Skip to main content
Download PDF
  • Systematic Review
  • Open access
  • Published:

Comparison of machine learning models with conventional statistical methods for prediction of percutaneous coronary intervention outcomes: a systematic review and meta-analysis

BMC Cardiovascular Disorders volume 25, Article number: 310 (2025) Cite this article

Abstract

Introduction

Percutaneous coronary intervention (PCI) has been the main treatment of coronary artery disease (CAD). In this review, we aimed to compare the performance of machine learning (ML) vs. logistic regression (LR) models in predicting different outcomes after PCI.

Methods

Studies using ML or deep learning (DL) models to predict mortality, MACE, in-hospital bleeding, and acute kidney injury (AKI) after PCI or primary PCI were included. Articles were excluded if they did not provide a c-statistic, solely used ML models for feature selection, were not in English, or only used logistic or LASSO regression models. Best-performing ML and LR-based models (LR model or conventional risk score) from the same studies were pooled separately to directly compare the performance of ML versus LR. Risk of bias was assessed using the PROBAST and CHARMS checklists.

Results

A total of 59 studies were included. Meta-analysis showed that ML models resulted in a higher c-statistic compared to LR in long-term mortality (0.84 vs. 0.79, P-value = 0.178), short-term mortality (0.91 vs. 0.85, P = 0.149), bleeding (0.81 vs. 0.77 P = 0.261), acute kidney injury (AKI; 0.81 vs. 0.75, P = 0.373), and major adverse cardiac events (MACE; 0.85 vs. 0.75, P = 0.406). PROBAST analysis showed that 93% of long-term mortality, 70% of short-term mortality, 89% of bleeding, 69% of AKI, and 86% of MACE studies had a high risk of bias.

Conclusion

No statistical significance existed between ML and LR model. In addition, the high risk of bias in ML studies and complexity in interpretation undermines their validity and may impact their adaption in a clinical settings.

Significance

What is already known on this topic

Many ML models are available for predicting adverse complications of PCI. However some methodology and performance concerns made it hard to choose between well-established statistical vs. ML models.

What this study adds

The overall ML and LR models c-statistics were comparable for short- and long-term mortality, bleeding, AKI, and MACE prediction. Our risk of bias assessment (using CHARMS and PROBAST checklists) identified a high risk of bias and applicability concerns.

How this study might affect research, practice, or policy

Future studies should consider the reporting checklists to improve their methodology.

Peer Review reports

Introduction

Percutaneous coronary intervention (PCI) has been the mainstay in the treatment of coronary artery disease (CAD) since it was introduced in 1977 [1]. Despite several advances in PCI technology, post-procedural complications such as acute kidney injury, bleeding, and mortality are not uncommon [2]. As such, several prediction models like the United States National Cardiovascular Data Registry Risk Score (NCDR-CathPCI risk score) [3, 4], Mehran Score [5, 6], and New York State Risk Score [7] have been developed to identify high-risk patients. Prediction models have also been used to assess patient prognosis. For instance, SYNTAX score II has been utilized in predicting long-term mortality after PCI [8]. All of these models have been based on conventional statistical methods like logistic regression (LR).

In contrast to statistical methods such as LR, ML refers to a set of computational techniques that automatically learn patterns from data to make predictions or decisions, rather than relying solely on explicitly programmed instructions. Unlike traditional statistical methods that often focus on hypothesis testing and inferring relationships between variables, machine learning is primarily concerned with prediction accuracy and pattern recognition [9].

While LR models the relationship between predictors and a binary outcome using an interpretable formula, ML models such as random forests or neural networks/deep learning models can capture complex, non-linear relationships in the data. This allows these models to potentially identify subtle interactions among variables that might be missed by traditional approaches. However, these advantages come with challenges: ML models may require larger datasets, careful tuning to avoid overfitting, and sometimes yield models that are less immediately interpretable than logistic regression [9]. There is also a widespread, difficult to navigate gap between achieving good performance metrics in internal/external validation (for ML models) and delivering clinical utility. Most studies on ML models in medicine use retrospective data, which limits the validity of their evidence; the adoption of randomized clinical trials as the gold standard for evaluating clinical utility is also lagging behind for ML models [9].

With the increasing adoption of artificial intelligence methods (including ML), the cardiology field as with many other areas of medicine has been promised better predictive accuracy, especially in scenarios with big complex datasets and non-linear relationships between the variables [10]. This has inevitably resulted in a huge surge in the number of papers using machine learning (ML) models to predict post-PCI complications and adverse events [2, 11]. However, as journals may lean towards accepting articles with better predictive performance (higher c-statistic), many ML models may inadvertently have overfitting problems caused by inappropriate methodology [12]. In addition, it is still unclear whether we should move on from well-established statistical models to ML in clinical practice and prognosis assessment of PCI patients. This makes a systematic review and critical appraisal of the literature imperative. Therefore, in the current investigation, we aimed to (1) critically review the available studies that used ML prediction models for post-PCI outcomes and (2) compare the pooled estimates of ML models and conventional risk scores or LR whenever possible.

Methods

The protocol for this systematic review and meta-analysis was registered in the international prospective register of systematic reviews (CRD42023494659). PRISMA 2020 statement was used for reporting this systemtic review and meta-analysis [13].

Patient and public involvement

Patients or the public were not involved in the design, or conduct, or reporting, or dissemination plans of our research.

