OBJECTIVES: Decision-making in chronic myeloid leukemia (CML) is based on risk scores that incorporate information collected at diagnosis. However, there is evidence that the rate of change in tumor biomarker levels after treatment initiation and adherence to treatment are two key factors that predict the risk of relapse among patients in remission. We developed a risk prediction model incorporating these two longitudinal features that may be key prognostic factors.

METHODS: We used EHR data from an integrated health system on 443 patients who were diagnosed with CML between 2007-2019 and monitored for relapse after remission. We fit a Cox model with time to relapse as the outcome. The predictors included log biomarker change between the first two collections (~3 months apart) after treatment initiation, medication adherence in the first year of treatment, race, sex, BMI, age, and smoking behavior. Medication adherence was measured approximately as the proportion of days in which a patient had access to medication in the first year, and dichotomized into high (>80%) versus low (<=80%) adherence. We evaluated the accuracy of our model using the area under the ROC curve (AUC).

RESULTS: Compared to patients who experienced no biomarker change, the risk of relapse was significantly higher in patients with increased biomarker levels and significantly lower in patients with decreased biomarker levels (HR=4.91, 95% CI (1.74,13.85); HR=0.45, 95% CI (0.22,0.91), respectively). Furthermore, the risk of relapse for patients with high medication adherence was 62% lower than patients with low adherence (HR: 0.38, 95% CI (0.20,0.72)). Our model was highly accurate at predicting relapse within 6 months post-remission, with an AUC of 0.80 (95% CI: 0.73-0.99).

CONCLUSIONS: We incorporated key longitudinal features to develop a high-performing risk prediction model in CML. Upon further external validation, this model may be used in clinical practice for identifying high-risk patients for targeted intervention.