Application of artificial intelligence in predicting the results of open-heart surgery: a scoping review

Abstract

Purpose

This scoping review aims to synthesize research on artificial intelligence (AI) in predicting open-heart surgery outcomes, evaluating AI model performance, and identifying gaps in data quality, algorithmic bias, and clinical applicability to guide future advancements in personalized surgical planning and patient outcomes.

Methods

Conducted using the PRISMA-ScR guideline, the review involved a systematic search across PubMed, Web of Science, IEEE, and Scopus. Articles were included if they focused on open-heart surgery, utilized AI methods, and were published in English. Exclusion criteria included non-relevance to open-heart surgery, non-original research, and lack of AI techniques. Data extraction included study details, AI methods, and performance metrics. Descriptive statistics were used for analysis.

Results

Of the 64 included studies, 89.06% were retrospective. The most frequently employed algorithm was logistic regression (n = 41), followed by random forest in 38 studies and XGBoost in 32 studies for data analysis. Most studies focused on predicting postoperative outcomes. Mortality, acute kidney injury, and complications were the outcomes that more studies concentrated on. XGBoost, used in 32 studies, exhibited the best performance in 11 of these studies. Deep learning and hybrid models were underutilized. Major limitations included inconsistent model validation, limited prospective data, and lack of diversity in patient populations.

Conclusion

AI demonstrates promising predictive capabilities in open-heart surgery, particularly through machine learning models. These models can already assist surgeons in real-world practice by supporting real-time risk stratification and personalized decision-making, such as identifying high-risk patients for targeted interventions. However, methodological limitations hinder clinical translation. Future work should emphasize prospective validation, explainable AI, and equitable data representation to enhance model reliability and applicability in real-world settings.

Supplementary information

The online version contains supplementary material available at 10.1186/s12911-025-03243-w.

Introduction

Open-heart surgery is a critical intervention for managing cardiovascular diseases, including coronary artery disease, valvular heart disease, congenital heart defects, and aortic aneurysms. This procedure significantly improves cardiac function, survival rates, and quality of life. Coronary Artery Bypass Grafting (CABG) is a common open-heart surgery that enhances survival and reduces complications in patients with coronary artery disease [1–3]. It also prevents life-threatening complications and reduces hospitalization duration, thereby contributing to the efficiency of the healthcare system [4].

Open-heart surgery is complex and carries risks, particularly hemorrhagic events and coagulation disorders [5]. Its intricate nature, involving individual, technical, and organizational factors, often leads to unpredictable outcomes, necessitating a deep understanding to optimize care and minimize complications [6]. Elderly patients face higher risks, with advanced age as a primary factor for postoperative complications, alongside hypertension, diabetes, smoking, renal dysfunction, and pre-existing cardiovascular conditions [7–9]. Severe complications, including cerebrovascular events, excessive bleeding, embolic brain injury, and acute kidney injury, highlight the high-risk nature of these procedures [10, 11]. Additionally, hospital care quality and surgical expertise significantly influence outcomes, with some institutions reporting elevated mortality rates [12].

Accurate prediction of open-heart surgery outcomes is essential for optimizing patient care and reducing complications. Predictive models allow tailored treatment plans for high-risk patients, such as those with liver cirrhosis or end-stage renal disease, supporting informed decision-making and efficient resource allocation [13–15]. These models prioritize high-risk patients for specialized care, enable early identification of complications, and improve care quality through targeted interventions [16–18]. They also evaluate surgeon and medical center performance, supporting continuous improvement programs to reduce complications [13]. Moreover, predictive models provide data for standardized clinical protocols, reducing variability in practice and updating techniques via evidence-based feedback [19].

Artificial intelligence (AI) has transformed healthcare by enabling analysis of complex medical data [20]. Machine learning algorithms enhance diagnostic accuracy and personalize treatment plans by analyzing medical images and diverse data [21–24]. AI-driven predictive analytics identify at-risk individuals, forecast health outcomes, and improve population health management [25]. By analyzing patient data trends, AI supports personalized preventive strategies, enhancing healthcare delivery [26].

AI holds significant potential to improve outcome prediction in open-heart surgery by enhancing risk stratification, decision-making, and postoperative outcome prediction. Machine learning models predict surgical risks with high accuracy, often surpassing traditional scoring systems [27–29]. These models excel in predicting short- and long-term outcomes, including survival and complications, in cardiac surgeries [30]. By integrating patient-specific data, AI enables personalized surgical planning and postoperative care, improving outcomes [31]. However, challenges such as data quality, algorithmic bias, and ethical concerns must be addressed to ensure safe and equitable AI-driven interventions.

Guided by this scoping review, we aimed to address three specific research questions: (1) What AI techniques and algorithms have been applied to predict outcomes in open-heart surgery? (2) How do these AI models perform regarding predictive accuracy and common evaluation metrics? (3) What limitations and gaps exist in current AI applications, particularly concerning data quality, algorithmic bias, and clinical applicability? These questions seek to synthesize existing literature, evaluate AI model effectiveness, and identify critical areas for improvement to advance personalized surgical planning and patient outcomes.

