A Lightweight Edge Deployable Deep Learning Framework for Speech-Based Pain Classification across Heterogeneous Datasets

Automatic pain assessment from speech is an emerging approach for non-invasive and objective healthcare monitoring. However, many existing methods are evaluated on single datasets or under controlled conditions, which limits their robustness under diverse real-world conditions. This paper presents a lightweight, end-to-end deep learning framework for speech-based pain level classification that explicitly targets multi-dataset robustness assessment. The proposed approach uses log-Mel spectrograms with an EfficientNetV2B0 backbone to learn discriminative acoustic features without relying on handcrafted feature engineering or multi-stage pipelines. The model is evaluated on heterogeneous datasets, including clinical recordings, controlled experimental data, and a merged dataset representing diverse conditions. Experimental results show competitive performance, achieving up to 77.8% accuracy in four-class classification, 71.7% in three-class classification, and 87.7% in binary pain detection on the merged dataset. Compared with previous works, the approach achieves competitive performance, with stronger results in clinical and reduced-class settings. The model was also converted using TensorFlow Lite to enable efficient edge deployment with minimal performance degradation and reduced computational cost, supporting future deployment in mobile and edge-based healthcare applications.