Stepwise Increasing Input to Machine Learning Models Predicting 30-Day Acute Myocardial Infarction or Death in...

This study evaluated the performance of neural network (NN) models with stepwise increasing input for identifying acute myocardial infarction (AMI) or death within 30 days among emergency department (ED) chest pain patients.

Data from 40,312 chest pain patients were used to train NN models with stepwise increasing input information: (1) age and sex; (2) previous diagnoses, redeemed medications, and coronary angiographies; (3) ECG features; (4) blood creatinine, hemoglobin, and glucose results; (5) the first high-sensitivity cardiac troponin T (hs-cTnT) result. The area under the receiver operating characteristic curve (AUROC) was evaluated for each step alone and for each step combined with the previous steps. For each such combination, we also counted the number of patients safely ruled out by the NN, with the sensitivity for 30-day AMI/death set to 100% for steps 1-4 and >99% for step 5.

AUROC scores differed significantly between the isolated input steps, but when combined, the added value of past medical history (step 2) and creatinine, hemoglobin, and glucose (step 4) was not significant. The model using only age and sex (step 1) ruled out 16% of the patients, and the model using all information (steps 1-5) ruled out 49% with an AUROC of 93%. When the combined models were applied sequentially to the patients not ruled out in the previous steps, 51% of patients were ruled out, but 99% sensitivity could not be achieved. ECG and hs-cTnT were the most important predictor variables. Adding all prior information to hs-cTnT alone did not significantly improve AUROC.

An NN model using age, sex, past medical history, ECG, and blood tests including hs-cTnT achieved an AUROC of 93%, ruling out 30-day AMI/death in 49% of the patients. When the NN models were applied sequentially, the desired sensitivity of 99% was not achieved.