Prognostic value of AI-enabled quantitative coronary CT angiography for major adverse cardiovascular events: A systematic review and meta-analysis Authors: Mushrin Malik, Pedro Batista, Larissa Araújo de Lucena, Marcela Vasconcelos Montenegro, Railla Raquel Albino Santos Silva, Caroline O Fischer-Bacca, Juliana Giorgi, Jordan Strom Published: 2026 ABSTRACT AbstractBackgroundArtificial intelligence–enabled quantitative coronary computed tomography angiography (AI-QCCTA) offers automated assessment of coronary plaque burden and morphology. Although AI-QCCTA has improved diagnostic consistency and downstream testing efficiency, its prognostic value for major adverse cardiovascular events (MACE) has not been comprehensively quantified.MethodsWe systematically searched PubMed, Embase, and Cochrane through October 2025 for studies evaluating AI–based plaque analysis in patients without prior MACE undergoing CCTA. Outcomes of interest were pooled using random-effects GLMM models, and prognostic associations were synthesized using inverse-variance random-effects meta-analysis of hazard ratios (HRs). The primary endpoint was MACE; secondary outcomes included myocardial infarction (MI), revascularization, angina, stroke, and mortality. Subgroup analysis was done to identify the association of different plaque characteristics in predicting MACE/MI/Death.ResultsTen studies (n = 20,195) were included. Across six cohorts (n = 18,804), pooled rates were: all-cause mortality 1.20% (95% CI 0.38–3.77%), cardiovascular mortality 0.32% (0.21–0.48%), MACE 5.07% (1.25–18.46%), MI 1.30% (0.41–3.99%), and revascularization 13.09% (6.57–24.40%). AI-enabled plaque burden predicted MACE (HR 1.95, 95% CI 1.29–2.94; I2 = 99%), consistent in sensitivity analysis as per same AI platform use (HR 1.88, 95% CI 1.15–3.07). Low-attenuation plaque showed the strongest association (HR 2.95, 95% CI 1.95–4.45).ConclusionsAI-QCCTA provides prognostic value beyond stenosis severity, with vulnerable plaque characteristics-particularly low-attenuation and non-calcified plaque most strongly predicting adverse cardiovascular outcomes. These findings support the integration of AI-enabled plaque analysis into contemporary risk stratification.Keywords: Coronary CT angiography, Artificial intelligence, Major adverse cardiovascular events Introduction Coronary computed tomography angiography (CCTA) has become the preferred first-line noninvasive imaging modality for evaluating suspected coronary artery disease (CAD), offering high diagnostic accuracy, reliable prognostic value, and the ability to guide individualized preventive therapies [1,2]. Beyond identifying obstructive stenosis, CCTA provides detailed characterization of atherosclerotic plaque burden and composition, yielding critical insights for cardiovascular risk stratification and targeted prevention of acute coronary syndromes (ACS) [3,4]. Nevertheless, conventional CCTA interpretation remains limited by interobserver variability, with less experienced readers prone to overestimating stenosis severity, potentially leading to redundant testing, inappropriate revascularization referrals, and delayed intensification of preventive therapy [1,4,5]. Artificial intelligence, enabled quantitative coronary CCTA (AI-QCCTA) has emerged as a transformative advance in cardiac imaging, enabling automated, reproducible quantification of total and noncalcified plaque volumes, lesion morphology, and ischemia-related features [[6], [7], [8], [9], [10]]. Early multicenter studies have shown that AI-QCCTA enhances diagnostic precision, improves reproducibility, and refines stenosis classification [8,10,11]. In real-world analyses, AI-guided plaque quantification increased clinician diagnostic confidence in 69% of cases, reduced planned myocardial perfusion testing by 63%, and decreased invasive angiography referrals by 45%, while promoting greater adherence to guideline-directed medical therapy [12]. These findings underscore AI-QCCTA's capacity to integrate anatomical and functional insights into a single, standardized assessment [8,[13], [14], [15], [16]]. However, while AI-QCCTA has demonstrated meaningful impact on diagnostic accuracy and treatment pathways, its prognostic