Given the increasing role of artificial intelligence (AI) in predicting outcomes of open-heart surgery, it is essential to conduct a scoping review to synthesize existing research, evaluate the performance of AI models, and identify gaps in data quality, algorithmic bias, and clinical applicability. By integrating findings on AI algorithms and techniques, prediction accuracy, and evaluation metrics, such a review can provide a comprehensive understanding of the potential and limitations of AI. This knowledge can guide future advancements in personalized surgical planning, postoperative care, and overall patient outcomes while informing ongoing research and innovation in the field of AI-driven surgical interventions.

Methods

This scoping review was conducted based on the PRISMA Extension for Scoping Reviews (PRISMA-ScR) guideline to synthesize studies on artificial intelligence (AI) for predicting open-heart surgery outcomes, an emergent and heterogeneous field. Unlike systematic reviews that focus on precise research questions and quantitative data synthesis through meta-analysis, scoping reviews are ideal for mapping existing evidence, identifying key concepts, and highlighting research gaps in a developing field. Due to the diversity in study designs, AI techniques, and outcome measures, formal data synthesis may be premature at this stage [32, 33]. The scoping approach allows for a comprehensive synthesis of existing research amid rapid advancements in AI and the wide variety of algorithms, datasets, and performance metrics used across studies (e.g., retrospective versus prospective designs, varying sample sizes, and inconsistent outcome reporting). This method avoids restrictive inclusion criteria that could exclude valuable insights. Aligning with the PRISMA-ScR guidelines, this approach facilitates the evaluation of AI model performance, algorithmic biases, data quality issues, and clinical applicability while charting future research directions in this interdisciplinary field. The methodology is detailed in the following four subsections.

Literature search

A systematic search of the literature was conducted in four databases – PubMed, Web of Science, IEEE, and Scopus – to retrieve relevant articles. The article search was conducted on September 23, 2024, and updated on September 7, 2025. It should be noted that no time restrictions were applied to the article search. A language restriction was applied to the search, and only articles published in English were included. The following keywords and their corresponding MeSH terms were used for the search: (“Thoracic Surgery” OR “Cardiac Surgery” OR “Heart Surgery” OR “open heart surgery” OR “open-heart surgery”) AND (“Artificial Intelligence” OR “Machine Intelligence” OR “Machine Learning” OR “Deep Learning” OR “Predictive Model”). Boolean operators were used to design and implement the search strategy in each database (supplementary 1).

Inclusion and exclusion criteria for study selection

Articles were included in this scoping review if they met the following inclusion criteria:

1. The studies focused on open-heart surgery.

2. The studies utilized at least one artificial intelligence method.

3. The studies were published in English.

Studies were excluded from this scoping review if they met the following exclusion criteria:

1. The title, abstract, or full text of the articles were not related to the field of open-heart surgery.

2. The articles were not original research articles (e.g., reviews, meta-analyses, letters to the editor, dissertations, etc.).

3. Artificial intelligence techniques were not employed.

4. Preprints or conference abstracts were not included in this study.

Data extraction phase

After defining the search strategy for each database, the searches were conducted, and the retrieved articles were imported into reference management software. Initially, duplicate entries were removed, and then the titles and abstracts of the articles were screened against the defined inclusion and exclusion criteria. This process was carried out by two of the authors. Studies that met the inclusion and exclusion criteria and appeared relevant had their full texts retrieved. The full texts of these articles were reviewed by three authors under the supervision of a supervisor, and any disagreements were resolved through discussion with the supervisor. In the next step, a data extraction form was designed in Microsoft Excel. The data extracted from the studies included the following: author, year, journal, study’s purpose, study design, sample size, age, gender, algorithms, techniques, and performance metrics.

Data analysis

EndNote software was used for reference management. The results are presented using descriptive statistics, including frequencies and percentages. Due to the heterogeneity and inconsistency of the study outcomes, a meta-analysis was not performed. Microsoft Excel software was used for quantitative data analysis.

Ethical considerations

We declare that we used an artificial intelligence-based tool (Google AI Studio) to translate this article.

Results

Study selection

The number of articles retrieved and reviewed during the search process, based on the PRISMA-ScR statement, is shown in Fig. 1. Initially, 4247 articles were retrieved by searching four databases: IEEE, Scopus, PubMed, and Web of Science. The second step involved removing duplicate articles, resulting in 2726 remaining articles. In the next step, the initial screening of the articles was performed, and the titles and abstracts of the remaining articles were reviewed. After this step, 302 articles remained, which entered the full-text review process. After reviewing the remaining 302 articles and matching them with the defined inclusion and exclusion criteria, finally, 64 articles remained, from which data were extracted.

Fig. 1.

