Development and validation of a clinical prediction tool for the diagnosis of tuberculous meningitis

Abstract

Introduction: Tuberculous (TB) meningitis is the most lethal and disabling form of TB. A disproportionate burden of TB meningitis is in resource-limited settings. There is considerable variation in mortality and neurological sequelae reported for TB meningitis across available studies, the reasons for which remain unclear. Delayed diagnosis and treatment, which is a risk factor for poor outcomes, is caused in part by lack of availability of diagnostic tests that are both rapid and accurate. Several attempts have been made to develop clinical prediction tools to fill this gap, but none have performed sufficiently well to be broadly implemented.Purpose: We aimed to (1) ascertain heterogeneity in TB meningitis outcomes; (2) develop and validate a clinical prediction tool for diagnosing TB meningitis; and (3) externally validate this clinical prediction tool to determine the overall accuracy of classification. Methods: We conducted two systematic reviews: one to identify studies reporting TB meningitis mortality and neurological sequelae and another to identify studies that undertook diagnostic testing for TB meningitis to obtain individual participant data (IPD) from. From the first systematic review, we conducted a meta-analysis of TB meningitis mortality and neurological sequelae from studies that met the inclusion criteria. We assessed heterogeneity in mortality by conducting stratified analyses by time of reported outcome, HIV status, geographic location, and year published. From the second systematic review, we contacted the authors and attained permission to use IPD from studies that met the inclusion criteria. We harmonized the data and imputed for missing values when possible. Three multivariate prediction model (MPM) development strategies were employed to develop the clinical prediction tool for TB meningitis cases. First, an IPD meta-analysis using a logistic regression MPM with stratified intercepts for each country was fitted with key predictors. Then, we developed classification and regression tree (CART) and random forest MPMs with machine learning methods. All three MPMs were internally validated and assessed for performance using all available data in a k-fold internal-external cross-validation (IECV) approach. In our final analysis, we externally validated all three MPMs in a dataset that was not used in the development stage. Results: In our first systematic review and meta-analysis, pooled six-month mortality was 24% and showed significant heterogeneity (I2 >95%; p<0.01). Physical disability was reported in 32% (95%CI; 22-43%) of TB meningitis survivors. The heterogeneity in mortality was partly explained by HIV status and geographic location. Mortality ranged from 2% to 67% in Asian studies and from 23% to 80% in sub-Saharan African studies. Mortality was significantly worse in HIV-positive persons and in persons from studies conducted in sub-Saharan Africa. In our second systematic review, we identified and obtained IPD from 15 studies with a total of 3,671 individual participants. All three MPMs indicated cerebrospinal fluid (CSF) white blood cell (WBC) count, WBC differential, CSF glucose, CSF cryptococcal antigen, and blood glucose as significant predictors of TB meningitis. IECV revealed significant heterogeneity in performance between IPD studies, which varied based on the prevalence of HIV in the IPD study. Overall, the machine learning MPMs were not superior in performance to the logistic MPM; however, random forest performed slightly better than the logistic MPM. In external validation, the logistic MPM outperformed both CART and random forest. Discussion: Results from these studies indicate the significant contribution HIV co-infection has on outcomes and clinical prediction tool performance for TB meningitis. MPMs based on clinical and lab values more readily accessible in resource-limited settings yield well-performing clinical prediction tools. The logistic MPM had the best performance and external validity in an HIV-prevalent setting for TB meningitis. Conclusion: Heterogeneity in TB meningitis outcomes and diagnostic performance persist. HIV-status and geographic location are major contributors to variation in TB meningitis outcomes. We were successful in developing a model that can better account for this heterogeneity. The logistic MPM poses a generalizable clinical prediction tool with the potential to reduce the delay in diagnosis, and subsequent poor outcomes, in TB meningitis.