A Machine Learning-Based Workflow for Predicting Transplant Outcome in Patients with Sickle Cell Disease

Introduction: Sickle cell disease (SCD) is a prevalent life-threatening condition with limited therapeutic options. A typical feature of the disease is the wide spectrum of clinical severity that will become more apparent as patients age. With the advent of curative but high-risk transplantation and genetic therapies, development of a risk stratification system to identify patients eligible for these curative treatments has become an urgent need. Of the curative therapies, allogeneic hematopoietic cell transplantation (HCT) remains the more established and relatively more accessible option for patients. In 2021, ASH published a guideline on HCT for SCD, highlighting an immediate need for predictive biomarkers of mortality associated with HCT. In prior work, we developed a machine learning (ML)-based algorithm for predicting mortality in patients with SCD that out-performed competing models for risk prediction. However, we recognized that SCD patients who underwent HCT were censored and not included in the analysis. Here, we applied a similar ML-approach to develop a risk stratification workflow to predict both mortality and transplant outcomes specifically in adults with SCD undergoing HCT.

Methods: The study included SCD patients that underwent HCT at the National Institutes of Health Clinical Center, between Sept 2004 and February 2020 (NCT#00061568, NCT#00977691, NCT#02105766 and NCT#03077542). 73 covariates (demographics, echo, electrocardiogram (EKG), vitals, laboratory, and pulmonary function tests) were considered as candidate risk factors for mortality and transplant failure after high missingness exclusion. The primary outcome was all-cause mortality ascertained by the date between consent date and the date of last follow-up, proxy interview, medical records, National Death Index, and Social Security Death Index search. We used a two-step statistical ML approach to select top influential covariates for mortality, including random survival forest and regularized-Cox regression. A risk score based on stepwise Cox regression was developed and further validated for transplant event prediction, which was defined as graft failure by the return of acute complication of SCD or death (whichever occurred first). The feature selection was based upon 95 patients and ML-model was evaluated on all patients with available final selected risk factors (N=111).

Results: A total of 111 SCD patients (median age [IQR]: 31.8 [24.7 – 37.8]) underwent HCT, 74 (67%) with a fully HLA-matched related donor. 17 deaths (15.3%) were observed with a median follow-up of 6.9 years. Multivariable-based regression identified 4 independent predictors of mortality: blood urea nitrogen (BUN), red cell distribution width (RDW), lactate dehydrogenase (LDH), and reticulocyte absolute count. The risk score was calculated by summing up the predictors with dichotomization and equal weight. Our prognostic risk score had superior performance with a C-statistic of 0.842 compared to 0.679 with the classic HCT-comorbidity index (HCT-CI) based on the 95 model construct cohort. The C-statistic of 0.815 on the cohort of 111 validated its robustness. The model stratified the cohort into four groups with significantly different 4-year mortality rates of 0%, 3.2%, 7.9% and 66.7%, corresponding to risk scores of 0, 1, 2, and 3, respectively. There were 27 graft failures with a higher mortality compared to those engrafted (40.7% vs. 7.1%). Besides the mortality stratification, the score could also identify patients’ risk of transplant failure with 4-year disease progression estimates of 9.1%, 22%, 32.8% and 74.1%, respectively.

Conclusion: Based on a cohort of 111 SCD patients, our ML workflow identified BUN, LDH, RDW, and absolute reticulocyte count, alongside the transplant donor type, that were associated with both mortality and transplant outcomes. The results highlighted the importance of red blood cell and renal health as risk factors for HCT. The proposed risk prediction workflow outperformed the established HCT-CI index. Given the heterogeneity in SCD severity and patient phenotypes among different centers, we recommend implementing the ML workflow, rather than the specific biomarkers identified, for tailored center-specific risk identification. The workflow is both versatile and adaptable for risk stratification and could be a useful decision-making tool in the selection of SCD candidates for HCT.