Tongue image analysis for accurate prediction of nutritional risk screening in cancer patients during radiochemotherapy: a feature selection network and aliasing attention mechanism approach

Background: The Nutritional Risk Screening 2002 (NRS2002) is a widely adopted tool for assessing nutritional risk in patients. This study introduces a novel, non-invasive, efficient, and accurate screening approach to complement the traditional NRS2002 assessment, addressing its inherent limitations. Methods: A dataset comprising 672 tongue images from 470 tumor patients was collected. A new predictive analysis system for NRS2002 was developed by integrating two model branches. Machine learning was employed for risk prediction, and a ResNet50 neural network was utilized to extract high-dimensional features from tongue images. This architecture was enhanced with Shuttle Attention mechanism. The final predictions were derived through the fusion of both model branches. Results: The fusion of the two branches significantly improved the model's ability to capture complex features. For both at-risk and risk-free cohorts, the system demonstrated optimal classification performance across three key metrics: AUC = 0.919, ACC = 0.927, and Recall = 0.888. In ablation studies, the SelectNet module improved ACC and AUC by 17.25 and 16.14%, respectively. Furthermore, the integration of the shuttle attention mechanism led to additional gains, with AUC and ACC increasing by 3.89 and 1.94%, respectively. Conclusion: We successfully developed and validated an NRS2002 nutritional risk prediction model based on tongue image characteristics. This tool has the potential to minimize human error, improve dynamic performance, and provide non-invasive, accurate nutritional risk screening. It represents a step forward in personalized medicine and holds substantial clinical value.