Machine learning-based detection of Parkinson's disease from facial expressions, hand movements, speech, and gait with general-purpose equipment: a systematic review.

BACKGROUND

Parkinson's disease (PD) diagnosis remains largely subjective, relying on clinical symptom assessment. The convergence of machine learning (ML) with low-cost, widely available digital devices creates new opportunities for objective, scalable, and accessible PD screening and monitoring.

OBJECTIVES

This systematic literature review aims to examine ML approaches for PD detection, early diagnosis, severity assessment, and stage classification using four key data modalities: facial expressions, hand movements, speech and voice, and gait employing general-purpose, low-cost equipment.

METHODS

A systematic literature review of 133 papers published between January 2020 and June 2025 was conducted following PRISMA 2020 guidelines. Data were extracted on feature extraction techniques, assessment protocols, computational frameworks, classifier architectures, performance metrics, and participant cohorts for each modality and for multimodal fusion studies. The quality of multimodal fusion studies was evaluated using the IJMEDI checklist.

RESULTS

Cross-modality comparison for PD vs. healthy control classification indicated a promising trend for speech and voice modality. Binary classification was the most common ML task (74.5% of studies). Support Vector Machines (SVM) and Random Forests (RF) were the predominant classifiers. While multimodal data fusion demonstrated a promising trend toward improved performance (observed in 75% of studies), this finding requires cautious interpretation due to the limited statistical validation in the source literature.

CONCLUSIONS

This review synthesizes the current landscape of ML-based PD assessment using accessible hardware. It demonstrates the technical viability and performance advantages of multimodal systems while providing a detailed compendium of methodologies for signal processing, feature engineering, and model selection. These findings support the development of practical, cost-effective tools for remote PD screening and monitoring.