EfficientNet-CICA: a sports image classification model based on causal intervention channel attention

Accurate recognition of sport and exercise contexts from static visual data can provide an automated annotation component for multimodal sport and exercise analysis, but it should be distinguished from direct physiological measurement. Such recognition is challenging because sports images often contain large variations in posture, equipment, viewpoint, occlusion, and environmental background, and visually similar activities may share overlapping movement cues. In this study, we developed an EfficientNet-based sports-context recognition framework with a Causal Intervention Channel Attention (CICA) mechanism. The proposed module uses class prototype information and counterfactual channel intervention to recalibrate feature responses toward category-discriminative regions, including athletes, equipment, body-action patterns, and representative exercise environments. Experiments were performed on a public 22-class sports image dataset containing 14,192 valid images. Across five random seeds, EfficientNet-CICA achieved an accuracy of 0.9456 ± 0.0060, a weighted F1-score of 0.9455 ± 0.0059, and a macro-average AUC of 0.9986 ± 0.0003. Paired statistical analysis showed consistent improvements over the EfficientNet baseline and conventional attention variants, whereas the margin over MedMamba was smaller and was therefore interpreted cautiously. Ablation experiments confirmed the contribution of prototype-guided channel intervention, while confusion matrices, ROC and precision-recall curves, t-SNE visualization, and Grad-CAM analysis showed improved feature separability and more interpretable attention to sport-relevant visual evidence. To address the physiological motivation more explicitly, we added two auxiliary analyses: a toy sport-aware rulerouting demonstration and a PAMAP2 heart-rate regression experiment, in which adding activity type to IMU features reduced MAE from 4.39 to 3.91 bpm. These analyses suggest that activity context can influence downstream physiological interpretation, while the proposed image model itself remains a static-image visual classifier. Our code are available in: https://github.com/xingshdlsjz-dev/CICA.