Machine Learning Models for Pre-Transplant Risk Stratification in Severe Aplastic Anemia

Background:

Severe Aplastic Anemia (SAA) is a life-threatening disease characterized by profound pancytopenia. Hematopoietic stem cell transplantation (HCT) has been a curative treatment since the 1970s [Blood.2012;120(6):1185-96] with the outcomes have tremendously improved over the past decades with current survival rates are approximately 82-92%. However, survival is much lower for patients with preexisting co-morbidities [Ann Hematol. 2020; 99(11):2529-2538]. This necessitates a pre-transplant risk stratification tool to personalize treatment and further improve patient outcomes. To our knowledge, there is no universally validated predictive tool for mortality in SAA patients undergoing HCT.

Objective:

Develop and evaluate machine learning models using pre-transplant features to predict mortality in SAA patients and identify the key features that played a vital role in the model's predictions for more personalized treatment.

Methods:

Data Collection & Preparation: An initial dataset of 545 SAA patients who underwent HCT between 1988 and 2018 was collected from the CIBMTR public datasets. The features included recipient, donor, disease, transplant- related variables, demographics, and lab results. The median age of patients was twenty-one years old (range 0-74), female-to- male ratio was 1:1.4, and mortality was 17.5%. No missing values were identified in the target variable of any patient. Missing values in remaining variables were imputed using the KNN imputer if more than 50% of the patients had missing data. Polynomial and interaction terms were generated to capture non-linear relationships. The data was split into an 80:20 training/validation set for hyperparameter fine-tuning.

Model Training: Machine learning models, such as RandomForest, Gradient Boosting, and XGBoost classifiers, were trained using a comprehensive preprocessing pipeline. Hyperparameter tuning was performed using GridSearchCV, and class imbalance was handled using SMOTE. Cross-validation was conducted using StratifiedKFold with five folds.

Model Evaluation: We assessed overall model discrimination using AUROC, and the balance of model sensitivity and positive predictive value using Area Under the Precision-Recall Curve (AUPRC) and F1 score. Feature importance was evaluated using SHAP (SHapley Additive exPlanations) value analysis. The robustness of the models was further verified through additional cross-validation strategies to ensure consistency in performance.

Results:

Among all models, the GradientBoosting model demonstrated high performance with a ROC AUC of 0.835 and AUPRC of 0.708. It showed high recall for class 0, live (0.95) with a good balance between precision and recall (F1-score of 0.92). Key predictors included sex, donor & recipient age, blood counts at diagnosis and graft type. Interactions between demographic (sex, age) and clinical features (blood counts at diagnosis, graft type) were significant.

Conclusion:

Our pre-transplant risk stratification model can potentially predict mortality in SAA patients undergoing HCT. Higher predictive power can be achieved with more comprehensive pre-transplant data. Future directions include validating the model on a separate dataset and in larger, multi-center prospective datasets. Continued research and collaboration are essential to advance these models and improve HCT outcome. Understanding the key predictors and their interactions can help clinicians identify high-risk patients and tailor treatment. We highlight the need for larger datasets with all relevant variables, including longitudinal dynamic ones, to further refine clinical practice.