Demographic subgroup-aware analysis of blood pressure predictive models

Blood pressure (BP) is an important cardiovascular health indicator, usually monitored non-invasively with a sphygmomanometer, which can be bulky and inconvenient for real-time monitoring. Blood pressure estimation using photoplethysmography (PPG) waveform analysis offers a promising alternative for intermittent or continuous cuff-less BP monitoring. However, changes in vascularization due to individual physiology and age can directly affect the morphology of the PPG waveform. Thus, prediction accuracy requires a demographic-aware approach. To address these challenges, we trained three conventional ML models for BP prediction and conducted a subgroup-aware analysis to evaluate model performance across individual and intersectional demographic groups. Additionally, we show that signal quality assessment is a pivotal step in the signal-processing pipeline before feature extraction to mitigate external perturbations (e.g., motion artifacts) in the PPG waveform. We achieve a mean absolute error (MAE) of 3.36 mmHg for systolic blood pressure (SBP) and 1.76 mmHg for diastolic BP (DBP) with models trained when signal quality assessment is incorporated.Author summary: Recent advancements in machine learning and artificial intelligence have led to the development of numerous predictive models to derive blood pressure estimates from PPG signals. Generally, the performance of blood pressure (BP) prediction models is reported as an average across the entire dataset without accounting for the distribution of different populations within the data. However, recent studies have highlighted significant disparities in the performance of machine learning models across diverse populations, raising concerns about equity and fairness. While many of these studies are not directly related to blood pressure prediction models, conducting such evaluations in this domain is crucial to ensure the robustness and equity of novel BP estimation methods that could be widely adopted. Dataset bias can arise from various distinct biases embedded in data-generating processes. Certain biases, including representation bias along demographic lines, can result in skewed outcomes. Performance disparities can stem from various factors, including differences in demographic characteristics, data quality, and model training methodologies. This paper aims to examine disparities in blood pressure prediction models, identifying key factors contributing to inconsistencies and enhancing model performance across diverse populations.