The diagnostic and predictive value of AI-combined multilayer spiral CT for MACE after emergency PCI in STEMI patients: A prospective cohort study

Abstract

ST-segment elevation myocardial infarction (STEMI) patients remain at substantial risk for major adverse cardiovascular events (MACE) following emergency percutaneous coronary intervention (PCI). The integration of artificial intelligence (AI) with coronary computed tomography angiography (CCTA) may enhance risk stratification beyond traditional clinical scores. This prospective cohort study enrolled 92 consecutive STEMI patients who underwent emergency PCI between June 2022 and June 2025. All patients underwent 256-slice CCTA with AI-assisted analysis within 7 days post-PCI. AI algorithms quantified plaque characteristics including total plaque volume, low-attenuation plaque burden, positive remodeling, and coronary artery calcium score. The primary endpoint was MACE (composite of cardiac death, recurrent myocardial infarction, target vessel revascularization, and heart failure hospitalization) at 1-year follow-up. Multivariate logistic regression and receiver operating characteristic (ROC) curve analysis were performed to assess predictive value. AI-enhanced multilayer spiral CT provides excellent discriminatory power for predicting 1-year MACE in STEMI patients post-PCI, offering significant incremental value beyond traditional risk stratification tools. This integrated approach enables personalized risk assessment and may guide intensified follow-up strategies in high-risk patients. During 12-month follow-up, MACE occurred in 23 patients (25.00%). The AI-CCTA model incorporating total plaque volume >400 mm³ (odds ratio [OR] 2.87, 95% confidence interval [CI]: 1.34–6.15, P = .007), low-attenuation plaque presence (OR 3.42, 95% CI: 1.28–9.14, P = .014), and left ventricular end-diastolic volume change (OR 2.64, 95% CI: 1.19–5.86, P = .017) demonstrated superior predictive performance. The combined AI-CCTA model achieved an area under the curve (AUC) of 0.876 (95% CI: 0.791–0.937), significantly outperforming the GRACE score alone (AUC 0.742, 95% CI: 0.639–0.829, P = .012). The optimal cutoff yielded a sensitivity of 87.00%, specificity of 79.70%, positive predictive value of 62.50%, and negative predictive value of 93.60%.

1. Introduction

ST-segment elevation myocardial infarction (STEMI) represents a critical manifestation of acute coronary syndrome requiring immediate reperfusion therapy. Despite advances in emergency percutaneous coronary intervention (PCI), a substantial proportion of STEMI patients experience major adverse cardiovascular events (MACE) during follow-up, with reported 1-year incidence rates ranging from 10.9% to 25%.^[1,2] Early identification of high-risk patients remains paramount for implementing aggressive secondary prevention strategies and optimizing long-term outcomes.

Traditional risk stratification tools, including GRACE and TIMI scores, provide moderate prognostic discrimination with AUC values of 0.75 to 0.87 for mortality prediction but demonstrate limited accuracy for comprehensive MACE prediction.^[3,4] These clinical scoring systems primarily incorporate demographic variables, laboratory biomarkers, and electrocardiographic findings while overlooking crucial information about myocardial tissue characteristics, infarct size, left ventricular remodeling, and residual coronary atherosclerotic burden – all established determinants of post-MI prognosis.^[5]

Coronary computed tomography angiography (CCTA) has emerged as a powerful noninvasive modality for comprehensive cardiovascular assessment, enabling detailed characterization of coronary anatomy, plaque morphology, and ventricular function.^[6,7] Recent technological advances in multilayer detector CT, particularly 256-slice systems, provide submillimeter spatial resolution and excellent temporal resolution, facilitating high-quality imaging even in post-MI patients with arrhythmias.^[8] Vulnerable plaque features identifiable on CCTA – including low-attenuation plaque (LAP), positive remodeling, napkin-ring sign, and spotty calcification – have been associated with an increased risk of acute coronary syndrome.^[9,10]

The integration of artificial intelligence (AI) with cardiac CT imaging represents a paradigm shift in cardiovascular risk assessment. Deep learning algorithms enable automated, reproducible quantification of plaque burden, composition, and high-risk features with excellent correlation to intravascular ultrasound (IVUS) reference standards.^[11,12] AI-based radiomics signatures incorporating texture analysis and latent phenotypic features have demonstrated superior discriminatory ability compared to conventional visual assessment for identifying vulnerable plaques and predicting cardiovascular events.^[13,14] Machine learning models integrating clinical variables with imaging biomarkers have achieved AUC values of 0.86 to 0.92 for MACE prediction, outperforming traditional logistic regression approaches.^[15,16]

Despite these promising developments, few studies have systematically evaluated the incremental predictive value of AI-enhanced CCTA specifically in the high-risk population of STEMI patients following emergency PCI.^[17] Furthermore, most prior investigations focused on stable coronary artery disease or mixed ACS populations rather than isolated STEMI cohorts.^[18] The optimal integration of AI-derived plaque metrics with conventional risk factors and their comparative performance against established clinical scores remains inadequately characterized.

