Clinical and CSF Measurements for Identification of Leptomeningeal Metastasis in Lung Adenocarcinoma

Background/Objectives: Because brain MRI and initial CSF cytology have limited sensitivity, suspected leptomeningeal metastasis (LM) may remain unresolved. We developed a model using clinical and CSF measurements available in source records to support identification of current LM in patients with lung adenocarcinoma undergoing CSF evaluation. Methods: This single-center retrospective cohort included 533 patients with lung adenocarcinoma and CNS metastases: 366 with a retrospectively verified reference diagnosis of current LM and 167 with brain metastases only. An eight-variable logistic model underwent repeated nested cross-validation. Conditional internal performance was estimated using outer-split thresholds selected by inner resampling at specificity >0.90. Results: Mean outer held-out AUROC was 0.960 (95% CI, 0.942-0.974), Brier score was 0.068 (0.054-0.084), calibration slope was 0.864 (0.708-1.106), sensitivity was 0.912 (0.883-0.938), and specificity was 0.901 (0.857-0.939). In a descriptive full-cohort analysis, the model flagged 157 of 170 (92.4%) current-LM patients not identified by the brain MRI strategy and 59 of 65 (90.8%) not identified by first CSF cytology for further evaluation. These were not independent head-to-head accuracy comparisons because MRI and cytology contributed to the reference diagnosis. Conclusions: In this selected cohort, the clinical and CSF measurements captured a multivariable pattern associated with a retrospectively verified reference diagnosis of current LM. The model showed strong conditional internal performance and may serve as an investigational adjunct to prompt further LM-directed evaluation. Independent multicenter validation is required before clinical use.