Eligibility criteria and outcome definition

Studies using ML or deep learning (DL) models to predict mortality, MACE, in-hospital bleeding, and acute kidney injury (AKI) after PCI or primary PCI were deemed eligible. Due to the small number of studies, no exclusion criteria were set for CAD type. However, studies evaluating patients with chronic total occlusion were not included. Articles were also excluded if they did not provide a c-statistic, solely used ML models for feature selection, were not in English, or only used logistic or LASSO regression models. Studies using similar datasets were included for risk of bias assessment, but only the investigation with a higher number of patients was considered in the meta-analysis.

Short-term mortality were considered as < 1 year follow-up, while long-term mortality were identified as ≥ 1 year of follow-up. MACE was defined if it was a combination of at least three of the following five components: death, myocardial infarction, coronary revascularization, stroke, and hospitalization because of heart failure. No prior definitions of bleeding and AKI were used to screen articles as studies used different criteria.

Search strategy

The search for the current study was conducted on PubMed, Embase, Web of Science, and Scopus from inception until December 11th, 2023. The search strategy comprised two components: (1) “machine learning” AND (2) “percutaneous coronary intervention”. The full search template is available in online Supplementary File S1.

Selection process and data gathering

Two independent reviewers (A.V and S.N) screened the articles first based on their title/abstract and then based on the full text. Inconsistencies were solved by consensus. Similarly, data collection was done independently by A.H and A.M.

Risk of Bias assessment

Critical appraisal and data extraction for systematic reviews of prediction modeling studies (CHARMS) checklist [14] and prediction model risk of bias assessment tool (PROBAST) [15] were used by two independent reviewers (S.N and A.H) to assess all of the included studies. The CHARMS and PROBAST Excel template developed by Fernandez-Felix et al. was utilized for this purpose [16]. Any discrepancies between the reviewers were solved by consensus. The assessment was only done for the best-performing ML model on the validation dataset.

Data analysis

C-statistics, the area under the receiver operating curves (AUC ROCs), were pooled using random effects meta-analysis. If the corresponding 95% confidence interval was not provided, it was calculated using the number of events and sample size based on the methods proposed by Hanley and McNeil [17]. Best-performing ML and LR-based models (LR model or conventional risk score) from the same studies were pooled separately to directly compare the performance of ML versus LR. This also ensured that heterogeneity stemming from study methodology and population would be limited. Furthermore, we performed secondary comparisons based on whether the LR models had the same number and type of features that ML models had. The pooled estimates were compared using the MedCalc online calculator which is based on the Hanley and McNeil method [17, 18]. All analysis was performed using R statistical software version 4.2.1 and metamisc package [19].

Results

Overall, 59 studies were included in the current systemic review from which 15 were on long-term mortality [20,21,22,23,24,25,26,27,28,29,30,31,32,33,34], 25 on short-term mortality [2, 10, 32, 33, 35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55], nine on bleeding [2, 10, 49, 51, 52, 56,57,58,59], 16 on AKI [2, 10, 30, 51, 52, 60,61,62,63,64,65,66,67,68,69,70], and seven on MACE [41, 71,72,73,74,75,76] (Fig. 1). Excluded articles in the full-text screening and reasons for exclusion are provided in the Supplementary Material 1.

Fig. 1
figure 1

PRISMA 2020 flow diagram of study selection

Full size image

Long-term mortality

Supplementary Table S1-2 (Supplementary Material 2) demonstrate the general characteristics of the studies. In brief, fifteen articles assessed the performance of ML on long-term mortality of which seven (46%) were included in the meta-analysis [20, 23, 24, 27, 29, 32, 34]. 40% (6/15) of the studies did not report an event per variable (EPV) [22,23,24, 31, 33, 34], while in the others this figure was from 1.1 to 28.8 with only one study [26] having an EPV > 10. No studies used multiple imputations for handling the missing values and 53% (8/15) studies did not report their methods for missing data [22, 24,25,26,27, 29, 31, 32]. Only 40% (6/15) studies reported model calibration [20, 25,26,27,28,29], and 26% (4/15) had an external validation dataset [20, 26, 28, 29].

One study (6%) had a low risk of bias [20], whereas all other studies, 93% (14/15), had a high risk of bias [21,22,23,24,25,26,27,28,29,30,31,32,33,34]. The majority of the risk of bias, 93%, was stemming from the analysis domain. Regarding applicability, 40% (6/15) had a low concern [22, 25, 26, 29, 30, 34], 46% (7/15) had a high concern [20, 21, 23, 24, 27, 28, 31], and 13% (2/15) were unclear [32, 33]. The detailed risk of bias assessment using CHARMS and PROBAST is provided in Excel in supplementary material 4 and Supplementary Figure S1 (Supplementary Material 3)

Meta-analysis showed that ML models resulted in a 5% higher c-statistic compared to LR (0.84 vs. 0.79, P-value = 0.178). This number was 3% when comparing similar features ML and LR, and 6% for different features ML and LR (0.83 vs. 0.77, P = 0.230). However, these differences were not statistically significant (Fig. 2, Supplementary Figure S2-3). Assessment of the funnel plots revealed no asymmetry (Supplementary Figure S4).

Fig. 2
figure 2

(A) Pooled c-statistic of the best performing ML (left) vs. LR/ risk scores (right) models for long-term mortality. (B) Pooled c-statistic of the best performing ML (left) vs. LR/ risk scores (right) models for short-term mortality

Full size image

Short-term mortality

Twenty-five studies which had developed models for predicting short-term mortality as an adverse effect of PCI were included in our review, 10 of which were assessed in the meta-analysis [2, 32, 36, 37, 40, 41, 43, 48, 53, 54]. EPV values were in the range 0.4–52, and only 5 studies had an EPV of > 10 [2, 40, 51, 52, 54]. A single study [54] utilized multiple imputation for handling missing data, 3 studies utilized single imputation [39, 50, 52], 8 studies utilized other procedures or provided unclear explanations on their approach to missing data [2, 10, 33, 40, 42, 47, 48, 55], and 13 studies didn’t provide any information on their approach to missing data [32, 35,36,37,38, 41, 43,44,45,46, 49, 51, 53]. (Supplementary Table S3-4).

