Machine learning for the diagnosis of Parkinson's disease: A systematic review

Jie Mei, Christian Desrosiers, Johannes Frasnelli

Abstract

Diagnosis of Parkinson's disease (PD) is commonly based on medical observations and assessment of clinical signs, including the characterization of a variety of motor symptoms. However, traditional diagnostic approaches may suffer from subjectivity as they rely on the evaluation of movements that are sometimes subtle to human eyes and therefore difficult to classify, leading to possible misclassification. In the meantime, early non-motor symptoms of PD may be mild and can be caused by many other conditions. Therefore, these symptoms are often overlooked, making diagnosis of PD at an early stage challenging. To address these difficulties and to refine the diagnosis and assessment procedures of PD, machine learning methods have been implemented for the classification of PD and healthy controls or patients with similar clinical presentations. To provide a comprehensive overview of data modalities and machine learning methods that have been used in the diagnosis and differential diagnosis of PD, this study conducted a systematic literature review of studies published until February 14, 2020. A total of 209 studies were included, extracted for relevant information, and presented with an investigation of their aims, sources of data, types of data, machine learning methods, and associated outcomes. These studies demonstrate a high potential for adaptation of machine learning methods and novel biomarkers in clinical decision making, leading to increasingly systematic, informed diagnosis of PD.

Introduction

Traditionally, Parkinson’s disease is diagnosed based on clinical observations of motor symptoms, such as tremors, rigidity, and difficulty initiating movement. However, this approach has significant drawbacks: it relies heavily on subjective human evaluation, subtle early-stage movements are easy to miss, and non-motor symptoms (like sleep disorders or loss of smell) are often overlooked because they overlap with other conditions.

Consequently, early and accurate diagnosis of PD remains a major clinical challenge. Machine learning (ML) offers a solution by automatically extracting complex, hidden patterns from patient data. Because the landscape of ML research in PD is vast and fragmented, this paper provides a comprehensive systematic review to map out exactly which data types and algorithms are actually working, aiming to bridge the gap between computer science research and clinical application.

Methodology

The authors conducted a systematic literature review following PRISMA guidelines, searching the PubMed and IEEE Xplore databases for papers published up to February 14, 2020. They filtered the results to include only studies that used ML to classify patients with PD against healthy controls (HC), or against patients with similar atypical Parkinsonian syndromes.

PRISMA flow diagram showing the study selection process.

Ultimately, 209 studies were extracted and analyzed. The authors categorized the studies based on the data modalities used (e.g., voice, gait, MRI) and the ML algorithms deployed (e.g., SVM, Neural Networks, Ensemble Learning, Decision Trees), evaluating them primarily on classification accuracy.

Data Sources and Sample Sizes

The review found a near-even split between studies using proprietary, locally collected clinical data and those utilizing open-source public repositories.

  • Public Repositories (51.7% of studies): The most popular included the UCI Machine Learning Repository (44 studies), the Parkinson's Progression Markers Initiative (PPMI) database (33 studies), PhysioNet (15 studies), the HandPD dataset (6 studies), and the mPower mobile database (4 studies).
  • Independent Recruitment (43.1% of studies): Data collected directly from human participants by the researchers.
Distribution of subject counts and data sources across the analyzed studies.

Sample sizes ranged drastically from as few as 10 subjects to as many as 2,289, with an average of about 185 subjects per study. As shown in the distribution above, a significant portion of studies relied on smaller cohorts, which can present challenges for model generalization.

Key Insights and Diagnostic Performance

Across all modalities that used accuracy as a metric, the average diagnostic accuracy was an impressive 89.9%. Several key findings emerged from the analysis:

  • Voice as a Powerful Biomarker: Voice data was the most frequently used modality (55 studies) and yielded highly accurate diagnostic results (90.9% average accuracy), proving that non-invasive, easily captured biometrics hold massive clinical potential.
  • The SVM Workhorse: Support Vector Machines (SVMs) were the undisputed heavyweight champions of PD classification. They were used in 130 of the 209 studies and consistently achieved top-tier performance across diverse data types.
  • Multimodal Synergy: Combining different data types (e.g., MRI scans with SPECT imaging) led to highly promising results, suggesting that the future of PD diagnosis lies in holistic, multi-sensor systems.
  • Performance by Modality:
    • SPECT imaging (14 studies) averaged 94.4% accuracy.
  • Movement/gait data (51 studies) averaged 89.1% accuracy.
  • MRI scans (36 studies) averaged 87.5% accuracy.

Limitations and Future Directions

The authors note several limitations in the current body of research. Poor reporting standards are common; many studies failed to detail hyperparameter tuning or cross-validation strategies, hindering reproducibility. Furthermore, class imbalance is a frequent issue that can artificially inflate simple accuracy scores.

For future work, the authors recommend using more robust metrics like F1-score, precision, and recall. To combat the "small data" bottleneck, they suggest combining local clinical cohorts with large public databases like PPMI to train more robust, generalizable models for real-world clinical decision-making systems.

Summary updated

Abstract

Diagnosis of Parkinson's disease (PD) is commonly based on medical observations and assessment of clinical signs, including the characterization of a variety of motor symptoms. However, traditional diagnostic approaches may suffer from subjectivity as they rely on the evaluation of movements that are sometimes subtle to human eyes and therefore difficult to classify, leading to possible misclassification. In the meantime, early non-motor symptoms of PD may be mild and can be caused by many other conditions. Therefore, these symptoms are often overlooked, making diagnosis of PD at an early stage challenging. To address these difficulties and to refine the diagnosis and assessment procedures of PD, machine learning methods have been implemented for the classification of PD and healthy controls or patients with similar clinical presentations (e.g., movement disorders or other Parkinsonian syndromes). To provide a comprehensive overview of data modalities and machine learning methods that have been used in the diagnosis and differential diagnosis of PD, in this study, we conducted a systematic literature review of studies published until February 14, 2020, using the PubMed and IEEE Xplore databases. A total of 209 studies were included, extracted for relevant information and presented in this systematic review, with an investigation of their aims, sources of data, types of data, machine learning methods and associated outcomes. These studies demonstrate a high potential for adaptation of machine learning methods and novel biomarkers in clinical decision making, leading to increasingly systematic, informed diagnosis of PD.
BibTeX
@inproceedings{mei2021machine,
  author = {Jie Mei and Christian Desrosiers and Johannes Frasnelli},
  title = {{Machine learning for the diagnosis of Parkinson's disease: A systematic review}},
  booktitle = {Frontiers in Aging Neuroscience},
  year = {2021},
  doi = {10.3389/fnagi.2021.633752},
}
Figure