Background Comorbidities increasingly complicate targeted therapy for chronic-phase chronic myeloid leukemia (CP-CML). This study evaluated three comorbidity scoring systems (CCI, ACE-27, and CIRS-G) to predict molecular responses in flumatinib-treated CP-CML patients. Methods Retrospective data from 559 patients (2018–2024) were analyzed. Machine learning algorithms, including XGBoost, were trained to predict 12-month major molecular response (MMR). SHapley Additive exPlanations (SHAP) analysis was employed to interpret non-linear associations and feature interactions. Results The XGBoost model demonstrated the highest predictive performance (AUC = 0.852). Integrating CIRS-G into the baseline clinical model provided the most substantial incremental value (ΔAUC = 0.078, p = 0.006), outperforming CCI and ACE-27. SHAP analysis revealed a non-linear threshold effect, suggesting that a CIRS-G score ≥ 8 may severely compromise therapeutic efficacy. Furthermore, interaction plots indicated an exploratory model-based association where advanced age and severe comorbidity clustered with lower predicted probabilities, while reduced dose interventions coincided with altered model predictions in this specific subpopulation. Conclusion CIRS-G showed the strongest incremental predictive value among the evaluated comorbidity scores in this single-center cohort. Machine learning and SHAP analysis provide exploratory insights for individualized risk stratification, though the reported predictive performance may overestimate prospective validity due to temporal variations and retrospective design.
Yang et al. (Thu,) studied this question.