Predictive Models for Discontinuation of Tyrosine Kinase Inhibitors in Chronic Myeloid Leukemia

Background: Tyrosine Kinase Inhibitors (TKI) have revolutionized the treatment of chronic myeloid leukemia (CML) and significantly improved survival rates for most patients, allowing a near-normal life expectancy. Treatment-free remission (TFR) has become a standardized management goal for CML patients, and TKI discontinuation could benefit patients experiencing adverse effects or financial strains. A lot of efforts have been put into seeking precise stop conditions. The existing recommendations suggest that patients treated with TKIs for at least three years and who have sustained a deep molecular response (DMR) for more than two years may consider discontinuing TKI therapy. However, around half of these patients lose their major molecular response (MMR) within the first six months after cessation of TKI treatment. Patient-specific factors might have complicated interactions with the outcomes of TKI discontinuation, and this article focuses on developing predictive models with these clinical factors to select patients with a higher probability of successful discontinuation.

Methods: A multi-center retrospective cohort comprising 540 CML patients across 10 Chinese centers between 1999 and 2020 was established. We extract demographic data and detailed treatment histories from clinical and laboratory records. The inclusion criterion was patients who had been treated successfully with any TKI for at least three years and maintained DMR for at least one year. This yielded a cohort of 493 patients who met the eligibility requirements for further analysis. The cohort is split into two sub-cohorts to develop and externally validate predictive models according to the centers. The endpoint was maintaining MMR within 12 months after TKI discontinuation. Using the clinical factors associated with the endpoint, we developed machine learning models predicting the success of TKI discontinuation.

Results: The median diagnosis age of the patients is 36.5 [interquartile range (IQR), 27.1-49.2] years, with 51.1% male, among whom 256 (51.9%) successfully maintained MMR 12 months after discontinuation. The median age at discontinuing TKI therapy is 45.3 [IQR, 34.0-56.3] years. The median duration for TKI therapy is 6.5 [IQR, 5.1-8.4] years. The median duration of DMR before stopping TKI is 5.0 [IQR, 3.4-6.7] years. The median duration of MMR before stopping TKI is 5.5 [IQR, 4.3-7.2] years. The EUTOS Long-term Survival Score distribution is as follows: 65.9% low risk, 14.5% intermediate risk, 3.9% high risk, and 15.7% data unavailable. The Sokal score distribution is as follows: 48.4% low risk, 22.6% intermediate risk, 11.0% high risk, and 18.0% data unavailable. The training dataset comprises 321 patients from six centers, and is randomly divided into five folds to mitigate overfitting, within each 20% of the data used as an internal validation set to select the best models. The final model is obtained by an ensemble of the models from these five-folds, and the performance of the model is evaluated on the external validation cohort containing 172 patients from the remaining four centers. The feature permutation analyses indicated that gene fluctuations, duration of maintaining DMR and reaching MMR, platelet and white blood cell count, and Sokal score significantly impact predictive accuracy.

Conclusion: Our ensembled machine learning model predicts the success of TKI discontinuation with high accuracy, a 0.812 (0.801-0.821) area under the receiver-operator characteristic curve (AUROC) on the external validation cohort. When probability 0.4 is used as the cutoff, our model achieves a sensitivity of 74.7%, specificity of 75.3%, and overall accuracy of 75.0% on the validation cohort. This accuracy is almost 10% higher than the previously suggested predictor for the safety of TKI discontinuation, i.e., DMR exceeded 3.1 years, with a sensitivity of 92.8%, specificity of 33.6%, and accuracy of only 64.4%. Adjusting DMR maintaining duration to account for the statistical difference, extending it to 5.1 years, can modestly increase prediction accuracy, reaching 66.3% on the validation cohort. However, this performance is still significantly lower than our ensembled model. The excellent performance of the proposed model shows the great potential to improve patient care. Future studies will focus on validating our models on larger multi-center cohorts and incorporating more factors.