Risk of bias for these studies is evaluated in supplementary material 5 and Supplementary Figure S4.

Pooled c-statistics were overall 6% higher for ML models vs. statistical models (Fig. 2); specifically 5% higher in the case of models with different numbers of features (Supplementary Figure S5), and 6% higher in the case of models with a similar number of features (Supplementary Figure S6). Notably, none of these differences were statistically significant (Table 1).

Table 1 Comparison of the pooled c-statistics of the ML and LR models
Full size table

Bleeding

Nine studies evaluated in-hospital bleeding with an average age of 60 to 77. In 33% (3/9) [49, 56, 57], the EPV was unknown. 11% (1/9) of the investigations used multiple imputations for handling the missing data [56] and 22% (2/9) used single imputation [2, 57]. No studies had an external validation dataset, while only one study (11%) reported results using cross-validation [57]. (Supplementary Table S5-6).

Risk of bias is available at Supplementary Material 6 and Supplementary Figure S7. Meta-analysis results showed a 4% net benefit for the ML models over LR without statistical significance (0.81 vs. 0.77, P = 0.261, Fig. 3).

Fig. 3
figure 3

(A) Pooled c-statistic of the best performing ML (left) vs. LR/ risk scores (right) models for bleeding. (B) Pooled c-statistic of the best performing ML (left) vs. LR (right) models for AKI. (C) Pooled c-statistic of the best performing ML (left) vs. LR/ Cox proportional hazard (right) models for MACE

Full size image

AKI

We included 16 ML studies of AKI prediction after PCI (Mean age: 62.5–70). The best-performing models had a c-statistic of 0.74–0.89. One study [61] used multiple and five studies [2, 30, 62, 63, 66] used single imputation methods for handling the missing data, half of the included studies (8/16) did not clearly state their approach to the missing data [51, 52, 60, 64, 65, 68,69,70]. Five studies (31%) validated their models on external datasets [61, 63, 65,66,67]. Eight studies examined the models’ calibration (50%) [2, 10, 62,63,64,65,66,67]. Five studies (31%) did not perform feature selection [2, 10, 60, 63, 65], five (31%) had insufficient or unclear information on choosing final variables [51, 52, 61, 64, 69], and one study (6%) used stepwise method [70]. Five studies reported EPV [41, 71,72,73, 76], and only one study had EPV > 10 [71] (Supplementary Table S7-8).

Ris of bias is demonstrated in Supplementary Material 7 and Supplementary Figure S8. Four studies were included in the meta-analysis to compare the overall ML and LR performance [2, 61, 65, 68], and the pooled c-statistics of ML and LR models were comparable (0.81 vs. 0.75, P = 0.373, Fig. 3). Our secondary analysis also reached almost identical results that there was no significant difference between ML and LR models, when different (0.78 vs. 0.73, P = 0.337) or similar features (0.81 vs. 0.75, P = 0.373) were used in the model development (Supplementary Figure S9-10).

MACE

Seven studies developed and validated prediction models of MACE in PCI patients. The c-statistics of the best ML ranged from 0.7 to 0.95 (Mean age: 60–69). Six (86%) studies developed prediction models on one to ten-year MACE [71,72,73,74,75,76], and one study investigated the in-hospital MACE [41]. Four studies embedded all of the candidate predictors into the final model [41, 71, 73, 74], one used the stepwise selection method [75], and two used random forest-based algorithms [72, 76]. Only two studies used external validation for their models [41, 76], and two measured calibration of the models [41, 76]. Two studies developed survival random forest models for MACE models [72, 73] (Supplementary Table S9-10).

Supplementary Material 8 and Supplementary Figure S11 show the risk of bias evaluation. Four studies were examined in the ML vs. LR meta-analysis [71, 73, 74, 76]. The pooled c-statistic of ML models were comparable to LR models (0.85 vs. 0.75, P = 0.406, Fig. 3).

Discussion

To our knowledge, this was the first systematic review and meta-analysis of the ML models in PCI for CAD patients. Our results revealed that ML models had a net benefit over LR in several outcomes including, mortality, MACE, AKI, and bleeding after PCI, however, there was no statistically significant difference.

The risk of bias analysis of the included studies identified multiple concerns. Several studies did not provide an external validation dataset which could result in overfitting. Overfitting happens when the model has memorized the training data too well including the noise rather than identifying the patterns. As internal validation has a similar source to training data, results in the internal validation may be too optimistic. Therefore, PROBAST recommends against the use of simple data splitting into train and validation sets, while encouraging using cross-validation [77,78,79]. However, even simple cross-validation in ML prediction studies may be problematic as studies risk data leakage. This is because many ML models require fine-tuning hyperparameters during the cross-validation, resulting in finding the optimal hyperparameters for the validation dataset that was supposed to be unseen data. The solution is a nest-cross validation approach in which in the outer loops the model performance is assessed, while in the inner loops (training data) the hyperparameters are optimized [80]. This is a critical issue that was routinely neglected in the evaluated studies.

A second source of data leakage occurs when data pre-processing, such as missing data imputation, feature selection, or data normalization, is performed before the data is split. Conducting these steps prior to partitioning the validation data may lead to data leakage, potentially resulting in overfitting. However, This too is difficult to assess in ML studies as it is rarely discussed in detail.

