Enhancing SpO2 Estimation through Deep Learning Approaches in Photoplethysmography

In an increasingly health-conscious world, the need for accurate and non-invasive monitoring of vital signs has led to remarkable advancements in medical technology. Among these innovations, photoplethysmography (PPG) stands out as a transformative method for assessing physiological parameters, particularly peripheral oxygen saturation (SpO2), which is vital for diagnosing and managing various health conditions such as respiratory and cardiovascular diseases. This thesis delves into the dynamic interplay between technology and biology, starting with a comprehensive investigation of the cardiovascular system and the underlying principles of PPG. It provides a foundation for understanding how light absorption may vary with blood flow and oxygenation and which could be the critical aspects of pulse oximetry, including wavelength selection and sensing modes. A key focus of this work is the potential of neural networks to improve the accuracy of SpO2 estimation. By addressing challenges such as individual physiological variability, these advanced techniques offer promising solutions for more reliable and precise monitoring systems. The research presented here includes a detailed neural network implementation based on the OpenOximetry Repository database, outlining the steps involved in data preprocessing, including noise filtering and the handling of outliers. Through careful parameter tuning and model validation, this thesis aims to establish a robust and clinically applicable method for non-invasive oxygen saturation monitoring. Furthermore, in an effort to extend the algorithm’s applicability to multiple wavelengths — a scenario only recently addressed in the literature — a dedicated data acquisition campaign was initiated. This campaign focuses on capturing PPG signals across various wavelengths, including red, infrared, and green. The innovative approach of this thesis lies in leveraging a broader spectrum of wavelengths beyond the traditional red and infrared methods. By utilizing additional wavelengths, such as green, the aim is to extract a richer set of information from the PPG signals. This wealth of data will allow for cross-referencing different wavelength responses, enabling the neural network to learn more effectively and improve the accuracy of SpO2 estimations. The expansion of this dataset is ongoing, with the intention of making it publicly available to encourage further research and development in the field.