AUTHOR=Gao Tianqi , Chen Fengxi , Li Man TITLE=Development of Two Diagnostic Prediction Models for Leptomeningeal Metastasis in Patients With Solid Tumors JOURNAL=Frontiers in Neurology VOLUME=13 YEAR=2022 URL=https://www.frontiersin.org/journals/neurology/articles/10.3389/fneur.2022.899153 DOI=10.3389/fneur.2022.899153 ISSN=1664-2295 ABSTRACT=Objectives

For accurate diagnosis of leptomeningeal metastasis (LM) and to avoid unnecessary examinations or lumber puncture (LP), we develop two diagnostic prediction models for patients with solid tumors.

Study Design, Setting, and Participants

This is a retrospective cohort study launched at the Second Affiliated Hospital of Dalian Medical University. In total, 206 patients who had been admitted between January 2005 and December 2021 with a solid tumor and clinical suspicion of LM were enrolled to develop model A. In total, 152 patients of them who underwent LPs for cytology and biochemistry were enrolled to develop model B.

Model Development

Diagnostic factors included skull metastasis, active brain metastasis, progressed extracranial disease, number of extracranial organs involved, number of symptoms, cerebrospinal fluid (CSF) protein, and CSF glucose. The outcome predictor was defined as the clinical diagnosis of LM. Logistic least absolute shrinkage and selection operator (LASSO) regression was used to identify relevant variables and fit the prediction model. A calibration curve and the concordance index (c-index) were used to evaluate calibration and discrimination ability. The n-fold cross-validation method was used to internally validate the models. The decision curve analysis (DCA) and the interventions avoided analysis (IAA) were used to evaluate the clinical application.

Results

The area under the curve (AUC) values of models A and B were 0.812 (95% CI: 0.751–0.874) and 0.901 (95% CI: 0.852–0.949). Respectively, compared to the first magnetic resonance imaging (MRI) and first LP, models A and B showed a higher AUC (model A vs. first MRI: 0.812 vs. 0.743, p = 0.087; model B vs. first LP: 0.901 vs. 0.800, p = 0.010). The validated c-indexes were 0.810 (95% CI: 0.670–0.952) and 0.899 (95% CI: 0.823–0.977). The calibration curves show a good calibrated ability. The evaluation of clinical application revealed a net clinical benefit and a reduction of unnecessary interventions using the models.

Conclusions

The models can help improve diagnostic accuracy when used alone or in combination with conventional work-up. They also exhibit a net clinical benefit in medical decisions and in avoiding unnecessary interventions for patients with LM. Studies focused on external validation of our models are necessary in the future.