utility in predicting major adverse cardiovascular events (MACE) remains to be established across multiple prospective cohorts. The ability of AI-derived plaque metrics and ischemia signatures to predict future myocardial infarction, death, or unstable angina would position AI-QCCTA as a transformative tool that links imaging, risk-directed prevention, and outcomes in CAD. Accordingly, we performed a systematic review and meta-analysis to evaluate the prognostic performance of AI-QCCTA for MACE among patients undergoing CCTA for suspected CAD. Methods Eligibility criteria Studies were eligible if they met the following criteria [1]: Included patients without a prior history of MACE [2] Included population who underwent CCTA [3] Assessment of coronary plaque characteristics was done using Artificial Intelligence (AI) or deep learning (DL) [4], Observational study designs, including prospective and retrospective cohort studies, and registry-based analyses. Studies utilizing any AI/DL based platform for automated plaque quantification, including commercial (e.g., Cleerly, HeartFlow) and research based models (e.g., CNN, YOLO architectures) were included, regardless of differences in algorithm design or training datasets. Studies were excluded if they had patients with a prior history of MACE, previous myocardial infarction, or included pediatric populations. Studies that did not utilize AI or DL for plaque characterization were excluded. Randomized clinical trials without extractable observational outcomes, case reports, case series, reviews, editorials, letters, and conference abstracts without complete data were also excluded. Search strategy and data extraction A systematic search was conducted in PubMed, Embase, and the Cochrane Library from inception to October 2025. The search strategy used is mentioned in detail with Boolean operators and search terms included in Supplementary material provided. We implemented no language restrictions while conducting the search. Two reviewers independently extracted data for this review (M.M. and R.S.), resolving any disagreements through discussion and consensus. In accordance with Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines, all references identified during the search process were systematically screened before extraction [17]. Titles and abstracts were initially assessed for relevance, and non-eligible studies were removed. Full-text articles of the remaining records were then retrieved and evaluated against the inclusion criteria. For studies meeting eligibility criteria, the following information was extracted: author details; study setting and design; sample size; follow-up period; type of imaging used; AI/DL tool used; participant age, sex, and comorbidities; and outcomes of interest. Statistical analysis All statistical analyses were conducted using RevMan 5.4 and RStudio with the meta and metafor packages. For proportion-based clinical outcomes, we pooled event rates using a random-effects generalized linear mixed model (GLMM), which incorporates both within- and between-study variability. Pooled estimates were reported as proportions with corresponding 95% confidence intervals (CIs). Heterogeneity was evaluated using Cochrane's Q statistic and the I2 statistic, with values > 50% considered indicative of substantial heterogeneity [18]. Forest plots were generated to illustrate individual study estimates and pooled effects for angina, coronary revascularization, myocardial infarction (MI), all-cause mortality, stroke, and overall MACE. A leave-one-out sensitivity analysis was performed for proportion, in which each study was sequentially removed and models were refitted using random-effects estimators. This approach assessed the influence of individual studies on the stability of pooled estimates. To further explore heterogeneity, subgroup analyses were performed based on plaque characteristics and AI platform. For studies reporting prognostic associations between AI-enabled plaque metrics and cardiovascular outcomes, we performed a meta-analysis of hazard ratios (HRs). Log-transformed HRs and standard errors were extracted, and pooled estimates were generated using inverse-variance random-effects models. All analyses