ESC Heart Failure (Dec 2020)

Multi‐modality machine learning approach for risk stratification in heart failure with left ventricular ejection fraction ≤ 45%

  • Gary Tse,
  • Jiandong Zhou,
  • Samuel Won Dong Woo,
  • Ching Ho Ko,
  • Rachel Wing Chuen Lai,
  • Tong Liu,
  • Yingzhi Liu,
  • Keith Sai Kit Leung,
  • Andrew Li,
  • Sharen Lee,
  • Ka Hou Christien Li,
  • Ishan Lakhani,
  • Qingpeng Zhang

DOI
https://doi.org/10.1002/ehf2.12929
Journal volume & issue
Vol. 7, no. 6
pp. 3716 – 3725

Abstract

Read online

Abstract Aims Heart failure (HF) involves complex remodelling leading to electrical and mechanical dysfunction. We hypothesized that machine learning approaches incorporating data obtained from different investigative modalities including atrial and ventricular measurements from electrocardiography and echocardiography, blood inflammatory marker [neutrophil‐to‐lymphocyte ratio (NLR)], and prognostic nutritional index (PNI) will improve risk stratification for adverse outcomes in HF compared to logistic regression. Methods and results Consecutive Chinese patients referred to our centre for transthoracic echocardiography and subsequently diagnosed with HF, between 1 January 2010 and 31 December 2016, were included in this study. Two machine learning techniques, multilayer perceptron and multi‐task learning, were compared with logistic regression for their ability to predict incident atrial fibrillation (AF), transient ischaemic attack (TIA)/stroke, and all‐cause mortality. This study included 312 HF patients [mean age: 64 (55–73) years, 75% male]. There were 76 cases of new‐onset AF, 62 cases of incident TIA/stroke, and 117 deaths during follow‐up. Univariate analysis revealed that age, left atrial reservoir strain (LARS) and contractile strain (LACS) were significant predictors of new‐onset AF. Age and smoking predicted incident stroke. Age, hypertension, type 2 diabetes mellitus, chronic kidney disease, mitral or aortic regurgitation, P‐wave terminal force in V1, the presence of partial inter‐atrial block, left atrial diameter, ejection fraction, global longitudinal strain, serum creatinine and albumin, high NLR, low PNI, and LARS and LACS predicted all‐cause mortality. Machine learning techniques achieved better prediction performance than logistic regression. Conclusions Multi‐modality assessment is important for risk stratification in HF. A machine learning approach provides additional value for improving outcome prediction.

Keywords