under the receiver operating characteristic curve (AUC), with the largest gains in early stage detection and differential diagnosis tasks. XAI methods, including SHapley Additive exPlanations (SHAP), gradient-weighted class activation mapping (Grad-CAM), and integrated gradients, were reported in only 31% of deep learning studies. This indicates a transparency gap with direct regulatory consequences. Across all modalities, only 22% of studies reported external validation, and multicenter prospective evidence remains scarce. These two deficits, rather than model architecture, are now the principal barriers to clinical adoption. This review provides an actionable roadmap for researchers, clinicians, and regulatory agencies working in the PD diagnostics field.