Categorical Versus Continuous Encoding of Routine CSF Variables for Diagnosing Tuberculous Meningitis: Development in Iran and External Validation in Peru

Abstract Background Diagnostic scores for tuberculous meningitis (TBM) based on routine cerebrospinal fluid (CSF) measurements often perform inconsistently across settings. We assessed whether representing routine CSF measurements as fixed, clinically motivated categories, rather than continuous values, may improve the transportability of a TBM diagnostic score across countries. Methods From four routine CSF variables (white cell count, protein, protein to glucose ratio, lymphocyte percentage), we developed a continuous logistic model and an integer score with fixed cut points in an Iranian cohort (69 TBM, 70 controls). Both were locked and evaluated without refitting in a Peruvian cohort (40 TBM; 62 bacterial, cryptococcal, or viral meningitis; mononuclear cell percentage replacing lymphocyte percentage), using harmonized multiple imputation; the study was reported according to TRIPOD + AI. Results In Iranian cross validation, discrimination was similar (AUC 0.822, 95% CI 0.744 to 0.886, continuous; 0.821, categorical). In Peru, the AUC was 0.583 (95% CI 0.473 to 0.691) for the continuous and 0.699 (95% CI 0.599 to 0.794) for the categorical score (paired difference 0.116, 95% CI 0.048 to 0.182); a three-predictor continuous model did not perform better (0.553). The advantage was evident against bacterial but not cryptococcal or viral meningitis. Calibration slopes were 0.471 (95% CI 0.163 to 0.779) and 0.150, both well below 1. Neither model discriminated well between TBM and neurobrucellosis (apparent AUC 0.71 and 0.66). Both cohorts were smaller than required (385 and 581). Conclusions In one external cohort, fixed categories showed higher discrimination than continuous values, but absolute discrimination was modest, results varied across mimics, and encoding was not isolated from other differences between the models. Fixed categories may improve transportability, but these findings are preliminary and warrant confirmation in larger, prospective, multicountry cohorts.