Conceptualization of Community (Advice)-based Health eCoach Recommendation Generation for Physical Activity Level Monitoring using an Explainable and Ethical Approach

Utilizing a digital platform for self-monitoring offers advantages for remote activity tracking and generating recommendations. An automatic coaching system, known as eCoach, can be a valuable tool. It consistently gathers individual health and wellness data for the purpose of creating lifestyle recommendations, either tailored to an individual or based on community input. While personal recommendations in eCoaching are well-received, there is a lack of research on community-based recommendations. This study introduces an effective community-based eCoach recommendation system for assessing daily activity levels and offering physical activity suggestions. The approach involves collecting data from remote participants using wearable activity sensors and categorizing the data into sedentary (0) and active (1) classes for binary classification. The dataset used for physical activity levels exhibits an imbalance in class distribution and unequal misclassification costs, making imbalanced classification challenging. To address this issue and improve the performance of traditional machine-learning models, the study employs various oversampling techniques, including the Synthetic Minority Oversampling Technique (SMOTE), to create balanced datasets by generating synthetic samples from the minority class. Third, we explain the model classifications with Shapley Additive Explanations (SHAP). Fourth, we use hyperparameter tuning methods to obtain the best classification performance. Fifth, we consider a preference, activity patterns, and demographic attributes, such as age, gender, height, and weight, for advice and community-based recommendations to meet activity goals. For experimentation, we collected activity data from 16 participants (Male: 12; Female: 4) using a wearable MOX2-5 activity sensor (CE-certified) and combined the dataset with the public Fibit-based PMData activity dataset of 15 participants (Male: 12; Female: 3) to avoid bias, and for verification and advice-based recommendation planning, we considered last week’s participant data. The system’s performance has been evaluated using metrics such as accuracy, precision, recall, F1-score, brier-score, and Mathew’s coefficient (MCC) and achieves 99.8% accuracy (F1 = 99.8%, Precision = 99.8%, Recall = 99.8%, and MCC = 99.8%) with DecisionTree Classifier on the imbalanced dataset and 99.8% accuracy (F1 = 99.8%, Precision = 99.8%, Recall = 99.8%, and MCC = 99.8%) with DecisionTree Classifier on the balanced dataset. Ethical AI involves more than just coding; it entails ethical considerations and decision-making throughout development. In this study, we evaluated our models for bias and fairness, concentrating on the recommendation generation within an eCoach system. Future research will focus on efficacy evaluation in controlled trials, extending beyond this theoretical investigation.