Machine Learning for Antibiotic Stewardship in the Treatment of Stapholycoccus Bacterial Infections

Antibiotic resistance is one of the leading issues in modern healthcare due to the inability to treat common infections with available antibiotics. Many of the mechanisms of resistance have been caused by the inappropriate prescription of antibiotics to treat illnesses such as the cold or flu or the over-prescription of broad-spectrum antibiotics. Epitomizing this problem is the Staphylococcus bacteria where certain strains have become resistant to penicillin-related drugs and Vancomycin, one of the treatments for MRSA. To address this, we developed machine learning models to predict antibiotic activity and susceptibility using a patients entire available electronic health record. We selected patients who were suspected of having a staph infection from the Medical Information Mart for Intensive Care III (MIMIC-III) data set and utilized their microbiological culture results to identify the number of patients that were prescribed an inappropriate antibiotic and then propose suitable alternatives. In our test set, we identified that empiric prescriptions had an efficiency rate of 40 percent (the rate at which an antibiotic that would provide activity was prescribed), and the other 60 percent of cases were not susceptible to the prescribed antibiotic or the antibiotic that they were given was not tested for susceptibility against their infection. Our best models identified antibiotic susceptibility with AUROCs up to 0.9 and raw specificity up to 0.7. The models were also able to propose suitable alternatives in all but 10 cases. Overall these results demonstrate the need for implementing clinical decision support systems advising clinicians during the prescription process, and our further work will address this issue.