Infectious Disease
Generalizability of machine-learned antibiotic treatment rules for children with watery diarrhea in low-resource settings Allison Olmsted* Allison Olmsted Olmsted Olmsted Olmsted Olmsted Olmsted Olmsted Department of Epidemiology, Emory University
Some children with watery diarrhea could benefit from antibiotic therapy not indicated by current guidelines. Data from the AntiBiotics for Children with severe Diarrhea (ABCD) randomized trial was used to develop personalized rules based on expected child-level benefits of azithromycin treatment with respect to diarrhea duration, re-hospitalization or death, and linear growth. To evaluate generalizability to a more representative population, we applied ABCD-learned rules for antibiotic treatment to the Enterics for Global Health (EFGH) study, a seven-site diarrhea surveillance study, and compared these to novel rules developed using EFGH data. Rules were developed using a robust ensemble machine learning-based procedure validated across 10 folds. Treatment was recommended if the child-level benefit exceeded a 7% reduction in day 3 diarrhea, 2% reduction in 90-day hospitalization or death, and 0.06 increase in length-for-age z-score (LAZ). We estimated the proportion recommended treatment by each rule and compared expected benefit by treatment recommendation. Overall treatment effects were 2.4% reduction (95% CI 0.2, 4.6) in day 3 diarrhea, 0.3% increase (95% CI -0.6, 1.2) in 90-day hospitalization or death, and a 0.01 increase (95%CI 0.0, 0.03) in 90-day LAZ. Comparing the ABCD to EFGH rule, the difference in the antibiotic effect among those recommended for treatment compared to not in terms of reduction in day 3 diarrhea was 4.3% (95%CI 1.0, 9.7) and 0.04% (95%CI -9.5, 8.6), respectively. The ABCD rule more often recommended treatment for children with bacterial etiology (59.7% vs. 40.3%) and acute malnutrition (median weight-for-age z-score -1.42 vs. -1.07). Neither rule identified heterogeneity in treatment effects for other outcomes. The ABCD rule successfully identified children in EFGH who would benefit from antibiotic treatment and better-identified heterogeneity in treatment effects against diarrhea duration compared to the EFGH rule.
