Artificial Intelligence-Enhanced Risk Stratification in Non-ST-Segment Elevation Myocardial Infarction: The Incremental Prognostic Value of the Modified Glasgow Prognostic Score in Predicting 36-Month Mortality

Objective: Current clinical guidelines for non-ST-segment elevation myocardial infarction (NSTEMI) emphasize the duration of dual antiplatelet therapy (DAPT) based on scores such as the Predicting Bleeding Complications in Patients Undergoing Stent Implantation and Subsequent Dual Antiplatelet Therapy (PRECISE-DAPT) score and the Dual Antiplatelet Therapy (DAPT) score. However, these anatomical and clinical models often overlook the "inflammatory gap,” namely the contribution of systemic inflammation to both thrombotic and hemorrhagic risk. Although traditional models focus primarily on anatomical complexity, systemic inflammation may serve as a silent driver of adverse outcomes. We aimed to evaluate, using an artificial intelligence (AI)-enhanced approach, the incremental prognostic value of the inflammation-based Modified Glasgow Prognostic Score (mGPS) over a 36-month follow-up period, with particular emphasis on capturing non-linear interactions between biological markers and long-term outcomes.
Method: This study included 456 NSTEMI patients who underwent percutaneous coronary intervention (PCI). mGPS, DAPT, and PRECISE-DAPT scores were calculated for all patients. Prognostic performance was assessed using a Gradient Boosting Machine (GBM)-based machine learning (ML) framework. Analyses focused on improvements in the concordance index (C-index), Net Reclassification Index (NRI), and Integrated Discrimination Improvement (IDI), as well as three-year mortality assessed using Cox proportional hazards modeling to identify residual inflammatory risk. Model calibration was evaluated using the Brier score.
Results: At 36 months, mGPS emerged as the strongest independent predictor of mortality. Compared with patients with mGPS 0, those with mGPS 2 demonstrated a 6.18-fold higher risk of mortality (hazard ratio 6.18; 95% confidence interval 2.95–12.94; P < 0.001). Observed mortality rates increased markedly from 3.2% in the mGPS 0 group to 57.1% in the mGPS 2 group (P < 0.001). While mGPS was not significantly associated with the ischemic DAPT score (P = 0.349), it was strongly associated with high bleeding risk (PRECISE-DAPT ≥ 25), with all patients in the mGPS 2 category classified as high bleeding risk (P < 0.001). Incorporation of mGPS into the baseline AI model significantly improved discriminative performance, increasing the C-index from 0.72 to 0.81 (P = 0.008). Additionally, the model correctly reclassified 46% of patients (NRI 0.46, P < 0.001) and demonstrated significant predictive improvement (IDI 0.08, P = 0.004) compared with traditional scoring systems.
Conclusion: Systemic inflammation, as quantified by mGPS, is a critical biological modifier of cardiovascular risk that may help bridge the "inflammatory gap" in current scoring systems. AI-driven integration of mGPS into conventional clinical scores significantly improves 36-month prognostic reclassification. mGPS appears to function as a biological recalibrator, identifying high-risk individuals who may be misclassified by conventional scoring systems. These findings suggest that mGPS-guided personalized antiplatelet strategies—particularly early DAPT de-escalation in patients with elevated mGPS—may help reduce the high mortality and bleeding risk observed in this vulnerable population by addressing biological fragility in addition to anatomical risk.

Keywords: DAPT, machine learning, mGPS, NSTEMI, PRECISE-DAPT, systemic inflammation, three-year mortality