Prediction of Event Free Survival in Hematopoietic Cell Transplantation: A Machine Learning Clustering and Proximity Analysis

Age and donor type have been shown to be key predictors of outcomes for hematopoietic stem cell transplantation (HCT) in sickle cell disease (SCD)(Brzauskas et al BLOOD 2020.) However, these broad categories, may hide subgroups with heterogeneous outcomes. impacting individualized treatment selection and shared-decision making. Traditional unsupervised clustering methods are less effective with registry datasets because they assign equal importance to all features, estimate proximity using algebraic distance measures that may not capture clinical data complexities, and fail to produce unique subtypes for each outcome of interest. To ameliorate this, we perform proximity analysis by analyzing the leaves of a machine learning Random Forest (RF) trained on the CIBMTR dataset with EFS as the target variable, which determines the most relevant features and produces a unique proximity matrix for each outcome of interest that better represents the actual similarity between samples. Clustering this proximity matrix gives us patient subtypes.

Methodology: The RF-GAP method, proposed by Rhodes et al, calculates the proximity between samples by analyzing their co-occurrence in the same leaf nodes across different trees. The principle is that samples frequently ending up in the same leaf node are more similar and, thus, will have higher relative proximity values. Additionally, RF-GAP considers the number of out-of-bag (OOB) occurrences of each sample and the frequency of the occurrence during bootstrapping, providing a comprehensive similarity measure.

To ensure that we included only the most informative features, we first used Recursive Feature Elimination with Cross-Validation, selecting the top 17 features based on their importance. The RF model's validity was confirmed through 10-fold cross-validation, yielding an AUC of 0.78, indicating acceptable predictive performance. We then trained the Random Forest on the entire dataset using these key features.

We created a proximity matrix where each sample's proximity to every other sample was determined using RF-GAP, and the matrix row-normalized to sum to one. This resulting square similarity matrix was treated as a directed graph. We applied a modified spectral clustering technique to this matrix to identify distinct subtypes for EFS.

We identified six optimal clusters and validated them using both quantitative metrics and qualitative assessment. The clusters were evaluated quantitatively using the elbow test and the silhouette score (0.36) and confirmed their clinical relevance.

Results: Cluster one (n=300; EFS 96% -low risk) and cluster two (n=316; 80% EFS-intermediate risk) consisted of patients receiving HCT from HLA identical sibling donors, the majority of whom were ≤ 10 years. Cluster two has a statistically significant greater proportion of ≥ 3 VOCs required hospitalization(p<0.001) and acute chest syndromes (ACS) requiring exchange transfusion( p<0.001), suggesting that for young patients with HLA identical donors, the presence of more severe disease symptoms may contribute to worse outcomes.

Cluster three (n=303; EFS-86% -intermediate risk) and cluster six(n=91; EFS 44% high risk) consisted of patients receiving HCT from HLA identical donors, mostly aged ≥10. Cluster six had a statistically significant greater proportion of non-myeloablative conditioning (p<0.001), peripheral blood graft type(p<0.001), ACS requiring exchange transfusion (p <0.001), suggesting a possible influence of graft source, conditioning intensity, and prior history of severe ACS in older patients with HLA identical sibling donors.

Cluster four (n=325; EFS 81%-intermediate risk) and cluster five (n=306; 44% EFS (high risk) all receiving HCT from alternate donors. Cluster 5 had a statistically significant greater proportion of patients ≤ 10 years (p<0.001), nonmyeloablative conditioning (p=0.001), peripheral blood graft type (p<0.001), ACS requiring intubation(p<0.001), VOCs needing hospitalization(p=0.01). We infer that outcomes may be impacted by graft type, conditioning intensity, and disease severity for patients with alternate donors.

Conclusion: We have identified prognostic subtypes within age & donor type for patients undergoing HCT for SCD. Identifying well-performing and poorly performing clusters within similar age and donor types with potential utility in shared decision-making about HCT for SCD.