Abstract 1106 Development and Validation of a Machine Learning Model to Aid Discharge Processes for Intensive Care Patients in the Cardiac Intensive Care Unit - A Novel Artificial Intelligence-Based Algorithm

Clin Res Cardiol (2023). https://doi.org/10.1007/s00392-023-02180-w
Development and Validation of a Machine Learning Model to Aid Discharge Processes for Intensive Care Patients in the Cardiac Intensive Care Unit - A Novel Artificial Intelligence-Based Algorithm
T. Thevathasan¹, F.-J. Krause¹, F. Balzer², U. Landmesser¹, C. Skurk¹
¹CC 11: Med. Klinik für Kardiologie, Charité - Universitätsmedizin Berlin, Berlin; ²Medizinische Informatik, Charité - Universitätsmedizin Berlin, Berlin;
Background: Particularly during the COVID-19 pandemic, the need for cardiac intensive care unit (ICU) beds remains high in order to monitor and treat emergency patients with severe cardiovascular diseases, such as acute coronary syndrome, cardiogenic shock or cardiac arrest. Therefore, timely discharge policies from the cardiac ICU to normal wards are crucial to meet the continuously increasing demand for cardiac ICU beds. ICU transfer to the normal ward is considered a high risk event which may results in worsening of the patient’s clinical condition with readmission to the ICU. Convolutional neural networks (CNN) may provide a timely accurate discharge decision support system for intensivists based on comprehensive, multi-dimensional ICU data. To the best of our knowledge, this is the first study to develop and validate a CNN on ICU readmission for patients in the cardiac ICU in the particularly strenuous environment of a cardiac arrest center. Methods: The study was carried out within a consecutively enrolled cohort of adult patients who were admitted to the cardiac ICU of a tertiary care center due to a primary cardiac or respiratory diagnosis between 2003 and 2021. CNN was utilized to predict ICU readmission within 72 hours after ICU discharge. The study cohort was randomly split into a training and validation cohort. The latter was used to test the trained CNN. Based on Delphi method and literature review, 127 variables were selected for inclusion in the CNN, including demographic data, ICU procedures, ICU medication, vital signs, admitting ward and main diagnoses. With an exploratory intent, the CNN was compared to a logistic regression model and the Stability and Workload Index for Transfer (SWIFT), an ICU readmission score. Results: 30,940 patients were included in this study. 31.7% patients were female and the average age was 66.8 (±13.4) years. Predictors of ICU readmission within 72 hours after ICU discharge were higher patient age, higher comorbidity level, use of Impella® micro-axial pump, endotracheal intubation, blood transfusions, renal replacement therapy and vasopressor use during the index ICU stay. The CNN had a good predicting performance: sensitivity 0.98, specificity 0.87, precision 0.99 and accuracy 0.97. CNN showed a better diagnostic accuracy compared to the logistic regression model and SWIFT: AUC 90.8% vs. 71.3% vs. 71.1%, respectively (see Figure). Conclusion: Artificial intelligence may be utilized to support timely and efficient ICU discharge strategies in order to facilitate patient safety, reduce ICU workload and shorten hospital length of stay.
https://dgk.org/kongress_programme/jt2023/aV2027.html