were conducted in RStudio (meta, metafor), with cross-validation in RevMan 5.4. Results were reported as pooled HRs with 95% confidence intervals (CIs). Heterogeneity was assessed using Cochrane's Q, I2, and Tau2 (REML). The primary analysis included studies reporting composite MACE, MI, or death using AI-enabled plaque burden assessment. Publication bias was assessed visually using a funnel plot of studies included. Endpoints and subgroup analyses The primary endpoint was major adverse cardiovascular events (MACE), defined as a composite of MI, coronary revascularization, unstable angina hospitalization, or cardiovascular death. Secondary endpoints included the individual components of MACE (MI, coronary revascularization, angina), as well as stroke and all-cause mortality. In addition to event pooling, we conducted a subgroup meta-analysis to assess the prognostic performance of different AI/DL-derived plaque characteristics, including, Total plaque volume (TPV), Percent atheroma volume (PAV), Non-calcified plaque volume (NCPV), Low-attenuation plaque volume (LAPV). These analyses evaluated the association of each plaque phenotype with the composite outcome of MACE/MI/death using separate random-effects models. Subgroup differences were assessed using χ2 tests for interaction. Additional sensitivity analysis was done as per AI model used. Quality assessment The risk of bias for all included prognostic studies was assessed using the Quality in Prognosis Studies (QUIPS) tool [19]. This framework assesses six methodological domains [1]: study participation [2], study attrition [3], prognostic factor measurement [4], outcome measurement [5], study confounding, and [6] statistical analysis and reporting. For each domain, two reviewers independently graded the risk of bias as low, moderate, or high according to established QUIPS criteria. Any discrepancies were resolved through discussion and consensus, with arbitration by a third reviewer when necessary. A summary table and domain-level judgments were generated to provide a transparent appraisal of study quality. Additionally, publication bias was assessed visually using funnel plot. Results Study selection Two independent reviewers (M.M. and L.L.) screened studies. In total, 210 records were retrieved from PubMed, 221 from Embase, and 23 from Cochrane. After removing duplicates, 332 unique records remained. Following the title and abstract evaluation, 263 records were excluded. Of the 69 full-text articles assessed for eligibility, 10 studies met all inclusion criteria. They were incorporated into the final quantitative synthesis, as illustrated in the PRISMA flow diagram shown in Fig. 1 [15,[20], [21], [22], [23], [24], [25], [26], [27], [28]]. Prisma flow diagram. Characteristics of included studies The 10 included studies comprised 20,195 participants (after excluding overlapping cohorts), predominantly middle-aged to older adults, with mean ages ranging from 54.7 to 66.2 years (Table 1). Across cohorts, 49–66% were male, and follow-up duration varied from 1 to 10.3 years. Cardiovascular risk factors were common, with hypertension present in 39–68%, hyperlipidemia in 27–73%, and smoking history in 12-55% of participants, while family history of CAD ranged from 9 to 53%. All studies employed CCTA-based imaging, supplemented in a select cohort by PET. AI- or DL-enabled plaque quantification platforms—including Cleerly Labs, YOLO-based models, hierarchical ConvLSTM networks, HeartFlow AI, and GE Deep Blue were used across studies. The prevalence of ≥50% CAD obstruction, when reported, ranged from 12.5% to 71%. Table 1Baseline characteristics of included studies.Table 1Number of ParticipantsStudyDesignCountryMethod of ImageAI/DL Tool usedMale, n (%)Follow-up, years (Mean or Median)Age, years (Mean or Median)Family history of CAD, n (%)Smokinga, n (%)Hypertension, n (%)Hyperlipidemia, n(%)≥50% CAD Obstruction, n (%)Bar et al. 2024201880ProspectiveCohortFinlandCCTAAI-Cleerly Labs1046 (55)6.9 [4.8–9.0]63.1 ± 9.5882 (47)633 (34)1059 (56)1218 (65)474 (26)Bar et al. 202422662ProspectiveCohortFinlandCCTA and PETAI-Cleerly Labs385 (58)6.2 [4.4-8.3]66.2 ± 8.1303 (46)272 (41)452 (68)481 (73)419 (66)Dundas et al. 