INTEGRATED MACHINE LEARNING AND SIR MODELING FOR PREDICTING RESPONSE IN CML
- 1Instituto Nacional de Câncer, Rio de Janeiro, RJ, Brazil
- 2Universidade Federal do Rio de Janeiro, Rio de Janeiro, RJ, Brazil
- 3Departamento de Hematologia, Hospital Universitário Clementino Fraga Filho, Universidade Federal do Rio de Janeiro, Rio de Janeiro, RJ, Brazil
Abstract
Chronic myeloid leukemia (CML) is a fitted cancer model for predictive machine learning (ML). Unlike most malignancies, CML is driven by a single molecular event, the BCR::ABL1 fusion gene. The introduction of tyrosine kinase inhibitors (TKIs), particularly imatinib, resulted in a well defined disease kinetics and transformed CML into one of the most successful cases of target therapy. Current goal is to achieve deep molecular response (DMR) sustained for over 2 years, allowing treatment-free remission (TFR). CML monitoring relies on standardized quantitative PCR (qPCR) reported on the International Scale over defined time sets. Although data driven ML models achieve accurate predictions, they may not fully capture the underlying disease dynamics encoded in longitudinal molecular responses. Herein, we integrate a mathematical model of CML dynamics with ML to improve prediction of DMR by combining mechanistic disease information with data-driven learning. This study aims to develop predictive models for DMR achievement and long-term maintenance in CML patients treated with first-line imatinib at the Brazilian National Cancer Institute, based on an adaptation of the SIR model integrated with ML as a clinical prediction tool. A retrospective cohort of 240 samples of 40 CML patients treated with first-line imatinib was analyzed using clinical, laboratory, and molecular data. BCR::ABL1 transcript levels were quantified by qPCR. Missing values were imputed using the median, and longitudinal molecular and cytogenetic measurements were summarized into statistical features. Multiple ML classifiers were trained to predict DMR using baseline and early treatment data. A mathematical model based on the SIR framework was integrated with the ML model by incorporating molecular response slopes and white blood cell counts. The hybrid model was compared with the data driven machine learning model to evaluate improvements in predictive performance. External validation was performed using a preliminary independent dataset comprising approximately 466 longitudinal molecular measurements. The integration of clinical data, ML, and mathematical modeling enabled the development of a hybrid predictive model for sustained DMR. Molecular variables were the strongest predictors, particularly BCR::ABL1 transcript levels at 3, 6, and 12 months, while molecular response slopes showed the highest relative importance, emphasizing the value of treatment kinetics. Incorporation of the SIR-based mathematical model into the logistic classifier preserved the discriminative performance of the previously optimized ML model based on 219 patients (AUC = 0.81 vs. 0.84) and substantially outperformed a ML model trained with the same variables alone (AUC = 0.81 vs. 0.62). The hybrid model also demonstrated superior calibration and overall accuracy, reflected by a 49.5% improvement in the Brier Skill Score, while providing interpretable estimates of each variable's contribution. Preliminary external validation using an independent dataset of longitudinal molecular measurements identified a limited number of false-positive and false-negative predictions, indicating the need for further calibration of model coefficients. Nevertheless, the model successfully identified patients with sustained molecular responses who are likely candidates for TFR, supporting the feasibility of integrating mechanistic mathematical modeling with ML for prognostic stratification in CML.
