A Multi-Model Machine Learning Approach for Predictive Analytics and Intelligent Decision Making

Abstract

Disease risk can be predicted and clinical decisions can be made early based on wearable sensor data of physiological and behavioral parameters. However, patterns in wearable data of disease development are mostly non-linear, time-dependent, and influenced by individual health variations. Therefore, it is challenging to get reliable results with a single machine learning model for disease risk prediction. This study introduces a multi-model framework using wearable sensor data to improve early disease risk prediction and support timely clinical decisions. First, the data is processed using a data cleaning, normalization, missing value handling, feature extraction and class balancing. Then, a variety of predictors such as Support Vector Machine, Random Forest, Extreme Gradient Boosting and Long Short-Term Memory networks are trained and tested. The prediction probabilities of the single models are integrated in a stacking-approach. An intelligent decision making component classifies users into different disease risk categories (low, moderate, high) based on predicted probabilities and on clinical decision making thresholds. The framework is complemented by Explainable AI to identify the most influential physiological parameters for individual user risk scores. Performance is measured by accuracy, precision, recall, F1-score, area under ROC curve, sensitivity, specificity and by calibration. In comparison to individual machine learning models, the proposed multi-model framework enables more robust and interpretable disease risk predictions and supports health monitoring, early clinical interventions as well as evidence-based clinical and healthcare decisions.