2024234430ObservationalStudyMulticenterCCTAAI-HeartFlow2918 (66)166.2 ± 10.3NA2251 (55)2635 (60)NA3156 (71)Kim et al. 202321747Post hoc analysisCONSERVE trialMulticenterCCTAAI-Cleerly Labs384 (51)1.1 ± 0.460 ± 12.267 (9)224 (30)427 (57)249 (33)97 (13)Kim et al. 202524408Retrospective CohortMulticenterCCTADL- YOLO-based model224 (55)2.559.4 ± 14.638 (9)137 (33)208 (51)110 (27)152 (37)Lin et al. 2022251611Retrospective CohortMulticenterCCTADL- Hierarchical ConvLSTM network780 (65)4·7 [4.0–5.7]65.2 ± 9.8194 (16)448 (37)650 (54)746 (62)Not reportedMiller et al. 2024261847ObservationalStudyMulticenterCCTADL- Hierarchical ConvLSTM network1072(58)4.8 and 2.5*59 and 65*755(40.8)317(17)719(39)942(51)Not reportedNurmohamed et al. 202315536Prospective cohortAmsterdamCCTAAI-Cleerly Labs297 (55)10.3 [8.6-11.5]58.6 ± 9.2284 (53)184 (35)250 (47)195 (37)193 (36)Rosendael et al. 2025271916Observational StudyMulticenterCCTAAI-Cleerly Labs937 (49)359 ± 16570 (28)229 (12)994 (52)802 (42)491 (26)Song et al. 2025288431Prospective cohortChinaCCTADL- GE Healthcare Deep Blue automated plaque analysis4065 (48)3.6854.7 ± 10.23397 (40)2346 (28)3658 (43)3172 (38)1056 (12.5)Values are expressed as counts (percentage), mean ± standard deviation, or median [interquartile range]. CCTA - coronary computed tomography angiography; PET - positron emission tomography; AI- Artificial Intelligence; DL - deep learning; CAD - coronary artery disease; a Current or past smoking; NA - not available. *- Two cohorts included in the study respectively. Baseline characteristics of included studies. Values are expressed as counts (percentage), mean ± standard deviation, or median [interquartile range]. CCTA - coronary computed tomography angiography; PET - positron emission tomography; AI- Artificial Intelligence; DL - deep learning; CAD - coronary artery disease; a Current or past smoking; NA - not available. *- Two cohorts included in the study respectively. Primary endpoint Pooled event rates of cardiovascular outcomes 18,804 patients were included in the analysis across six cohorts. As shown in Fig. 2, the pooled all-cause mortality rate was 1.20% (95% CI 0.38–3.77%). Cardiovascular mortality was substantially lower, at 0.32% (95% CI 0.21–0.48%). Angina prevalence varied widely across studies, with a pooled estimate of 21.9% (95% CI 1.59 -- 82.56%), while coronary revascularization occurred in 13.09% (95% CI 6.57 -- 24.40%) of patients. MACE occurred in 5.07% (95% CI 1.25 -- 18.46%), and MI occurred in 1.30% (95% CI 0.41 -- 3.99%). Forest plot of AI-based plaque quantification for predicting cardiovascular outcomes per patient. Forest plot summarizing the pooled incidence of major adverse cardiovascular events (MACE), myocardial infarction (MI), coronary revascularization, and all-cause mortality assessed with AI-enabled quantitative plaque analysis on CCTA. Leave-one-out sensitivity analyses for patient-level outcomes (Supplementary Table 1) again demonstrated stability of the pooled estimates, with no disproportionate influence from any individual study. Prognostic value of AI-enabled plaque burden analysis for MACE, MI, or death The primary pooled analysis evaluating AI-derived plaque burden for prediction of composite cardiovascular outcomes is depicted in Fig. 3. Across seven independent studies, AI-enabled plaque burden was significantly associated with incident MACE, yielding a pooled HR of 1.95 (95% CI 1.29–2.94) with substantial heterogeneity (I2 = 99%, p < 0.00001) [15,21,23,[25], [26], [27], [28]]. A prespecified sensitivity analysis limited to studies explicitly using same AI platform (Cleerly Labs) for Plaque analysis (Fig. 4) demonstrated a pooled HR of 1.88 (95% CI 1.15–3.07) with high heterogeneity (I2 = 92%) [15,20,21,27]. Forest Plot of AI enable plaque burden analysis on MACE. Forest Plot of Sensitivity Analysis performed on AI enable plaque burden analysis for prediction of MACE as per similar AI platform used. AI-enabled stenosis or AI-QCCTA analysis for MACE, MI, or death Some studies have reported AI-enabled coronary artery stenosis assessment for the prediction of MI/Death, with a pooled analysis shown in Fig. 5, yielding a pooled HR of 1.56 (95% CI, 0.84–2.92; p = 0.16) [24,26]. Forest plot of AI-enable stenosis