Acute myocardial infarction prognosis prediction with reliable and interpretable artificial intelligence system

Abstract Objective Predicting mortality after acute myocardial infarction (AMI) is crucial for timely prescription and treatment of AMI patients, but there are no appropriate AI systems for clinicians. Our primary goal is to develop a reliable and interpretable AI system and provide some valuable insights regarding short, and long-term mortality. Materials and methods We propose the RIAS framework, an end-to-end framework that is designed with reliability and interpretability at its core and automatically optimizes the given model. Using RIAS, clinicians get accurate and reliable predictions which can be used as likelihood, with global and local explanations, and “what if” scenarios to achieve desired outcomes as well. Results We apply RIAS to AMI prognosis prediction data which comes from the Korean Acute Myocardial Infarction Registry. We compared FT-Transformer with XGBoost and MLP and found that FT-Transformer has superiority in sensitivity and comparable performance in AUROC and F1 score to XGBoost. Furthermore, RIAS reveals the significance of statin-based medications, beta-blockers, and age on mortality regardless of time period. Lastly, we showcase reliable and interpretable results of RIAS with local explanations and counterfactual examples for several realistic scenarios. Discussion RIAS addresses the “black-box” issue in AI by providing both global and local explanations based on SHAP values and reliable predictions, interpretable as actual likelihoods. The system’s “what if” counterfactual explanations enable clinicians to simulate patient-specific scenarios under various conditions, enhancing its practical utility. Conclusion The proposed framework provides reliable and interpretable predictions along with counterfactual examples.


A. FEATURE ENGINEERING AND DATA PROCESSING
We excluded certain samples due to loss of follow-up, resulting in a dataset of 15,629 samples used to predict in-hospital mortality, 14,613 samples for 6-month mortality, and 14,260 samples for 12-month mortality.The KAMIR dataset has originally 614 features, but most of them are considered irrelevant to or hindered the mortality prediction tasks by clinical experts.Therefore, we excluded the unnecessary features, leaving 114 features.We identified several outliers caused by potential data collection issues and converted them to missing values for imputation later.Then, we excluded features with a missing rate of 40% or more, leaving 97 features.We impute the remaining missing values using a powerful statistical technique called MICE [1].The data processing flowchart is shown in Fig. A1.

C. EXPECTED CALIBRATION ERROR (ECE)
The Expected Calibration Error (ECE) [2] quantifies the discrepancy between a model's predicted confidence (i.e.likelihood) and its true accuracy.To compute the ECE, predictions are grouped into M bins of equal sizes based on their confidence, and the di↵erence between the average accuracy and average confidence for each bin is determined.Formally, the ECE is given by: where Bm is the set of indices of samples whose prediction confidence falls within interval Im = ( m 1 M , m M ], |Bm| represents the number of predictions in the m th bin, n denotes the total number of samples, and acc and conf denote the average accuracy and average confidence of each bin, respectively.She had 3-vessel disease and was suboptimally treated with only medical treatment for the culprit vessel of LCX, resulting in a final good LVEF of 55%.Initially, she had an elevated creatinine of 3.0mg/dL, peaking at 3.3mg/dL, indicating pre-existing diabetic or hypertensive chronic kidney disease.She was not prescribed a statin, beta-blockers, or any ACEi/ARBs as discharge medication.She was predicted to have a 71% chance of mortality after 12 months using our system, primarily due to complications during hospitalization, elevated peak creatinine level, not using statins, and low body weight.The clinician wants to verify whether the usage of SGLT2 inhibitor can make the patient survive or not.Therefore, the clinician uses the RIAS and the system concludes that the usage of 14.859mg of dapagliflozin, one of SGLT2 inhibitors, contributes to the decrease in the likelihood of death.As shown in Figure A3, administering dapagliflozin decreases the likelihood from 71% to 38%.Dapagliflozin was not included as a variable correlated with any mortality event after AMI in this algorithm.The cardioprotective role of SGLT2 inhibitors in diabetes are well known in high risk of atherosclerotic cardiovascular disease including old myocardial infarction [3,4].Furthermore, SGLT2 inhibitors have emerged as the gold standard treatment for heart failure, extending beyond patients with reduced ejection fraction, now encompassing those with preserved ejection fraction exceeding 40%, regardless of diabetes [5][6][7][8].Regarding the acute phase of myocardial infarction, for individuals at high risk of Kim et al.

D. THE SEARCH RANGE OF HYPERPARAMETERS FOR EACH MODEL
developing heart failure, several ongoing randomized trials are expected to o↵er conclusive evidence on this matter.The RIAS system has predicted that even AMI patients with diabetes and preserved LVEF would benefit from SGLT2 inhibitors during the acute phase of AMI in terms of reducing mortality.This generates hypotheses and suggestions from the algorithm.Similarly, an analysis of claims data using propensity matching showed that the use of SGLT2 inhibitors during the acute phase of AMI significantly reduced all-cause mortality and hospitalizations for heart failure (the predefined primary endpoint) during a 2.1-year follow-up period [9].In the findings of this study, the group receiving SGLT2 inhibitors achieved a 34% reduction in the primary endpoint and a 45% decrease in all-cause mortality.In this counterfactual analysis, the system predicted a 33% reduction in likelihood of mortality when using dapagliflozin for one year.The RIAS system provides valuable insights into unresolved issues, such as this scenario, with outcomes closely aligned with real-world data.
Fig. A3: The change in mortality without and with SGLT2 inhibitor prescription when LVEF is above 40%.

Table A1 .
A summary of the risk level using the GRACE score

Table A2 .
Hyperparameter search space for XGBoost

Table A3 .
Hyperparameter search space for FT-Transformer

Table A4 .
Hyperparameter search range for MLP