Predicting grade≥2 chemotherapy-induced myelosuppression in older patients with colorectal cancer receiving first-line chemotherapy: a multi-algorithm consensus approach using 9 routine parameters to develop and validate an explainable boosting machine

Older patients with colorectal cancer undergoing first-line chemotherapy are prone to myelosuppression, which can lead to treatment delays, dose adjustments, and adverse outcomes such as infection and bleeding; however, individualized risk prediction tools with external validation remain lacking. This study aimed to identify routine clinical predictors using a multi-algorithm consensus strategy, construct and compare risk prediction models for grade ≥2 chemotherapy-induced myelosuppression in this population, and evaluate the external generalizability and clinical utility of the leading model.

In this retrospective cohort study leveraging data from two hospitals, we adopted a predictive model development and independent external validation design. A total of 553 older patients receiving first-line chemotherapy were enrolled, including 442 from Guang’anmen Hospital (development cohort) and 111 from Gansu Provincial Cancer Hospital (external validation cohort). The primary endpoint was grade ≥2 myelosuppression per Common Terminology Criteria for Adverse Events version 5.0 (CTCAE v5.0). Nineteen candidate predictors underwent consensus selection via least absolute shrinkage and selection operator logistic regression, the Boruta algorithm, and random forest permutation importance, with final core predictors determined by bootstrap inclusion frequency and prespecified clinical retention rules. Nine models were constructed: conventional logistic regression, LASSO logistic regression, elastic net logistic regression, random forest, extreme gradient boosting, categorical boosting, radial basis function support vector machine, linear discriminant analysis, and explainable boosting machine (EBM). Five repeated 10-fold stratified cross-validation generated out-of-fold predictions in the development cohort, and locked models were applied to the external validation cohort. Model performance was assessed by area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), Brier score, calibration intercept, calibration slope, Hosmer-Lemeshow test, and decision curve analysis.

The incidence of grade ≥2 myelosuppression was 26.2% (116/442) in the development cohort and 27.0% (30/111) in the external validation cohort (P>0.05). Nine core predictors were selected: albumin, carcinoembryonic antigen, carbohydrate antigen 19-9, carbohydrate antigen 125, body surface area, tumor location, chemotherapy cycle, chemotherapy regimen, and age-adjusted Charlson comorbidity index. The EBM was identified as the leading model based on balanced discrimination, calibration, and interpretability. In the development cohort, EBM achieved an out-of-fold AUROC of 0.720 (95%CI: 0.676 to 0.772), AUPRC of 0.421, Brier score of 0.176, calibration intercept of 0.035, and calibration slope of 0.778, with no significant miscalibration (Hosmer-Lemeshow P=0.087). In external validation, EBM yielded an AUROC of 0.714 (95%CI: 0.595 to 0.834), AUPRC of 0.641, Brier score of 0.164, calibration intercept of 0.155, and calibration slope of 0.778, though calibration remained suboptimal (P<0.001). At the optimal threshold of 0.28 derived from the development cohort, sensitivity, specificity, positive predictive value, negative predictive value, and F1 score in the external validation cohort were 0.667, 0.741, 0.488, 0.857, and 0.563, respectively. Decision curve analysis indicated superior net benefit over treat-all or treat-none strategies across threshold probabilities of 0.06 to 0.48 (development cohort) and 0.18 to 0.60 (external validation cohort). Interpretation revealed albumin as the strongest contributor, followed by carcinoembryonic antigen, carbohydrate antigen 19-9, and carbohydrate antigen 125; lower albumin, smaller body surface area, ≥5 chemotherapy cycles, and the FOLFIRI regimen were associated with increased risk, while tumor markers exhibited nonlinear associations with predicted risk.

Nine routine clinical parameters selected by a multi-algorithm consensus strategy enable construction of a prediction model for grade ≥2 myelosuppression during first-line chemotherapy in older patients with colorectal cancer. The finalized EBM model demonstrates moderate discrimination and clinical net benefit in both development and external validation cohorts but exhibits residual calibration drift externally. It may serve as an adjunctive tool for clinical risk stratification rather than a standalone decision-making instrument at this stage.