This project developed a non-invasive blood pressure monitoring system using PPG signals and machine learning techniques. The initial concept involved a wrist-based device but shifted to a finger-based sensor for improved accuracy. The PPG signals were processed through a deep learning pipeline to estimate systolic and diastolic blood pressure.
Download Project Report (PDF)Hardware Development: A custom-built PCB was designed for the finger-based PPG sensor, featuring green LEDs and a photodetector for signal acquisition. The system transmitted data via WiFi to a local server for analysis.
Signal Processing: PPG signals were preprocessed to reduce noise, and relevant features were extracted for analysis.
Machine Learning: A hybrid CNN-LSTM model was used to classify and estimate blood pressure. The system achieved an average accuracy of 88% for signal classification.
Results
Diastolic blood pressure estimation achieved a mean absolute error (MAE) of ±7.94 mmHg, showing promising results. Systolic blood pressure estimation faced challenges, with an MAE of ±14.78 mmHg, indicating areas for further refinement.
PCB Design and Hardware Development
The heart of the system relied on a custom-designed PCB, built with a focus on optimising signal acquisition for high-quality PPG data. The Analog Front End (AFE), the ADPD188GG, was chosen for its compactness and low power consumption, integrating photodetectors and LEDs within a 3.8mm x 5mm package. The PCB also included voltage regulators and logic-level converters to accommodate the system's varied voltage requirements (1.8V, 3.3V, and 5V) across different components. Grounding strategies, noise mitigation techniques, and precise LED current management were implemented for optimal signal quality during operation.
Challenges and Future Improvements
This project was undertaken as an individual effort during my third year with a strict budget of £150 and limited time. While similar research projects typically involve large teams and several years of research with access to hospital datasets or clinical trials, I independently designed, developed, and tested the system. The custom hardware met the project’s technical requirements despite these constraints, achieving promising diastolic blood pressure estimations. However, the systolic estimation posed a greater challenge due to noise and limited data availability. Future iterations of this project could benefit from more sophisticated signal processing techniques and access to larger datasets, ideally through collaborations with medical professionals.Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
PPG Signal Quality
The PPG signal data was categorised based on signal quality, with classifications ranging from excellent to unsuitable. The waveform plots represent the variations in amplitude and signal clarity, critical in determining the accuracy of blood pressure estimations. The system’s performance relies heavily on acquiring clean signals, particularly in dynamic environments where noise and motion artifacts can interfere with data capture.
The PPG sensor’s hardware is housed in a custom-designed 3D-printed casing, which was engineered for ease of use and portability. The modular design allows for easy access to the internal components while maintaining a compact and user-friendly form factor. This casing was designed to accommodate the PCB and support stable signal acquisition during use.
The printed circuit board (PCB) was designed to accommodate the Analog Front End (AFE) and support components. The design minimises signal noise and interference while optimising power efficiency. The PCB layout includes multiple signal paths and ground planes to maintain signal integrity, an important factor in achieving reliable PPG data for blood pressure estimation.