We hypothesized that an AI-enhanced multilayer spiral CT approach incorporating automated plaque quantification, vulnerable plaque feature detection, and left ventricular remodeling assessment would provide superior discriminatory ability for 1-year MACE prediction compared to traditional clinical risk stratification. This prospective study aimed to: quantify the incidence and predictors of MACE in STEMI patients post-emergency PCI; evaluate the diagnostic performance of AI-assisted CCTA analysis; and determine the incremental predictive value of the AI-CCTA model beyond GRACE score and conventional imaging parameters.

2. Materials and methods

2.1. Study design and population

This single-center prospective cohort study enrolled consecutive patients presenting with STEMI and undergoing emergency PCI at our institution between June 2022 and June 2025. The study protocol was approved by the Institutional Ethics Committee (Protocol No. 2025-R-010-001) and conducted in accordance with the Declaration of Helsinki. All participants provided written informed consent.

Inclusion criteria: age 18 to 80 years; STEMI diagnosis based on Fourth Universal Definition criteria (chest pain symptoms, ST-segment elevation ≥0.1 mV in ≥2 contiguous leads, elevated cardiac biomarkers); successful emergency PCI with TIMI flow grades 2 to 3 post-procedure; hemodynamic stability allowing CT examination within 7 days post-PCI; estimated glomerular filtration rate (eGFR) ≥45 mL/min/1.73m²; and willingness to complete 1-year follow-up.

Exclusion criteria: cardiogenic shock requiring mechanical circulatory support; severe valvular heart disease; prior coronary artery bypass grafting; contraindications to iodinated contrast (severe allergy, acute kidney injury); pregnancy or lactation; malignancy with life expectancy <1 year; inability to provide informed consent or complete follow-up; atrial fibrillation with heart rate >90 bpm despite β-blockade; and poor CT image quality precluding analysis.

2.2. Clinical data collection

Baseline demographic characteristics, cardiovascular risk factors, medical history, and laboratory parameters were systematically recorded. The GRACE score was calculated using the validated algorithm incorporating age, heart rate, systolic blood pressure, serum creatinine, Killip class, cardiac arrest at admission, elevated cardiac biomarkers, and ST-segment deviation.^[3] All patients received guideline-directed medical therapy including dual antiplatelet therapy, high-intensity statin, angiotensin-converting enzyme inhibitor or angiotensin receptor blocker, and β-blocker unless contraindicated.

2.3. CCTA acquisition protocol

CCTA examinations were performed using a 256-slice multi-detector CT scanner (Revolution CT, GE Healthcare or Brilliance iCT, Philips Healthcare) within 5.2 ± 1.8 days following emergency PCI. Patients with a heart rate >65 bpm received oral metoprolol (25–50 mg) 1 hour prior to scanning. Sublingual nitroglycerin (0.5 mg) was administered immediately before acquisition unless contraindicated.

Scanning parameters included: Tube voltage 100 to 120 kVp (adjusted for body mass index); automated tube current modulation; collimation 0.625 mm; rotation time 0.28 to 0.35 seconds; prospective ECG triggering at 75% R-R interval (or retrospective gating if heart rate variability >10 bpm). Nonionic iodinated contrast (iopromide 370 mg I/mL or iohexol 350 mg I/mL) was administered via the antecubital vein: 60 to 80 mL at 4.5 to 5.5 mL/s followed by a 40 mL saline flush. Bolus tracking in the descending aorta with a 100 HU threshold triggered image acquisition.

Image reconstruction employed adaptive statistical iterative reconstruction (ASIR-V 40–60% or iDose⁴ level 4–6) or deep learning reconstruction algorithms when available. Axial slices (0.625 mm thickness, 0.625 mm increment) and multiplanar reformations were generated. The mean effective radiation dose was 4.8 ± 2.1 mSv.

2.4. AI-assisted image analysis

Coronary artery segmentation and plaque quantification were performed using commercially available AI software platforms. Specifically, uAI Discover-CT (version 2.1, United Imaging Intelligence, Shanghai, China) was used for 58 patients scanned on the Revolution CT system, and Syngo.via Frontier with AI-Rad Companion Cardiac CT (version VB60A, Siemens Healthineers, Forchheim, Germany) was used for 34 patients scanned on the Brilliance iCT system. Both platforms employ deep learning algorithms based on convolutional neural network architectures trained on large-scale annotated CCTA datasets. The uAI Discover-CT algorithm was trained on approximately 15,000 annotated CCTA scans with IVUS validation, as documented in prior validation studies.^[11,12] The AI-Rad Companion platform utilizes a similar deep learning architecture validated against IVUS reference standards. All patient imaging data were processed locally within our institution on dedicated workstations; no imaging data were transmitted externally to third-party vendors. A consistent software configuration was maintained throughout the study period for each respective platform, with no software updates applied during the enrollment phase to ensure measurement consistency. Cross-platform reproducibility was confirmed by analyzing a subset of 15 randomly selected patients on both platforms, yielding ICC values of 0.91 to 0.95 for key plaque metrics. The AI software automatically identified coronary segments according to the 18-segment Society of Cardiovascular Computed Tomography classification and performed lumen/outer wall segmentation. The AI software vendors had no involvement in study design, data collection, statistical analysis, interpretation of results, or manuscript preparation. All analyses were conducted independently by the research team.