Flow diagram of the literature search and study selection

General characteristics of the included studies

The general characteristics of the studies included in this scoping review are summarized in Table 1. The number of published studies per year is presented in Fig. 2. The earliest study was from 2006, and the latest was from 2025. The trend of publishing articles in the field of artificial intelligence application in this area has been increasing over the years, with the highest number of articles published in 2024 (19 studies). Regarding the data collection method, whether retrospective or prospective, it was mentioned in 63 out of the 64 included studies in this scoping review. Among these, 57 (89.06%) studies used a retrospective approach, 5 studies used a prospective approach, and one study utilized both methods for data collection. In the retrospective studies, data were collected from databases and electronic health records. The smallest sample size was 72, and the largest was 662,772. Also the median sample size was 1980.

Table 1.

General characteristics of included studies in this scoping review

#	Author	Year	Journal	study’s purpose	Study Design	Sample Size	Age	Gender	Algorithms	Techniques	Performance Metrics
1	J. Nilsson [34]	2006	The Journal of thoracic and cardiovascular surgery	To develop a method to select risk variables and predict mortality	Prospective	18362	Average age: 62.6 ± 10.7	Female: 5194, Male:13168	• NN • LR	Cross-validation	Best performance: NN ROC area: 0.81
2	A. Albert [35]	2007	Journal of Cardiothoracic Surgery	Determination of the role of possible determinants for mid-term survival, postoperative gradients and QoL after AVR	Prospective	587	Mean age: 75	N/M	• LR • Cox regression	Features Selection	C-indexes > 0.7
3	G. Meyfroidt [36]	2011	BMC medical informatics and decision making	To develop a predictive model for ICU discharge after non-emergency cardiac surgery	Retrospective	961	N/M	N/M	• Gaussian processes • LR • DT • RF • SVM	Cross-validation	Best performance: Gaussian processes Brier Score: 0.179
4	M. T. Nouei [37]	2014	Journal of medical systems	Developing an expert system for the risk assessment of mortality	Retrospective	1811	Mean age: 61.82 ± 10.72	N/M	• LGFAS • LR • MLP	Feature selection	Best performance: LGFAS AUC: 0.910
5	A. K Kumar [38]	2017	Annals of Cardiac Anaesthesia	To identify the incidence, motoric subtypes, and risk factors associated with development of delirium	Prospective	120	18–80	Male: 77 Female:43	• LR	—	History of hypertension RR: 6.7857 Carotid artery disease RR: 4.5000 Noninvasive ventilation use RR: 5.0446 ICU stay more than 10 days RR: 3.1630 poor postoperative pain control RR: 2.4958
6	A. Meyer [39]	2018	The Lancet Respiratory Medicine	To predict severe complications during critical care in real time after cardiothoracic surgery	Retrospective	9269	67.6	More than 1 number	• RNN	Cross-validation	AUC ≥ 87 for all models
7	H. C. Lee [40]	2018	Journal of clinical medicine	To compare the performance of machine learning approaches to predict acute kidney injury after cardiac surgery	Retrospective	2010	64 (56–71)	Female: 553	• DT • RF • XGBoost • SVM • NN • Deep learning • LR	Feature selection	Best performance: XGBoost AUC: 0.78
8	A. Kilic [41]	2020	The Annals of thoracic surgery	To evaluate the predictive utility of a machine learning algorithm in estimating operative mortality risk in cardiac surgery	Retrospective	11190	67 ± 11	69% man 31% woman	• XGBoost	Cross-validation	Mean average precision: 0.221 C-index: 0.808 Calibration: 0.993 F1 score: 0.281
9	X. Zeng [42]	2021	Scientific reports	To develop a model for accurately predicting complications after pediatric congenital heart surgery	Retrospective	1964	Median age: 11 months	Male: 958	• K-means • XGBoost	Cross-validation	Best performance: XGBoost AUC: 0.839
10	Z. Zhong [43]	2021	Scientific reports	To predict 30-days mortality, and 3 complications including after open-heart surgery	Retrospective	6844	66	Female: 2144, Male: 4700	• XGBoost • RF • NN • LR	Cross-validation	Best performance: XGBoost AUC:0.98
11	Y. Zhou [44]	2021	International Journal of Cardiology	To establish a risk-prediction model for assessing prognosis of HTx	Retrospective	381	43.783 ± 16.453	Female: 108, Male: 273	• RF • GB • SVM • LR • XGBoost • AdaBoost • NN	Cross-validation	Best performance: RF AUC: 0.801
12	P.Hu [45]	2021	Journal of cardiac surgery	Predicting postoperative AKI in patients of advanced age undergoing cardiac surgery.	Retrospective	848	≥60	Male: 460	• LR	Features Selection, Cross-validation	Best AUC: 0.801
13	H. Jiang [46]	2021	Frontiers in Cardiovascular Medicine	To identify critical preoperative and intraoperative variables and predict the risk of several severe complications after cardiac valvular surgery	Retrospective	1488	18–75	Female: 907 Male: 581	• CAT • LGBM • MLP • SVM • LR • RF • GB • KNN • AdaBoost • NB • XGBoost	Features Selection, Cross-validation	Best performance: XGBoost AUC: 0.90
