Antibiotic resistance of bacterial infections is a major public health threat. When a patient is presented with a bacterial infection, there is a built-in uncertainty in a doctor’s decision of prescribing antibiotics before laboratory confirmed results are available. A key piece of information is missing in this decision-making process - a quantitative estimate of the causal effect of antibiotics on future antibiotic resistance of bacterial infections. We apply advanced causal inference and machine learning methods to a highly detailed dataset of electronic medical records of over 40,000 Israeli hospitalized patients with bacterial infections.
Antibiotic resistance frequencies are dynamic and constantly changing due to the rapid evolutionary response of bacteria to environmental and individual-level changes. Our first step towards understanding this time-varying system is to identify predictors with constant effect on antibiotic resistance across time, despite the dynamic nature of the system. These predictors are themselves causal, or are surrogates of stable causal relationships.
Time-varying associations of the predictors along with 95% confidence intervals. Local estimation of time-varying coefficient models is used to identify predictors with time-fixed effects. The left panel shows a constant association (up to noise) while the right panel shows substantial changes in the association through time.