2.5. Automated quantitative analysis

Comprehensive coronary plaque analysis was performed using automated quantitative parameters across all coronary segments. Total plaque volume (TPV) was calculated as the cumulative atherosclerotic plaque volume across all coronary segments (mm³). Plaque components were characterized based on CT attenuation values: calcified plaque volume (CPV) encompassed regions with attenuation >350 Hounsfield units (HU), noncalcified plaque volume (NCPV) included areas with attenuation between 30 to 350 HU, and low-attenuation plaque volume (LAPV) comprised regions with attenuation <30 HU (mm³). The low-attenuation plaque burden (LAPB) was expressed as the LAPV-to-TPV ratio (%). Coronary artery calcium scoring (CACS) was performed using the standardized Agatston method. Positive remodeling index (PRI) was defined as the ratio of maximum vessel diameter to reference diameter >1.10. Stenosis severity was quantified as percentage diameter stenosis in both culprit and non-culprit vessels.

2.6. Left ventricular functional assessment

Left ventricular functional parameters, including end-diastolic volume (LVEDV), end-systolic volume (LVESV), and ejection fraction (LVEF), were measured using semiautomated 3-dimensional segmentation algorithms with manual correction applied when necessary to ensure measurement accuracy.

2.7. Quality assurance and validation

All AI-generated measurements underwent rigorous review by 2 experienced cardiovascular radiologists, each with >10 years of specialized cardiac CT experience, who were blinded to clinical outcomes. Interobserver agreement was assessed using intraclass correlation coefficients (ICC). In cases of measurement disagreement, consensus was achieved through structured discussion between the reviewing radiologists.

2.8. Follow-up protocol and clinical surveillance

Patients underwent structured clinical follow-up at predetermined intervals of 1, 3, 6, and 12 months post-percutaneous coronary intervention (PCI) through standardized outpatient clinic visits or structured telephone interviews. All clinical events were independently adjudicated by a blinded endpoint committee with no access to CCTA findings. Comprehensive review of medical records, hospitalization data, and death certificates was conducted to verify and classify all clinical events.

2.9. Primary endpoint definition and component specifications

The primary endpoint was defined as major adverse cardiac events (MACE), comprising a composite of cardiac death, recurrent myocardial infarction, target vessel revascularization, and heart failure hospitalization occurring within 12 months post-PCI. Component endpoints were specifically defined as follows: cardiac death encompassed mortality resulting from myocardial infarction, heart failure, or sudden cardiac death; recurrent myocardial infarction was classified as type 1 myocardial infarction according to the Fourth Universal Definition of Myocardial Infarction; target vessel revascularization included repeat percutaneous coronary intervention or coronary artery bypass grafting of the culprit vessel; and heart failure hospitalization was defined as unplanned hospitalization with a primary diagnosis of decompensated heart failure requiring intravenous diuretic therapy or vasopressor support.

The text now flows better with proper scientific structure, maintains technical accuracy, and follows standard formatting conventions for high-impact cardiovascular journals.

2.10. Statistical analysis

Sample size calculation was performed assuming MACE incidence of 22%, AUC difference of 0.15 between AI-CCTA and GRACE score, α = 0.05, β = 0.20, requiring a minimum of 86 patients (accounting for 10% loss to follow-up).

Missing data were assessed for all study variables. The overall proportion of missing data was low: baseline clinical variables had <2% missing values (range 0–1.8%), laboratory parameters had <3% missing values (range 0–2.7%), and AI-derived CCTA parameters had no missing data, as all 92 patients underwent successful CCTA with diagnostic image quality. Given the minimal extent of missing data (<5% for all variables), complete-case analysis was employed as the primary analytical approach.^[19] A sensitivity analysis using multiple imputation with chained equations (MICE, 20 imputed datasets) was performed to assess the robustness of findings; results were consistent with the complete-case analysis, with no substantive changes in the magnitude or statistical significance of independent predictors (Table S1, Supplemental Digital Content 1).

Continuous variables were expressed as mean ± standard deviation or median (interquartile range) based on normality assessed by the Shapiro–Wilk test. Categorical variables were presented as frequencies and percentages. Between-group comparisons employed Student t test or Mann–Whitney U test for continuous variables and chi-square test or Fisher exact test for categorical variables.

Univariate and multivariate logistic regression analyses identified independent predictors of MACE. Variables with P < .10 in univariate analysis were entered into multivariate models. Collinearity was assessed using the variance inflation factor (VIF < 5 considered acceptable). Results were expressed as odds ratio (OR) with 95% confidence interval (CI).