14	L. Luo [47]	2022	Journal of the American Heart Association	To construct a model to identify patients at high risk of early mortality after surgery for infective endocarditis	Retrospective	476	Mean age: 43.7	Female: 82 (Training data) Male:194	• XGBoost • GBDT • LGBM • RF • Extra Tree • LR	Features Selection, Cross-validation	Best performance: XGBoost AUC: 0.812
15	R. S. Molina [48]	2022	The Annals of Thoracic Surgery	To develop a ML–based preoperative score to predict cardiac surgical operative mortality	Retrospective	2786	> 75	N/M	• LR • NB • MLP • SVM • RF • GB	Feature selection	Best performance: GB ROC:0.755
16	B. Nistal-Nuño [49]	2022	Journal of clinical monitoring and computing	Investigating performance prediction for ML methods for 12-hour ICU and in-hospital mortality for individual adult patients in the CSRU and CCU	Retrospective	11059	≥16	N/M	• Tree Ensemble • RF • XGBoost Tree Ensemble • NB • BN	Feature selection	Best performance: XGBoost Tree Ensemble Accuracy:0.875 AUROC:0.926
17	A. Orfanoudaki [50]	2022	Journal of Cardiac Surgery	Evaluating postoperative MVS outcomes and design mortality and morbidity risk calculators to supplement the STS risk score	Retrospective	383550	Mean age:64.76	Male: 53.85%	• LR • RF • XGBoost • OCT • L-OCT	Cross-validation	Best performance: XGBoost AUC:0.826
18	E. Sherman [51]	2022	The Annals of Thoracic Surgery	Leveraging ML to predict 30-day hospital readmission after cardiac surgery	Retrospective	4924	61.1	Female: 1843, Male: 3081	• RF • XGBoost • SVM • LR	Cross-validation	Best performance: RF AUC: 0.76
19	Y. Yu [52]	2022	Frontiers in Cardiovascular Medicine	To predict long-term mortality and identify risk factors in unselected patients’ post-cardiac surgery	Retrospective	7368	> 18	Female: 2220, Male: 5148	• LR • NN • NB • GB • RF • AdaBoost • Bagged trees • XGBoost	Cross-validation, Feature selection	Best performance: AdaBoost AUC :0.801
20	H. Zhang [53]	2022	Journal of Translational Medicine	To present a series of models for predicting AKI after cardiac surgery	Retrospective	1457	60	Female: 609, Male: 848	• XGBoost, • RF • Deep forest • LR	Cross-validation,Feature selection	Best performance: Deep Forest AUC: 0.881
21	X. Zhao [54]	2022	Frontiers in Cardiovascular Medicine	To explore the data of clinical features and outcomes to provide individualized care for patients with low cardiac output syndrome	Retrospective	1205	55.5	Female: 454, Male: 751	• Clustering	N/M	N/M
22	S. A. Lee [55]	2022	The Annals of Thoracic Surgery	To investigate the relationship between sarcopenia and postoperative outcomes in older patients undergoing surgical AVR	Retrospective	874	72	Male: 464 (53.1)	• Linear regression • LR • Cox regression	—	N/M
23	J. Li [56]	2022	Frontiers in Medicine	To establish a ML prediction model for ARF occurrence in AAS patients	Retrospective	1,637	mean = 50	Female: 301 (22.8%)	• XGBoost • RF • AdaBoost • DT • SVM • Nu-SVM • XGBoost + RF • XGBoost + DT	Cross-validation, Features Selection	Best performance: XGBoost AUC: 0.82
24	S. Konar [57]	2022	Health and Technology	To predict the ICU stay, hospital stay, and survival outcome of cardiac surgical patients	Retrospective	1077	46.47_+15.58	Male: 644 Female: 432	• DT • KNN • RF • Adaboost • GB • XGBoost • GNB • BNB • LDA • LR • GNB + LR • Linear Regression • Lasso Regression • Ridge Regression • K Neighbors Regressor • DT Regressor • RF Regressor • GB Regressor • XGBoost Regressor • SVR • Huber Regressor	Feature selection -feature engineering	Best performance for survival analysis: GNB + LR with AUC of 0.72 Best performance for hospital stay: GB with R2-score (0.023) Best performance for ICU stay: XGBoost regressor with R2-score (0.125)
25	X-Q. Luo [58]	2023	Journal of medical Internet research	To predict the development of CSA-AKI in the pediatric population	Retrospective	3863	1	Male: 1997, Female:1866	• XGBoost • KNN • NB • SVM • RF • NN	Cross-validation	Best performance: XGBoost AUROC: 0.890
26	C. S. Ong [59]	2023	The Journal of thoracic and cardiovascular surgery	To predict the risk of operative mortality for an expanded set of cases	Retrospective	7745	Median age:64 and 67	Female: 3064, Male:4681	• LR • SVM • RF • XGBoost	Cross-validation, Features Selection	Best performance: RF AUC:0.83
27	J. C. Penny-Dimri [60]	2023	Plos one	Development and benchmarking of a novel ML approach to consider cardiac surgery risks	Retrospective	151,078	65.61 69.57	N/M	• LR • XGBoost • UAN	Cross-validation	Best performance: UAN AUC: 0.82
28	J. C. Penny-Dimri [61]	2023	Frontiers in Cardiovascular Medicine	To compare the performance of two models and identify key predictors of long-term mortality using the superior model	Retrospective	144536	65.6	Female: 41,261, Male: 112,683	• DT • RF • GB	Cross-validation, Features Selection	Best performance: GB C-index: 0.803 (0.002)
29	S. Sinha [62]	2023	European Journal of Cardio-Thoracic Surgery	Risk prediction of in-hospital mortality	Prospective	227,087	67.5 and 70.8	Female: 61,795, Male: 165,292	• LR • RF • NN • XGBoost • weighted SVM	Cross-validation	Best performance: XGBoost AUC: 0.834
