Improving PPG Signal Classification with Machine Learning: The Power of a Second Opinion

Photoplethysmography (PPG) is a non-invasive technique that uses light to measure blood volume changes in tissues. It is widely used in clinical settings for monitoring vital signs such as heart rate, blood oxygen saturation, and blood pressure. To extract relevant information from PPG signals, such as peak detection and signal quality assessment, careful processing is required. Although traditional machine learning-based methods are capable of extracting useful information from PPG signals, they do not provide a measure of their confidence. In contrast, probabilistic machine learning approaches, such as Bayesian networks, can quantify uncertainty and provide estimates of prediction confidence. This would be beneficial for clinical decision-making and lead to more transparent and interpretable results. This paper proposes a new confidence-aware framework for Cuff-Less prediction of blood pressure using Monte Carlo Dropout (MCD) and Bayesian optimization techniques. Our results demonstrate that our approach outperforms simple MCD, providing more reliable predictions.