Joint estimation of multiple cardiovascular parameters from blood pressure waveforms using multi-task learning

Abstract Estimation of cardiovascular parameters using blood pressure waveforms has been widely investigated for cardiovascular health assessment. However, previous studies have focused mainly on the independent estimation of vascular age, stroke volume, or peripheral resistance, while joint estimation within a unified framework remains underexplored. Given that these parameters jointly shape blood pressure waveform morphology, we formulated their estimation as a multi-task learning problem. Yet, parameter-specific temporal effects may cause negative transfer in naive multi-task architectures. Thus, we proposed a cross-talk layer, termed residual neural dimension reductor with temporal and channel attention (RNDR-TCA), which models nonlinear inter-task relations while preserving a main task feature through residual connections. Additionally, temporal and channel attention modules enhance intra-task temporal features and inter-task feature interactions, respectively, thereby improving task-specific feature extraction and selective inter-task information sharing. Experimental validation was conducted using a physiologically credible cardiovascular simulator dataset, with independently controlled parameter values treated as classification targets. RNDR-TCA achieved the highest average Macro F1 score of 0.836, outperforming separate single-task learning models (0.784), a multi-class classifier (0.771), shared-trunk architectures (0.781-0.816), and existing cross-talk architectures (0.792-0.824). Ablation studies confirmed the contributions of both attention modules, and RNDR-TCA also showed the best performance under a class-imbalanced setting.