Two machine learning models using clinical and echocardiographic variables demonstrated high negative predictive value for identifying low-risk acute heart failure patients suitable for ward admission.
Observational (n=908)
No
Do machine learning models using clinical data predict adverse outcomes (death or ICU transfer) in patients presenting to the ED with acute heart failure?
Machine learning models using routine clinical and echocardiographic data can identify low-risk acute heart failure patients in the ED with high negative predictive value, aiding in safe triage to the general ward.
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.
Shchekochikhin et al. (Thu,) conducted a observational in Acute heart failure (n=908). Machine learning risk stratification models was evaluated on Death from any cause during hospitalization or transfer to an intensive care unit (ICU) for using inotropes/vasopressors. Two machine learning models using clinical and echocardiographic variables demonstrated high negative predictive value for identifying low-risk acute heart failure patients suitable for ward admission.