TL;DR
This paper reviews 209 studies applying machine learning (ML) to diagnose Parkinson's Disease (PD). It highlights ML's potential to overcome the subjectivity and challenges of traditional PD diagnosis by analyzing diverse data types, from voice recordings to MRI scans, achieving high accuracy. The review aims to provide a comprehensive overview of ML methods and data sources, guiding future research and clinical applications for more systematic and informed PD diagnosis.
What Problem Does This Paper Address?
Diagnosing Parkinson's Disease (PD) is tricky. Traditional methods rely on observing motor symptoms, which can be subtle and difficult to classify, leading to potential misdiagnosis. Early non-motor symptoms are often mild, non-specific, and easily overlooked, making early detection a significant challenge. This subjectivity and difficulty in early detection hinder timely intervention and treatment. The paper addresses this by exploring how machine learning (ML) can offer a more objective, systematic, and potentially earlier approach to diagnosing PD and differentiating it from similar conditions.
Methods
This paper is a systematic literature review. The authors searched PubMed and IEEE Xplore databases for studies published up to February 14, 2020, using specific boolean search strings (detailed in Table 1). They included studies that used ML for PD diagnosis (vs. healthy controls, vs. other Parkinsonian syndromes, or subtyping PD) and excluded those not focused on PD diagnosis, lacking sufficient methodological detail, or not in English. A total of 209 studies were included. Data extraction covered objectives, diagnosis type, data source, data type, number of subjects, ML methods, splitting strategies, cross-validation, outcomes, year, and references. Model evaluation primarily focused on accuracy, but other metrics like sensitivity, specificity, and AUC were also considered.
Datasets
The review analyzed studies that utilized a variety of data sources and types. Key public repositories and databases mentioned include:
- University of California at Irvine (UCI) Machine Learning Repository: Frequently used for voice and other data.
- Parkinson's Progression Markers Initiative (PPMI): A major source for clinical and imaging data.
- PhysioNet: A repository for physiological signals.
- mPower database: Mobile health data collected via smartphone apps.
- HandPD dataset: Specifically for handwriting analysis.
Many studies also collected their own data from human participants, often combining it with public datasets.
Key Insights and Ideas
- Diverse Data, High Accuracy: ML models can achieve impressive diagnostic accuracy across a wide range of data modalities, including voice, movement, handwriting, and neuroimaging (MRI, SPECT). This suggests that subtle patterns, often missed by human observation, can be effectively captured by ML.
- Beyond Motor Symptoms: ML's ability to analyze non-motor data (like voice patterns or even handwriting) opens doors for earlier and more comprehensive PD detection, potentially before significant motor deficits appear.
- The Power of Combination: Combining multiple data types (e.g., MRI and SPECT) often leads to improved diagnostic performance, highlighting the benefit of a holistic approach.
- Methodology vs. Clinical Application: The review identified two main types of studies: "methodology" studies focusing on developing new ML models or techniques, and "clinical application" studies validating existing methods or exploring new data modalities in real-world settings. A gap exists in translating promising methodology studies into validated clinical applications.
- Standardization is Key: The lack of standardized reporting for ML methods, datasets, and validation strategies across studies makes direct comparison and replication challenging, hindering progress.
Key Results
- High Average Accuracy: Across various data types, ML models achieved an average diagnostic accuracy significantly above chance, often ranging from 85% to over 94%. For instance, voice recordings yielded an average accuracy of 90.9%, while SPECT imaging averaged 94.4%.
- Popular ML Models: Support Vector Machines (SVM), neural networks, and ensemble learning were the most frequently used and often achieved the highest per-study accuracy.
- Sample Size Matters (but not always): While larger sample sizes generally correlate with better model performance, some studies with smaller datasets still reported high accuracies, suggesting that well-chosen features and robust models can be effective.
- Data Modalities: Voice recordings, movement data, and MRI were the most commonly used data modalities.
Limitations and Future Work
- Data Imbalance and Leakage: The paper acknowledges the challenge of imbalanced datasets (more healthy controls than PD patients, or vice versa) and the risk of data leakage when data from the same subject is used across training, validation, and testing sets, which can inflate accuracy.
- Reproducibility Issues: Insufficient or inaccurate reporting of methods, datasets, and results in many studies makes replication difficult. The authors emphasize the need for more transparent reporting standards.
- Limited Scope of Review: The review excluded non-English articles and conference abstracts, potentially introducing publication bias. It also noted a lack of large-scale, multi-centric studies on PD subtyping and severity assessment, which are crucial areas for future research.
- Small Sample Sizes in Specific Modalities: For data types like PET and CSF, the number of studies was too small to draw generalizable conclusions about ML performance.
- Future Directions: The authors suggest combining data from local cohorts with public repositories to overcome small sample size limitations. They also highlight the need for greater transparency in reporting ML methodologies to facilitate replication and clinical adoption. Further research into subtyping and severity assessment using ML is also recommended.
Summary updated