Personalized azithromycin treatment rules for children with watery diarrhea using machine learning

Abstract

We use machine learning to identify innovative strategies to target azithromycin to the children with watery diarrhea who are most likely to benefit. Using data from a randomized trial of azithromycin for watery diarrhea (NCT03130114), we develop personalized treatment rules given sets of diagnostic, child, and clinical characteristics, employing a robust ensemble machine learning-based procedure. This procedure estimates the child-level expected benefit for a given set of covariates by combining predictions from a library of statistical models. For each rule, we estimate the proportion treated under the rule and the average benefits of treatment. Among 6692 children, treatment under the most comprehensive rule is recommended on average for one third of children. The risk of diarrhea on day 3 is 10.1% lower (95% CI: 5.4, 14.9) with azithromycin compared to placebo among children recommended for treatment (NNT: 10). For day 90 re-hospitalization and death, risk is 2.4% lower (95% CI: 0.6, 4.1; NNT: 42). While pathogen diagnostics are strong determinants of azithromycin effects on diarrhea duration, host characteristics may better predict benefits for re-hospitalization or death. This suggests that targeting antibiotic treatment for severe outcomes among children with watery diarrhea may be possible without access to pathogen diagnostics.

Fig. 1. Comparison of average treatment benefit among those recommended for treatment and not recommended for treatment under each rule.

A Comparison of average treatment benefit among those recommended and not recommended for treatment for diarrhea on day 3; B Comparison of average treatment benefit among those recommended and not recommended for treatment for re-hospitalization or death by day 90; C Comparison of average treatment benefit among those recommended and not recommended for treatment for change in linear growth. In all panels, the proportion recommended for treatment is denoted in parentheses after the label for each rule. Error bars represent a 95% confidence interval, and the sample size for each treatment benefit estimation is determined by the proportion of n = 6692 treated under each rule. CI Confidence interval, LAZ Length-for-age Z-score. Source data are provided as a Source Data File on Github.

Fig. 2. Concordant correlation coefficient between child-level expected benefits from the comprehensive and alternative treatment rules.

A CCC of child-level expected benefits for diarrhea on day 3; B CCC of child-level expected benefits for re-hospitalization or death by day 90; C CCC of child-level expected benefits for change in linear growth. In all panels, the 95% confidence interval for the CCC is provided with sample size n = 6692. The dotted grey line represents perfect correlation (CCC = 1). CCC Concordant correlation coefficient, CI Confidence interval, LAZ Length-for-age Z-score. Source data are provided as a Source Data File on Github.

Fig. 3. Proportion treated and average benefit among those recommended for azithromycin under the comprehensive rule with varying thresholds of clinical benefit to define the rule.

A Proportion treated and average benefit across thresholds for diarrhea on day 3; B Proportion treated and average benefit across thresholds for re-hospitalization or death by day 90; C Proportion treated and average benefit across thresholds for change in linear growth. The 95% confidence intervals are represented by the grey dotted lines with shaded bands, and the sample size for each treatment benefit estimation is determined by the proportion of n = 6692 treated under each threshold. RD Risk difference, CI Confidence interval, LAZ Length-for-age Z-score. Source data are provided as a Source Data File on Github.