Multi-datasets transfer multitask learning for simultaneous blood glucose and blood pressure monitoring using common PPG features

The simultaneous monitoring of both blood glucose level (BGL) and blood pressure (BP) has rarely been studied directly. The exploitation of physiological interactions between them will advance the learning of either task. However, the lack of available datasets with labels of both targets presents an obstacle. Therefore, in this paper, we propose three methods for multi-dataset (MD) learning. First, we extract PPG features from three datasets: a source dataset comprising diabetes mellitus (DM) and hypertension (HTN) classes labels and two target datasets for each BP and BGL. Subsequently, we select a common merged feature set for all datasets tasks. This study experiments with the three proposed multi-dataset methods: 1) transfer learning (MD-TL) to transfer knowledge from a source task-DM and HTN single-task classifier, to a target task, BGL and BP single-task regression, respectively, 2) multitask learning (MD-MTL) of both targets via a shared two-way TL, and 3) combined TL with MTL (MD-TL-MTL) to transfer knowledge from a source multitasking classifier to the target tasks in the MD-MTL network. The final proposed MD-TL-MTL achieves an MAE±SD of 2.53 ± 3.73 for SBP, 1.47 ± 1.84 for DBP, and MAE of 1.36 for BGL. Clarke error grid analysis shows that 99.86 % of samples fall into zone A. Overall, the MD-TL-MTL improves performance in all tasks compared to baseline models. An interpretability analysis using Shapley Additive Explanations (SHAP) and permutation importance is conducted to facilitate the clinical understanding behind predictions.