Machine learning-based cost prediction in pediatric allogeneic bone marrow transplantation: Development and validation of an interactive clinical decision support tool

Abstract

Abstract Background: Allogeneic bone marrow transplantation (BMT) costs exhibit substantial variability, yet robust prediction models for healthcare resource planning remain limited. Accurate cost forecasting is critical for institutional budgeting, family counseling, and clinical decision-making in pediatric transplant centers. Methods: We performed a comprehensive cost analysis of 166 consecutive pediatric patients with malignant disorders undergoing allogeneic BMT at King Hussein Cancer Center (2016-2022). Costs were stratified by transplant period and service category. Advanced machine learning algorithms (XGBoost, Random Forest, Neural Networks) were trained for both binary classification (high-cost ≥75,000 JD) and continuous cost prediction. Clinical predictors were identified through univariate and multivariable logistic regression. Results: Total first-year costs averaged 88,445 JD (median 75,009 JD, IQR 61,681-95,000 JD, range 18,995-349,612 JD) with perfect cost distribution balance (n=83 high-cost vs n=83 low-cost patients). Cost concentration was heavily skewed toward early transplant period: Day 0-100 contributed 73,332 JD (78% of total costs) versus Day 101-365 at 21,076 JD (22%). Service-specific costs revealed pharmacy as the dominant component (38,082 JD, 43% of total), followed by room charges (28,951 JD, 33%), laboratory/blood bank (19,906 JD, 22%), and radiology (1,481 JD, 2%). Graft acquisition costs varied dramatically by source: umbilical cord blood (35,000 JD), bone marrow (2,719 JD), and peripheral blood stem cells (2,154 JD). Multivariable analysis identified MDS/myeloproliferative disorders as the strongest cost predictor (OR 16.32, 95% CI 3.02-114.21, p=0.002), and AML (OR 2.85, 95% CI 1.17-7.25, p=0.024) while related donors provided substantial cost protection (OR 0.22, 95% CI 0.09-0.53, p=0.001). Age demonstrated linear cost escalation (OR 1.13 per year, 95% CI 1.03-1.24, p=0.009), and high-grade chronic GVHD tripled costs (OR 3.10, 95% CI 1.16-8.62, p=0.025). Umbilical cord blood transplants exceeded PBSC costs across all services: laboratory (+19.7%), pharmacy (+8.9%), and room charges (+10.7%). Machine learning models achieved excellent performance: XGBoost demonstrated optimal classification accuracy (AUC 0.828, 82.8% discriminative ability) and regression precision (R² 0.421, explaining 42.1% cost variance). Cost predictions maintained clinical utility with mean absolute error of ±21,467 JD (27.6% MAPE) and 55.9% of predictions within 20% accuracy threshold. Conclusions: Pediatric BMT costs are highly predictable using readily available clinical parameters, with diagnosis type, donor relationship, and GVHD risk as primary determinants. The 78% cost concentration in the first 100 days enables targeted resource allocation and budget planning. MDS/MPD and AML patients require 16-fold higher cost preparation compared to ALL patients, while related donors provide 78% cost reduction. Machine learning models demonstrate clinical-grade accuracy suitable for real-time decision support.