A machine learning-based fall risk prediction model for Parkinson's disease considering ophthalmic disorders

The association between fall risk and ophthalmic disorders has not been well understood. We used an open dataset to examine predictive models of fall risk at one year among the patients with Parkinson's disease (PD) using machine learning.

A dataset from participants of the Parkinson's Progression Markers Initiative (PPMI) study, a US-based open-access database was used. The analysis data comprised 543 individuals with PD, with data on baseline and 1-year follow-up fall history, neurological symptoms, and ocular findings. Predictive models were examined by categorical data analysis program (CATDAP) and random forest, with the presence or absence of falls at 1 year as the response variable and clinical symptoms as explanatory variables. We performed stratified random data split on the analysis data, with 80% used as learning set and 20% as test set. The prediction model was constructed from the learning set, and the model performance was evaluated using the test set.

The random forest model with feature selection using CATDAP achieved an MCC of 0.456, accuracy of 82.6%, PPV of 50%, NPV of 91.8%, sensitivity of 63.2%, specificity of 86.7%, and F1-score of 0.558 in the test set. The variables with the greatest impact on fall risk prediction performance were "fall history", followed by "glaucoma", "arising from chair", and "gait".

Glaucoma is a risk factor for falls in addition to fall history and neurological symptoms. The potential contribution was indicated to prevent falls in the patients with PD from the ophthalmologic field in clinical practice.