PT Unknown
AU Wuerich, C
TI Non-invasive continuous blood pressure measurement using machine learning
PD 06
PY 2023
DI 10.17185/duepublico/81920
LA en
DE Medical Engineering; Machine Learning; Neural Networks; Non-invasive Monitoring; Vital Signs; Blood Pressure; Remote Photoplethysmography; Feature Engineering; Remote Sensing; Contactless Measurement
AB Blood pressure is an important clinical parameter that is used to assess cardiovascular health. Deviations from the normal range can lead to severe organ damage and increased mortality, while sudden changes can indicate cardiovascular events. Therefore, a regular ambulatory blood pressure measurement as well as a reliable monitoring in clinical settings are essential for an accurate diagnosis and treatment of such cardiovascular diseases and events. However, the commonly used cuff-based measurement devices exhibit several limitations regarding accuracy, temporal resolution and measurement comfort. Therefore, this thesis investigates new methods for camera-based beat-to-beat blood pressure monitoring. Such methods rely on the extraction of remote Photoplethysmogram (rPPG) signals from skin pixels of the video. To overcome illumination and movement artifacts as well as an insufficient Signal-to-Noise Ratio (SNR) for darker skin tones, various colour model transformations and rPPG extraction algorithms are evaluated with respect to the application for remote blood pressure measurement. The experiments show that Plane Orthogonal to Skin (POS) performs best under difficult measurement conditions and that considering skin pixels of the palm as opposed to the face significantly improves SNR for darker skin. Moreover, a method based on hand-crafted rPPG features and a Random Forest Regression (RFR) model is proposed. To obtain a light-weight model and increase prediction accuracy, a Sequential Forward Selection (SFS) is performed. The prediction accuracy could be improved to an Mean Absolute Error (MAE) +/-Standard Deviation (SD) of 11.91 +/-9.66 mmHg for Systolic Blood Pressure (SBP) and 7.92 +/-6.02 mmHg for Diastolic Blood Pressure (DBP) on a wide range of blood pressure values outperforming comparable studies. An analysis of the feature selection results is provided to enhance model interpretability for medical applications and aid future developments. Next, based on the ResNet-50 architecture, a more complex Convolutional Neural Network (CNN) is developed to automatically extract features from the raw rPPG signals. Due to the small size of the rPPG data set, the model is developed and pre-trained on the MIMIC III waveform data base. This Photoplethysmogram (PPG)-based beat-to-beat prediction method reaches an MAE +/-SD of 8.73 +/-7.36 mmHg for SBP and 8.07 +/-6.86 mmHg for DBP which is comparable to related studies that rely on longer PPG sections or an additional ECG signal. Further analyses underline the potential of model personalisation and the importance of a balanced fine-tuning data set, since the results of personalisation strongly depend on the selected tuning data and is prone to overfitting when using sequential tuning samples. Therefore, different strategies are derived for balancing the tuning data set in real-world applications. Finally, transfer-learning from the PPG domain to the rPPG domain is assessed and shows encouraging results on the rPPG data set with prediction errors close to the feature-based method.
ER