Genomic-Based Machine Learning Towards Prediction of the Etiology of Bone Marrow Failure Syndromes

Genetic testing has been increasingly used to assist with differential diagnosis of acquired vs inherited bone marrow failure syndromes (IBMFS), a group of rare and heterogeneous diseases. However, the assay is still costly and not routinely available for many hematologists. To improve decision-making for genetic testing, we developed a genomic-based machine-learning model based on a two-step data-driven clustering and classification process to predict the likelihood of BMF patients having either an acquired or inherited disease based on 27 clinical and laboratory variables recorded at initial clinical encounter.

Clinical records from two independent cohorts of patients screened for pathogenic variants in genes associated with IBMFS were included in this study: the NIH cohort with 441 consecutive patients followed at the NHLBI and NCI, and the USP cohort with 172 consecutive patients from the Medical School of Ribeirão Preto/USP. In a binary target classification, cases were labeled as inherited if they had a pathogenic/likely pathogenic disease-causing variant and as acquired when they had benign or likely benign variants or negative genetic test, regardless of patients' clinical diagnoses. K-means clustering was first applied to resolve our highly dimensional data into two main clusters (Clusters A and B). An optimized bootstrap aggregation ensemble Cluster A specific was trained with cases from the NIH cohort (n=359). The model was then validated with Cluster A cases from the external USP cohort (n=127). The binary classification task was utilized to predict the etiology of BMF cases, labeled as acquired or inherited depending on patients' genomic data.

At first, unsupervised clustering separately grouped datasets into Cluster A, the largest group mostly represented by aplastic anemia (AA), and Cluster B, those underrepresented in our cohort including some classical IBMFS at early disease onset. The ensemble model Cluster A-specific was accurate to predict the BMF etiology in 88% of cases, correctly predicting inherited and likely immune BMF in 72% and 92% of cases, respectively. Out of the 27 initial clinical variables included in the model, 25 were found to be important for prediction. Telomere length (TL), age, and clinical variables were most important for the model's predictive accuracy, highlighting that a comprehensive history and physical examination encompassing all organ systems is imperative. Based on our model, genetic testing must be considered for patients in Cluster A predicted to have inherited disease and also for patients in Cluster B as no specific model was available but they were more likely to have IBMFS in comparison to Cluster A (50% vs 30%). We also recommend genetic screening in patients from Cluster A predicted to have acquired disease who are children (age <18 years who may not have clinical signs of IBMFS), have consanguinity in the family, have a diagnosis of myelodysplastic syndromes with or without suspicion for familial predisposition to myeloid malignancies (all cases where the model had limited prediction). A model without TL, an assay that can also be limited in low-resource centers, underperformed for prediction of inherited cases with sensitivity of 55%, highlighting the importance of TL measurement for the model's performance.

Our machine-learning model reproduced the clinical knowledge used by clinicians specialized in BMF and accurately predicted BMF etiology in 88% of cases. The model was particularly accurate for differential diagnosis of immune AA in adults, which may allow for selections of patients in whom rapidly starting immunosuppression rather than waiting weeks for genetic results is preferable. Clinical variables were strong predictors and adult patients with severe AA rarely had an inherited disease without a positive family history, a suggestive phenotype of IBMFS, or consanguinity being present. The generalizability of our model indicates that this tool can be used by hematologists not specialized in BMF to prioritize patients that would benefit from genetic testing. TL was a top predictor and a key variable for this model's accuracy. Implementation of TL measurement may be critical for differential diagnosis of BMF, especially in low-resource centers where genetic testing is not feasible or readily available. We plan to continue adding to the model to better predict IBMFS cases that were underrepresented in the current cohort.