Digital Twins for Stress Management Utilizing Synthetic Data

In the era of Medical 4.0, technologies such as big data, wearables, and Machine Learning (ML) are being deployed for predictive healthcare delivery. In this regard, digital twins have been adopted in healthcare to enhance diagnosis and personalized treatment. Health Digital Twins (HDTs) are virtual representations of patients' data, mirroring the health state of patients to provide insights. Despite its promise, the existing works on HDTs relied on large historical data to train ML models. These historical data may be difficult to obtain due to privacy concerns of data fiduciaries and subjects. In this paper, we propose a Digital Twin for Stress Management (DTSM) that employs generative models to learn the distribution of patients' data retrieved from a wearable device for stress management score prediction. To obtain a virtual replica of a patient's data, we used synthetic data generative models such as Conditional Tabular Generative Adversarial Network (CTGAN), Tabular Variational Autoencoder (TVAE), Gaussian Copula, and Large Language Models (LLM) (REaLTabFormer and GReaT). The best result came from REaLTabFormer which accurately learns the distributions of the real data with a data quality score of approximately 93%. Furthermore, four well-known ML models trained on the synthetic data obtained a mean absolute error (MAE) of less than 5% in the prediction of stress score. Our experimental results show that the proposed DTSM can be used for the prediction of stress management scores.