ARTIFICIAL intelligence can be used as a strategy to predict the overall survival outcome of patients with colorectal cancer using routine blood indices according to a new Chinese study.
AI and Colorectal Cancer
Patients with advanced stage colorectal cancer have less favourable survival rates despite strong progress in screening and treatment. The aim of this study was to develop and validate explainable machine learning models using blood indices to predict prognosis in patients with colorectal cancer.
Researchers pooled a training cohort of 850 patients with colorectal cancer, along with two external validation cohorts. Several machine learning approaches were tested, and the random survival forest (RSF) model demonstrated the strongest overall predictive performance.
Predicting Colorectal Cancer prognosis
The RSF model showed good discriminatory ability in the primary test cohort, with AUC values of 0.768, 0.775, and 0.731 for predicting one-, two, and three-year overall survival, respectively. The AUC (area under the curve) reflects how well the model distinguishes between patients who survive and those who do not, with higher values indicating better predictive accuracy.
Performance remained consistent in external validation cohorts, including Hefei (AUCs: 0.820, 0.805, 0.775) and Shihezi (0.651, 0.706, 0.747), indicating reproducibility of model performance in external datasets.
Explainability analysis highlighted CEA, CA125, age, MPV, CA19-9, INR, and monocyte count as the most influential predictors in the RSF model.
Conclusion
Researchers called for more prospective clinical trials to validate these models given the study’s limitations.
The authors noted that bias might be inevitable given that the data was retrospectively collected via an electronic medical system. They also highlighted that only baseline clinical data at admission were used, meaning changes in patient status over time were not captured and could affect predictive performance. In addition, differences in laboratory instruments, assays, and reference ranges across centres over the 10-year study period may have introduced inter-site variability.
Overall, the findings suggest that routinely available blood indices, combined with explainable machine learning models, may help stratify overall survival risk in patients with colorectal cancer and support more personalised clinical decision-making.
Reference
Tian S et al. Artificial intelligence for the prediction of prognosis in colorectal cancer patients using routine blood indices. npj Digit. Med. 2026; DOI:10.1038/s41746-026-02781-5
Featured Image: volha_r on Adobe Stock
- Author:
