Non-invasive methods for continuous blood pressure estimation from sensor data

Hypertension represents a significant health risk, potentially leading to serious complications, making it a primary cause of death around the world. Continuous blood pressure (BP) monitoring is considered essential for ensuring reliable diagnosis. However, invasive methods, even though a gold standard for BP monitoring, can lead to serious complications. Non-invasive techniques are commonly used but cannot monitor BP continuously. In this context, it is crucial to develop a non-invasive, safe, and comfortable method for continuous BP monitoring during daily activities. Alternatively, photoplethysmography (PPG), a portable optical sensor that continuously detects volumetric changes in blood circulation, offers a promising solution for non-invasive and continuous BP monitoring. However, BP estimations derived exclusively from PPG signals frequently suffer from limited accuracy and fail to meet clinical standards. To overcome these limitations, recent approaches have focused on integrating PPG with electrocardiographic (ECG) signals, using Pulse Transit Time (PTT) based method and artificial intelligence (AI) based techniques. Despite their promising performance, PTT-based methods require subject-specific calibration, whereas AI-based models are often validated in an overoptimistic manner due to data leakage. In this thesis work, the estimation of continuous BP through PTT using regression techniques is investigated, evaluating models from literature and training them with a limited number of individual BP values per subject. Subsequently, various machine learning (ML) models are assessed using a large set of features extracted from ECG and PPG, with a subject-wise validation strategy to ensure generalizability and avoid data leakage. This work involves the analysis and processing of PPG, ECG and arterial blood pressure waveform extracted from the MIMIC II online database. These data were used as input for three relevant PTT-based methods, characterized by 2, 3, and 4 coefficients, and the models were calibrated using a limited number of BP values. As a further step, a set of features was extracted from the ECG and PPG to enable the application of different ML techniques. The dataset was partitioned by subject into 40% for training, 30% for validation, and 30% for testing to ensure subject-independent evaluation. Among the various model and training data combinations evaluated, 2 coefficient model trained with 5 BP values achieved the best performance, with a mean absolute error (MAE) of 5.309 ±4.561 mmHg. This configuration balances model simplicity with minimal training. Furthermore, the distribution of MAE reveals that approximately 60% of subjects attain a MAE of 5 mmHg or less and about 30% fall between 5 and 10 mmHg. Gaussian Process Regression emerged as the best-performing method during validation and achieved a mean absolute error of 8.613 ± 8.315 mmHg on the test set. Moreover, MAE distribution reveals that approximately 50% of individuals attain a MAE of 5 mmHg or less and around 30% fall between 5 and 10 mmHg. In conclusion, regression model offers high predictive accuracy but its dependence on subject-specific calibration represents a significant limitation. In contrast, ML approach, although slightly less accurate, offers the advantage of being calibration-free and more generalizable across individuals. These promising results suggest that the algorithm could be further optimized and mark an initial step toward its future integration into an automatic wearable device.