A Machine Learning Model to Successfully Predict Future Diagnosis of Chronic Myelogenous Leukemia With Retrospective Electronic Health Records Data

Abstract

Background Chronic myelogenous leukemia (CML) is a clonal stem cell disorder accounting for 15% of adult leukemias. We aimed to determine if machine learning models could predict CML using blood cell counts prior to diagnosis. Methods We identified patients with a diagnostic test for CML (BCR-ABL1) and at least 6 consecutive prior years of differential blood cell counts between 1999 and 2020 in the largest integrated health care system in the United States. Blood cell counts from different time periods prior to CML diagnostic testing were used to train, validate, and test machine learning models. Results The sample included 1,623 patients with BCR-ABL1 positivity rate 6.2%. The predictive ability of machine learning models improved when trained with blood cell counts closer to time of diagnosis: 2 to 5 years area under the curve (AUC), 0.59 to 0.67, 0.5 to 1 years AUC, 0.75 to 0.80, at diagnosis AUC, 0.87 to 0.92. Conclusions Blood cell counts collected up to 5 years prior to diagnostic workup of CML successfully predicted the BCR-ABL1 test result. These findings suggest a machine learning model trained with blood cell counts could lead to diagnosis of CML earlier in the disease course compared to usual medical care.

Resistance to Tyrosine Kinase Inhibitors in Chronic Myeloid Leukemia—From Molecular Mechanisms to Clinical Relevance

“…In particular, the ML models could predict CML using blood cell counts prior to diagnosis. These findings indicate that a ML model trained with blood cell counts can lead to diagnosis of CML earlier in the disease course as compared to usual medical care [234]. Other authors have recently developed a leukemia artificial intelligence program (LEAP) using the Extreme Gradient Boosting (XGBoost) decision tree method for the optimal treatment recommendation of tyrosine kinase inhibitors (TKIs) in patients with CML-CP.…”

“…A recent study that also used ML models was able to predict future diagnosis of CML based on the analysis of data from retrospective electronic health records [234]. In particular, the ML models could predict CML using blood cell counts prior to diagnosis.…”

A Review of Artificial Intelligence Applications in Hematology Management: Current Practices and Future Prospects

“…In the model selection process, 103 clinical variables were used to train 4 classification algorithms, random forest (RF), decision tree, support vector machine (SVM), and linear regression, to distinguish relapses from nonrelapses in the 3 clinically predefined risk categories: standard-, intermediate-, and high-risk levels. While Pan et al [ 131 ] built a model to predict disease relapse, Hauser et al [ 144 ] studied the possibility of predicting CML prior to diagnosis using only CBC test results and ML algorithms such as XGBoost and LASSO algorithms on 1623 patients with a definitive CML status. The variables used in the study included laboratory CBC test results, patient demographic features such as their age and gender, and patient encounter information (the number of patient visits to outpatient clinics, etc.).…”

“…However, the suggested approach is considered insufficient for internal validation that requires at least 50 repeats [ 145 ]. By contrast, the chosen data set in [ 144 ] was divided into 2 distinct groups: train/validation and test groups. While the latter split-sample validation approach seems reasonable and justifiable to use in this case given the large sample size, potential drawbacks may arise, and several aspects still need attention throughout the application.…”

“…After excluding 74 patients who died and 373 who initiated their treatment before the prediction point, the study cohort’s final size corresponded to 3729 patients. In contrast to Hauser et al [ 144 ], Agius et al [ 55 ] employed stratified sampling to maintain class distributions and compensate for the 52% International Prognostic Index for CLL (CLL-IPI) missing variables, by dividing the data set into 65% training set and 17.5% for each of the test and internal validation sets. Using 7288 features resulting from a collection of variables from different sources, comprising baseline variables at the time of diagnosis including age, gender, etc.…”