30	A. J. Weiss [63]	2023	JTCVS open	To develop and compare an ML model (using multi-modal EHR) with STS models.	Retrospective	6392	Median age: 64.7	Female: 2320, Male: 4072	• XGBoost • SVM • RF • LR	Feature selection	Best performance: XGBoost AUROC: 0.978
31	B. L. Wisotzkey [64]	2023	Pediatric transplantation	Enhancing the prediction of 1-year allograft loss in pediatric heart transplant patients	Retrospective	3787	4.9	Female: 1894, Male: 1893	• Cox regression • GB • RF	Cross-validation	Best performance: RF C-statistic: 0.74
32	R. Zea-Vera [65]	2023	The Annals of thoracic surgery	To predicts postoperative outcomes and costs after cardiac surgery	Retrospective	4874	Median age: 65	Female: 1451, Male: 3423	• XGBoost	Cross-validation	Best accuracy: 95%
33	K. Zhu [66]	2023	Journal of Cardiovascular Development and Disease	To identify risk factors after cardiac valvular surgery, and construct prediction models to provide new ideas for mortality risk assessment	Retrospective	7163	Mean age: 69.8	Female: 3224, Male: 3939	• Adaboost • BNB • DT • GB • KNN • LDA • SVM • LR • RF • SGD • XGBoost • LASSO-LR	Cross-validation, Feature selection	Best performance: LASSO-LR AUC: 0.739
34	C. Zürn [67]	2023	Interdisciplinary Cardiovascular and Thoracic Surgery	To improve postoperative risk assessment in congenital heart surgery	Retrospective	1765	between 1 day and 18 years	N/M	• LR • RF	Cross-validation	Best performance: LR AUC: 94.86
35	N. Allou [68]	2023	The Journal of Thoracic and Cardiovascular Surgery	To compare a new mortality prediction model with EuroSCORE II and two ML models for cardiac surgery decision-making.	Retrospective	165640	Mean age: 66.98 ± 12.42	Male: 117,868 (71.2)	• RF • XGBoost • TabBERT	Cross-validation	Best performance: TabBERT AUC: 0.834
36	Q. Li [19]	2023	Journal of Clinical Medicine	To develop and validate a predictive ML model for cardiac surgery associated with AKI	Retrospective	6495	18–70	Male: 4515 Female: 1980	• LR • SVM • KNN • NB • DT • RF • GB • XGBoost • LGBM • CAT • AdaBoost • ExtraTree	Cross-validation, Features Selection	Best performance: CAT ROC-AUCs: 0.85, 0.67, and 0.77
37	A. Abbasi [69]	2023	The Journal of Thoracic and Cardiovascular Surgery	To predict post-cardiac surgery outcomes and determine important variables	Retrospective	662772	Mean age: 64.9 years ±11.8	Male: 459,864 (69.4)	• Deep learning feed forward neural network	—	Best AUC: 0.97
38	V. Hui [70]	2023	European Journal of Cardio-Thoracic Surgery	To predict bleeding after cardiac surgery	Retrospective	2000	> 18 years	N/M	• LR • RF • XGBoost, • Deep Neural Networks • Ensemble Voting	Cross-validation	Best performance: Ensemble Voting AUPRC: 0.310, AUROC: 0.738
39	M. Nagy [71]	2024	Pediatric Nephrology	To predict moderate to severe CS-AKI at postoperative day 2	Retrospective	402	< 21	Female: 225, Male:177	• LGBM	Cross-validation	AUROC: 0.88
40	M. Tadege [72]	2024	BMC pediatrics	To assess the prevalence of death due to cardiac disease and its risk factors	Retrospective	1520	Average age: 24.44	Female: 842, Male: 678	• LR • NN • CHAID • Decision List • BN • C5 • C&R Tree • Quest • Discriminant	Feature selection	Best performance: LR AUC: 0.961
41	C. Tong [73]	2024	International Journal of Surgery	To evaluate the performance of five ML algorithms for predicting four major APOs after pediatric congenital heart surgery	Retrospective	23000	Median age: 11.6	Female: 10,586, Male: 12,414	• LGBM • LR • SVM • RF • CAT	Cross-validation	Best performance: LGBM AUC: 0.970
42	H. Zhang [74]	2024	Journal of Translational Medicine	To develop and external validate seven ML models, for predicting ARDS after cardiac surgery	Retrospective	1996	Mean age: 61.9 Mean age validation: 63.1	Male: 1175	• DT • GBDT • AdaBoost • XGBoost • LGBM • RF • Deep Forest	Feature selection	Best performance: Deep Forest AUC: 0.882
43	Q. Li [75]	2024	International Journal of Medical Informatics	Develop and comprehensively externally validate a ML model to estimate RBC transfusion in cardiac surgery with CPB	Retrospective	11201	Age : 18–70	Male: 7654	• LR • SVM • KNN • NB • DT • GB • XGBoost • LGBM • AdaBoost • Extra Tree • CAT • RF	Cross-validation, Feature selection	Best performance: LR AUROC: 0.829
44	Q. Li [76]	2024	Journal of Cardiothoracic Surgery	To develop and validate the performances of ML for PMI with different cut-off values in cardiac surgery with CPB	Retrospective	2983	Median age: 55.1	Male: 1841 Female: 1142	• LR • SVM • KNN • NB • DT • RF • GB • XGBoost • LGBM • CAT • AdaBoost • Extra Tree	Features Selection	Best performance: RF & CAT AUROC: 0.68
45	R. E. Freundlich [77]	2024	Journal of Clinical Anesthesia	Explore validation of a model to predict patients’ risk of failing extubation,	Retrospective	1642	Median: 62	Male: 1145	• LR	—	Optimism corrected c-statistic = 0.77
