A systematic comparison of short-term and long-term mortality prediction in acute myocardial infarction using machine learning models.

Abstract

Background and objective: The machine learning (ML) models for acute myocardial infarction (AMI) are considered to have better predictive ability for mortality compared to conventional risk scoring models. However, previous ML prediction models have mostly been short-term (1 year or less) models. Here, we established ML models for long-term prediction of AMI mortality (5 years or 10 years) and systematically compare the predictive capabilities of short-term models versus long-term models across varying survival time periods.

Methods: An observational retrospective study was conducted to analyse mortality prediction in patients with varying survival times. A total of 4,173 patients were enrolled from two different hospitals in China. The dataset was allocated into three groups and an external test set based on their survival duration: the 1-year group (n = 3,626), the 5-year group (n = 2,102), the 10-year group (n = 721), and the external test set (n = 545). A comprehensive set of 53 variables was collected and utilized for model development. Mortality prediction was analysed using oversampling and feature selection methods coupled with machine learning algorithms. SHapley Additive exPlanations (SHAP) values were utilized to quantify the feature importance of AMI risk. The best-performing models from each group were selected for a systematic comparison of predictive accuracy using the external test set with follow-up exceeding 10 years but with varying survival times.

Results: For the 1-year model, the RF model achieved the best performance on the test dataset, with an F1 score of 97.81% using only oversampling without feature selection. Conversely, in the case of the 5-years, the combination of LASSO and RF yielded the best performance, achieving F1 scores of 91.35% with both feature selection and oversampling. The best model of 10-years group was SVM with only oversampling without feature selection, yielding an F1 score of 80.7%. Age, BNP, and the Killip classification of AMI were consistently identified as robust predictors across all three groups. This underscores aging as a critical AMI risk factor contributing to mortality. However, despite the model's success, when examining the actual survival times of the 545 patients, of which 64% survived beyond 5 years and 37% beyond 10 years, the 1-year model failed to distinguish between these patients, predicting all as low risk. This highlights the limitation of short-term models, indicating their inability to accurately predict actual long-term survival times despite being commonly used in AMI mortality prediction.

Conclusions: The study identifies Age, BNP, and Killip classification as consistent predictors of AMI mortality across all groups, with Age being the most significant factor. CBC parameters and renal biomarkers were pivotal in short-term models, while therapeutic interventions gained prominence over time. The 10-year group emphasised disease severity and treatment history, indicating survivorship bias. Short-term models, typically relying on data spanning 1 year or less, commonly established as predictive models for AMI risk, demonstrate limited capability in accurately predicting actual long-term survival times. To effectively issue early warnings for genuine long-term mortality risks, it is imperative to collect longer-term patient information and establish ML prediction models tailored to long-term outcomes. Further research is warranted to validate these findings in diverse populations.