Related
- An integrated evolutionary optimization and machine learning framework for predicting pharmacokinetics of nanocarrier-encapsulated drugs
- Advanced machine learning strategies for predicting therapy response in preclinical glioblastoma using longitudinal MRI
- Predicting atropine treatment response heterogeneity in pediatric myopia using explainable machine learning: an on-treatment longitudinal monitoring framework
- Development and Validation of an Interpretable Machine‐Learning Model for Predicting Suboptimal Clinical Response After Sleeve Gastrectomy: A Multicenter Study
- Definitive Hypofractionated Radiotherapy in Elderly Patients With Non-melanoma Skin Cancer: Predicting Complete Response Using a Machine Learning-based Approach
BibTeX
@article{sousa2026integrated,
title = {{INTEGRATED MACHINE LEARNING AND SIR MODELING FOR PREDICTING RESPONSE IN CML}},
author = {Marianne Camile Silva SOUSA and Gustavo Trevizani STELZER and Simone BONECKER and Bruna Souza Sabioni GONÇALVES and Ilana ZALCBERG and Luciana Mayumi GUTIYAMA},
journal = {Hematology, Transfusion and Cell Therapy},
year = {2026},
month = oct,
doi = {10.1016/j.htct.2026.107513},
url = {https://doi.org/10.1016/j.htct.2026.107513},
note = {CC BY},
}
RIS
TY - JOUR
TI - INTEGRATED MACHINE LEARNING AND SIR MODELING FOR PREDICTING RESPONSE IN CML
AU - Marianne Camile Silva SOUSA
AU - Gustavo Trevizani STELZER
AU - Simone BONECKER
AU - Bruna Souza Sabioni GONÇALVES
AU - Ilana ZALCBERG
AU - Luciana Mayumi GUTIYAMA
PY - 2026
DA - 2026/10/23
JO - Hematology, Transfusion and Cell Therapy
DO - 10.1016/j.htct.2026.107513
UR - https://doi.org/10.1016/j.htct.2026.107513
AB - Chronic myeloid leukemia (CML) is a fitted cancer model for predictive machine learning (ML). Unlike most malignancies, CML is driven by a single molecular event, the BCR::ABL1 fusion gene. The introduction of tyrosine kinase inhibitors (TKIs), particularly imatinib, resulted in a well defined disease kinetics and transformed CML into one of the most successful cases of target therapy. Current goal is to achieve deep molecular response (DMR) sustained for over 2 years, allowing treatment-free remission (TFR). CML monitoring relies on standardized quantitative PCR (qPCR) reported on the International Scale over defined time sets. Although data driven ML models achieve accurate predictions, they may not fully capture the underlying disease dynamics encoded in longitudinal molecular responses. Herein, we integrate a mathematical model of CML dynamics with ML to improve prediction of DMR by combining mechanistic disease information with data-driven learning. This study aims to develop predictive models for DMR achievement and long-term maintenance in CML patients treated with first-line imatinib at the Brazilian National Cancer Institute, based on an adaptation of the SIR model integrated with ML as a clinical prediction tool. A retrospective cohort of 240 samples of 40 CML patients treated with first-line imatinib was analyzed using clinical, laboratory, and molecular data. BCR::ABL1 transcript levels were quantified by qPCR. Missing values were imputed using the median, and longitudinal molecular and cytogenetic measurements were summarized into statistical features. Multiple ML classifiers were trained to predict DMR using baseline and early treatment data. A mathematical model based on the SIR framework was integrated with the ML model by incorporating molecular response slopes and white blood cell counts. The hybrid model was compared with the data driven machine learning model to evaluate improvements in predictive performance. External validation was performed using a preliminary independent dataset comprising approximately 466 longitudinal molecular measurements. The integration of clinical data, ML, and mathematical modeling enabled the development of a hybrid predictive model for sustained DMR. Molecular variables were the strongest predictors, particularly BCR::ABL1 transcript levels at 3, 6, and 12 months, while molecular response slopes showed the highest relative importance, emphasizing the value of treatment kinetics. Incorporation of the SIR-based mathematical model into the logistic classifier preserved the discriminative performance of the previously optimized ML model based on 219 patients (AUC = 0.81 vs. 0.84) and substantially outperformed a ML model trained with the same variables alone (AUC = 0.81 vs. 0.62). The hybrid model also demonstrated superior calibration and overall accuracy, reflected by a 49.5% improvement in the Brier Skill Score, while providing interpretable estimates of each variable's contribution. Preliminary external validation using an independent dataset of longitudinal molecular measurements identified a limited number of false-positive and false-negative predictions, indicating the need for further calibration of model coefficients. Nevertheless, the model successfully identified patients with sustained molecular responses who are likely candidates for TFR, supporting the feasibility of integrating mechanistic mathematical modeling with ML for prognostic stratification in CML.
SN - 2531-1379
SN - 2531-1387
ER -