Machine learning framework for predicting the presence of high-risk clonal haematopoiesis using complete blood count data: a population-based study of 431,531 UK Biobank participants

BackgroundClonal haematopoiesis (CH), the disproportionate expansion of a haematopoietic stem cell and its progeny, driven by somatic DNA mutations, is a common age-related phenomenon that engenders an increased risk of developing myeloid neoplasms (MN). At present, CH is identified by targeted sequencing of peripheral blood DNA, which is impractical to apply at population scale. The complete blood count (CBC) is an inexpensive, widely used clinical test. Here, we explore whether machine learning (ML) approaches applied to CBC data could predict individuals likely to harbour CH and prioritise them for DNA sequencing. MethodsThe UK Biobank was filtered to identify 431,531 participants with paired CBC and whole exome sequencing (WES). Somatic mutations were previously identified from blood WES using Mutect2 to classify individuals with CH driver mutations. Using 18 CBC indices/features and basic demographics (age and sex), we trained a range of tree-based ML classifiers to infer as binary output, the presence/ absence of CH. FindingsUsing Random Forest (RF) classifiers, we predicted the presence/absence of CH driven by mutations in one of five genes known to confer a high-risk of incident MN (JAK2, CALR, SF3B1, SRSF2 and U2AF1). We subsequently developed a unified, optimised RF classifier for high-risk CH driven by any of these genes and assessed its performance (median AUC 0.85). However, the low prevalence of high-risk CH implies that our model cannot be generalised to population scale without compromising its sensitivity (20.1% using stringent cutoff probability score). InterpretationWe showcase a proof-of-concept that the presence of high-risk CH can be inferred from CBC perturbations using RF classifiers. The future integration of raw blood cell analyser data can help improve the performance of our model and facilitate its application at scale. FundingCancer Research UK. Research in context Evidence before this studyWe searched PubMed for articles published, in English, between database inception and 5th of June 2024, using the terms "clonal hematopoiesis" AND ("machine learning" OR "artificial intelligence"). We additionally searched for the terms "clonal hematopoesis" AND "complete blood count". We found 18 research articles: one article used ML approaches (XGBoost classifiers) to differentiate clonal haematopoiesis "driver" mutations from "passenger" mutations, but none linked machine learning frameworks to complete blood count data for predicting the presence of clonal haematopoiesis. Progression from clonal haematopoiesis to myeloid neoplasia is known to be associated with several blood count parameters; two recent publications developed clonal haematopoiesis risk stratification tools that incorporated blood count indices in their final risk prediction models (Gu et al. Nature Genetics, Weeks et al. NEJM Evidence). However, we found no study assessing whether blood count indices could be used to infer the presence of clonal haematopoiesis. Added value of this studyHere we show that CH driven by mutations in genes associated with high risk of progression to myeloid neoplasia can be reliably differentiated using ML approaches applied on peripheral blood indices; however, low-risk forms of CH (driven by mutations in the DNMT3A or TET2 genes) cannot be reliably inferred from CBC indices. While optimising the model we identified challenges in upscaling its applicability; we propose that the integration of single-cell resolution "raw" blood analyser data might overcome these issues. Previous efforts to enhance the scalability of CH screening focused on reducing DNA sequencing costs. Here, we provide a proof-of-concept that an extensively used clinical test, the CBC, can, using machine learning approaches, predict individuals more likely to harbour high-risk CH, who should be prioritised for genetic testing. Implications of all the available evidenceOur study proposes a model for predicting high-risk CH mutations by applying a Random Forest classifier on CBC indices; this represents an important step towards scalable screening for identifying individuals at high risk of developing myeloid neoplasia in the future. This is an attractive approach, as it relies solely on a routine, inexpensive test. Despite good sensitivity, the low prevalence of high-risk CH leads to a low positive predictive value that precludes the use of the predictive model as a population-wide pre-screening tool. To overcome this, we propose the future integration of raw blood analyser data into models like ours to improve the performance and scalability of this approach.