A simple blood test analysed by artificial intelligence could revolutionise early detection of primary vitreoretinal lymphoma (PVRL), a rare and aggressive eye cancer often mistaken for inflammation. Researchers developed a machine learning model using routine complete blood count data from 255 PVRL patients and 292 controls, offering a noninvasive screening tool where invasive biopsies currently dominate.
Superior accuracy beats eye fluid tests
The six-feature random forest model achieved an area under the curve (AUC) of 0.85 in the discovery cohort, with consistent validation across groups at AUC 0.80–0.83. This outperformed traditional intraocular biomarkers like the interleukin-10/interleukin-6 ratio, which scored only 0.65–0.78. PVRL typically masquerades as uveitis with blurred vision, floaters, or hazy sight, delaying diagnosis by months or years due to nonspecific symptoms.
Real-world success flags hidden cases
In a hospital prospective cohort of 100,526 people, the model identified 38 PVRL cases among 66 high-risk individuals, plus 2 more among 83,610 low-risk patients. This delivered 95.0% sensitivity, 99.97% specificity, 57.6% positive predictive value, and 99.99% negative predictive value. A community cohort of 515,326 screened 22 as high risk, confirming 13 PVRL cases (PPV 59.1%), proving scalability for population screening.
Web tool enables rapid triage
The study highlights PVRL’s fatality when it spreads to the brain, often bilateral and affecting older adults with painless vision loss. A free web application now allows clinicians to input blood results for instant risk scores, prioritising urgent eye exams or vitrectomy. This blood-based strategy could save vision and lives in low-resource settings, bypassing costly imaging or lumbar punctures, with experts calling it a breakthrough for this underdiagnosed malignancy.
Reference
Li S et al. A noninvasive machine learning model using a complete blood count for screening of primary vitreoretinal lymphoma. Nature Communications. 2025;16:10667.
