DETECTION OF VENTILATOR-ASSOCIATED EVENTS AMONG ADULTS USING NAIVE BAYES CLASSIFER

Abstract

TOPIC: Education, Research, and Quality Improvement TYPE: Original Investigations PURPOSE: Patients on mechanical ventilation are at risk for ventilator-associated pneumonia (VAP). As a potential patient safety measure, the CDC established surveillance definitions known as ventilator-associated events (VAEs). Under the VAE continuum, mechanically ventilated patients can progress from a ventilator-associated condition (VAC), to an infection-related ventilator-associated complication (IVAC), and finally to possible ventilator-associated pneumonia (PVAP). The aim of this study was to develop a screening tool to detect mechanically-ventilated patients at risk for VAEs. METHODS: Using standard procedures to construct a classifier, the dataset was split 7:3 to create samples for training and testing. Considering there are nested classifications in the VAE continuum, we assured VACs, IVACs, and PVAPs, were properly represented as evenly as possible in each sample. From the training set, the classifier calculates the conditional probabilities for each factor in the model. After the proposed model is trained, it is validated against the testing data sample. To appraise the model’s performance, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), area under the curve (AUC), and Matthew’s Correlation coefficient were calculated. All analyses were performed using R 3.6.1 (R Foundation for Statistical Computing, Vienna, Austria) and an alpha level of 0.05 was considered statistically significant. RESULTS: We reported 186 VAE cases (99 VACs, 47 IVACs, and 40 PVAPs) and matched 1:1 with non-VAE cases. After splitting the dataset, there were 260 patients in the training sample and 112 patients in the testing sample. The naïve Bayes classifier demonstrated an overall accuracy result of 83.0% (95% CI; 74.8% – 89.5%) The model was able to correctly classify 80.4% of VAEs (sensitivity) and 85.7% of non-VAEs (specificity). The positive predictive value and the negative predictive value was 81.4% and 84.9%, respectively. The AUC was 0.831 (95% CI; 0.751 – 0.911) and the Matthew’s Correlation coefficient was 0.662. CONCLUSIONS: We present a classification model that potentially could be used to screen our patients at the time of intubation. As this was a single-center retrospective study, the results may be difficult to generalize to other hospitals. We would anticipate that variables that have been implicated in previous studies, such as CHF, to be a ubiquitous factor. Additionally, there may be different factors that contribute to the classification of mechanically ventilated patients at risk for VAEs that are institution-unique. For instance, we chose to include a specific time-dependent variable, as our median time-to-event is 5 ± 3 ventilator days. Likewise, we found chronic kidney disease covaried with CHF, while positive smoking history covaried with COPD. While the addition of both predictors did not enhance the quality of the model, both were subsequently removed. CLINICAL IMPLICATIONS: The model can provide insight on how to anticipate VAP and VAE. This model can ultimately prevent adverse outcomes related to infections and ventilated associated complications. The model needs to be verified and implemented in a prospective setting to see effects on patient outcomes. DISCLOSURES: No relevant relationships by Ragheb Assaly, source=Web Response No relevant relationships by William Barnett, source=Web Response No relevant relationships by Nithin Kesireddy, source=Web Response No relevant relationships by Waleed Khokher, source=Web Response No relevant relationships by Fadi Safi, source=Web Response