Xgboost model using diagnostic CBC to predict early mortality in newly diagnosed AML patients

Abstract

Abstract Background: Even among clinically fit patients, early mortality after intensive induction chemotherapy remains a challenge in newly diagnosed acute myeloid leukemia (AML), particularly in resource-limited settings. In the era of emerging less intensive Venetoclax-based regimens, identifying risk factors for early mortality is essential to guide safer, risk-adapted therapies. In this context, machine learning (ML) approaches may offer enhanced predictive accuracy compared to traditional statistical models. Methods: We conducted a retrospective cohort study including all 230 adults with newly diagnosed AML at our public center in Brazil between 2014 and 2025, focusing on those who received intensive induction chemotherapy (n=101). Demographic, clinical and laboratory data were collected from electronic medical records. The primary outcome was early mortality, defined as death within 30 days of induction chemotherapy. Alongside logistic regressions, Decision Tree models were applied to identify relevant predictors of early mortality. These analyses highlighted diagnostic complete blood count (CBC) parameters as the most influential predictors in our dataset. Based on these findings, using Python scikit-learn v1.6.1 and xgboost v2.1.4 we developed Support Vector Machine (SVM) and Extreme Gradient Boosting (XGBoost) models using only diagnostic CBC data. To address class imbalance, we applied Synthetic Minority Over-sampling Technique (SMOTE) and Random Undersampling. Optimal hyperparameters were determined using a GridSearch strategy, model robustness was assessed via K-Fold cross-validation (k=5) and performance was measured by AUROC on a hold-out test. Finally, SHapley Additive exPlanations (SHAP) was used to interpret the XGBoost model.Results: The final analysis included all patients newly diagnosed with AML fit for intensive chemotherapy (n=101). All cases received standard “7+3” induction (Cytarabine 100 mg/m² and Daunorubicin 60 mg/m²). The median age was 51 years (range 38-62), 55.4% were female and 39.6% were classified as adverse risk by ELN 2022. For consolidation, 41.6% received intermediate-dose Cytarabine and 50.5% underwent allogeneic transplantation. The mean follow-up was 21.6 months. The primary outcome of early mortality occurred in 12.8% of the cases, mostly for infectious complications (76.9%). No significant associations were found in univariate analysis with the outcome and baseline parameters such as age (OR 0.98, p=0.094), Charlson Comorbidity Index (OR 0.92, p=0.395), time from symptom onset to diagnosis (OR 1.00, p=0.718), hemoglobin (OR 0.88, p=0.204), leukocyte count (OR 1.00, p=0.055), or creatinine (OR 1.12, p=0.557). Following this analysis, we tested several Decision Tree models using different combinations of demographic, clinical, and laboratory variables to identify predictors of early mortality. The best performance consistently came from models based solely on diagnostic CBC parameters. To address class imbalance in our modeling process, we applied SMOTE sampling (ratio=0.5) to the entire cohort, followed by random undersampling (ratio=0.9). Based on these findings, we developed XGBoost, Decision Tree and SVM models using diagnostic CBC data to predict early mortality. The XGBoost model showed the best performance with an AUROC of 0.823, sensitivity (Sen) of 74%, and specificity (Spe) of 74%, with a <9% AUROC gap between training and hold-out sets. It outperformed both the Decision Tree model (AUROC 0.768, Sen 76%, Spe 63%) and SVM model (AUROC 0.731, Sen 60%, Spe 66%). Performance was consistent across sex, age, and race/ethnicity groups, supporting model fairness. SHAP analysis of the XGBoost model identified hemoglobin, monocyte count, and basophil count as the three most influential CBC parameters to predict early mortality.Conclusions: In this single-center preliminary study, a ML XGBoost model built solely on diagnostic CBC parameters demonstrated good predictive accuracy for early mortality following intensive induction chemotherapy. This approach could help identify patients who might benefit from risk-adapted strategies, such as the use of less intensive regimens or closer follow-up. Models built on rapid and low-cost tests are particularly valuable in resource-limited settings. However, external validation using larger multicenter cohorts is essential to reduce potential overfitting and minimize reliance on synthetic data, leading to more robust and generalizable models.