Utilisation of machine learning algorithms to predict relapse risk following allogeneic stem cell transplantation (alloSCT) for patients with myeloid malignancies

Abstract

Abstract Background: Disease relapse remains the primary cause of treatment failure following alloSCT for acute myeloid leukaemia (AML) and myelodysplasia/myeloproliferative neoplasms (MDS/MPN). Early mixed donor chimerism is frequently associated with relapse risk at a population level, yet its predictive accuracy is modest when considered in an individual patient. We hypothesised that a more nuanced analysis of peripheral blood CD3+/CD3- chimerism dynamics in combination with key recipient, donor and transplant-related variables could improve relapse prediction, providing a valuable tool to guide post-alloSCT surveillance and pre-emptive intervention decision making. Methods: We conducted a retrospective study using data from 259 adult patients who underwent first alloSCT for AML or MDS/MPN between 2016 and 2023 at the Royal Melbourne Hospital and Peter MacCallum Cancer Centre, Australia. A total of 45 clinical, demographic and transplant-related variables were extracted, including CD3+/CD3- chimerism at days 30, 60, and 100 post-alloSCT. Chimerism analysis was performed by short tandem repeat testing by polymerase chain reaction followed by fragment analysis, with a sensitivity of 1-5%. Derived features such as the rate and direction of donor chimerism change between timepoints (e.g., day 60 to 100 post-alloSCT) were empirically determined. We developed and compared multiple machine learning models, including Random Forest, Support Vector Machine, Logistic Regression, gradient-boosted trees, k-Nearest Neighbours and Naive Bayes classifiers. To address the class imbalance between relapsed and non-relapsed patients, we employed Synthetic Minority Oversampling Technique and adjusted class weights. Feature selection was performed using recursive feature elimination and univariate analysis. Hence, a total of 100 different feature extraction and machine learning model combinations were tested. Performance was evaluated via cross-validation using area under the receiver operating characteristic curve (AUC), F1-score, sensitivity, specificity and calibration curves. Results: Patients consisted predominantly of males (60%) with AML (70%) at a median age of 59 (range: 17-73). Reduced intensity conditioning was used in 63% of patients and matched unrelated donors were the primary donor source (52%). Graft-versus-host disease prophylaxis was predominantly calcineurin inhibitor-based (85%) with anti-thymocyte globulin used in 52%. Morphological relapse was encountered in 76 patients (29%) at a median of 224 days (interquartile range: 466). The most predictive features included CD3+ donor chimerism at day 60 and 100 post-alloSCT, changes in chimerism between these timepoints, recipient age, conditioning intensity and disease subtype. A Random Forest classifier was the best performing model — achieving an AUC of 0.83, sensitivity of 0.78, specificity of 0.80 and an F1-score of 0.74. Calibration analysis demonstrated that predicted relapse probabilities were well-aligned with relapse events, suggesting that the model provides accurate risk estimates for disease relapse. Conclusion: Using a unique machine learning algorithm, we have identified that changes in CD3+ donor chimerism, when analysed alongside key established transplant variables, enables highly accurate prediction of post-alloSCT relapse and may guide clinicians with making timely individualised interventions to prevent relapse.