Development and Temporal Validation of a Three-Metabolite Cerebrospinal Fluid Model for the Diagnosis of Leptomeningeal Metastasis from Solid Tumors

Background/Objectives: Leptomeningeal metastasis (LM) is a severe complication of solid tumors. Cerebrospinal fluid (CSF) cytology is only about 70% sensitive on a single examination, and objective markers are needed. Methods: CSF from 49 patients with LM and 16 with intracranial disease without LM was analyzed by targeted LC–MS in a training batch (n = 30) and a temporally independent validation batch (n = 35). Metabolites selected by both LASSO and random forest entered a logistic regression model, which was applied unchanged to the validation batch and evaluated by repeated and nested cross-validation on all 65 samples. Results: Hydroxyphenyllactic acid, argininosuccinic acid and norepinephrine were selected by both algorithms. The model reached an AUC of 0.954 in the training batch and 0.839 (95% CI 0.687–0.991) in the validation batch. In repeated cross-validation on all 65 samples, the AUC was 0.790 (95% CI 0.655–0.924), with sensitivity 0.714 and specificity 0.938; nested cross-validation gave 0.747. Conclusions: A three-metabolite CSF panel discriminates LM from non-LM and retains discrimination under cross-validation and in a temporally independent batch. It is a candidate adjunct to cytology and imaging whose specificity, based on 16 controls, requires confirmation in larger cohorts.