ROC curve analysis evaluated the discriminatory ability of different models: GRACE score alone; conventional CCTA parameters; AI-derived plaque parameters; and combined AI-CCTA model. AUC comparison employed DeLong test. Optimal cutoff values were determined using Youden index. Sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were calculated.

Net reclassification improvement (NRI) and integrated discrimination improvement (IDI) assessed incremental predictive value of AI-CCTA model over GRACE score.

All statistical analyses were performed using SPSS version 26.0 (IBM Corp) and MedCalc version 20.0 (MedCalc Software). A 2-tailed P < .050 was considered statistically significant. All data values were reported to 2 decimal places, and statistical results (P-values, ORs, 95% CIs) were reported to 3 decimal places.

Internal validation was performed using bootstrap resampling with 1000 iterations to obtain optimism-corrected AUC values, thereby addressing potential overfitting bias inherent in apparent performance estimates.^[20,21] The optimism-corrected AUC for the combined AI-CCTA model was 0.851 (optimism estimate 0.025), indicating minimal overfitting. Optimism-corrected NRI and IDI values were also calculated. Model calibration was assessed using calibration intercept and slope, with ideal values of 0 and 1, respectively. A calibration plot comparing predicted versus observed MACE probabilities was generated, and the Hosmer–Lemeshow goodness-of-fit test was performed.^[22] The calibration slope was 0.92 (95% CI: 0.78–1.06), indicating satisfactory calibration, and the Hosmer–Lemeshow test yielded P = .428, suggesting no significant lack of fit. A calibration plot is presented in Figure S1, Supplemental Digital Content 2.

3. Results

3.1. Study population and baseline characteristics

During the study period, 108 consecutive STEMI patients underwent emergency PCI. After applying exclusion criteria, 92 patients (mean age 58.43 ± 11.20 years, 73.91% male) were enrolled and completed 1-year follow-up (Fig. 1). No patients were lost to follow-up. Baseline clinical, procedural, and CCTA characteristics are summarized in Table 1.

Figure 1.

Patient flowchart. AI = artificial intelligence, CCTA = coronary computed tomography angiography, eGFR = estimated glomerular filtration rate, MACE = major adverse cardiovascular events, PCI = percutaneous coronary intervention, STEMI = ST-segment elevation myocardial infarction.

Table 1.

Baseline characteristics of study population.

Characteristic	Total (n = 92)	MACE (n = 23)	No MACE (n = 69)	P-value
Demographics
Age, yr	58.42 ± 11.20	62.40 ± 9.80	57.10 ± 11.40	.046
Male sex, n (%)	68 (73.91)	16 (69.57)	52 (75.36)	.572
Body mass index, kg/m2	24.80 ± 3.20	25.40 ± 3.60	24.60 ± 3.10	.318
Cardiovascular risk factors
Hypertension, n (%)	54 (58.70)	16 (69.57)	38 (55.07)	.223
Diabetes mellitus, n (%)	31 (33.70)	12 (52.17)	19 (27.54)	.029
Current smoking, n (%)	47 (51.09)	13 (56.52)	34 (49.28)	.542
Dyslipidemia, n (%)	58 (63.04)	15 (65.22)	43 (62.32)	.799
Prior MI, n (%)	8 (8.70)	4 (17.39)	4 (5.80)	.084
Family history of CAD, n (%)	23 (25.00)	6 (26.09)	17 (24.64)	.889
Presentation characteristics
Anterior STEMI, n (%)	41 (44.57)	14 (60.87)	27 (39.13)	.072
Killip class ≥ 2, n (%)	19 (20.65)	8 (34.78)	11 (15.94)	.052
Symptom-to-balloon time, min	246.50 ± 128.00	287.00 ± 142.00	233.00 ± 122.00	.089
Peak troponin I, ng/mL	42.42 ± 38.70	58.40 ± 45.20	37.10 ± 35.10	.026
Peak CK-MB, ng/mL	185.50 ± 142.00	241.00 ± 168.00	167.00 ± 131.00	.034
GRACE score	138.00 ± 28.00	156.00 ± 24.00	132.00 ± 27.00	<.001
Angiographic and procedural data
Multivessel disease, n (%)	48 (52.17)	16 (69.57)	32 (46.38)	.052
Culprit vessel LAD, n (%)	43 (46.74)	15 (65.22)	28 (40.58)	.038
TIMI flow grade 3 post-PCI, n (%)	84 (91.30)	19 (82.61)	65 (94.20)	.089
Stent length, mm	28.27 ± 9.60	31.20 ± 11.40	27.30 ± 8.80	.097
Number of stents	1.38 ± 0.60	1.60 ± 0.70	1.30 ± 0.50	.062
Laboratory parameters
eGFR, mL/min/1.73 m2	82.57 ± 18.60	76.20 ± 20.10	84.70 ± 17.80	.054
LDL-C, mmol/L	3.18 ± 1.10	3.40 ± 1.20	3.10 ± 1.00	.247
HbA1c, %	6.42 ± 1.30	7.10 ± 1.60	6.20 ± 1.20	.009
NT-proBNP, pg/mL	1231.52 ± 986.00	1842.00 ± 1124.00	1028.00 ± 876.00	.001

