We present the first random forest (RF) machine learning model to predict the likelihood of leptomeningeal metastasis (LMM) based on patient characteristics, clinical data, and cerebrospinal fluid (CSF) findings.

LMM occurs in 5-20% of patients with cancer and often portends poor prognosis, owing to delayed diagnosis. Neuroimaging and CSF testing are limited by poor sensitivity and specificity; the latter is significantly affected by disease burden and sample acquisition. Improving the diagnostic accuracy for LMM could allow for earlier treatment and potentially better outcomes.

3,518 sequential patients who underwent lumbar puncture (LP) with CSF cytological evaluation (totaling 4,197 LPs, 3.8% positive for LMM) between 2017 and 2022 across multiple hospitals in the Cleveland Clinic system were included. A two-stage RF was developed on 70% of the dataset, holding out 30% for testing. The first stage included static/slowly changing factors (e.g., malignancy history, co-morbidities) to generate a baseline risk. The second stage used the baseline risk in conjunction with dynamic variables (e.g., CSF and laboratory values). Class weighting, multiple imputation by chained equations, and model optimization using Shapley additive explanations were used to ensure model robustness.

The final model included 13 parameters (9 clinical, 3 CSF, 10 serum) and predicted LMM in the independent testing set with an area under the curve of 0.86 (sensitivity: 89%, specificity: 71%). The strongest predictors of LMM were BMI, prior malignancy type and duration, CSF corrected total nucleated cell count, and CSF lymphocyte percentage. Amongst model-predicted false positive cases, 14.4% were found by manual review to have ultimately been diagnosed with LMM by imaging or clinical course.

Clinical features predict the likelihood of LMM with acceptable sensitivity and specificity using a RF. Further multi-institutional cohorts will establish generalizability and facilitate creation of a risk scoring tool for clinical use.