analysis on predicting MI/death. Subgroup analysis: predictive value of specific plaque characteristics Subgroup analyses (Fig. 6) revealed differential prognostic effects across AI-quantified plaque components. ●Total plaque volume: HR 1.70 (95% CI 0.99–2.92) Total plaque volume: HR 1.70 (95% CI 0.99–2.92) ●Percent atheroma volume: HR 2.78 (95% CI 1.88–4.13) Percent atheroma volume: HR 2.78 (95% CI 1.88–4.13) ●Non-calcified plaque volume: HR 2.15 (95% CI 1.12–4.11) Non-calcified plaque volume: HR 2.15 (95% CI 1.12–4.11) ●Low-attenuation plaque volume: HR 2.95 (95% CI 1.95–4.45) Low-attenuation plaque volume: HR 2.95 (95% CI 1.95–4.45) Forest plot of Subgroup analysis performed on different plaque characteristics predicting MACE. The pooled subgroup HR was 2.25 (95% CI 1.68–3.01), suggesting that AI derived plaque phenotyping provides strong prognostic value across multiple plaque domains, with low-attenuation plaque demonstrating the strongest association with future events. Risk of bias assessment Risk of bias was assessed using the QUIPS tool across six domains. Overall, all the included studies except for Kim et al. (2023) demonstrated low risk of bias (Table 2). Table 2Risk of bias of included studies.Table 2Author, yearStudy ParticipationStudy AttritionPrognostic Factor MeasurementOutcome MeasurementStudy ConfoundingStatistical Analysis and ReportingOverallBar et al., 2024LowLowLowLowModerateLowLowBar et al., 2024 (small cohort)ModerateLowLowLowModerateLowLowDundas et al., 2024LowLowLowModerateModerateLowLowKim et al., 2023ModerateLowLowModerateModerateModerateModerateKim et al., 2025ModerateModerateLowLowModerateLowLowLin et al., 2022ModerateLowLowLowModerateLowLowMiller et al., 2024LowLowLowModerateModerateLowLowNurmohamed et al., 2023LowLowLowLowModerateLowLowRosendael et al., 202LowLowLowLowModerateLowLowSong et al., 2025LowLowLowLowModerateLowLow Risk of bias of included studies. Kim et al. (2023), Kim et al. (2025), and Lin et al. (2022) demonstrated moderate risk in study participation or attrition, likely due to cohort selection or follow-up variability. Outcome measurement was generally strong, although moderate risk was observed in studies such as Dundas et al. (2024), Kim et al. (2023), and Miller et al. (2024) due to variability in endpoint definitions. The primary limitation across studies was moderate risk of confounding, reflecting incomplete adjustment for clinical covariates. Statistical analysis and reporting were largely appropriate, with only Kim et al. (2023) demonstrating moderate concerns. Taken together, the methodological quality was acceptable, supporting the validity of the pooled findings, although residual confounding should be considered when interpreting results. Publication bias assessment The funnel plot assessing the association between AI-enabled CCTA and hazard ratio (HR) of MACE is shown in Supplementary Fig. 2. The plot demonstrates a partially asymmetric distribution of studies around the pooled effect estimate. Visually, larger studies (lower standard error, near the top of the plot) are clustered close to the pooled estimate, indicating relatively stable and precise effect sizes. In contrast, smaller studies (higher standard error, toward the bottom) are more widely dispersed and tend to demonstrate larger effect estimates, predominantly on one side of the funnel. Notably, there is a relative paucity of smaller studies reporting neutral or negative associations (i.e., HR closer to 1) on the left side of the funnel. This pattern suggests the presence of mild asymmetry, which may reflect small-study effects, including potential publication bias, whereby smaller studies with non-significant findings are less likely to be published. Discussion In this systematic review and meta-analysis, we demonstrate that patients without prior MACE undergoing CCTA experience a non-trivial burden of subsequent cardiovascular events and that AI/DL-enabled quantitative plaque analysis provides meaningful prognostic information beyond stenosis assessment. Across more than 20,000 patients from 10 cohorts, pooled analyses revealed appreciable rates of MACE, myocardial infarction, angina, and coronary revascularization, underscoring residual risk in