Explainable pain level classification using a natural visibility graph-driven adaptive dilated recurrent unit with speech signals

Abstract

Pain detection is essential in healthcare because it provides clinicians with information for diagnosis, treatment, and patient monitoring. However, it has always depended on biased and variable self-reporting. Automatic pain assessment technologies aim to address these problems by developing objective, non-invasive, and scalable methods. In this paper, we propose a speech-based pain-level classification system that employs an explainable Natural Visibility Graph (NVG) and an Adaptive Dilated Recurrent Unit (ADRU) to learn temporal dependencies and structural relationships from speech signals. The experiment used a publicly available TAME-Pain dataset comprising over 7,000 speech samples annotated with pain intensity ratings from a cold pressor task. The model takes mel spectrogram features, converts them into NVGs, and then uses adjacency-aware neighbor features in ADRU to sort them. To make things clearer, we conducted an explainability analysis by following gradients and using occlusion tests on the raw waveform to find the time periods that had the biggest impact on the model's pain-level predictions. The method could distinguish between pain and no pain 84.51% of the time, between warm and cold 87.30% of the time, and between different levels of pain severity 84.43% of the time. The NVGguided ADRU consistently yielded more balanced evaluations compared to conventional machine learning techniques and a minimal CNN baseline. This suggests that it may be an effective and comprehensible approach for assessing pain through speech. The explainability analysis showed that the proposed model used speech features like prosodic bursts and transitions to make its predictions.