Explainable machine learning for chronic lymphocytic leukemia treatment prediction using only inexpensive tests.

Area of Science:

• Oncology
• Medical Informatics
• Machine Learning

Background:

• Chronic lymphocytic leukemia (CLL) is a heterogeneous cancer common in the elderly.
• Current prognostic tools like CLL-IPI do not predict treatment necessity.
• Accurate prediction of treatment need is crucial for personalized patient management.

Purpose of the Study:

• To develop a machine learning (ML) model for predicting the need for CLL treatment within two years of diagnosis.
• To utilize only demographic and routine laboratory data for accessibility in low-resource settings.
• To create a cost-effective and reliable tool for clinical decision support.

Main Methods:

• A single-center study included 109 adult CLL patients diagnosed between 2009-2019.
• Demographic, clinical, and laboratory data were extracted from medical records.
• Multiple ML models were evaluated, including Gradient Boosting (GBM) and Generalized Linear Models (GLM), using cross-validation.
• SHapley Additive exPlanations (SHAP) and decision trees were used for model interpretability.

Main Results:

• A GBM model achieved an AUPRC of 0.7686 (±0.0837) using only basic laboratory and demographic data, surpassing CLL-IPI.
• Red blood cell count was identified as a key predictor for treatment necessity.
• A decision tree model indicated that low hemoglobin and low Neutrophil to Lymphocyte Ratio (NLR) predicted a high likelihood of requiring treatment.

Conclusions:

• ML models can accurately predict the need for CLL treatment using inexpensive, routine data.
• The developed model offers a valuable, accessible tool for clinical decision-making, especially in resource-limited areas.
• Further validation on larger patient cohorts is recommended to confirm these findings.