Your Next Blood Test Could Predict Leukemia Risk Years Before Symptoms Appear

A routine complete blood count—one of medicine's most common tests—may soon double as a powerful early warning system for blood cancer. Researchers have developed PRISM, a machine learning model that analyzes standard blood test parameters to identify adults harboring clonal hematopoiesis, a silent precursor condition that can progress to leukemia and other blood cancers.

The breakthrough, published in eClinicalMedicine in 2025, demonstrates how artificial intelligence can extract hidden risk signals from data already collected in millions of annual checkups, potentially transforming blood cancer prevention without requiring expensive new tests.

Detecting the Silent Threat

Clonal hematopoiesis affects approximately 20% of adults over 65, occurring when blood stem cells acquire mutations that allow them to outcompete normal cells. While most carriers face minimal cancer risk, a small subset harbors dangerous mutations that progress to myeloid neoplasms at rates of 0.5–1.0% annually. The challenge has been identifying this high-risk minority without sequencing everyone.

PRISM addresses this by analyzing 14 parameters from routine complete blood counts—including red blood cell counts, lymphocytes, monocytes, neutrophils, hemoglobin levels, and platelet measurements—to flag patients who warrant genomic testing. The balanced random forest model was trained on 369,260 UK Biobank participants aged 40–69 and validated across three diverse cohorts totaling over 109,000 individuals.

The model achieved an area under the curve of 0.89 in UK Biobank testing, with sensitivity ranging from 0.73 to 0.89 across validation cohorts. Critically, PRISM captured 100% of high-risk cases defined by the clonal hematopoiesis risk score across all three populations, while maintaining a negative predictive value above 99%.

From Universal Screening to Precision Triage

The clinical efficiency gains are substantial. PRISM identified one high-risk case per 247 people flagged, compared to one per 2,500 under universal sequencing approaches. This translates to potential reductions in unnecessary genomic testing of 14–55%, depending on the population screened.

Among PRISM-positive individuals, 10-year myeloid neoplasia incidence reached 1.44%, compared to just 0.11% in those testing negative—a hazard ratio of 8.81. The model also correlated with cardiovascular risk, with 10-year CV risk of 15.7% in the positive class versus 4.6% in PRISM-negative patients.

External validation demonstrated portability across clinical contexts, including 2,233 Dana-Farber Cancer Institute clinic patients and 14,674 All of Us Research Program participants representing diverse populations. Performance remained robust despite differences in patient demographics and clinical settings.

Implications for Cancer Prevention

For healthcare providers, PRISM offers a scalable triage pathway using universally available laboratory data. The high negative predictive value provides clinical confidence to reassure patients with negative results, reducing anxiety and avoiding overdiagnosis in the approximately 90% of clonal hematopoiesis carriers who face minimal cancer risk.

The model creates opportunities for risk-stratified surveillance, enabling earlier detection and monitoring of high-risk clones before progression to overt leukemia or myelodysplastic syndrome. Potential applications extend beyond screening to transplant donor evaluation and hematology clinic triage.

The research, led by Nandi, Zhao, Pershad and colleagues from Dana-Farber Cancer Institute and multiple academic centers, exemplifies how machine learning can unlock hidden clinical intelligence from routine medical data, transforming common tests into precision medicine tools without requiring new infrastructure or patient procedures.

Sources

1. Routine Blood Counts Could Reveal Who Carries a Hidden Blood Cancer Risk — Scienmag