Deep Learning-Based Quantification of Spontaneous Pain Behaviours in Mice

Accurate evaluation of pain-related behaviours in rodent models is crucial for elucidating pain mechanisms and assessing the efficacy of novel analgesic compounds. Traditional nociceptive assays primarily measure evoked responses, which have limited translational relevance since spontaneous pain is the predominant symptom in pain patients. Rodent models of spontaneous pain reveal innate behaviours, including paw licking and flinching, but their manual quantification is often subject to observer bias, is error-prone and labor-intensive.

We have developed a deep-learning-based algorithm designed to automatically detect and quantify spontaneous pain behaviours in mice from standard 2D bottom-up video recordings. Focusing on paw licking and flinching in the capsaicin pain model, and validated in the formalin test and a model of cancer-evoked pain, our approach provides an objective, high-throughput and reproducible alternative to traditional manual scoring.

The pipeline, combining DeepLabCut-based keypoint tracking with a transformer classifier, achieved 92% balanced accuracy in identifying licking, flinching and baseline behaviours in the capsaicin model. Automated quantification showed an excellent correlation with expert human annotations (Pearson's r = 0.99), while frame-by-frame comparisons with blinded observers showed high agreement and minimal bias. Performance was maintained on an independently generated external dataset (Pearson's r = 0.98), supporting reproducibility across experimental settings. Validation in inflammatory and cancer pain models further demonstrated robust generalizability across distinct pathological conditions.

By reducing observer-dependent variability and capturing subtle temporal and spatial dynamics, this framework provides a standardized and accessible tool for the objective assessment of ongoing nociceptive responses, facilitating comparisons across distinct experimental models and institutional settings.

Our findings highlight the potential of deep learning to advance preclinical pain research by enabling standardized, unbiased and clinically relevant assessment of spontaneous pain in rodent models.