Risk score for predicting mortality including urine lipoarabinomannan detection in hospital inpatients with HIV-associated tuberculosis in sub-Saharan Africa: Derivation and external validation cohort study

Abstract

Background: The prevalence of and mortality from HIV-associated tuberculosis (HIV/TB) in hospital inpatients in Africa remains unacceptably high. Currently, there is a lack of tools to identify those at high risk of early mortality who may benefit from adjunctive interventions. We therefore aimed to develop and validate a simple clinical risk score to predict mortality in high-burden, low-resource settings.

Methods and findings: A cohort of HIV-positive adults with laboratory-confirmed TB from the STAMP TB screening trial (Malawi and South Africa) was used to derive a clinical risk score using multivariable predictive modelling, considering factors at hospital admission (including urine lipoarabinomannan [LAM] detection) thought to be associated with 2-month mortality. Performance was evaluated internally and then externally validated using independent cohorts from 2 other studies (LAM-RCT and a Médecins Sans Frontières [MSF] cohort) from South Africa, Zambia, Zimbabwe, Tanzania, and Kenya. The derivation cohort included 315 patients enrolled from October 2015 and September 2017. Their median age was 36 years (IQR 30-43), 45.4% were female, median CD4 cell count at admission was 76 cells/μl (IQR 23-206), and 80.2% (210/262) of those who knew they were HIV-positive at hospital admission were taking antiretroviral therapy (ART). Two-month mortality was 30% (94/315), and mortality was associated with the following factors included in the score: age 55 years or older, male sex, being ART experienced, having severe anaemia (haemoglobin < 80 g/l), being unable to walk unaided, and having a positive urinary Determine TB LAM Ag test (Alere). The score identified patients with a 46.4% (95% CI 37.8%-55.2%) mortality risk in the high-risk group compared to 12.5% (95% CI 5.7%-25.4%) in the low-risk group (p < 0.001). The odds ratio (OR) for mortality was 6.1 (95% CI 2.4-15.2) in high-risk patients compared to low-risk patients (p < 0.001). Discrimination (c-statistic 0.70, 95% CI 0.63-0.76) and calibration (Hosmer-Lemeshow statistic, p = 0.78) were good in the derivation cohort, and similar in the external validation cohort (complete cases n = 372, c-statistic 0.68 [95% CI 0.61-0.74]). The validation cohort included 644 patients between January 2013 and August 2015. Median age was 36 years, 48.9% were female, and median CD4 count at admission was 61 (IQR 21-145). OR for mortality was 5.3 (95% CI 2.2-9.5) for high compared to low-risk patients (complete cases n = 372, p < 0.001). The score also predicted patients at higher risk of death both pre- and post-discharge. A simplified score (any 3 or more of the predictors) performed equally well. The main limitations of the scores were their imperfect accuracy, the need for access to urine LAM testing, modest study size, and not measuring all potential predictors of mortality (e.g., tuberculosis drug resistance).

Conclusions: This risk score is capable of identifying patients who could benefit from enhanced clinical care, follow-up, and/or adjunctive interventions, although further prospective validation studies are necessary. Given the scale of HIV/TB morbidity and mortality in African hospitals, better prognostic tools along with interventions could contribute towards global targets to reduce tuberculosis mortality.

Fig 1. Study profile.

LAM, lipoarabinomannan; MSF, Médecins Sans Frontières; TB, tuberculosis.

Fig 2. Risk score calculation to predict mortality.

TB, tuberculosis; TB-LAM, Determine TB LAM Ag.

Fig 3. Distribution of risk scores and mortality in the derivation dataset.

Distribution of risk scores for mortality stratified by outcome at 2 months (stacked bar chart) and mortality risk (percent, shown by blue line) for (A) the full risk score (based on the regression coefficients) and (B) the simplified risk score. Mortality risks and absolute numbers in each category are presented in S2 Table.

Fig 4. Observed mortality risk by risk score category in the derivation and validation cohorts.

Observed mortality risk (A) at 56 days, (B) during inpatient stay, and (C) post-discharge in the derivation and validation cohorts, stratified by risk score category (derivation and validation cohorts) and simplified risk score category (derivation cohort). Numbers on bars represent absolute mortality risk; error bars represent 95% confidence intervals. For the full risk score, low risk was defined as 10 points or fewer, medium risk as 11 to 20 points, and high risk as more than 20 points. For the simplified risk score, the low-risk group had a predictor score of 0 or 1 point, the medium-risk group had a predictor score of 2 points, and the high-risk group had predictor score of ≥3 points. p-Values based on chi-squared tests between groups for derivation and validation cohorts, respectively, are (A) p < 0.001 and p < 0.001, (B) p = 0.001 and p = 0.001, and (C) p = 0.015 and p = 0.003.

Fig 5. Survival curves stratified by clinical risk score category in the derivation dataset.

Survival curves and risk tables with number at risk for the (A) derivation cohort and (B) validation cohort stratified by risk group using the simplified clinical risk score. The low-risk group (blue line) had a risk score of 0 or 1 point, the medium-risk group (red line) had a risk score of 2 points, and the high-risk group (green line) had a risk score of ≥3 points. Log-rank test p < 0.001.