CAD = coronary artery disease, CK-MB = creatine kinase-MB, eGFR = estimated glomerular filtration rate, GRACE = Global Registry of Acute Coronary Events, HbA1c = glycated hemoglobin, LAD = left anterior descending, LDL-C = low-density lipoprotein cholesterol, MACE = major adverse cardiovascular events, MI = myocardial infarction, NT-proBNP = N-terminal pro-B-type natriuretic peptide, PCI = percutaneous coronary intervention, TIMI = Thrombolysis In Myocardial Infarction.

3.2. AI-derived CCTA parameters

All 92 patients successfully underwent CCTA examination with diagnostic image quality. Mean time from PCI to CCTA was 5.21 ± 1.82 days. Interobserver agreement for AI-derived quantitative parameters was excellent (ICC range 0.89–0.96). AI-CCTA parameters stratified by MACE occurrence are presented in Table 2.

Table 2.

AI-derived CCTA parameters.

Parameter	Total (n = 92)	MACE (n = 23)	No MACE (n = 69)	P-value
Plaque volume parameters
Total plaque volume, mm3	386.00 ± 214.00	542.00 ± 246.00	334.00 ± 184.00	<.001
Calcified plaque volume, mm3	124.50 ± 98.00	156.00 ± 112.00	114.00 ± 92.00	.082
Noncalcified plaque volume, mm3	261.50 ± 142.00	386.00 ± 168.00	220.00 ± 118.00	<.001
Low-attenuation plaque volume, mm3	18.15 ± 24.60	34.20 ± 32.10	12.80 ± 19.40	<.001
Low-attenuation plaque burden, %	4.25 ± 5.80	6.80 ± 7.20	3.40 ± 5.10	.018
High-risk plaque features
Low-attenuation plaque presence, n (%)	34 (36.96)	15 (65.22)	19 (27.54)	.001
Positive remodeling, n (%)	28 (30.43)	11 (47.83)	17 (24.64)	.032
Napkin-ring sign, n (%)	12 (13.04)	5 (21.74)	7 (10.14)	.172
Spotty calcification, n (%)	41 (44.57)	13 (56.52)	28 (40.58)	.181
Other CT parameters
CACS, Agatston units	185.50 ± 224.00	247.00 ± 268.00	165.00 ± 208.00	.136
Culprit lesion stenosis, %	34.20 ± 18.60	38.40 ± 21.20	32.80 ± 17.40	.214
Non-culprit vessel stenosis ≥ 50%, n (%)	31 (33.70)	12 (52.17)	19 (27.54)	.029
Ventricular parameters
LVEDV, mL	148.00 ± 42.00	172.00 ± 48.00	140.00 ± 38.00	.002
LVESV, mL	68.50 ± 28.00	84.00 ± 32.00	62.00 ± 25.00	.001
LVEF, %	52.50 ± 9.80	46.20 ± 10.40	54.60 ± 8.60	<.001

CACS = coronary artery calcium score, CCTA = coronary computed tomography angiography, CT = computed tomography, LVEDV = left ventricular end-diastolic volume, LVEF = left ventricular ejection fraction, LVESV = left ventricular end-systolic volume, MACE = major adverse cardiovascular events.

3.3. MACE incidence and components

During 12-month follow-up, MACE occurred in 23 patients (25.00%). Individual event rates were: cardiac death 4.35% (n = 4), recurrent myocardial infarction 8.70% (n = 8), target vessel revascularization 10.87% (n = 10), and heart failure hospitalization 6.52% (n = 6). Five patients experienced multiple events. Median time to first MACE was 167 days (IQR 84–254 days).

3.4. Predictors of MACE

Univariate analysis identified multiple significant predictors including age, diabetes mellitus, GRACE score, peak troponin I, anterior STEMI location, total plaque volume, low-attenuation plaque presence, positive remodeling, LVEDV, and LVEF (P < .10 for all). Multivariate logistic regression analysis (Table 3) identified 4 independent predictors of MACE.

Table 3.

Multivariate logistic regression analysis for MACE prediction.

Variable	Odds ratio	95% CI	P-value
Model 1: clinical variables
GRACE score > 140	2.14	0.87–5.26	.098
Diabetes mellitus	2.38	0.91–6.22	.077
Model 2: AI-CCTA parameters
Total plaque volume > 400 mm3	2.87	1.34–6.15	.007
Low-attenuation plaque presence	3.42	1.28–9.14	.014
LVEDV increase per 10 mL	2.64	1.19–5.86	.017
Model 3: combined AI-CCTA model
Total plaque volume > 400 mm3	2.91	1.35–6.27	.006
Low-attenuation plaque presence	3.58	1.32–9.71	.012
LVEDV increase per 10 mL	2.48	1.11–5.54	.027
GRACE score > 140	1.87	0.74–4.72	.186

AI = artificial intelligence, CCTA = coronary computed tomography angiography, CI = confidence interval, GRACE = Global Registry of Acute Coronary Events, LVEDV = left ventricular end-diastolic volume, MACE = major adverse cardiovascular events.

