RELIABLE CUFFLESS BLOOD PRESSURE MONITORING USING MULTIPLE ARTIFICIAL INTELLIGENCE MODELS

Cardiovascular diseases remain the leading cause of death globally, with a life lost every three seconds in the U.S. Early detection and effective hypertension management are crucial to reducing its impact. However, traditional cuff-based blood pressure (BP) monitors, while accurate, pose usability challenges due to their size, cost, and complexity, particularly for in-home monitoring. Although cuffless BP monitors offer simplicity, their need for frequent calibration against traditional devices limits widespread adoption. This study aimed to develop artificial intelligence (AI) models capable of accurately estimating cuffless BP without the need for periodic calibration.

The research utilized data from 147 participants using the Avidhrt Sense device, incorporating demographic variables such as BMI, age, and gender, alongside physiological signals like ECG and PPG captured from the fingertips to enhance predictive accuracy. Additionally, novel features—including SpO₂ filtering, skin temperature, environmental temperature, core body temperature, and ECG classification—were integrated to further enhance model performance, particularly for diastolic BP estimation. The study employed Multiple Linear Regression, XGBoost, Feedforward Neural Networks, and a Hybrid model combining Convolutional Neural Networks with Recurrent Neural Networks. Among these models, XGBoost achieved the highest accuracy, with a Mean Squared Error of 6.15 and a Mean Error of -0.67 ± 2.39 for systolic BP, and a Mean Squared Error of 10.03 with a Mean Error of 0.44 ± 3.14 for diastolic BP. These results represent one of the best performances reported in cuffless BP measurement research.

The findings indicate that AI-enhanced cuffless BP monitoring, when augmented with additional physiological features, can achieve accuracies meeting ANSI/AAMI standards, making it a viable alternative to traditional BP monitors. Furthermore, excluding socioeconomic factors and race from model inputs reduced potential biases, thereby enhancing the model’s generalizability across diverse populations. Future research should focus on expanding the dataset, exploring continuous monitoring, and integrating real-time feedback systems to further enhance clinical applicability.