A NOVEL PREDICTION MODEL FOR RESPIRATORY FAILURE AFTER CARDIAC REVASCULARIZATION SURGERIES

Abstract

SESSION TITLE: Wednesday Abstract Posters SESSION TYPE: Original Investigation Posters PRESENTED ON: 10/23/2019 09:45 AM - 10:45 AM PURPOSE: Machine learning techniques are used in application fraud protection for over 30 years. It has been used for medical outcomes for 10 years for prediction of clinical medical outcomes in clinical outcomes over the last decade. Researchers have developed algorithms for detection of clinical outcome but the clinical significant is uncertain. Respiratory Failure (RF) after revascularization surgery is associated with significant morbidity/mortality, increased length of stay, and readmissions. Few reports have evaluated the risk factors for RF, and no prediction model is available. We used data mining and machine learning techniques to stratify the risk factors and build a model to predict RF using NIS database, and test the predictive ability of the model. METHODS: Data was obtained from National Inpatient Sample (NIS), the first three quarters of 2015 were considered to establish the learning model. We used the R programming language and SQLite to extract data for all patients that underwent cardiac revascularization surgeries; we identified 27103 patients, 5035 of which developed RF in their hospital stay. We identified the most common diagnoses that were recorded in the database before RF using routine algorithmic techniques. These diagnoses were stratified and using Support Vector Machines and Random Forests, a predictive model was built to predict RF. Given that the data was unbalanced (∼20% develop RF), Synthetic Minority Over-sampling was used to increase the sensitivity of the predictive model. RESULTS: Our model identified more than 60 diagnoses that had statistically significant correlation with RF. The only type of revascularization that had significant impact on RF was single IMA bypass, which was protective. We chose 20 most impactful (highest odd ratio), among which were number of chronic disease, kidney failure, sepsis, elective vs. emergent, Hypokalemia, and obesity, to build our learning model using the techniques described. We tested our first run of the model had accuracy of 90%, sensitivity of 64% and specificity of 99%. Given that we felt in this case, sensitivity is the most important factor, we identified another model using minority oversampling, with sensitivity of 95%, but it came at the cost of decreasing specificity to 92%. CONCLUSIONS: The novelty of this technique is using machine learning to build a predictive model for RF in patients undergoing revascularization surgery. The advantage of this learning method is that the predictive accuracy increases with the size of training sample. We believe that if more years are incorporated in our learning model, the accuracy, sensitivity, and specificity will increase. Furthermore, with more granular identifiers (e.g. potassium level instead of hypokalemia), the predictive ability will improve. CLINICAL IMPLICATIONS: Such techniques can be used in quality patient care and peer review outcomes measures for coding and reimbursement. DISCLOSURES: No relevant relationships by Siavash Bolourani, source=Web Response No relevant relationships by Frank Manetta, source=Web Response No relevant relationships by Alexandra Renzi, source=Web Response No relevant relationships by Ping Wang, source=Admin input