Abstract:
The article examines the effectiveness of the cuffless blood pressure (BP) estimator based on photoplethysmography (PPG) signals, patient demographics, and machine learning. A prototype of IoT was created on the basis of an ESP32 microcontroller and a sensor of optical measurements of the maximal30102 to obtain physiological measurements. It was deployed with a full pipeline like signal processing, morphology and nonlinear features, feature selection (Mutual Information) with Recursive Feature Elimination (RFE) and SHAP analysis, and, finally, training of some machine learning models. CatBoost and Random Forest were optimized by randomizing some values. Although the approach was rigorous, predictive accuracy was lower than clinical. The highest performing models were those with a mean absolute error (MAE) of over 7 mmHg and R2R 2 of less than 0.25. These results reveal that superior machine learning algorithms are unable to combat low input quality. The makeup of the sensor, high sensitivity (maximum 15-15 ohms) and restricted waveform fidelity of the MAX30102 sensor was insufficient to capture the morphology features needed for accurate BP determination. Also, the sample size was not adequate, and the sample was not demographically balanced Abbott et al., 2019), which may further undermine the generalizability. The novelty of the work is the evidence of the limitation: revealing that cheap PPG sensors like the MAX30102 cannot be casually used to estimate BP. To enhance cuffless BP monitoring, future investigations need to utilize better quality sensors, larger, more heterogeneous datasets, and consolidated validation systems.
