Optimizing diagnostic algorithms for prompt detection of childhood tuberculosis
Meredith Brooks, PhD, MPH
Summary
BACKGROUND: One million children fall sick with tuberculosis (TB) each year. Children are a uniquely vulnerable and under-diagnosed population because of difficulty obtaining bacteriological confirmation of TB, high risk of infection due to prolonged and intense exposure by a co-habiting adult with infectious TB, and rapid progression to disease once infected.
GAP: Due to non-specific symptoms and difficulty confirming disease in children, TB in children is often missed, overlooked, or delayed.
HYPOTHESIS: Using machine learning will inform accurate diagnostic algorithms in children to classify TB disease and sub-clinical TB infection. Diagnostic algorithms will have high discriminatory properties when applied prospectively to children being screened for TB in Lima, Peru.
METHODS: A random forest model will be generated using data from a large, prospective cohort study of TB-affected households, inclusive of almost 6,000 children who underwent TB screening and diagnostics over a year-long follow-up period in Lima, Peru (2009-2012). The validity of these diagnostic algorithms will be assessed through a prospective implementation study in which approximately 1,200 children in Lima, Peru will be screened for TB disease and sub-clinical TB infection.
RESULTS: Of 6,001 children who were household contacts of adults with TB disease (1,789 0-4, 1,468 5-9, 1,340 10-14, and 1,404 15-19 years old), 304 (5.1%) were diagnosed with TB disease during a year of follow-up (0-4: 4.5%, 5-9: 3.4%, 10-14: 5.8%, 15-19: 6.8%). Having more than one individual in the child’s household with TB disease is the most important predictor for all age groups, with specificities ranging from 93.2% to 94.9% across age groups and odds ratios of 9.2 (95% CI: 5.4-15.6), 13.5 (95% CI: 7.3-24.9), 7.9 (95% CI: 4.5-13.7), and 7.1 (95% CI: 4.3-11.6), respectively, for each increasing age group.
IMPACT: This research may improve screening and evaluation efforts in high-TB burden communities through identifying important characteristics that improve the accuracy of classifying children with TB. These diagnostic algorithms can serve as an integrative decision support system to enable programs to target specific tools for use in children at highest risk and increase efficiency in resource utilization.