Cerebrospinal Fluid Chloride as an Independent Prognostic Biomarker in Cryptococcal Meningitis: A Dual-Center Externally Validated Machine Learning Study with SHAP-Based Interpretability

<title>Abstract</title> <p>Background Cryptococcal meningitis (CM) carries a high 10-week mortality rate, necessitating accurate early prognostic tools. While cerebrospinal fluid chloride (CSF-Cl) is a routine test, its independent prognostic value in CM remains ill-defined. Furthermore, existing machine learning models lack external validation and clinical interpretability. This study aimed to validate CSF-Cl as an independent biomarker and to develop an externally validated, interpretable prognostic model for CM. Methods We enrolled 253 treatment-naive CM patients from two tertiary hospitals. Univariate and multivariate logistic regression identified independent prognostic risk factors across the full cohort. Patients were stratified into a training cohort (The Fourth People’s Hospital of Nanning, n = 194) and an independent external test cohort (Jiangxi Provincial Chest Hospital, n = 59). In the training set, nine machine learning classifiers were developed following feature selection via LASSO-recursive feature elimination (LASSO-RFE) and sample balancing via synthetic minority oversampling technique (SMOTE). Paired models with and without CSF-Cl were compared to quantify the incremental value of CSF-Cl using ΔAUC, net reclassification index (NRI), integrated discrimination improvement (IDI), and decision curve analysis (DCA). SHapley Additive exPlanations (SHAP) was deployed to dissect dose-response relationships and feature interactions, with subgroup analyses to define risk thresholds. Results Multivariate regression confirmed CSF-Cl as an independent risk factor for unfavorable CM outcomes (OR = 1.059, 95% CI 1.011–1.110, P = 0.017). CatBoost demonstrated the best generalizability among the nine models evaluated, yielding an AUC of 0.712 and precision of 0.722 in the external test set. Incorporation of CSF-Cl increased the external AUC from 0.651 to 0.712 (ΔAUC = 0.061), with an NRI of 0.132 and IDI of 0.040. SHAP analysis uncovered a positive dose-response association between CSF-Cl and adverse prognosis, alongside synergistic interactions with intracranial pressure (ICP) and HIV status. The model retained robust performance in HIV-positive patients and those aged &lt; 50 years (AUC = 0.752 and 0.759, respectively). A CSF-Cl threshold of ≥ 115 mmol/L for risk stratification was derived, and an online interpretable prediction tool was implemented. Conclusions Elevated baseline CSF-Cl is independently linked to 10-week unfavorable prognosis in CM and substantially augments the predictive power of conventional clinical markers. Our dual-center externally validated CatBoost model with SHAP-enabled interpretability supports early, precise risk stratification for CM patients.</p>