A recurrent issue identified in the reviewed literature was the inadequacy of the EPV. PROBAST guideline advises a minimum EPV of 10 for traditional modeling approaches; however, this figure may be insufficient for machine learning (ML) techniques [81]. Research indicates that ML models, including RF, SVM, and ANN, might require an EPV that is at least 10 times higher than that of the LR [81]. This was often overlooked in the reviewed articles with numerous studies reporting EPVs below 10, far from the suggested threshold of 200 [77, 81].

The objective of many articles was to compare ML models with LR or traditional risk scores [2, 20, 34, 61, 66]. However, a discrepancy often existed in the number of features used; LR models typically included fewer features compared to the more extensive feature set selected for ML models. This disparity led to overly favorable results for ML models. Additionally, numerous studies benchmarked the performance of ML models against established risk scores such as the GRACE or SYNTAX scores. Given that ML models are developed datasets similar to those they are tested against, it is plausible to anticipate superior performance over traditional risk scores, which are derived from different datasets and incorporate a more limited number of features. To address this issue, we ensured that when possible the ML models were compared using a feature set analogous to that of the LR models. This was done alongside the broader comparison encompassing all models to provide a balanced evaluation. Neither of the analyses showed statistical significance, however, in long-term mortality, the net benefit of ML models with a similar feature set to LR was only 3% in comparison to 6% when comparing ML models with more features than LR.

Another significant issue identified in the reviewed articles was the absence of model calibration reporting. Calibration refers to the alignment of model outcomes with the actual likelihood of an event’s occurrence. This is crucial, especially when precise predicted probabilities are needed alongside a binary outcome. Such detailed information can greatly aid in clinical decision-making, as it allows a clinician to understand the likelihood of an event occurring, rather than simply receiving a binary yes or no answer [82].

In our study, outcomes such as MACE and long-term mortality were associated with time-to-event data. This implies that, in addition to determining the probability of an outcome, it is crucial to predict the timing of its occurrence. Despite this importance, our review found that none of the long-term mortality studies and only two MACE studies employed time-to-event models. To remedy this gap, various machine learning and deep learning-based time-to-event models, such as random survival forests and DeepSurv [83] exist.

Our findings are in line with the previous research. A study by Dhiman et al. identified that the methodological conduct of ML studies in oncology was substandard in several domains including, sample size, handling of missing data, model development, and model availability for evaluation [84]. A study by Mortazavi et al. suggested that ML models may only improve performance when trained using appropriate features that do not reduce the information [57]. The validation study by Shi et al. demonstrated that the ML-based PRAISE score overestimated the risk of 1-year mortality, Bleeding, and recurrent acute myocardial infarction. Furthermore, the AUC for GRACE 2.0 score was 0.81 compared to 0.75 for PRAISE [29].

Limitations

The current study has some limitations. Different ML studies have various methodologies which may lead to heterogeneity. To overcome this hurdle, we only compared studies that provided both LR and ML models. In addition, when possible we provided secondary analysis of studies that used an analogous feature set for both ML and LR models. Additionally, For some outcomes like bleeding, only a limited number of articles provided the relevant data for meta-analysis which could lead to lower statistical power. Finally, we included only English articles which may introduce some publication bias to the current review. Nevertheless, it was the first study evaluating ML investigations in PCI and we provided a critical review of the articles in addition to statistical analysis.

Conclusion

No statistical significance was observed between ML and LR models. Methodological assessment of the articles revealed concerns such as small sample size, lack of external validation, possible data leakage, and overfitting. While ML models may perform better with much larger datasets, there was the black-box nature of ML models may make the LR models more useful for clinical adaption for now.

We recommend that future studies ensure clearer reporting of methodologies, adhere to PROBAST and CHARMS guidelines, employ nested cross-validation, achieve high values, utilize appropriate methods for handling missing data (such as multiple imputation), and incorporate external validation cohorts. These steps will enable a more robust and reliable comparison between ML and LR models.

Data availability

The data of the paper are presented in the main text and the supplementary files.

References

  1. Serruys PW, Rutherford JD. The birth, and evolution, of percutaneous coronary interventions: A conversation with Patrick Serruys, MD, phd. Circulation. 2016;134(2):97–100.

    Article  PubMed  Google Scholar 

  2. Niimi N, Shiraishi Y, Sawano M, et al. Machine learning models for prediction of adverse events after percutaneous coronary intervention. Sci Rep. 2022;12(1):6262.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  3. Rao SV, McCoy LA, Spertus JA, et al. An updated bleeding model to predict the risk of post-procedure bleeding among patients undergoing percutaneous coronary intervention: a report using an expanded bleeding definition from the National cardiovascular data registry CathPCI registry. JACC Cardiovasc Interv. 2013;6(9):897–904.

    Article  PubMed  Google Scholar 

  4. Peterson ED, Dai D, DeLong ER, et al. Contemporary mortality risk prediction for percutaneous coronary intervention: results from 588,398 procedures in the National cardiovascular data registry. J Am Coll Cardiol. 2010;55(18):1923–32.

    Article  PubMed  PubMed Central  Google Scholar 

  5. Mehran R, Aymong ED, Nikolsky E, et al. A simple risk score for prediction of contrast-induced nephropathy after percutaneous coronary intervention: development and initial validation. J Am Coll Cardiol. 2004;44(7):1393–9.

    PubMed  Google Scholar 

  6. Nikolsky E, Mehran R, Lasic Z, et al. Low hematocrit predicts contrast-induced nephropathy after percutaneous coronary interventions. Kidney Int. 2005;67(2):706–13.

