AutoML-Driven Insights into Patient Outcomes and Emergency Care During Romania's First Wave of COVID-19.

Abstract

Background: The COVID-19 pandemic severely impacted healthcare systems, affecting patient outcomes and resource allocation. This study applied automated machine learning (AutoML) to analyze key health outputs, such as discharge conditions, mortality, and COVID-19 cases, with the goal of improving responses to future crises.

Methods: AutoML was used to train and validate models on an ICD-10 dataset covering the first wave of COVID-19 in Romania (January-September 2020).

Results: For discharge outcomes, Light Gradient Boosted models achieved an F1 score of 0.9644, while for mortality 0.7545 was reached. A Generalized Linear Model blender achieved an F1 score of 0.9884 for "acute or emergency" cases, and an average blender reached 0.923 for COVID-19 cases. Older age, specific hospitals, and oncology wards were less associated with improved recovery rates, while mortality was linked to abnormal lab results and cardiovascular/respiratory diseases. Patients admitted without referral, or patients in hospitals in the central region and the capital region of Romania were more likely to be acute cases. Finally, counties such as Argeş (South-Muntenia) and Brașov (Center) showed higher COVID-19 infection rates regardless of age.

Conclusions: AutoML provided valuable insights into patient outcomes, highlighting variations in care and the need for targeted health strategies for both COVID-19 and other health challenges.