# Development and validation of a machine learning model for post-PCI exercise intolerance in patients with coronary artery disease via electronic medical records

## TL;DR

This study creates a machine learning model using electronic medical records to predict exercise intolerance after heart procedures, helping doctors identify at-risk patients early.

## Contribution

A novel machine learning model using EMR data to predict post-PCI exercise intolerance in CAD patients is developed and validated.

## Key findings

- The MLP model achieved an AUC–ROC of 0.911 and high specificity and NPV for predicting exercise intolerance.
- Eight key variables, including age, BMI, and hemoglobin, were identified as significant predictors.
- The model's clinical utility was confirmed via calibration plots and decision curve analysis.

## Abstract

Exercise intolerance after percutaneous coronary intervention (PCI) is a common yet often overlooked condition in patients with coronary artery disease (CAD), associated with impaired cardiopulmonary recovery and poor prognosis. However, an accurate and easily applicable non-exercise-based model for predicting post-PCI exercise intolerance remains lacking. This study aimed to develop and validate such a model using electronic medical record (EMR) data.

Between June 2020 and June 2024, clinical data were retrospectively collected from Quanzhou First Hospital. Forty-five variables were considered as candidate predictors, and seven machine learning algorithms were developed to estimate the risk of post-PCI exercise intolerance. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC–ROC), area under the precision–recall curve (AUC–PRC), accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1 score. Calibration and clinical utility were assessed via calibration plots, Brier score, Hosmer–Lemeshow (H–L) goodness-of-fit test, and decision curve analysis. Model interpretability was examined using Shapley additive explanations, and an interactive web-based calculator was deployed for clinical use.

A total of 575 patients were included, with an incidence of exercise intolerance of 22.0%. Eight key variables were selected: age, sex, BMI, smoking status, diabetes status, hemoglobin level, red blood cell count, and resting heart rate. The multilayer perceptron (MLP) model achieved the best performance (threshold = 0.30): an AUC–ROC of 0.911 (0.854–0.956), an AUC–PRC of 0.706 (0.548–0.846), an accuracy of 0.87, a sensitivity of 0.82, a specificity of 0.88, a PPV of 0.67, and an NPV of 0.94 (Brier = 0.108; H–L test p = 0.493).

The proposed EMR-based model effectively identifies patients at high risk of post-PCI exercise intolerance, supporting early screening and targeted clinical interventions.