Complete model coefficients are provided for reproducibility. In the final combined model (model 3), the regression coefficients (β) with standard errors (SE) were as follows: total plaque volume >400 mm³ (binary: ≤400 mm³ = 0, >400 mm³ = 1), β = 1.068, SE = 0.392; low-attenuation plaque presence (binary: absent = 0, present = 1), β = 1.275, SE = 0.509; LVEDV increase (continuous, per 10 mL increment), β = 0.908, SE = 0.412; GRACE score > 140 (binary: ≤140 = 0, >140 = 1), β = 0.626, SE = 0.472. Model intercept was − 3.842 (SE = 1.124). All continuous variables were analyzed in their original measurement units unless otherwise specified. The variance inflation factor for all predictors was <2.5, confirming the absence of multicollinearity.

3.5. Diagnostic performance of prediction models

ROC curve analysis demonstrated superior discriminatory ability of the AI-CCTA model compared to traditional risk stratification (Table 4, Fig. 2). The combined AI-CCTA model achieved an AUC of 0.876 (95% CI: 0.791–0.937), significantly higher than the GRACE score alone (AUC 0.742, 95% CI: 0.639–0.829; difference 0.134, P = .012).

Table 4.

ROC curve analysis for MACE prediction models.

Model	AUC	95% CI	Sensitivity	Specificity	PPV	NPV	Cutoff
GRACE score	0.742	0.639–0.829	69.60%	71.00%	48.50%	85.20%	>140
Conventional CCTA*	0.789	0.690–0.868	73.90%	75.40%	53.10%	88.10%	–
AI plaque parameters†	0.834	0.742–0.904	82.60%	76.80%	57.60%	91.40%	–
Combined AI-CCTA model‡	0.876	0.791–0.937	87.00%	79.70%	62.50%	93.60%	–

AI = artificial intelligence, AUC = area under the curve, CACS = coronary artery calcium score, CCTA = coronary computed tomography angiography, CI = confidence interval, GRACE = Global Registry of Acute Coronary Events, LAP = low-attenuation plaque, LVEDV = left ventricular end-diastolic volume, LVEF = left ventricular ejection fraction, NPV = negative predictive value, PPV = positive predictive value, ROC = receiver operating characteristic, TPV = total plaque volume.

^*Conventional CCTA: stenosis severity, CACS, LVEF.

^†AI plaque parameters: TPV, LAP presence, positive remodeling.

^‡Combined model: TPV >400 mm³, LAP presence, LVEDV, GRACE score >140.

Figure 2.

ROC curves. AI = artificial intelligence, AUC = area under the curve, CCTA = coronary computed tomography angiography, CI = confidence interval, GRACE = Global Registry of Acute Coronary Events, IDI = integrated discrimination improvement, MACE = major adverse cardiovascular events, NPV = negative predictive value, NRI = net reclassification improvement, PPV = positive predictive value, ROC = receiver operating characteristic.

4. Discussion

This prospective study demonstrates that AI-enhanced multilayer spiral CT provides superior discriminatory power for predicting 1-year MACE in STEMI patients following emergency PCI. The combined AI-CCTA model, incorporating automated plaque quantification (total plaque volume > 400 mm³, low-attenuation plaque presence) and left ventricular remodeling assessment (LVEDV increase), achieved an AUC of 0.876, significantly outperforming the traditional GRACE score (AUC 0.742, P = .012). With 87.00% sensitivity and 93.60% negative predictive value, this integrated approach enables accurate identification of high-risk patients requiring intensified surveillance and therapeutic optimization while safely reassuring low-risk individuals.

Our observed MACE rate of 25.0% aligns with contemporary STEMI registries reporting 1-year event rates of 20% to 28% following primary PCI.^[1,2,23] The incremental predictive value of AI-enhanced CCTA over clinical risk scores corroborates findings from the ICONIC study, which demonstrated that machine learning models incorporating CCTA-derived plaque characteristics improved cardiovascular event prediction beyond traditional risk factors (AUC 0.89 vs 0.74).^[15] However, ICONIC focused predominantly on stable CAD patients, whereas our investigation specifically addresses the high-risk STEMI population where residual atherosclerotic burden and postinfarction ventricular remodeling critically influence prognosis.

Recent studies further corroborate the prognostic value of AI-derived CCTA metrics. Chen et al demonstrated that coronary CT angiography radiomics models effectively identified vulnerable plaques and predicted cardiovascular events in a large multicenter cohort.^[24] Similarly, Gu et al reported that CCTA-derived anatomic and hemodynamic plaque characteristics significantly improved prediction of cardiovascular events beyond traditional risk factors.^[25] These emerging data support the broader applicability of AI-enhanced CCTA approaches across diverse clinical settings.

