Cluster-Driven PPG BP Estimation and Systematic Review

Abstract

Background: Hypertension remains a global health challenge, affecting over 1.28 billion individuals and contributing significantly to cardiovascular diseases. Traditional BP monitoring methods, such as cuff-based devices, provide only intermittent readings, limiting their ability to capture real-time fluctuations. Wearable, non-invasive BP monitoring technologies have emerged as a promising alternative, offering continuous measurement capabilities. However, these methods face persistent challenges, including motion artifacts, sensor inaccuracies, and calibration instability, which hinder their real-world applicability. This dissertation presents an innovative ML framework to enhance BP estimation accuracy, addressing critical limitations in current wearable BP monitoring solutions. Objectives: The primary objectives of this study are: I. To evaluate the effectiveness of wearable BP monitoring devices through a systematic review and meta-analysis, assessing their accuracy, usability, and compliance with clinical standards. II. To develop a novel clustering-based ML framework that improves BP estimation accuracy by integrating feature selection, dimensionality reduction, and subgroup-specific calibration techniques. III. To validate the proposed model using real-world PPG-based datasets (MIMIC and WHeMoBoard) and assess its applicability to wearable BP monitoring. Methods: A systematic review was conducted to analyze wearable BP estimation techniques, focusing on their accuracy and limitations in clinical and real-world settings. Following this, a novel clustering-driven ML framework was developed, incorporating K-means clustering, recursive feature elimination (RFE), and t-distributed Stochastic Neighbor Embedding (t-SNE) to enhance model performance. This framework was trained and validated on large-scale datasets, including the MIMIC intensive care unit (ICU) database and the WHeMoBoard wearable dataset, utilizing Random Forest, Gradient Boosting, and Support Vector Machines as predictive models. Results: The proposed clustering-based ML framework significantly improved BP prediction accuracy compared to traditional models. Key findings include: I. Mean Absolute Error (MAE) reduction from 3.41 mmHg to 2.94 mmHg for SBP and from 2.34 mmHg to 2.23 mmHg for DBP. II. Increased prediction accuracy within ±5 mmHg from 78.8% to 82.5% for SBP and from 88.2% to 89.1% for DBP. III. Reduced need for frequent recalibration, enhancing usability for long-term wearable applications. Conclusion: By integrating clustering, feature selection, and ML-based calibration, this study bridges the gap between generalized and personalized BP estimation models. The results highlight the potential of AI-driven wearable BP monitoring for continuous, real-time health tracking, particularly in managing chronic conditions like hypertension. Future research should focus on refining real-time adaptive clustering, multi-sensor fusion, and hybrid calibration models to further improve the robustness and clinical applicability of non-invasive BP monitoring technologies. These advancements will pave the way for scalable, AI-powered BP monitoring solutions, transforming preventive healthcare and remote patient management.