First, we investigated the causal relationships between lifestyle habits, health awareness, and health status to establish foundational insights. We applied the NOTEARS algorithm for causal discovery on data collected from wearable devices and self-assessment questionnaire. By performing subgroup analyses based on gender, BMI, and health awareness, our findings revealed causal structures across cohorts. These results suggest that a multifaceted healthcare approach is essential for promoting healthier behaviors. Furthermore, the uncovered causal structures provide an explainable framework for constructing a healthcare knowledge graph, facilitating the advancement of personalized precision healthcare.

Then, to enhance the reliability and interpretability of multivariate health data analysis, we proposed a two-stage Discovery-and-Validation framework designed for robust causal inference from time-series health data. The first stage employs Conditional Transfer Entropy (CTE) for discovering of potential causal relationships and their corresponding time lags, while mitigating the effects of confounding variables. The second stage validates each relationship discovered in stage one. This involves quantifying the causal effect size and confirming the results' robustness through a placebo refutation test. In our experiment, we used a dataset that we collected via wearable devices, comprising 16 health indicators. The results show that there are seven pairs of time-lagged causal relationships identified from the discovery stage. Remarkably, all seven relationships were validated in the validation stage, each passing the refutation test with a p-value > 0.05, and the causal effects were quantified.

These contributions resulting in two peer-reviewed papers, which published in IEEE international conferences. We extend our sincere gratitude to Waseda University for the Special Research Project funding that made these work possible.