    Article  PubMed  Google Scholar 

  7. Hannan EL, Farrell LS, Walford G, et al. The new York state risk score for predicting in-hospital/30-day mortality following percutaneous coronary intervention. JACC Cardiovasc Interv. 2013;6(6):614–22.

    Article  PubMed  Google Scholar 

  8. Vroegindewey MM, Schuurman AS, Oemrawsingh RM, et al. SYNTAX score II predicts long-term mortality in patients with one- or two-vessel disease. PLoS ONE. 2018;13(7):e0200076.

    Article  PubMed  PubMed Central  Google Scholar 

  9. Ellis RJ, Sander RM, Limon A. Twelve key challenges in medical machine learning and solutions. Intelligence-Based Med. 2022;6:100068.

    Article  Google Scholar 

  10. Kulkarni H, Amin AP. Artificial intelligence in percutaneous coronary intervention: improved risk prediction of PCI-related complications using an artificial neural network. BMJ Innovations. 2021;7(3):564.

    Article  Google Scholar 

  11. Quer G, Arnaout R, Henne M, Arnaout R. Machine learning and the future of cardiovascular care: JACC State-of-the-Art review. J Am Coll Cardiol. 2021;77(3):300–13.

    Article  PubMed  PubMed Central  Google Scholar 

  12. Singh K, Beam AL, Nallamothu BK. Machine learning in clinical journals: moving from inscrutable to informative. Circ Cardiovasc Qual Outcomes. 2020;13(10):e007491.

    Article  PubMed  PubMed Central  Google Scholar 

  13. Page MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ. 2021;372:n71.

    Article  PubMed  PubMed Central  Google Scholar 

  14. Moons KG, de Groot JA, Bouwmeester W, et al. Critical appraisal and data extraction for systematic reviews of prediction modelling studies: the CHARMS checklist. PLoS Med. 2014;11(10):e1001744.

    Article  PubMed  PubMed Central  Google Scholar 

  15. Wolff RF, Moons KGM, Riley RD, et al. PROBAST: A tool to assess the risk of bias and applicability of prediction model studies. Ann Intern Med. 2019;170(1):51–8.

    Article  PubMed  Google Scholar 

  16. Fernandez-Felix BM, López-Alcalde J, Roqué M, Muriel A, Zamora J. CHARMS and PROBAST at your fingertips: a template for data extraction and risk of bias assessment in systematic reviews of predictive models. BMC Med Res Methodol. 2023;23(1):44.

    Article  PubMed  PubMed Central  Google Scholar 

  17. Hanley JA, McNeil BJ. The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology. 1982;143(1):29–36.

    Article  CAS  PubMed  Google Scholar 

  18. MedCalc Software Ltd Comparison of AUC of independent ROC curves. April 23. 2024; 22.023:https://www.medcalc.org/calc/comparison_of_independentROCtest.php

  19. R Core Team R: A Language and Environment for Statistical Computing. 2022; https://www.R-project.org/

  20. Ninomiya K, Kageyama S, Shiomi H, et al. Can machine learning aid the selection of percutaneous vs surgical revascularization?? J Am Coll Cardiol. 2023;82(22):2113–24.

    Article  PubMed  Google Scholar 

  21. Bai Z, Lu J, Li T, et al. Clinical Feature-Based machine learning model for 1-Year mortality risk prediction of ST-Segment elevation myocardial infarction in patients with hyperuricemia: A retrospective study. Comput Math Methods Med. 2021;2021:7252280.

    Article  PubMed  PubMed Central  Google Scholar 

  22. Liu N, Liu Q. A comparative study of the prognostic value of independent SYNTAX score, GRACE score and a new scoring system combining SYNTAX and GRACE for patients with ACS undergoing PCI. J Arrhythmia. 2019;35:601.

    Google Scholar 

  23. Huang YC, Chen KY, Li SJ, Liu CK, Lin YC, Chen M. Implementing an ensemble learning model with feature selection to predict mortality among patients who underwent Three-Vessel percutaneous coronary intervention. Appl SCIENCES-BASEL 2022;12(16).

  24. Feng XX, Zhang C, Huang X, et al. Machine learning improves mortality prediction in three-vessel disease. ATHEROSCLEROSIS. 2023;367:1–7.

    Article  CAS  PubMed  Google Scholar 

  25. Liu SY, Yang SW, Xing AL, et al. Machine learning-based long-term outcome prediction in patients undergoing percutaneous coronary intervention. Cardiovasc DIAGNOSIS THERAPY. 2021;11(3):736–43.

    Article  Google Scholar 

  26. D’Ascenzo F, De Filippo O, Gallone G, et al. Machine learning-based prediction of adverse events following an acute coronary syndrome (PRAISE): a modelling study of pooled datasets. Lancet. 2021;397(10270):199–207.

    Article  PubMed  Google Scholar 

  27. Călburean PA, Grebenișan P, Nistor IA, et al. Prediction of 3-year all-cause and cardiovascular cause mortality in a prospective percutaneous coronary intervention registry: machine learning model outperforms conventional clinical risk scores. Atherosclerosis. 2022;350:33–40.

    Article  PubMed  Google Scholar 

  28. Burrello J, Gallone G, Burrello A et al. Prediction of All-Cause mortality following percutaneous coronary intervention in bifurcation lesions using machine learning algorithms. J Pers Med 2022;12(6).

  29. Shi B, Wang HY, Liu J, et al. Prognostic value of Machine-Learning-Based PRAISE score for ischemic and bleeding events in patients with acute coronary syndrome undergoing percutaneous coronary intervention. J Am Heart Assoc. 2023;12(7):e025812.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  30. Song L, Li Y, Nie S et al. Using machine learning to predict adverse events in acute coronary syndrome: A retrospective study. Clin Cardiol. 2023.

