Combination of Explainable AI with Uncertainty Quantification for Arrhythmia Detection in PPG-Based Diagnostics

Photoplethysmography (PPG) is a non-invasive optical technique enabling the measurement of various physiological parameters. With its versatility, ease of acquisition, and cost-effectiveness, PPG has become an increasingly promising tool for modern diagnostic systems. Different pathologies can be unveiled by PPG signals and arrhythmia is one of these. Given the asymptomatic nature and wide variability of arrhythmias, continuous monitoring paired with highly accurate detection of multiple arrhythmia types may offer a winning strategy for dealing with it. Artificial intelligence (AI) algorithms support the demand for high performance, even in complex, multi-label tasks. Deep Learning (DL), a powerful AI approach, has shown potential to achieve human-comparable performance. However, the adoption of AI tools in clinical settings remains challenging due to their "black box" nature, which limits the transparency of the generated outputs. Explainable Artificial Intelligence (XAI) is an emerging field that aims to unveil the inner workings of these algorithms, enhancing interpretability. Another critical challenge is the reliability of predictions. To address this, Uncertainty Quantification (UQ) techniques have been developed to assess the confidence levels associated with AI predictions. In this work, a combination of Monte Carlo Dropout (MCD) and Gradient-weighted Class Activation Mapping (GradCAM) techniques is employed to enhance the performance of a VGG-16 network trained on multi-label PPG segments sourced from an open-access GitHub repository. The initial phase of the study focused on calibrating the dropout probability, a key parameter in the MCD method. Based on the entropy distributions of Correctly Classified (CC) and Misclassified (MC) samples across different dropout probability values, the selected criterion was to determine the dropout probability that most effectively distinguishes the CC and MC distributions. Hence, two information are extracted: first, the entropy of the predicted baseline class, averaged over the MCD predictions, quantifies prediction uncertainty, that reflects the uncertainty related to the output. Second, GradCAM assigns relevancy to the features right before the prediction, unveiling what the network is focused on before its decision. To bridge these two aspects, the heatmap information was converted into a measure of uncertainty, by calculating the Spearman correlation between the baseline heatmap and each of the MCD dropout heatmap. By combining these two uncertainty measures and calibrating confidence thresholds on the validation set, the testing phase aims to identify samples with low confidence that can be excluded and high-confidence samples suitable for reliable predictions. The investigation aimed to explore a potential relationship between these two levels of uncertainty and to assess whether their combined use could enhance the algorithm’s performance by identifying the type of uncertainty most impacting predictions. The results indicate that applying a single entropy threshold to define two levels of uncertainty can improve model performance, though it results in a significant reduction of retained samples. In contrast, the application of both an entropy threshold and a correlation threshold allows for defining four levels of uncertainty, yielding a more conservative approach to sample retention while providing a more modest performance increase.