Abstract
<title>Abstract</title> <p>Digital twin systems require reliable sensing, efficient computation, and adaptive modeling to support continuous health monitoring. These requirements are especially challenging in LMIC environments, where cost, power, and hardware resources are limited. This paper presents a simulation-based design and validation of an adaptive multimodal biosensing digital twin framework for resource-constrained deployment. The system integrates EEG, PPG, GSR, IMU, and temperature signals to monitor stress, sleep, and emotion. Signals were denoised using bandpass and notch filtering with moving-average smoothing, producing stable data streams for analysis. Lightweight machine learning models, specifically CNNs and Random Forest classifiers, were evaluated on synthetic physiological data within a Raspberry Pi class edge environment. The experiments demonstrated real-time inference with end-to-end latency consistently below 200 ms, while maintaining strong classification performance. For stress detection, Random Forest achieved 99.7% accuracy compared to 96.7% for CNN. Sleep classification reached a perfect 100% accuracy with Random Forest. Emotion recognition showed similarly high results, with Random Forest at 99.7% and CNN at 99.3%. The CNN training curves converged smoothly, stabilizing at a loss of approximately 0.2 and an accuracy of about 0.95 by the twentieth epoch.. Overall, simulation-driven development provides a robust foundation for digital-twin healthcare monitoring in LMIC contexts, achieving high accuracy and low latency under constrained conditions. Future work will validate the framework with real physiological signals to ensure robustness beyond synthetic datasets.</p>