46	Q. Li [78]	2024	BMC Cardiovascular Disorders	Develop and validate a ML-based prediction model for POD in cardiac valve surgery patients	Retrospective	507	Mean age: 55.7	Female: 300	• RF • LR • SVM • KNN • GNB • Gradient Boosting Decision Tree • Perceptron	Features Selection	Best performance: RF AUC: 0.92
47	F. Barbieri [79]	2024	Journal of clinical medicine	Develop a biomarker-based risk score by means of a neuronal network	Retrospective	3079	More than 1 number	More than 1 number	• NN	—	AUC: 0.922
48	Y.Liu [80]	2024	Scandinavian Cardiovascular Journal	Prediction of early mortality in patients undergoing cardiac surgery for infective endocarditis	Retrospective	357	47.1 ± 14.5	Male: 245 Female:112	• LR	—	AUC: 0.88
49	R. S. Loomba [81]	2024	Pediatric cardiology	Acute Effects of Aminophylline on Hemodynamic Parameters and Fluid Balance	Retrospective	72	3 Months	N/M	• RF	Feature selection	Accuracy: 80%
50	D. Mauricio [82]	2024	Journal of clinical medicine	Maximizing survival in pediatric congenital cardiac surgery using machine learning	N/M	565	N/M	N/M	• RF • BN • DT • GB	Cross-validation	Best performance: GBM Accuracy: 95.79%
51	C. Han [83]	2024	iScience	To predict delirium after cardiac surgery using intraoperative biosignals and clinical data	Retrospective & Prospective	2114	Median: 66	Male: 62.5%	• RF • XGBoost • LR • SVM • NN • GB • LGBM • Extra Tree • ENS	—	Best performance: ENS AUROC: 0.887
52	E. D. Omar [84]	2024	International journal of nephrology and renovascular disease	To identify the best-performing algorithm for predicting AKI necessitating dialysis	Retrospective	1741	N/M	N/M	• LR • RF • SVM • GBDT	Feature selection	Best performance: GBDT Accuracy: 88.66%
53	S. Hur [85]	2024	International Journal of Medical Informatics	Develop and validate a machine learning algorithm for personalized RBC demand prediction	Retrospective	7843	Training set: 60.5 ± 13.2 Test set: 61.2 ± 13.4	Male: 5391 Female: 2452	• NN • XGBoost • RF • GPR	Cross-validation	Best performance: XGBoost RMSE: 3.203
54	L. A. Kapsner [86]	2024	Diagnostics	To identify risk factors for mortality after surgery	Retrospective	1302	Mean age: 402.92 ± 562.31 days	Male: 56.14% Female: 43.86%	• XGBoost • RSF	Cross-validation Feature selection	Best performance: RSF C-indices: 0.85
55	H. Li [87]	2024	JMIR Medical Informatics	To develop and validate a Patient Similarity Network to provide decision support for outcomes after surgery	Retrospective	5030	Median: 12.0 (map) 22.1 (test) month	Male: 2447	• KNN • KNN+LR • k-Random+LR	—	Best performance: KNN+LR AUC: 0.926
56	B. Gaye [88]	2024	BMJ Open	To identify prevalent ascending aortic dilatation in patients with bicuspid aortic valve and tricuspid aortic valve	Prospective	1034	BAV: 60.4 ± 12.4; TAV: 70.4 ± 9.1	Male: 708	• RF • NN	Cross-validation	Best performance: RF AUC: 0.88
57	R. Florquin [89]	2024	Journal of Anesthesia	To predict patients at risk of severe postoperative complications	Retrospective	1364	Median: 9.6 months	Male: 57%	• LR • GNB • DT • GB • RF • SVM	Cross-validation	Best performance: LR AUC:
58	H. Bai [90]	2025	Cardiovascular Diagnosis and Therapy	To identify risk factors for Major Adverse Cardiovascular Events	Retrospective	480	MACE group: 60.5 ± 11.7; No MACE: 55.2 ± 12.1	Male: 189 Female: 291	• Elastic Net Regression	Cross-validation	AUC: 0.92
59	B. Lau [91]	2025	Mathematical and Computational Applications	Predicting Red Blood Cell Transfusion in Elective Cardiac Surgery	Retrospective	1036	Median:73 for the with-transfusion group and 69 for the withouttransfusion group	Female: 413 Male: 623	• NN • XGBoost	Cross-validation	Best performance: NN AUROC: 0.798
60	S. Leiler [92]	2025	Interdisciplinary CardioVascular and Thoracic Surgery	To explore the relationship between preoperative interatrial block (IAB) and the occurrence of POAF after cardiac surgery	Retrospective	1350	Median: 68 years (IQR: 62–76)	Female : 30.2%	• LR	—	Odds Ratio for IAB: 2.64 (95% CI: 2.02–3.46, p < 0.001)
61	J. A. Cedeno [93]	2025	The Brazilian Journal of Infectious Diseases	Early prediction of 30-day mortality in patients with surgical wound infections following surgery	Retrospective	1713	Mean age: 60.4 years	Female: 55%	• LR • NN	Cross-validation	Best performance: NN Accuracy: 90%
62	A. B. Birlik [94]	2025	PLOS Digital Health	To enhance mortality prediction in patients undergoing coronary artery bypass grafting and/or valve surgery	Retrospective	543	Mean: 63.87 ± 11.64	Female: 33.52%	• LR • RF • SVM • MLP • DT • XGBoost • XGBoost + RF	Cross-validation Feature selection	Best performance: Stacking ensemble (RF + XGBoost) AUC: 0.9395
63	Y. Pei [95]	2025	Cardiovascular diabetology	To evaluate the relationship between the postoperative SHR index and all-cause mortality	Retrospective	3848	68 ± 12	Male: 2669 Female: 1179	• AdaBoost • XGBoost • SVM • NB • LR • GB	Cross-validation Feature selection	Best performance: NB AUC: 0.7936
64	T. J. Miles [96]	2025	The journal of thoracic and cardiovascular surgery	To assess risk of acute kidney injury after cardiac surgery	Retrospective	1224	Median age: 66	Male: 70%	• LR	—	N/M

