Key result
Random forest and modified Zwolle score improve STEMI risk reclassification, expanding low-risk early-discharge candidates to ~74%.
Why the study?
The Zwolle Risk Score for identifying early discharge candidates after STEMI treated with PCI lacks contemporary prognostic markers like frailty and post-PCI LVEF, losing precision in ageing, complex populations.
Does integrating frailty and post-PCI LVEF into the Zwolle Risk Score (M-ZRS) or using a machine learning Random Forest model improve the prediction of in-hospital mortality and identification of low-risk candidates for early discharge compared to the original Zwolle Risk Score in STEMI patients surviving primary PCI?
Observational (n=585)
No
Does integrating frailty and post-PCI LVEF into the Zwolle Risk Score (M-ZRS) or using a machine learning Random Forest model improve the prediction of in-hospital mortality and identification of low-risk candidates for early discharge compared to the original Zwolle Risk Score in STEMI patients surviving primary PCI?
Integrating frailty and post-PCI LVEF into the Zwolle Risk Score and applying Random Forest machine learning algorithms improve high-risk STEMI detection and safely expand the proportion of patients eligible for early discharge after primary PCI.
May support expanding early discharge eligibility in low-risk STEMI; extends ZRS but leaves open prospective validation.
Introduction Accurate risk stratification in ST-segment elevation myocardial infarction (STEMI) after percutaneous coronary intervention (PCI) is vital to identify early discharge candidates and optimize resources. The Zwolle Risk Score (ZRS) lacks contemporary prognostic markers like frailty and post-PCI left ventricular ejection fraction (LVEF), losing precision in ageing, complex populations. Objectives To evaluate the prognostic improvement of integrating frailty and post-PCI LVEF into the original ZRS (M-ZRS), and to compare this updated bedside tool against a machine learning Random Forest (RF) approach, seeking the optimal identification of low-risk candidates for safe early discharge. Methods A prospective, single-center, observational study of 585 STEMI patients who survived primary PCI (July 2023-January 2026) was performed. The sample was divided chronologically into derivation (n=449) and validation (n=136) cohorts. Frailty (Clinical Frailty Scale ≥4) and post-PCI LVEF were assessed. Independent predictors of in-hospital mortality were identified via multivariable logistic regression. M-ZRS was compared to ZRS and an RF model (trained via downsampling using age, KILLIP, anterior MI, 3-vessel disease, ischemia time, post-PCI TIMI flow, Clinical Frailty Scale, and categorical LVEF) using C-statistics and categorical Net Reclassification Improvement (NRI). Results In-hospital mortality was 6.5% (derivation 6.5% vs validation 6.6%, p=0.952). Multivariable analysis confirmed ZRS (OR 1.22 per point; p<0.001), post-PCI LVEF <40% (OR 2.71; p=0.029), and Frailty (OR 3.09; p=0.008) as independent predictors. Comparing B-coefficients, LVEF <40% (B=0.996) and Frailty (B=1.129) were ≈5 and ≈6 times more impactful than a 1-point ZRS increase (B=0.199). Thus, 5 and 6 points were respectively added to ZRS to create the M-ZRS. In the validation cohort, discriminative capacity was highest for RF (C-statistic: 0.894), followed by M-ZRS (0.832) and ZRS (0.758); p>0.05 by DeLong due to low event rate (Figure). Reclassification showed potentially great clinical value (Table), M-ZRS yielded an NRI of 22.2% vs ZRS (event NRI=22.2%, non-event NRI=0%). RF achieved an NRI of 36.4%, matching M-ZRS in identifying mortality cases (event NRI=22.2%) while correctly downgrading survivors to low-risk (non-event NRI=14.2%). With a ≤4 cut-off, M-ZRS identified 61.0% (83/136) as low risk (1.2% mortality; NPV 98.8%, Sensitivity 88.9%). Using its optimal threshold, RF safely expanded the low-risk pool to 74.3% (101/136) (1.0% mortality; NPV 99.0%, Sensitivity 88.9%). Conclusions Integrating frailty and LVEF into the ZRS substantially improves high-risk STEMI detection at the bedside. Furthermore, applying RF algorithms provides exceptional reclassification (NRI 36.4%), safely expanding the pool of true low-risk candidates for early discharge. Updating traditional scores and integrating AI modernizes and expedites safe discharge protocols.Figure 1.AUC of Predictive ModelsTable 1.
No takes yet. Share an insight, caveat, or question.
Santiago et al. (2026) conducted an observational in ST-segment elevation myocardial infarction (STEMI) (n=585). Modified Zwolle Risk Score (M-ZRS) and Random Forest (RF) machine learning model vs. Original Zwolle Risk Score (ZRS) was evaluated on In-hospital mortality (C-statistic 0.894 (RF) vs 0.832 (M-ZRS) vs 0.758 (ZRS), p=>0.05). Random Forest and Modified Zwolle Risk Score improved in-hospital mortality reclassification over ZRS (NRI 36.4% and 22.2%), expanding low-risk STEMI candidates to 74.3% (1.0% mortality).
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: