Heart Rate Prediction from Photoplethysmography Signals Using Machine Learning Approaches

Sammanfattning

Continuous cardiovascular monitoring is increasingly important for supporting preventive healthcare, fitness tracking, and early detection of physiological abnor- malities. One promising non-invasive approach for achieving this in everyday set- tings is the use of wearable sensors based on Photoplethysmography (PPG), of- ten combined with accelerometer sensors to support more robust monitoring during physical activity. However, while PPG signals can provide reliable measurements during resting conditions, physical activity introduces motion artifacts that make ac- curate heart rate estimation challenging. This thesis explores and compares different approaches for heart rate estima- tion using both clean resting-state signals and noisy signals recorded during physical activity. The study evaluates traditional machine learning methods based on hand- crafted features alongside deep learning approaches using Convolutional Neural Net- works (CNNs). Different preprocessing and noise-reduction techniques were also examined using PPG and accelerometer data. The evaluation was conducted using both research data collected during resting conditions and an open dataset containing recordings from physical activities with varying intensity levels. The results show that traditional feature-based methods perform well for clean and stable signals during rest (1.97 BPM MAE) but become less reliable in highly dynamic conditions with strong motion noise (22.02 BPM MAE), even when using advanced preprocessing techniques and sensor fusion based on PPG and accelerom- eter data. The CNN model demonstrated more stable performance during physical activity and handled noisy multi-channel sensor data more effectively, although the error rates remained relatively high due to the challenging nature of motion-corrupted signals. These findings suggest that additional datasets, more diverse experimental conditions, and further development of preprocessing and modeling techniques are needed to improve and validate heart rate estimation performance in real-world wear- able scenarios.