populations frequently classified as low risk by conventional metrics. Quantitative AI-derived plaque burden was consistently associated with adverse outcomes, nearly doubling the risk of composite MACE, MI, or death, with stronger associations observed for hard endpoints. Among individual plaque features, non-calcified and low-attenuation plaque volumes demonstrated the most strong and reproducible prognostic value, supporting their role as imaging markers of plaque vulnerability. In contrast, AI-based stenosis severity alone showed limited prognostic discriminative ability, highlighting the shortcomings of a lumen-centric evaluation. Despite substantial heterogeneity, extensive sensitivity analyses confirmed the stability of these findings. Collectively, these data support AI-enabled plaque quantification as a clinically relevant approach for enhanced risk stratification and individualized preventive strategies. Our findings extend and quantitatively consolidate prior observational and trial-level evidence suggesting that coronary plaque burden and composition convey prognostic information beyond angiographic stenosis severity [29,30]. Earlier studies using visual or semi-quantitative CCTA assessment demonstrated associations between high-risk plaque features and future events, however, their clinical applicability has been limited by interobserver variability and lack of standardized thresholds [31,32]. By aggregating data from contemporary cohorts using fully automated or AI-enabled plaque quantification, our analysis confirms and strengthens these observations at scale, demonstrating consistent risk gradients across total plaque burden and specific vulnerable phenotypes [7,15,33,34]. Notably, low-attenuation and non-calcified plaque volumes emerged as the most reproducible predictors of adverse outcomes. These findings are concordant with prior mechanistic and intravascular imaging studies linking lipid-rich plaque to instability and thrombosis [31,35]. In contrast, the comparatively weaker prognostic performance of AI-derived stenosis severity aligns with growing evidence that luminal narrowing alone inadequately captures atherosclerotic risk, particularly in earlier disease stages. Together, these results position AI-enabled plaque analysis as an essential methodological advance that bridges high-resolution coronary imaging with clinically actionable risk stratification [15,30,33]. Prior meta-analyses of artificial intelligence in coronary imaging have primarily focused on diagnostic performance rather than prognostic utility. A meta-analysis of AI-based CAC scoring (25 studies) demonstrated strong agreement with manual assessment (kappa 0.83), while meta-analysis of AI-assisted CCTA have shown excellent diagnostic accuracy, with pooled AUCs of 0.93–0.96 and high sensitivity for detecting ≥50% stenosis and another meta-analysis confirmed that AI-based CCTA demonstrated superior diagnostic performance compared to clinicians in identifying ≥50% stenosis (sensitivity 0.92 vs 0.85; AUC 0.96 vs 0.91) [[36], [37], [38], [39]]. Although these findings establish AI as a reliable diagnostic tool, they do not address whether AI-derived imaging metrics predict future cardiovascular events. In contrast, our meta-analysis provides the first comprehensive synthesis of the prognostic value of AI-enabled CCTA based plaque analysis, demonstrating significant associations with MACE, MI, and death. These results extend AI applications beyond diagnosis to risk stratification, supporting its role in the evolving paradigm of precision cardiovascular prevention. The present findings have important clinical implications for CCTA guided risk assessment. By systematically quantifying total plaque burden and vulnerable plaque phenotypes, particularly non-calcified and low-attenuation plaque, AI-QCCTA may support more precise cardiovascular risk stratification and inform earlier, personalized preventive strategies [15,30,40]. However, while the prognostic value of AI-QCCTA is increasingly well established, several practical considerations must be addressed before routine clinical implementation. The economic landscape for AI-QCCTA remains a significant barrier. The research and development costs