Enabling Population-Scale Screening of Myeloid Neoplasms Using Blood Cell Count Dynamics

INTRODUCTION

During myeloid transformation, hematopoietic stem cells accrue genetic alterations that alter their resilience leading to clonal hematopoiesis of indeterminate potential (CHIP). This progression predisposes patients to the development of myeloid neoplasms, including acute myeloid leukemia (AML), myelodysplastic syndrome (MDS), or myelofibrosis (MF).

Recent studies have sought to identify pre-neoplastic states by analyzing standard blood count values and genetic information (Weeks 2023 NEJM, Gu 2023 Nature Genetics). Given the impracticality of scaling genome sequencing to all patients with abnormal blood counts, we hypothesized that laboratory values, particularly their temporal trends, could enhance the detection of individuals at high risk for developing myeloid neoplasms.

METHODS

We extracted 24 common blood count values from the Helsinki University Hospital for patients later developing de novo AML (n=460 patients), MDS (n=843), or MF (n=250). As controls, we included laboratory values for subjects without blood cancer history (n=11,683) but examined or diagnosed for a hematological condition such as anemia (n=6591), thrombocyte disorders (n=786), or coagulation disorders (n=794). The follow-up time varied from 1 to 309 months.

We created 6 dynamic blood count features to model both the current laboratory values and their trajectories, reflecting changes 60-36 months, 36-12 months, and 12-0 months prior, resulting in a total of 32,611,656 data points. Laboratory values were normalized for age and gender. We trained an Extreme Gradient Boosting (XGBoost) model with survival hazard for each disease and validated these models using a hold-out test set. Lastly, we evaluated the performance of our models with the area under the receiver operating curve (AUC) and precision-recall (PR) metrics.

RESULTS

The XGBoost models detected pre-MF with high reliability (AUC 0.90, PR 0.55) suggesting that blood counts are impaired in most patients years before diagnosis. We were also able to identify most patients developing MDS (AUC 0.75, PR 0.21) and AML (AUC 0.87, PR 0.16) despite a larger rate of false positive predictions. Platelet count and its temporal development represented one of the most essential predictors in all three models. Moreover, white blood cell dynamics predicted the occurrence of AML, while red blood cell metrics - such as cell volume, hemoglobin content, and distribution width - were associated with the future development of MDS and MF.

The addition of dynamic variables improved model performance for both MF and AML (MF +55.0% and AML +55.8% in PR values). However, we could not observe any clear improvement for MDS suggesting higher heterogeneity in pre-neoplastic blood counts. Consistent with our expectations, laboratory results from 12 months prior to diagnosis had a greater influence on the modeling outcomes compared to earlier timepoints. This resulted in an increase in the model AUC of 10.5% for AML, 19.5% for MDS, and 8.7% for MF.

Integrating molecular genetics information revealed associations between pre-diagnostic laboratory data with genomic and cytogenetic alterations. Pre-AML neutropenia and monocytopenia co-occurred with NRAS/KRAS mutations (n=29/144, p<0.001). Instead, TP53 mutations (n=20/144) were less commonly associated with pre-AML thrombocytopenia and neutropenia (p<0.001), suggesting these emerge closer to disease onset. Opposite pre-MF blood counts were associated in patients with normal (n=87/147) and monosomal karyotype (n=6/147). We observed lower hemoglobin levels but higher MCV values in patients with normal karyotype (p<0.001). Instead, MF patients with monosomal karyotype had higher platelet (p=0.002) and neutrophil counts (p<0.001).

CONCLUSION

The results establish the feasibility of screening for myeloid neoplasms using routine, readily available laboratory values. We demonstrated the importance of distinct laboratory tests and associate these with genomic and cytogenetic alterations observed at disease onset. Taken together, this could enable automated, early detection of myeloid neoplasms by cell counters in routine clinical care. Model validation in an independent dataset is currently ongoing.

Porkka:Novartis: Research Funding; Incyte: Research Funding; Roche: Research Funding. Brück:Pfizer: Research Funding; Amgen: Consultancy; GSK: Consultancy; Roche: Consultancy; Sanofi: Consultancy; Novartis: Consultancy; Gilead Sciences: Research Funding; Hematoscope: Current equity holder in private company.