Services on Demand
Journal
Article
Indicators
Cited by SciELO
Related links
Similars in SciELO
Share
Revista argentina de cardiología
On-line version ISSN 1850-3748
Abstract
GAMBARTE, MARIA JIMENA et al. Prognostic Comparison between Risk Scores and Neural Networks to Predict Short- and Mid-Term Mortality in Acute Heart Failure. Rev. argent. cardiol. [online]. 2021, vol.89, n.5, pp.435-446. Epub Oct 01, 2021. ISSN 1850-3748.
Background:
Heart failure (HF) risk scores to assess all-cause mortality during the first year have areas under the ROC curve (AUC) ranging between 0.59 and 0.80
Objective:
To develop and validate a neural network (NN) algorithm-based model to improve traditional scores’ performance for predicting short- and mid-term mortality of patients with acute HF.
Methods:
A prospective clinical database was analyzed including 483 patients admitted with diagnosis of acute HF in a coronary care unit community hospital of Buenos Aires, between June 2005 and June 2019. Among 181 demographic, laboratory, treatment and follow-up variables, only 25 were selected to calculate five acute heart failure risk scores aimed to predict 30-day, 6-month and 1-year mortality: EFFECT, ADHERE, GWTG-HF, 3C-HF, and ACUTE-HF.
Results:
Mean age was 78 ± 11.1 years, 58% were men, 35% had ischemic necrotic HF and median left ventricular ejection fraction was 52% (35-60). At 30 days, the EFFECT score (AUC:0.68) and the 3C-HF score (AUC: 0.68) showed better performance than the ACUTE-HF score (AUC: 0.54). At 6-month and 1-year follow-up, the EFFECT score (ROC: 0.69 and 0.69) outperformed the ADHERE score (AUC: 0.53 and 0.56), and EFFECT (AUC: 0.69 and 0.69), GWRG-HF (AUC = 0.68 and 0.66), and 3C-HF (AUC:0.67 and 0.67) scores outperformed the ACUTE-HF score (AUC:0.53 and 0.56). The best results with NN algorithms were obtained with a two-hidden layer multilayer perceptron. A 24-9-7-2-layer architecture NN was used with the following results: AUC: 0.82, negative predictive value (NPV) 93.2% and positive predictive value (PPV) 66.7% for 30-day mortality; AUC: 0.87, NPV: 89.1% and PPV: 78,6% for 6-month mortality; and AUC: 0.85, NPV: 85.6% and PPV: 78.9% for 1-year mortality. In terms of discrimination, NN algorithms outperformed all the traditional scores (p <0.001). For this algorithm, the most influential factors in descending order that scored ≥50% normalized importance to predict 30-day mortality were serum creatinine, hemoglobin, respiratory rate, blood urea nitrogen, serum sodium, age and systolic blood pressure. Also, NYHA functional class III-IV and dementia added prognostic capacity to 6-month mortality, and heart rate and chronic kidney disease to 1-year mortality.
Conclusions:
The models with NN algorithms were significantly superior to traditional risk scores in our population of patients with HF. These findings constitute a working hypothesis to be validated with a larger and multicenter sample of cases.
Keywords : Heart failure; Prognosis; Mortality; Risk score; Deep learning; Artificial intelligence.