Stepwise increasing input to machine learning models predicting 30-day AMI or death in emergency department...

Introduction 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. Methods Data from 40 312 chest-pain patients were used to train NNs with increasing information in five steps. Step 1: age and sex; Step 2: past medical history, redeemed medications, and past coronary angiography results; Step 3: ECG features; Step 4: creatinine, hemoglobin, and glucose blood results; Step 5: the first high-sensitivity cardiac troponin T (hs-cTnT) results. The area under the receiver operating characteristic curve (AUROC) was evaluated at each step, both in isolation and combined with previous steps. For each combination, we also counted the number of patients safely ruled out by the NN, with set sensitivities for 30-day AMI/death of 100% for steps 1-4 and >99% for step 5. Results AUROC scores differed significantly for each step in isolation, but the added value of steps 2 and 4 was not significant. Using only age and sex, 16% of patients could be ruled out, while at step 5, 49% were ruled out. Adding all prior information to hs-cTnT alone did not significantly improve AUROC. Conclusion A model using clinical variables achieved an AUROC of 92.8%, ruling out nearly 50% of patients, while a simpler model with age and sex ruled out 16%. Sensitivity fell below prespecified criteria, especially when models were used sequentially. ECG and hs-cTnT were the most important predictors.