Explainable AI Reveals the Primacy of Early Troponin Kinetics in Identifying Intramyocardial Hemorrhage in Reperfused ST-elevation MI Patients

Background: Intramyocardial Hemorrhage (IMH) is a marker of severe microvascular injury after ST-elevation MI (STEMI) patients and is linked to major adverse clinical events [1]. Prior work proposed that post-PCI troponin levels can identify hemorrhagic myocardial infarction (MI) when validated against CMR-defined IMH, with thresholds derived in a discovery cohort and confirmed in a validation cohort (CMR 48–72 h) [2]. Herein, we utilize explainable AI (XAI) to study the interplay between troponin kinetics and IMH in reperfused STEMI patients.

Methods: We analyzed 183 patients with serial troponin determinations (immediately post-PCI (0) – 48 h post PCI) and CMR-derived IMH occurrence (prevalence 81/183, 44%). We studied troponin kinetics features from the troponin time series (Fig 1A). Illustrative examples are shown in (Fig 2). We used the Superposable Neural Networks (SNN) XAI approach (Fig 1B) to produce an explainable predictive model of IMH occurrence from troponin kinetics features [3,4]. To further validate the XAI results, we trained three MST models [5] using (Early-phase features, Late-phase features, and All features) and compared their performance using 5-fold stratified cross-validation ROC-AUC average (Fig 3A).

Results: Across 5-fold stratified cross-validation, the All-feature model achieved AUC=0.9037, the Early-phase model 0.9015, and the Late-phase model 0.8909. The absolute difference between All-feature and Early-phase model was 0.0022, indicating that nearly all discriminative power is captured by early-phase features, with only marginal incremental gain from late-phase kinetics. Global explainability (mean AUC loss under feature permutation) identified AUC at 0–12h and Peak-value as the dominant contributors, followed by AUC at 12–24 h, AUC at 0–6 h, and max-rise-slope. Consistent with this, the minimal optimal subset identified by the SNN-XAI approach comprised AUC at 0–12h, AUC at 0–6 h, and max-rise-slope as the optimal feature subset for predicting IMH. The SNN model achieved an AUC=0.9003, essentially matching the performance of the state-of-the-art MST model that leverages the full feature set (ΔAUC≈0.003).

Conclusion: Explainable AI on serial troponin demonstrated that early-phase kinetics is the primary driver of IMH discrimination. While late-phase troponin kinetics carry predictive signal for IMH, adding late-phase features produced no meaningful gain over the early set (ΔAUC≈0.002). A compact SNN using only three early-phase features matched the full-feature MST benchmark (AUC≈0.90). By demonstrating that nearly all discriminative signal resides within the first 12 hours, this work demonstrated that hemorrhagic myocardial infarctions can be identified using early-phase troponin readouts as early as 6 hours after PCI.

Figure 1. Panel A: Troponin kinetics feature maps. Example troponin time series (blue curve) with annotated features used in the study. Early-phase features: peak value, time-to-peak, rise 10–90% duration and max rise slope, AUC of 0–6h and AUC of 0–12h, and the 1 h/peak ratio. Late-phase features: post-peak log-linear decay rate, AUC of 12–24 h, and AUC of 24–48 h. These engineered features are the inputs to the model used for explainability analyses for IMH prediction. Panel B. Superposable Neural Network Architecture (SNN). Illustration of the SNN architecture XAI model used in this work.
Figure 2. Illustrative examples: this figure shows two illustrative examples of reperfused STEMI patients without IMH (-) and with IMH (+) . Left Column: Troponin kinetics feature maps. Middle columns: short-axis CMR at 48–72 h—LGE and T2*. Top row (IMH −Ve): no hypointense myocardial region on T2*. Bottom row (IMH +Ve): hypointense region (arrowheads) consistent with intramyocardial hemorrhage. The IMH+ example shows larger early and late exposure, illustrating how most discriminative signal is captured by early-phase features while late-phase kinetics remain biologically informative.
Figure 3. Results. A. Five-fold stratified Cross-Validation ROC curves comparing three models: All features (AUC = 0.9037), Early phase features (AUC = 0.9015), and Late phase features (AUC = 0.8909). The near overlap between all features and Early features indicates that most discriminative power is available within the early-phase features, with limited incremental gain from late-phase kinetics. B. Optimized SNN model identified AUC of 0–12h, AUC of 0–6h, and Max rise slope as the optimal feature combination for predicting IMH. The model had an AUC of 90.03%, virtually matching the state-of-the-art MST model that uses all the features. C. Global feature importance (average AUC loss under feature permutation) for the top 5 contributors: AUC of 0–12h and Peak value dominate, followed by AUC of 12–24h, AUC of 0–6h, and Max rise slope. D. Optimal feature subset: AUC of 0–12h, AUC of 0–6h and Max rise slope were identified as the optimal combination of features for predicting IMH, consistent with early-phase dominance and late-phase redundancy.