Are smartphones and machine learning enough to diagnose tremor?

Area of Science:

• Neurology
• Biomedical Engineering
• Machine Learning

Background:

• Essential tremor (ET), Parkinson's disease (PD), and dystonic tremor (DT) present overlapping symptoms, complicating accurate diagnosis.
• Objective classification of these tremor types is crucial for effective patient management.

Purpose of the Study:

• To develop and evaluate machine learning models using smartphone accelerometer data for automated tremor classification.
• To assess the utility of clinical features alongside accelerometer data in improving classification accuracy.
• To explore unsupervised learning for identifying natural patient groupings and their correlation with clinical diagnoses.

Main Methods:

• Trained a supervised machine learning classifier on data from 78 tremor patients using leave-one-out cross-validation.
• Validated the classifier on an independent cohort of 27 patients.
• Compared classifiers trained on accelerometer data versus those trained on clinical metrics for a subset of 48 patients.

Main Results:

• Achieved 74.4% accuracy for trinary classification (PD, ET, DT) with an AUC of 0.904.
• Demonstrated high accuracy (97% specificity, 84% sensitivity) in distinguishing PD from ET/DT.
• Showed moderate accuracy for ET (88%) but poor accuracy for DT (29%).
• Independent cohort validation yielded poorer performance; accelerometer and clinical data classifiers performed similarly.

Conclusions:

• Machine learning models show promise for classifying PD but struggle with ET and particularly DT.
• The limited accuracy highlights the diagnostic challenges posed by overlapping clinical and neuropathological features in these tremor disorders.
• Findings suggest that ET, PD, and DT may represent overlapping clinical syndromes rather than distinct entities.