Exploratory Analysis of Machine Learning Methods for the Prognosis of Falls in Elderly Care Based on Accelerometer Data

This paper investigates the feasibility of employing machine learning techniques to categorize individuals into fall-risk and non-fall-risk groups based solely on accelerometer data. The research utilizes a publicly available movement monitoring dataset, containing accelerometer data from a diverse group of individuals. The study pursues three primary objectives. First, it develops a preprocessing pipeline to prepare raw accelerometer data, which includes noise reduction, data cleaning, and identification of walking segments and the extraction of over twenty gait-related features. The second objective is to systematically explore the influence of these features on machine learning model performance. Gait stability-related parameters, known from medical literature, are of particular interest. To fulfil this objective, different machine learning algorithms are evaluated using an automated exploration framework. The third objective centres on finding a balanced combination of features and lightweight machine learning models suitable for embedded systems, which typically have limited computational resources. The emphasis here is on computational efficiency, an original aspect of this study. The results indicate that gradient boosting algorithms, such as XGBoost, LightGBM, and CatBoost, outperform other models, achieving promising performance results, including an area under the curve (AUC) score of up to 0.949.