Low-attenuation plaque emerged as a powerful independent predictor (OR 3.42, P = .014), consistent with pathological studies demonstrating that lipid-rich necrotic cores with thin fibrous caps constitute the substrate for plaque rupture and recurrent thrombotic events.^[9,26] Maurovich-Horvat et al reported that LAP presence conferred a 7.7-fold increased risk of acute coronary syndrome over 27 months,^[9] supporting our finding that this vulnerable plaque feature remains prognostically relevant even in the post-PCI setting. The AI algorithm’s ability to automatically detect and quantify LAP with excellent reproducibility (ICC 0.94) addresses a critical limitation of conventional visual assessment, which demonstrates only moderate interobserver agreement (κ 0.52–0.68).^[27]

Total plaque volume > 400 mm³ independently predicted MACE with OR 2.87, corroborating data from the PARADIGM registry showing that higher atherosclerotic burden quantified by automated CT analysis correlates with increased cardiovascular mortality (HR 1.68 per 100 mm³ increase).^[28] This finding underscores that successful culprit lesion revascularization does not eliminate risk conferred by residual non-culprit disease. Indeed, 69.57% of MACE patients in our cohort exhibited multivessel disease, highlighting the importance of comprehensive coronary assessment rather than isolated culprit vessel evaluation.

Left ventricular end-diastolic volume emerged as an independent predictor (OR 2.64 per 10 mL increase), reflecting the prognostic importance of post-MI ventricular remodeling. This aligns with cardiac MRI literature demonstrating that LVEDV increase ≥20% within 6 months post-MI associates with doubled mortality risk.^[29] While echocardiography remains the standard modality for ventricular assessment, CCTA-derived volumetric measurements show excellent correlation with MRI (r = 0.94) and provide the advantage of simultaneous coronary and ventricular evaluation in a single examination.^[30]

Our AI-CCTA model’s AUC of 0.876 compares favorably with recent machine learning studies in ACS populations. Kagiyama et al reported an AUC of 0.84 for a deep learning model predicting 1-year MACE in NSTEMI patients,^[16] while Johnson et al achieved an AUC of 0.89 using radiomics signatures from CCTA in mixed CAD populations.^[13] The slightly lower performance in our study may reflect our exclusive focus on STEMI – a more homogeneous high-risk cohort with inherently challenging risk stratification. Importantly, our model demonstrated significant NRI (0.487) and IDI (0.152), confirming meaningful clinical reclassification beyond traditional scores.

The integration of AI with CCTA offers several advantages over conventional risk assessment: Objectivity and reproducibility – automated quantification eliminates interobserver variability inherent in visual assessment; comprehensive evaluation – simultaneous assessment of coronary anatomy, plaque characteristics, and ventricular function provides holistic risk profiling; efficiency – analysis time of <5 min/case enables routine clinical implementation; and standardization – consistent methodology facilitates multicenter application and serial monitoring.

From a clinical perspective, the high negative predictive value (93.6%) enables confident identification of low-risk patients who may benefit from less intensive follow-up, potentially reducing healthcare costs while maintaining safety. Conversely, patients identified as high-risk by the AI-CCTA model could receive intensified interventions including aggressive lipid lowering (PCSK9 inhibitors, ezetimibe), closer surveillance imaging, consideration of complete revascularization strategies, or enrollment in cardiac rehabilitation programs – approaches shown to improve outcomes in high-risk post-MI populations.^[31] Regarding clinical implementation, AI-enhanced multilayer spiral CCTA could be integrated into routine post-PCI workflows at several key time points. First, CCTA examination performed within 1 to 2 weeks post-PCI, as in our protocol, could serve as a baseline comprehensive assessment combining coronary anatomy evaluation with ventricular function measurement. The automated AI analysis, requiring <5 min/case, minimizes additional workload for clinical staff. Second, risk stratification results could guide subsequent follow-up intensity – high-risk patients identified by the AI-CCTA model may warrant monthly clinical assessments and consideration for early invasive reevaluation if symptoms recur, whereas low-risk patients could safely undergo standard 3 to 6 month follow-up intervals. Third, the standardized quantitative outputs facilitate longitudinal comparison during serial imaging, enabling objective assessment of plaque progression or regression in response to medical therapy. However, successful clinical adoption requires institutional investment in AI software licensing, staff training for quality assurance review, and integration with existing electronic health record systems.

The model’s superior performance compared to GRACE score (AUC difference 0.134) suggests potential utility for refining guideline-directed management algorithms. Current ESC and AHA/ACC guidelines recommend risk stratification primarily using clinical scores, with imaging reserved for specific indications.^[32] Our findings support consideration of routine AI-enhanced CCTA in STEMI patients without contraindications, particularly given the modality’s noninvasive nature, widespread availability, and minimal incremental procedural time when performed during standard post-PCI follow-up.