QoL: Quality of Life; AVR: aortic valve replacement; AKI: acute kidney injury; CPB: cardiopulmonary bypass; POD: postoperative delirium; LR: Logistic Regression; KNN: K-nearest Neighbors; RF: Random Forest; SVM: Support Vector Machine; RNN: Recurrent deep neural network; DT: Decision Tree; NN: Neural Network; NB: Naive Bayes; BN: Bayesian network; UAN: Uncertainty-aware attention network; CPH: Cox Proportion Hazards; EuroSCORE II: European System for Cardiac Operative Risk Evaluation; AUPRC: area under the precision-recall curve; LGBM: Light Gradient Boosting Machine; CAT: CatBoost Classifier; LGFAS: Lookup Genetic Fuzzy Annealing System; RR: Relative Risk; GBDT: gradient boosting decision trees; GB: Gradient Boosting; GNB: Gaussian Naive Bayes; BNB: Bernoulli Naïve Bayes; LDA: Linear Discriminant Analysis; SVR: Support Vector Regressor; SGD: stochastic gradient descent; TabBERT: Tabular Bidirectional Encoder156 Representations from Transformers; OCT: Optimal Classification Trees; L-OCT: multivariate logistic regression with optimal classification trees; RSF: Random Survival Forest; GPR: Gaussian process regression; ENS: Soft Voting Ensemble; MVS: mitral valve surgery; ARF: Acute renal failure; AAS: acute aortic syndrome; ARDS: Acute respiratory distress syndrome; EHR: Electronic Health Records; STS: Society of Thoracic Surgeons; SHR: Stress Hyperglycemia Ratio; POAF: postoperative atrial fibrillation; ML: Machine Learning; RBC: Red Blood Cell

