An Interpretable Machine Learning Framework for ICU Mortality Risk Prediction in COVID-19 Patients

View at publisher Summarise with VIRAT
Authors: Rutuja Ingole, Disha Moharle, Vrushali Mahalle, Achal Bangade, Rajendra M. Rewatkar
Cited by: 0
References: 10

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

It is critically important to have dependable tools available to help identify inpatients during admission for those patients at highest risk for in-hospital death due to COVID19. The COVID19 Pandemic has created an unprecedented amount of demand on intensive care units (ICUs) around the world; consequently, it is vital to initiate clinical intervention and allocate limited intensive care unit (ICU) resources as quickly as possible by early risk stratification among patients, particularly in an intensive care unit (ICU) setting. This article combines routinely-collected clinical data routinely available at the time of admission to the intensive care unit (ICU) in order to provide a predictive analytic framework for evaluating the mortality of patients who have been admitted into an intensive care unit (ICU) due to COVID19 infection. The final data set consisted of 1,250 patients admitted into the intensive care unit (ICU) from a tertiary care medical center, of these patients, approximately 28% had died as a direct result of their COVID infection. A variety of supervised learning techniques were tested including, Logistic Regression, Gradient Boosting Machines, Extra Trees, XGboost, and Random Forests; Frequently loaded datasets from Methodological Evaluation were used as the source for developing and validating the supervised machine learning models listed above. The Random Forest Model selected as the Optimal Classifier Model had excellent classification capabilities and inherent interpretable features. The Random Forest classifier model had an ROC-AUC of 0.866, an F1 of 0.566, and a Recall of 0.439 when evaluated on a holdout Test Dataset; Features considered important for predicting clinical outcomes consistent with adapting clinical management based on current clinical knowledge were: Age, Oxygen Saturation Levels, Inflammatory Markers, and Comorbidities. In light of the fact that moderate recall highlights the potential for false negative results, the framework should be viewed as a means to aid clinical judgement rather than replace it. Overall, the proposed strategy provides an open, evidence-based decision support system for use in ICU triage and resource allocation during pandemics.

Fields of Study & Subjects

Machine Learning in Healthcare COVID-19 Clinical Research Studies Artificial Intelligence in Healthcare

Publication details

Journal / source
International Conference on Image Processing and Capsule Networks (ICIPCN)
Pages
1803–1808
Publication date
27 January 2026
Language
EN