Development and Temporal Validation of a Machine Learning Based Prognostic Model for Predicting Outcome of Patients with Tuberculous Meningitis

Abstract

Mortality prediction in adult tuberculous meningitis has a strong clinical rationale and a modest but coherent modeling base. The best established prognostic work is not machine learning. It comes from regression based models built on large prospective cohorts. The dynamic landmark model using time updated Glasgow Coma Scale and plasma sodium. Across cohorts, the most reproducible mortality signals are older age, impaired consciousness or low Glasgow Coma Scale, advanced disease stage, focal neurological deficit, hydrocephalus, HIV coinfection, low CD4 count in HIV positive disease, and sodium derangement, with stronger evidence for serial sodium values than for a single baseline value. ICU studies add mechanical ventilation and organ failure scores, yet these variables represent a limited scope of care and are probably not useful when integrated into a general TBM model.

There is a genuine absence of TBM-specific machine learning. The most recent studies are focused on a single geography, have a small sample size, and most of them go for predicting composite or ordinal outcomes rather than mortality. The best cases have used a combination of MRI and clinical data and have demonstrated that imaging can provide additional information, but they are restricted to relatively small patient cohorts, rely solely on internal validation, and do not provide a reliable benchmark for predictive mortality models. For a new study that is based on an original dataset, the most justifiable contribution is probably a mortality model that is well-constructed on standardized TBM definitions, along with internal and external validation that are disciplined, plus a limited approach to missing data. Set any machine learning approach to the regression standards that are already noted in the field, rather than using the established ones.