Objective: This study aims to evaluate the potential of a wearable sensor-based walking and balance assessment to predict the risk of falling in individuals with Parkinson’s Disease (PD) over a five-year period.

Background: PD is associated with an increased risk of falling,  but there is no accepted objective measure that can provide accurate fall risk estimations. This study aims to investigate the effectiveness of a short, sensor-based assessment in accurately predicting fall risk among PD patients.

Method: 104 PD patients participated in the study. They performed a 2-minute walk and a 30-second standing sway task while wearing six Inertial Measurement Unit sensors. Participants were followed up for 2 years with in-clinic visits and then for another 3 years via phone-calls and fall-record surveys. A preliminary feature selection process was conducted to identify the most informative features, including baseline clinico-demographic data, that could contribute to predicting the onset of falling. Class imbalances were corrected prior to model training. Three Machine Learning classifiers (Random Forest, Support Vector Machine and Elastic Net) were benchmarked to predict the onset of falling within 5 years. All models were trained with and without a preliminary feature selection  and evaluated with a 5-fold cross validation.

Results: The Random Forest model trained on the entire set of features demonstrated superior performance in predicting fall risk  (Area Under the Curve [SP1] 0.94 at 5 years (95% CI 0.84-0.99)). Walking variability features were the most important contributors to the model’s performance. While clinico-demographic variables showed significant differences between fallers and non-fallers, they failed to enhance the performance of the model.

Conclusion: Wearable sensors and Machine Learning can accurately predict fall risk in individuals with PD over five years. Walking variability is a crucial fall risk predictor, and often missed by human observation. The relatively low weighting of the clinico-demographical features to the model performance highlights the significance of sensor data in fall risk prediction among individuals with PD. The current study paves the way for integrating wearable sensor-based assessments into regular clinical practice  for falls risk assessment.

« Back to 2024 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/prediction-of-falls-in-pd-over-5-years-using-kinematic-data-and-machine-learning/