for AI plaque analysis platforms often run into millions of dollars, and per analysis costs can reach several hundreds to thousands of dollars, substantially exceeding the reimbursement for the base CCTA examination, which is typically in the low hundreds of dollars in most state funded systems [[41], [42], [43]]. The concentration of AI-QCCTA platforms in high resource academic centers raises concerns about equitable access. Most included studies in our meta-analysis originated from well resourced institutions in North America, Europe, and East Asia, with limited representation from low and middle income countries. As noted in the AHA Scientific Statement on AI in cardiovascular imaging, AI models are prone to performance degradation when encountering out-of-distribution data from changes in imaging protocols, equipment, or target populations [44]. Prospective external validation across geographically and ethnically diverse cohorts, using standardized acquisition protocols and multiple scanner platforms, is essential before widespread adoption. Future research should prioritize prospective, multicenter studies integrating AI/DL-derived plaque metrics into clinical decision-making pathways to determine whether imaging-guided risk stratification improves hard clinical outcomes. Harmonization of AI/DL methodologies, plaque definitions, and reporting standards will be essential to facilitate reproducibility and regulatory translation [44,45]. Additionally, integration of AI/DL-enabled plaque analysis with clinical risk scores, circulating biomarkers, and longitudinal imaging may enable dynamic risk modeling across the atherosclerotic disease continuum. Randomized trials evaluating therapy escalation guided by AI/DL-derived plaque vulnerability represent a critical next step toward precision prevention in coronary artery disease [15,30]. Limitations Several limitations merit consideration. First, despite comprehensive inclusion criteria and sensitivity analyses, substantial between-study heterogeneity was observed, reflecting differences in study design, population risk profiles, follow-up duration, endpoint definitions, and AI/DL algorithms used for plaque quantification. Although leave-one-out analyses confirmed the stability of pooled estimates, residual confounding cannot be excluded. Second, most included studies were observational, limiting causal inference and susceptibility to unmeasured confounders, including downstream management changes triggered by CCTA findings. Third, variability in plaque feature definitions, acquisition protocols, and AI platforms may affect generalizability and preclude the identification of universal prognostic thresholds. Fourth, patient-level analyses were reported inconsistently across studies, limiting direct comparisons between outcome reporting frameworks. Fifth, some included studies originated from overlapping source cohorts (Cedars-Sinai, SCOT-HEART, St. Francis), introducing the possibility of non-independent samples. Although we minimized this by selecting only the most comprehensive dataset per cohort for pooling, the potential for residual overlap remains a limitation. Sixth, there is possibility of potential publication bias, however, interpretation of funnel plot asymmetry should be made with caution. The observed pattern may also be influenced by between-study heterogeneity, differences in AI algorithms, population risk profiles, and variability in outcome definitions across studies. Additionally, the relatively limited number of included studies reduces the reliability of visual and statistical assessment of publication bias. Finally, individual patient-level data were unavailable, precluding time-to-event harmonization and more granular subgroup analyses. These limitations underscore the need for prospective standardization and validation across diverse populations. Conclusions In this systematic review and meta-analysis, AI/DL-enabled quantitative coronary plaque analysis demonstrated strong prognostic value for future cardiovascular events in patients without prior MACE. Vulnerable plaque phenotypes, particularly non-calcified and low-attenuation plaque volumes, were consistently associated with increased risk of adverse