Fig. 1: Conceptual framework and validation of AI-bRNN.

a Schematic diagram of the experimental approach. AAV1-hsyn-GCaMP6s was injected into S1. b Imaging was performed for 2 min at each time point (before and 1–3 min after formalin injection). The time points for imaging were selected based on the levels of nociceptive behavior after formalin injection in freely moving animals. c A representative image of S1 neurons identified by semiautomated ROI analysis (top). Example Ca²⁺ traces from each ROI (bottom). The scale bar represents 50 μm. d Heatmaps showing the activity of S1 neurons. The line traces below each heatmap indicate the averaged values of all ROIs. The periods of mouse locomotion identified by the motion tracking analysis are overlaid on the line traces using sky-blue shading. e Architecture of AI-bRNN. The Ca²⁺ traces extracted from each ROI were averaged subject by subject to train the neural network. In the test session, the Ca²⁺ traces from individual ROIs were separately applied to the deep learning model for testing. f The predictions of AI-bRNN regarding whether the subject was experiencing pain. On the x-axis, ‘B’ indicates the time before injection. Saline (s.c.) group (n = 14 mice); formalin 5% (s.c.) group (n = 13 mice); formalin 1% (s.c.) group (n = 8 mice); formalin 5% (s.c.) + ketoprofen (100 mg/kg, i.p.) group (n = 7 mice); formalin 5% (s.c.) + 2% lidocaine (10 μl, s.c.) group (n = 3 mice). g The classification performance for formalin pain conditions based on the S1 neuronal signals. Scatter plots indicate individual data. Bars indicate the mean ± SEM; N.S., nonsignificant; ***P < 0.001, *P < 0.05 compared to the pre-injection period (Wilcoxon test).