RiM: Record, Improve and Maintain Physical Well-being using Federated Learning

TL;DR

RiM addresses privacy-sensitive wellness support for students by combining federated learning with a lightweight mobile recommendation engine. The approach pre-trains an MLP on a large synthetic dataset and privately fine-tunes on real on-device data via Flower, using both FedAvg and FedPer to balance global learning and personalization. A hybrid recommender integrates ML-predicted deficits with rule-based risk scoring to surface the top personalized lifestyle recommendations. Experimental results on 10 clients show FedAvg achieving higher accuracy (60.71%) and lower MAE (0.91) than FedPer (46.34%, 1.19), illustrating the privacy-preserving potential for real-world wellness interventions; Android API support was limited to version 13, with planned enhancements including on-device inference and Ditto-based bi-level FL.

Abstract

In academic settings, the demanding environment often forces students to prioritize academic performance over their physical well-being. Moreover, privacy concerns and the inherent risk of data breaches hinder the deployment of traditional machine learning techniques for addressing these health challenges. In this study, we introduce RiM: Record, Improve, and Maintain, a mobile application which incorporates a novel personalized machine learning framework that leverages federated learning to enhance students' physical well-being by analyzing their lifestyle habits. Our approach involves pre-training a multilayer perceptron (MLP) model on a large-scale simulated dataset to generate personalized recommendations. Subsequently, we employ federated learning to fine-tune the model using data from IISER Bhopal students, thereby ensuring its applicability in real-world scenarios. The federated learning approach guarantees differential privacy by exclusively sharing model weights rather than raw data. Experimental results show that the FedAvg-based RiM model achieves an average accuracy of 60.71% and a mean absolute error of 0.91--outperforming the FedPer variant (average accuracy 46.34%, MAE 1.19)--thereby demonstrating its efficacy in predicting lifestyle deficits under privacy-preserving constraints.

Figures (5)

• Figure 1: The figure illustrates the distributions of simulated features used for pre-training the MLP model. Simulated step counts follow a negative binomial distribution, while distance traveled and weight are drawn from log-normal distributions. In addition, meal consumption and gender are generated using Bernoulli distributions, and age is modeled based on an empirical distribution. The figure also illustrates the normal distributions for sleep hours and height.
• Figure 2: The figure presents the distribution of meal consumption—breakfast, lunch, and dinner—where a value of 1 indicates the meal was taken and 0 indicates it was skipped. It also shows the gender distribution within the simulated dataset. All variables follow a Bernoulli distribution.
• Figure 3: Architecture of the proposed MLP model, which consists of five hidden layers and outputs sleep and distance deficits that feed into the recommendation system to generate personalized recommendations.
• Figure 4: Training loss and validation accuracy curves plotted across pre‑training iterations of the MLP model.
• Figure 5: The figure illustrates the FedPer weight‐update process: orange neurons denote the shared root layers that are sent to the server averaged via FedAvg, while blue neurons denote the personal head layers that remain on the client device to preserve personalization.