Diagnostic performance of traditional troponin-based predictive models and explainable ML-models in patients with suspected AMI

Background

High sensitivity cardiac troponin (hs-cTn) is the gold standard biomarker for the diagnosis of acute myocardial infarction (AMI). Diagnostic accuracy of hs-cTn varies across patient subgroups, with individual characteristics influencing traditional models and thresholds.

Purpose

The aims of this study were to compare the machine learning (ML)-diagnostic performance of hs-cTn for AMI risk with traditional models in specific subgroups and to assess if explainable ML models may assist in better interpretation of the interactions of hs-cTn.

Methods

Data from a large multi-centric cohort that enrolled consecutive patients with suspected AMI were used. The adjudicated AMI diagnosis according to the universal MI definition was the outcome. We evaluated multiple state-of-the-art ML models, tuning the hyper-parameters via random search (≈7000 total fits) within the recent TabArena benchmarking framework and used ROC AUC as the performance metric. Models like explainable boosting machine (EBM) performed competitively to the ML foundation models (TabPFN-v2, TabM, RealMLP) and traditional models (tree-based methods, logistic regressions) (Figure 1). Subgroup-specific EBM models were compared with a baseline log. regression model assessing baseline hs-cTn via ROC-AUC. EBM models were trained separately for each subgroup using stratified 80/20 splits. Shape functions from these models visualize how each feature influences AMI risk on the log-odds scale.

Results

A total of 1818 patients were included. Median age was 61.4 (standard deviation, SD 13.5) years, 314 (17.2%) presented with atrial fibrillation (AF), mean glomerular filtration rate (GFR) was 79.0 (SD 21.4) and 413 (22.7%) patients had an AMI event. All evaluated ML methods indicated a superior predictive performance of baseline hs-cTn for AMI compared with traditional linear regression (AUC 0.98 vs 0.96) (Figure 1). Similarly, among patients of different sex, and with at least one comorbidity, EBM-Models compared with traditional interpretation of hs-cTN, showed a higher predictive performance demonstrating a more accurate diagnostic approach.

In patients with AF at presentation, increased baseline hs-cTn contributed more strongly to the prediction of AMI risk compared with those without AF (Figure 1A). Baseline hs-cTn thresholds for diagnosing AMI were higher in patients with AF, while in those with reduced GFR hs-cTn predicted AMI only at elevated thresholds (Figure 2B).

Conclusion

Interpretable models like EBM achieved robust predictive performance for identifying AMI, not only in a general patient cohort, but also in specific subgroups. These results indicate that employing ML-based diagnostic models for troponin on detecting AMI could enable more personalized diagnostic approaches.

ML-models can facilitate the understanding of the interactions and diagnostic capacity of hs-cTn in complex clinical scenarios such presence of AF or impaired renal function.