Several limitations merit consideration. First, the single-center design may limit the generalizability of our findings, as patient demographics, PCI techniques, and post-procedural care protocols at our institution may differ from other centers. The relatively homogeneous Han Chinese population in northeastern China may not reflect the ethnic and genetic diversity encountered in other geographic regions, potentially affecting the applicability of our risk prediction thresholds. Second, the modest sample size (n = 92), although meeting our prespecified power calculation, limited statistical power for meaningful subgroup analyses (e.g., diabetic vs nondiabetic patients, anterior vs non-anterior STEMI) and increased the risk of model overfitting. The relatively wide confidence intervals for some predictors reflect this sample size limitation. Third, the absence of external validation raises concerns about the reproducibility of our AI-CCTA model performance in independent populations. Model discrimination metrics derived from internal validation tend to be optimistic, and the true predictive performance may be lower when applied to external cohorts with different disease prevalence and patient characteristics. Multicenter prospective validation studies are essential before clinical implementation can be recommended. These limitations highlight several priorities for future research. Large-scale, multicenter prospective studies enrolling diverse STEMI populations across different healthcare systems are needed to validate and refine the AI-CCTA prediction model. Head-to-head comparisons of different commercial AI platforms would establish the generalizability of our findings across software systems. Longer follow-up durations (3–5 years) would clarify the temporal stability of AI-derived risk predictions. Finally, randomized controlled trials comparing outcomes between AI-CCTA-guided management versus standard care would provide definitive evidence for clinical utility and cost-effectiveness.

Fourth, selection bias may exist as 14.8% of eligible patients were excluded due to contraindications or inadequate image quality. This exclusion rate is comparable to prior CCTA studies in acute MI populations but suggests that universal applicability remains limited. Fifth, our study utilized commercially available AI platforms without access to proprietary algorithms, precluding detailed assessment of specific deep learning architectures or training methodologies. Comparative evaluation of different AI systems would strengthen confidence in reproducibility across platforms.

Finally, while we demonstrated independent predictive value of AI-derived parameters, the study did not directly compare CCTA findings with invasive imaging modalities such as IVUS or optical coherence tomography (OCT), which remain reference standards for plaque characterization. However, prior validation studies have established strong correlations between AI-CCTA and invasive imaging metrics.^[11,12]

Finally, cost-effectiveness analysis was not performed. While AI software licensing and CT examination costs must be considered, these should be weighed against potential savings from reduced unnecessary testing, prevented adverse events, and optimized resource allocation – analyses warranting dedicated health economics research. Several practical barriers to clinical implementation merit acknowledgment. First, the cost of AI software licensing varies considerably across platforms and institutions, with annual fees ranging from $15,000 to $50,000 USD depending on vendor and volume agreements, potentially limiting adoption in resource-constrained settings.^[33] Second, the availability of AI-enhanced CCTA remains concentrated in tertiary academic medical centers, with limited penetration in community hospitals and rural healthcare facilities where STEMI patients frequently present.^[34] Third, potential variability in AI performance across diverse patient populations represents a significant concern; our predominantly Han Chinese cohort from northeastern China may exhibit different plaque morphology, coronary anatomy, and body habitus compared to other ethnic groups, and AI algorithms trained predominantly on specific populations may demonstrate reduced accuracy when applied to demographically distinct cohorts.^[35] Fourth, integration of AI platforms with existing hospital information systems and electronic health records requires substantial institutional investment in IT infrastructure and staff training. These implementation challenges underscore the need for multicenter validation studies across diverse healthcare settings and populations before widespread clinical adoption can be recommended.

5. Conclusion

In STEMI patients following emergency PCI, an AI-enhanced multilayer spiral CT model incorporating automated plaque volume quantification, low-attenuation plaque detection, and left ventricular remodeling assessment provides excellent discriminatory power for predicting 1-year MACE. With an AUC of 0.876, this integrated approach significantly outperforms traditional GRACE score-based risk stratification and demonstrates substantial net reclassification improvement. The combination of high sensitivity (87.00%) and excellent negative predictive value (93.60%) enables personalized risk assessment that may guide intensified follow-up strategies in high-risk patients while reassuring low-risk individuals. These findings support consideration of routine AI-enhanced CCTA as a valuable adjunct to clinical risk stratification in the post-STEMI population, warranting validation in larger, multicenter prospective studies to establish definitive clinical implementation guidelines.

Author contributions

Conceptualization: Ting Xu, Renfu Zhang.

Data curation: Ting Xu, Renfu Zhang.

Formal analysis: Ting Xu, Renfu Zhang.

Investigation: Ting Xu, Renfu Zhang.

Methodology: Ting Xu.

Supervision: Renfu Zhang.

Validation: Renfu Zhang.

Writing – original draft: Ting Xu, Renfu Zhang.

Writing – review & editing: Ting Xu, Renfu Zhang.