DSAI-03 DIAGNOSIS AND PROGNOSIS OF LEPTOMENINGEAL METASTASIS: A MACHINE LEARNING APPROACH

Le résumé fourni par la source

Abstract Objectives We present a prognostic machine learning (ML) model to predict, among patients undergoing lumbar puncture (LP), the likelihood of leptomeningeal metastasis (LMM). Background LMM, seen in 5-20% of patients with cancer, is challenging to diagnose and historically carries a poor prognosis, often owing to delayed diagnosis. Standard clinical workup is limited by poor sensitivity and specificity; better diagnostic accuracy for LMM may facilitate improved outcomes, and prognostic understanding could inform management. Methods 3,515 sequential patients who underwent LP with CSF cytology (totaling 4,192 LPs, 4.8% positive for LMM) between 2017 and 2022 across multiple hospitals in the Cleveland Clinic system were included. A two-stage random forest (RF) model was developed on 70% of patients in the dataset, with 30% held for independent validation. The first stage included slowly changing factors (e.g., malignancy history, comorbidities) to generate a baseline risk of LMM. The second stage refined the risk prediction using dynamic variables (e.g., CSF and laboratory values). Results The final model used 24 parameters (9 clinical, 5 CSF, 10 serum) to predict risk of LMM with an area under the curve of 0.87 (sensitivity: 80%, specificity: 83%) in the independent validation cohort. The strongest predictors of LMM were BMI, prior malignancy type and duration, CSF corrected total nucleated cell count, and CSF lymphocyte percentage. Although not designed as a prognostic model, given LMM’s association with prognosis and the need to identify high-risk patients, Cox proportional hazard modeling was conducted. For patients predicted as LMM-positive, there was a markedly increased risk of death (HR 2.26, 95% CI 1.86-2.76, p < 0.001). Conclusion Clinical features combined with laboratory values in a ML framework predict LMM with higher sensitivity and specificity than standard approaches. Our model is highly prognostic, informing management decisions. Multi-institutional validation is underway to establish generalizability.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.