  31. Li YH, Lee IT, Chen YW, Lin YK, Liu YH, Lai FP. Using text content from coronary catheterization reports to predict 5-Year mortality among patients undergoing coronary angiography: A deep learning approach. Front Cardiovasc Med. 2022;9:800864.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  32. Agasthi P, Ashraf H, Chao CJ et al. Machine learning helps predict short and Intermediate-term risk for all cause mortality in patients undergoing percutaneous coronary intervention. Circulation 2020;142(SUPPL 3).

  33. Iftikhar A, Bond RR, Rjoob K et al. Machine Learning to Predict 30 Days and 1-Year Mortality in STEMI and Turndown Patients. Paper presented at: Computing in Cardiology2020.

  34. Ngew KY, Tay HZ, Yusof AKM. Development and validation of a predictive models for predicting the cardiac events within one year for patients underwent percutaneous coronary intervention procedure at IJN. BMC Cardiovasc Disord. 2023;23(1):545.

    Article  PubMed  PubMed Central  Google Scholar 

  35. Tourassi GD, Xenopoulos NP. An artificial neural network to predict mortality in patients who undergo percutaneous coronary interventions. Paper presented at: IEEE International Conference on Neural Networks - Conference Proceedings1997.

  36. Freeman RV, Eagle KA, Bates ER, et al. Comparison of artificial neural networks with logistic regression in prediction of in-hospital death after percutaneous transluminal coronary angioplasty. Am Heart J. 2000;140(3):511–20.

    Article  CAS  PubMed  Google Scholar 

  37. Tindale A, Cretu I, Meng H, Panoulas V. Complete revascularization is associated with higher mortality in patients with ST-elevation myocardial infarction, multi-vessel disease and shock defined by hyperlactataemia: results from the Harefield shock registry incorporating explainable machine learning. Eur Heart J Acute Cardiovasc Care. 2023;12(9):615–23.

    Article  PubMed  PubMed Central  Google Scholar 

  38. Sladojevic M, Tadic S, Pavlovic K, et al. Data mining approach for in-hospital mortality prediction of patients presented with STEMI after primary PCI. Eur J Prev Cardiol. 2014;21(1):S125.

    Google Scholar 

  39. Limprasert S, Phu-Ang A. Data modeling using vital sign dynamics for In-hospital mortality classification in patients with acute coronary syndrome. Healthc Inf Res. 2023;29(2):120–31.

    Article  Google Scholar 

  40. Al’Aref SJ, Singh G, van Rosendael AR, et al. Determinants of In-Hospital mortality after percutaneous coronary intervention: A machine learning approach. J Am Heart Assoc. 2019;8(5):e011160.

    Article  PubMed  PubMed Central  Google Scholar 

  41. Resnic FS, Ohno-Machado L, Selwyn A, Simon DI, Popma JJ. Simplified risk score models accurately predict the risk of major in-hospital complications following percutaneous coronary intervention. Am J Cardiol. 2001;88(1):5–9.

    Article  CAS  PubMed  Google Scholar 

  42. Resnic FS, Popma JJ, Ohno-Machado L. Development and evaluation of models to predict death and myocardial infarction following coronary angioplasty and stenting. Proc AMIA Symp. 2000:690–693.

  43. Hsieh MH, Lin SY, Lin CL, et al. A fitting machine learning prediction model for short-term mortality following percutaneous catheterization intervention: a nationwide population-based study. Ann Transl Med. 2019;7(23):732.

    Article  PubMed  PubMed Central  Google Scholar 

  44. Sladojevic M, Pavlovic K, Velicki L, et al. In-hospital mortality prediction for STEMI patients submitted to primary PCI. Eur Heart J. 2013;34:85.

    Article  Google Scholar 

  45. Sladojevic M, Sladojevic S, Dejanovic J, et al. Predictive model for in-hospital outcome in ST elevated myocardial infarction complicated with acute heart failure. Eur J Heart Fail. 2016;18:131–2.

    Google Scholar 

  46. Sladojevic M, Sladojevic S, Tadic S, Stefanovic M. In-hospital outcome predictions for acute coronary sindrome patients after coronry angioplasty by mining echocardiography parameters data. J Hypertens. 2017;35:e118.

    Article  Google Scholar 

  47. Senecal C, Zack CJ, Kinar Y et al. Leveraging machine learning techniques to forecast patient prognosis after percutaneous coronary intervention. JACC Cardiovasc Interv 2016;134.

  48. Chen P, Wang B, Zhao L, et al. Machine learning for predicting intrahospital mortality in ST-elevation myocardial infarction patients with type 2 diabetes mellitus. BMC Cardiovasc Disord. 2023;23(1):585.

    Article  PubMed  PubMed Central  Google Scholar 

  49. Bansal A, Kapadia S, Ellis S, MACHINE LEARNING TECHNIQUES TO PREDICT IN-HOSPITAL CARDIOVASCULAR OUTCOMES IN ELDERLY PATIENTS PRESENTING WITH ACUTE MYOCARDIAL INFARCTION, et al. J Am Coll Cardiol. 2020;75(11):3603.

    Article  Google Scholar 

  50. Deng L, Zhao X, Su X, Zhou M, Huang D, Zeng X. Machine learning to predict no reflow and in-hospital mortality in patients with ST-segment elevation myocardial infarction that underwent primary percutaneous coronary intervention. BMC Med Inf Decis Mak. 2022;22(1):109.

    Article  Google Scholar 

  51. Hamilton D, Seth M, Sukul D, Gurm HS. A novel tool for highly reliable and accurate prediction of multiple complications in patients undergoing percutaneous coronary intervention. Circulation. 2021;144(SUPPL 1).

