Prognostic Markers of Adverse Outcomes in Acute Heart Failure: Use of Machine Learning and Network Analysis with Real Clinical Data

Dmitri Shchekochikhin, Kristina Charaya, Alexandra Shilova, Alexey Nesterov, Ekaterina Pershina, Andrei Sherashov, Sergei Panov, Shevket Ibraimov, Alexandra Bogdanova, Alexander Suvorov, Olga Trushina, Zarema Bguasheva, Nina Rozina, Alesya Klimenko, Varvara Mareyeva, Natalia Voinova, Alexandra Dukhnovskaya, Svetlana Konchina, Eva Zakaryan, Philipp Kopylov, Abram Syrkin, Denis Andreev

Background: Acute heart failure (AHF) is one of the leading causes of admissions to the emergency department (ED). There is a need to develop an easy-to-use score that can be used in the ED to risk-stratify patients with AHF and in hospitalization decisions regarding cardiac wards or intensive care units (ICUs). Methods: A retrospective observational study was conducted at a city hospital. The data from the presentation of AHF patients at the ED were collected. The combined primary endpoint included death from any cause during hospitalization or transfer to an intensive care unit (ICU) for using inotropes/vasopressors. Feature selection was performed using artificial intelligence. Results: From August 2020 to August 2021, 908 patients were enrolled (mean age: 71.6 ± 13 years; 500 (55.1%) men). We found significant predictors of in-hospital mortality and ICU transfers for inotrope/vasopressor use and built two models to assess the need for ICU admission of patients from the ED. The first model included SpO2 < 90%, QTc duration, prior diabetes mellitus and HF diagnosis, serum chloride concentration, respiratory rate and atrial fibrillation on admission, blood urea nitrogen (BUN) levels, and any implanted devices. The second model included left ventricular end-diastolic size, systolic blood pressure, pulse blood pressure, BUN levels, right atrium size, serum chloride, sodium and uric acid concentrations, prior loop diuretic use, and pulmonary artery systolic blood pressure. Conclusions: We developed two models that demonstrated a high negative predictive value, which allowed us to distinguish patients with low risk and determine patients who can be hospitalized and sent from the ED to the floor. These easy-to-use models can be used at the ED.