outcomes. Despite heterogeneity across studies, findings were stable and reproducible, supporting the role of AI/DL-enhanced plaque characterization as a clinically meaningful tool for refined risk stratification. These findings are hypothesis generating and warrant further prospective validation before considering integration into routine cardiovascular care. Consent to participate Not applicable. Consent to publish Not applicable. Ethics approval Not applicable. This is a meta-analysis. The Sirio Libanes Research Ethics Committee has confirmed that no ethical approval is required. Code availability Not applicable. Data availability statement The data used in this meta-analysis were extracted from publicly available sources, including published observational and randomized controlled trials. All relevant data supporting the findings of this study are included within the manuscript and supplementary materials. Due to the nature of this analysis, no new primary data were generated. Additional details is available at https://doi.org/10.5281/zenodo.18041122. Editing support The authors used Grammarly (Grammarly Inc.) to assist with grammar, spelling, and clarity during manuscript preparation. All interpretations and conclusions are solely those of the authors. Funding The authors declare that no funds, grants, or other support were received during the preparation of this manuscript. CRediT authorship contribution statement Mushrin Malik: Conceptualization, Investigation, Supervision, Writing – original draft, Writing – review & editing. Pedro Batista: Methodology. Larissa Araújo de Lucena: Methodology, Writing – original draft. Marcela Vasconcelos Montenegro: Investigation, Writing – original draft. Railla Raquel Albino Santos Silva: Writing – original draft. Caroline O. Fischer-Bacca: Writing – original draft. Juliana Giorgi: Writing – original draft, Writing – review & editing. Jordan Strom: Supervision, Writing – review & editing. Declaration of competing interests All authors certify that they have no affiliations with or involvement in any organization or entity with any financial or non-financial interest in the subject matter or materials discussed in this manuscript. Footnotes Supplementary data to this article can be found online at https://doi.org/10.1016/j.ijcrp.2026.200660. Contributor Information Mushrin Malik, Email: mushrinmalik@gmail.com. Pedro Batista, Email: pedro.batista@academico.ufpb.br. Larissa Araújo de Lucena, Email: larissa.js@hotmail.com. Marcela Vasconcelos Montenegro, Email: marcela.montenegro@upe.br. Railla Raquel Albino Santos Silva, Email: raillaalbino@gmail.com. Caroline O. Fischer-Bacca, Email: carolinefischerbacca@yahoo.com.br. Juliana Giorgi, Email: julianagiorgi@hotmail.com. Jordan Strom, Email: jstrom@bidmc.harvard.edu. Abbreviations: Major Adverse Cardiovascular Events Myocardial Infarction Coronary Computed Tomography Angiography coronary artery disease Acute Coronary Syndrome Artificial Intelligence–Guided Quantitative Coronary Computed Tomography Angiography Deep learning Convolutional Neural Networks You only look once Hazard Ratio Confidence Interval Percent Atheroma Volume Total Plaque Volume Non Calcified Plaque Volume Low-Attenuation Plaque Volume Fractional Flow Reserve Derived From Computed Tomography Area Under the Curve Positron Emission Tomography Myocardial Perfusion Imaging Coronary Artery Calcium Coronary Artery Disease–Reporting and Data System Preferred Reporting Items for Systematic Reviews and Meta-Analyses Generalized Linear Mixed Model Quality in Prognosis Studies Supplementary data The following are the Supplementary data to this article. Multimedia component 1 Multimedia component 2 References Associated Data Supplementary Materials Multimedia component 1 Multimedia component 2 Data Availability Statement The data used in this meta-analysis were extracted from publicly available sources, including published observational and randomized controlled trials. All relevant data supporting the findings of this study are included within the manuscript and supplementary materials. Due to the nature of this analysis, no new primary data were generated. Additional details is available at https://doi.org/10.5281/zenodo.18041122.