  52. Hamilton DE, Albright J, Seth M, Sukul D, Gurm HS. Merging Machine Learning and Patient Preference: A Contemporary, Comprehensive, Patient-Centered Tool for Risk Prediction Prior to Percutaneous Coronary Intervention. CIRCULATION. 2022;146.

  53. Lee G, Gurm HS, Syed Z. Predicting complications of percutaneous coronary intervention using a novel support vector method. J Am Med Inf Assoc. 2013;20(4):778–86.

    Article  Google Scholar 

  54. Niedziela JT, Cieśla D, Wojakowski W, et al. Is neural network better than logistic regression in death prediction in patients after ST-segment elevation myocardial infarction? Kardiol Pol. 2021;79(12):1353–61.

    Article  PubMed  Google Scholar 

  55. Bhattacharya A, Sadasivuni S, Chao CJ et al. Multi-modal fusion model for predicting adverse cardiovascular outcome post percutaneous coronary intervention. Physiol Meas 2022;43(12).

  56. Hurley NC, Desai N, Dhruva SS et al. A dynamic model to estimate evolving risk of major bleeding after percutaneous coronary intervention. In:2021.

  57. Mortazavi BJ, Bucholz EM, Desai NR, et al. Comparison of machine learning methods with National cardiovascular data registry models for prediction of risk of bleeding after percutaneous coronary intervention. JAMA Netw Open. 2019;2(7):e196835.

    Article  PubMed  PubMed Central  Google Scholar 

  58. Rayfield C, Agasthi P, Mookadam F, et al. Machine learning on high-dimensional data to predict bleeding post percutaneous coronary intervention. J Invasive Cardiol. 2020;32(5):E122–9.

    Article  PubMed  Google Scholar 

  59. Zhao XY, Wang JM, Yang JA et al. Machine learning for prediction of bleeding in acute myocardial infarction patients after percutaneous coronary intervention. THERAPEUTIC Adv CHRONIC DISEASE 2023;14.

  60. Chen P, Liu Y, Chen S, Xian Y, Chen J, Tan N. A novel tool for pre-procedural risk stratification for contrast-induced nephropathy and associations between hydration volume and clinical outcomes following coronary angiography at different risk levels. J Am Coll Cardiol 2018;71(11).

  61. Choi H, Choi B, Han S et al. Applicable machine learning model for predicting Contrast-induced nephropathy based on Pre-catheterization variables. Intern Med 2023.

  62. Gurm HS, Seth M, Kooiman J, Share D. A novel tool for reliable and accurate prediction of renal complications in patients undergoing percutaneous coronary intervention. J Am Coll Cardiol. 2013;61(22):2242–8.

    Article  PubMed  Google Scholar 

  63. Huang C, Li SX, Mahajan S, et al. Development and validation of a model for predicting the risk of acute kidney injury associated with contrast volume levels during percutaneous coronary intervention. JAMA Netw Open. 2019;2(11):e1916021.

    Article  PubMed  PubMed Central  Google Scholar 

  64. Huang C, Murugiah K, Annapureddy A et al. Automating risk assessment of acute kidney injury for patients undergoing percutaneous coronary intervention. Circulation. 2021;144(SUPPL 1).

  65. Huang C, Murugiah K, Li X et al. Effect of removing race correction factor in glomerular filtration rate Estimation on predicting acute kidney injury after percutaneous coronary intervention. In:2022.

  66. Huang C, Murugiah K, Mahajan S, et al. Enhancing the prediction of acute kidney injury risk after percutaneous coronary intervention using machine learning techniques: A retrospective cohort study. PLoS Med. 2018;15(11):e1002703.

    Article  PubMed  PubMed Central  Google Scholar 

  67. Kuno T, Mikami T, Sahashi Y, et al. Machine learning prediction model of acute kidney injury after percutaneous coronary intervention. Sci Rep. 2022;12(1):749.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  68. Ma X, Mo C, Li Y, Chen X, Gui C. Prediction of the development of contrast–induced nephropathy following percutaneous coronary artery intervention by machine learning. Acta Cardiol. 2023;78(8):912–21.

    Article  PubMed  Google Scholar 

  69. Tsutsui R, Johnston J, Felix C, et al. A supervised machine learning approach for predicting acute kidney injury following percutaneous coronary intervention. J Am Coll Cardiol. 2019;74(13):B604–604.

    Article  Google Scholar 

  70. Yuan N, Ebinger J. A new multivariate model for safe contrast limits to prevent contrast induced nephropathy after percutaneous coronary intervention. Circulation 2019;140.

  71. Dehdar Karsidani S, Farhadian M, Mahjub H, Mozayanimonfared A. Intelligent prediction of major adverse cardiovascular events (MACCE) following percutaneous coronary intervention using ANFIS-PSO model. BMC Cardiovasc Disord 2022;22(1).

  72. Farhadian M, Dehdar Karsidani S, Mozayanimonfared A, Mahjub H. Risk factors associated with major adverse cardiac and cerebrovascular events following percutaneous coronary intervention: a 10-year follow-up comparing random survival forest and Cox proportional-hazards model. BMC Cardiovasc Disord. 2021;21(1):38.

    Article  PubMed  PubMed Central  Google Scholar 

  73. Jenab Y, Hedayat B, Karimi A, Taaghi S, Ghorashi SM, Ekhtiari H. Effects of opium use on one-year major adverse cardiovascular events (MACE) in the patients with ST-segment elevation MI undergoing primary PCI: a propensity score matched - machine learning based study. BMC Complement Med Ther. 2023;23(1):16.

