Machine Learning in the ICU: Predicting Mortality in Patients with Carbapenem-Resistant Gram-Negative Bacilli Bloodstream Infections

Comprehensive Summary

The rise of antibiotic-resistant bacteria has become a more pressing issue in recent years, as more “superbug” infections occur that are now resistant to last line carbapenem antibiotics leading to high mortality rates. These bacteria include Pseudomonas aeruginosa, Klebsiella pneumoniae, and Acinetobacter spp. The aim of this retrospective study was to develop a machine learning tool to help predict mortality in patients with these diseases to better prepare the patient and their family for treatment options or end-of-life discussions. For this study, survival was determined as alive at the 15-day mark. 10 different machine learning artificial intelligence tools were used, and their respective outcomes were measured on accuracy, precision, sensitivity, F1 score, Brier score, and Mattews correlation coefficient. A total of 197 were included in this study, with the average age of 65 and the average length of hospital stay of 42 days. The best learning model with LightGBM. Overall, machine learning tools can help assist in the best progression of treatment given a patient’s specific vitals and test results, however, although it showed promising potential, there is still a lot of work to be done to be able to accurately predict mortality rates in patients with severe antibiotic-resistant bacteria.

Outcomes and Implications

With the help of machine learning tools, patient treatment plans, specifically patients in the ICU, could eventually be maximized for effectiveness and safety through the use of antibiotics, doses, and frequencies, and choose what is most appropriate given the specific health and trajectory of the patient. This has the potential to save many lives through the benefit of switching antibiotics sooner and at the best interval and help prepare families and patients for end-of-life care for severe, deadly antibiotic-resistant bacteria. The machine learning models accurately predicted the deadliest outcomes by identifying specific markers such as coagulopathy, septic shock, SOFA score, platelet count, and CRP levels, so this knowledge could provide physicians and families with help in making tough clinical decisions about likely outcomes in these subgroups of patients. It can also halt the progress of antibiotic resistance by limiting the overprescription of antibiotics. Stopping ineffective antibiotics earlier can lead to less resistance for other strains and improve survival outcomes for those infected with both antibiotic resistant and antibiotic sensitive bacteria.