A noninvasive machine learning model using a complete blood count for screening of primary vitreoretinal lymphoma

Abstract

Primary vitreoretinal lymphoma (PVRL) is a rare and aggressive intraocular malignancy that is frequently misdiagnosed because of its nonspecific early manifestations and the lack of effective screening tools. We conduct a multicentre case–control study including 255 PVRL patients and 292 controls to develop a machine learning–based screening model using complete blood count data. A six-feature random forest model demonstrates high diagnostic accuracy in the discovery cohort (area under the curve [AUC] = 0.85) and validates across all cohorts (AUC = 0.80–0.83), outperforming intraocular biomarkers such as the interleukin-10/interleukin-6 ratio (AUC = 0.65–0.78). Model performance further validates in a hospital-based prospective cohort (n = 100,526), where 38 PVRLs are identified among 66 individuals classified as high risk, and 2 additional cases are identified among 83,610 individuals classified as low risk, yielding a sensitivity of 95.0%, specificity of 99.97%, positive predictive value (PPV) of 57.6%, and negative predictive value of 99.99%. In the community cohort (n = 515,326), 22 individuals are flagged as high risk, 13 of whom are confirmed as having PVRL (PPV = 59.1%). This study presents the noninvasive and scalable blood-based screening strategy for detection of PVRL, with a web application enabling timely triage and population-level risk stratification.