    Article  PubMed  PubMed Central  Google Scholar 

  74. Ke B, Gong R, Shen A, et al. Risk stratification algorithm for clinical outcomes in anemic patients undergoing percutaneous coronary intervention. Ann Med. 2023;55(2):2249200.

    Article  PubMed  PubMed Central  Google Scholar 

  75. Park JS. Machine learning for risk prediction of future clinical events in patients with acute coronary syndrome undergone percutaneous coronary intervention. Circulation 2019;140.

  76. Wang XB, Cui NH, Liu X. A novel 6-metabolite signature for prediction of clinical outcomes in type 2 diabetic patients undergoing percutaneous coronary intervention. Cardiovasc Diabetol. 2022;21(1):126.

    Article  PubMed  PubMed Central  Google Scholar 

  77. Moons KGM, Wolff RF, Riley RD, et al. PROBAST: A tool to assess risk of bias and applicability of prediction model studies: explanation and elaboration. Ann Intern Med. 2019;170(1):W1–33.

    Article  PubMed  Google Scholar 

  78. Ramspek CL, Jager KJ, Dekker FW, Zoccali C, van Diepen M. External validation of prognostic models: what, why, how, when and where? Clin Kidney J. 2021;14(1):49–58.

    Article  PubMed  Google Scholar 

  79. Riley RD, Ensor J, Snell KI, et al. External validation of clinical prediction models using big datasets from e-health records or IPD meta-analysis: opportunities and challenges. BMJ. 2016;353:i3140.

    Article  PubMed  PubMed Central  Google Scholar 

  80. Cawley GC, Talbot NL. On over-fitting in model selection and subsequent selection bias in performance evaluation. J Mach Learn Res. 2010;11:2079–107.

    Google Scholar 

  81. van der Ploeg T, Austin PC, Steyerberg EW. Modern modelling techniques are data hungry: a simulation study for predicting dichotomous endpoints. BMC Med Res Methodol. 2014;14:137.

    Article  PubMed  PubMed Central  Google Scholar 

  82. Malleson N. Calibration of simulation models. In: Bruinsma G, Weisburd D, editors. Encyclopedia of criminology and criminal justice. New York, NY: Springer New York; 2014. pp. 243–52.

    Chapter  Google Scholar 

  83. Wang Z, Lee JW, Chakraborty T et al. Survival modeling using deep learning, machine learning and statistical methods: A comparative analysis for predicting mortality after hospital admission. arXiv preprint arXiv:240306999. 2024.

  84. Dhiman P, Ma J, Andaur Navarro CL, et al. Methodological conduct of prognostic prediction models developed using machine learning in oncology: a systematic review. BMC Med Res Methodol. 2022;22(1):101.

    Article  PubMed  PubMed Central  Google Scholar 

Download references

Funding

None.

Author information

Author notes

  1. Sepehr Nayebirad and Ali Hassanzadeh contributed equally to this work.

Authors and Affiliations

  1. Tehran Heart Center, Cardiovascular Diseases Research Institute, Tehran University of Medical Sciences, Tehran, Iran

    Sepehr Nayebirad, Akbar Shafiee & Farzad Masoudkabir

  2. Diabetes Research Center, Endocrinology and Metabolism Clinical Sciences Institute, Tehran University of Medical Sciences, Tehran, Iran

    Ali Hassanzadeh

  3. School of Medicine, Tehran University of Medical Sciences, Tehran, Iran

    Amir Mohammad Vahdani & Shayan Forghani

  4. Rheumatology Research Center, Tehran University of Medical Sciences, Tehran, Iran

    Aida Mohamadi

  5. Cardiac Primary Prevention Research Center, Cardiovascular Diseases Research Institute, Tehran University of Medical Sciences, Tehran, Iran

    Akbar Shafiee & Farzad Masoudkabir

Authors
  1. Sepehr Nayebirad
    View author publications

    You can also search for this author inPubMed Google Scholar

  2. Ali Hassanzadeh
    View author publications

    You can also search for this author inPubMed Google Scholar

  3. Amir Mohammad Vahdani
    View author publications

    You can also search for this author inPubMed Google Scholar

  4. Aida Mohamadi
    View author publications

    You can also search for this author inPubMed Google Scholar

  5. Shayan Forghani
    View author publications

    You can also search for this author inPubMed Google Scholar

  6. Akbar Shafiee
    View author publications

    You can also search for this author inPubMed Google Scholar

  7. Farzad Masoudkabir
    View author publications

    You can also search for this author inPubMed Google Scholar

Contributions

S.N and A.H wrote the manuscript. S.N and A.V helped in screening. A.H and A.M contributed in data collection. A.S and F.M conceptualized the article.

Corresponding author

Correspondence to Sepehr Nayebirad.

Ethics declarations

Conflict of interest

None.

Ethics approval statement

Not applicable.

Patient consent statement

Not applicable.

Permission to reproduce material from other sources

Not applicable.

Registration

PROSPERO (CRD42023494659).

Consent for publication

Not applicable.

Additional information

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Electronic supplementary material

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Nayebirad, S., Hassanzadeh, A., Vahdani, A.M. et al. Comparison of machine learning models with conventional statistical methods for prediction of percutaneous coronary intervention outcomes: a systematic review and meta-analysis. BMC Cardiovasc Disord 25, 310 (2025). https://doiorg.publicaciones.saludcastillayleon.es/10.1186/s12872-025-04746-0

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doiorg.publicaciones.saludcastillayleon.es/10.1186/s12872-025-04746-0

Keywords

BMC Cardiovascular Disorders

ISSN: 1471-2261

Contact us