Interpreting Neural Network Predictions for Cancer Diagnostics via SHAP Values

Early detection remains a critical challenge to improve cancer survival rates. This study presents a machine learning framework for multi-class cancer classification using a panel of blood-based protein biomarkers. Using the NIHMS982921 dataset, which comprises 1,820 patient samples in eight distinct tumor types and a normal control group, we developed a binary classifier model based on 39 protein features. Our results demonstrate the ability of the model to identify specific proteomic signatures for different cancers. SHAP analysis validated the clinical relevance of the model by highlighting biomarkers as the top-ranking features in the classification process. This research underscores the potential of combining Explainable AI with proteomic profiling to enhance the transparency and reliability of non-invasive cancer screening tools.