Background
High sensitivity cardiac troponins (hs-cTn) are the gold standard biomarkers to diagnosis acute myocardial infarction (AMI).In real-life clinical practice, different patient characteristic may lead to complex interactions that affect the traditional models and thresholds in use. The aims of this study were to assess if explainable machine learning (ML) models can identify the shift in feature importance and to compare the diagnostic performance of hs-cTn for AMI risk with traditional-based models in specific subgroups.
Methods
Data from a multi-centric cohort that prospectively enrolled consecutive patients with suspected AMI was used. The clinical outcome of interest was adjudicated AMI. We evaluated multiple state-of-the-art ML models, tuning the hyperparameters via random search (≈7000 total fits) withing the recent TabArena benchmarking framework and used ROC AUC as the performance metric. Overall, interpretable 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 (incorporating all features) were compared against a baseline logistic regression model assessing only baseline hs-cTn via ROC-AUC analysis. Finally, individual feature importance for each subgroup was assessed using EBM's term importance scores, calculated as the mean absolute contribution of each feature's learned shape function across all predictions.
Results
A total of 1818 patients were included in the analyses. Further details about baseline charactersitics are available at Keller, JAMA, Dec,2011. All evaluated ML methods showed a superior performance compared with traditional linear regression analysis for baseline hs-cTN in predicting AMI (AUC 0.98 for EBM vs 0.96,) (Figure 1). These findings were consistent across subgroups divided according to sex, presence of AF, DM, or renal impairment, suggesting that integration of ML with traditional interpretation of hs-cTN could help achieve a more individualized diagnostic approach. Figure 2 depicts the importance of individual characteristics within the prediction model, according to the subgroup. 3h and 6h hs-cTn, combined with peak troponin, were more relevant in females and males, respectively (Figure 2A). In patients presenting with AF the relevance of delta troponin (0-3h) was higher compared with patients without AF (Figure 2B). Delta troponin emerged as more relevant in the feature ranking for patients with DM (Figure 2C), whereas for patients with impaired renal function baseline hs-cTn values were more important than the 3h and 6h hs-cTn (Figure 2D).
Conclusion
Using an interpretable ML model (EBM) we identified subgroup-specific variations in the importance of patient characteristics, hs-cTn measurement timing, and their dynamic changes across different patient subgroups with suspected AMI. Interpretable models like EBM achieved robust predictive performance for identifying AMI in a general patient cohort and in specific subgroups. These findings suggest that using ML based diagnostic models for identification of AMI may lead to more individualized diagnostic strategies.
