Abstract
This study introduces PRISM-CRC, a novel deep learning framework designed to improve the diagnosis and prognosis of colorectal cancer (CRC) by integrating histopathology, radiology, endoscopy and clinical data. The model demonstrated high accuracy, achieving a concordance index of 0.82 for predicting 5-year disease-free survival and an AUC of 0.91 for identifying microsatellite instability (MSI) status. A key finding is the synergistic power of this multimodal approach, which significantly outperformed models using only a single data type. The PRISM-CRC risk score proved to be a strong, independent predictor of survival, offering more granular risk stratification than the traditional TNM staging system. This capability has direct clinical implications for personalizing treatment, such as identifying high-risk Stage II patients who might benefit from adjuvant chemotherapy. The study acknowledges limitations, including a modest performance decrease due to "domain shift" and classification errors in morphologically ambiguous cases, highlighting the need for future prospective trials to validate its clinical utility.