Journal: Nature communications
This study addresses the challenge of early detection in primary vitreoretinal lymphoma (PVRL), a rare and aggressive intraocular cancer often misdiagnosed due to nonspecific symptoms and lack of effective screening.
Researchers developed a machine learning-based screening model using complete blood count data from a multicenter case-control study involving:
- 255 PVRL patients
- 292 controls
A six-feature random forest model showed strong diagnostic performance with an AUC of 0.85 in the discovery cohort and maintained robust validation across multiple cohorts (AUC 0.80–0.83), outperforming traditional intraocular biomarkers like the IL-10/IL-6 ratio.
Prospective validation in a hospital-based cohort of over 100,000 individuals demonstrated:
- High sensitivity of 95.0%
- High specificity of 99.97%
- Positive predictive value of 57.6%
- Negative predictive value of 99.99%
Further community-level screening similarly identified high-risk individuals with a positive predictive value of 59.1%.
This noninvasive, scalable blood-based model provides an effective tool for population-level PVRL risk stratification and timely clinical triage, supported by a web application for practical implementation.