Machine learning helps predict long-term mortality and graft failure in patients undergoing heart transplant

Gen Thorac Cardiovasc Surg. 2020 Dec;68(12):1369-1376. doi: 10.1007/s11748-020-01375-6. Epub 2020 May 7.

Abstract

Objective: We aimed to develop a risk prediction model using a machine learning to predict survival and graft failure (GF) 5 years after orthotopic heart transplant (OHT).

Methods: Using the International Society of Heart and Lung Transplant (ISHLT) registry data, we analyzed 15,236 patients who underwent OHT from January 2005 to December 2009. 342 variables were extracted and used to develop a risk prediction model utilizing a gradient-boosted machine (GBM) model to predict the risk of GF and mortality 5 years after hospital discharge. After excluding variables missing at least 50% of the observations and variables with near zero variance, 87 variables were included in the GBM model. Ten fold cross-validation repeated 5 times was used to estimate the model's external performance and optimize the hyperparameters simultaneously. Area under the receiver operator characteristic curve (AUC) for the GBM model was calculated for survival and GF 5 years post-OHT.

Results: The median duration of follow-up was 5 years. The mortality and GF 5 years post-OHT were 27.3% (n = 4161) and 28.1% (n = 4276), respectively. The AUC to predict 5-year mortality and GF is 0.717 (95% CI 0.696-0.737) and 0.716 (95% CI 0.696-0.736), respectively. Length of stay, recipient and donor age, recipient and donor body mass index, and ischemic time had the highest relative influence in predicting 5-year mortality and graft failure.

Conclusion: The GBM model has a good accuracy to predict 5-year mortality and graft failure post-OHT.

Keywords: Graft failure; Heart transplant; Machine learning; Mortality.

MeSH terms

  • Heart Failure* / surgery
  • Heart Transplantation* / adverse effects
  • Humans
  • Machine Learning
  • Registries
  • Retrospective Studies