Fig. 2.

Distribution of studies based on the publication year of articles

An examination of the journals in which the included articles were published revealed that “The Journal of thoracic and cardiovascular surgery” and “The Annals of thoracic surgery” each featured five studies, while “Journal of clinical medicine” and “Frontiers in Cardiovascular Medicine” each published four studies, indicating the highest engagement with research related to the application of artificial intelligence in cardiac surgery.

Clinical outcomes predicted by the models

The predicted outcomes in the studies are presented in Table 2. These outcomes were divided into three main categories: pre-operative, intra-operative, and post-operative. Perioperative red blood cell transfusions and ascending aortic dilatation were two outcomes predicted in the preoperative phase. None of the studies included in this scoping review were conducted in the intraoperative phase. Most studies focused on predicting postoperative outcomes. Mortality, acute kidney injury, and complications were the outcomes that more studies concentrated on.

Table 2.

Clinical outcomes predicted by the models

Category	Outcomes	References
Pre-operative	Perioperative red blood cell transfusions	[75, 85, 91]
	Identify ascending aortic dilatation	[88]
Intra-operative	No study
Post-operative	Mortality	[34, 37, 41, 43, 44, 47–50, 52, 55, 59, 61–63, 65, 66, 68, 72, 80, 82, 86, 87, 93–95]
	Survival	[35, 57, 67, 79]
	Postoperative gradients	[35]
	Quality of Life	[35]
	ICU discharge	[36]
	Postoperative delirium	[38, 78, 83]
	Complications	[39, 42, 43, 46, 87, 89]
	Acute kidney injury	[19, 40, 45, 53, 56, 58, 71, 73, 84, 96]
	Postoperative MVS outcomes	[50]
	Morbidity	[50, 65]
	30-day hospital readmission	[51]
	Low cardiac output syndrome	[54, 73]
	Adverse events	[55, 90]
	ICU stay	[57, 87]
	Hospital stay	[57, 87]
	Surgery event	[60]
	1-year allograft loss	[64]
	Hospitalization cost	[65]
	Hemorrhage	[69, 70]
	Venous thromboembolism	[69]
	Stroke	[69]
	Pneumonia	[73]
	Deep venous thrombosis	[73]
	Acute respiratory distress syndrome	[74]
	Myocardial injury	[76]
	Risk of reintubation	[77]
	Effects of Aminophylline on Hemodynamic Parameters and Fluid Balance	[81]
	Mechanical ventilation duration	[87]
	Atrial fibrillation	[92]

Artificial intelligence techniques used for data analysis

Figure 3 presents a mind map illustrating the artificial intelligence methods employed in the included studies. Notably, the majority of studies incorporated more than one algorithm. The most frequently employed algorithm was logistic regression (n = 41), followed by random forest in 38 studies and XGBoost in 32 studies for data analysis. Furthermore, SVM, GB, DT, NN and AdaBoost algorithms were also utilized more frequently compared to other algorithms.

Fig. 3.

Artificial intelligence techniques used for data analysis

Performance of artificial intelligence algorithms

The performance of the employed algorithms was reported in 61 studies. The performance of these algorithms was presented using various metrics, including C-indices, brier score, AUC, F1 score, precision, accuracy, sensitivity, and specificity. Figure 4 illustrates the number of studies that utilized each algorithm, as well as the number of studies in which each algorithm demonstrated the best performance. In each surveyed study, multiple AI algorithms were employed for a specific task and their performances were compared against each other. In this research, from the algorithms used in each study, we extracted the one that demonstrated the best performance when compared to the others within that same study. XGBoost, used in 32 studies, exhibited the best performance in 11 of these studies compared to other algorithms. The AdaBoost Classifier, utilized in 11 studies, achieved the best performance in 1 of them. Furthermore, the deep forest algorithm was employed in 2 studies and demonstrated the best performance in both when compared to other algorithms. The Gradient Boosting algorithm also showed the best performance in 4 out of 15 studies, logistic regression in 4 out of 41 studies, and random forest in 7 out of 38 studies.

Fig. 4.

Distribution of studies based on the algorithms used with the best performance

Discussion

The reviewed studies, encompassing 64 articles, predominantly focused on predicting mortality (e.g., in-hospital and long-term outcomes) and postoperative complications, such as acute kidney injury (AKI) and other complications like delirium. A growing emphasis was observed on optimizing postoperative care, including predicting outcomes like readmission, and comparing artificial intelligence models against traditional risk scoring systems. However, the limited application of these models in specific populations, such as pediatric and elderly patients, highlights the need for customized algorithms to address their unique needs and challenges.

Our findings align with the prevalent trend in artificial intelligence-driven medical research. Our review revealed that 89.06% of studies employed a retrospective design, consistent with the 96.4% reported in intensive care unit studies [97]. This reliance on retrospective designs, driven by the greater accessibility and lower cost of historical datasets, has been noted in prior research [98]. However, the dominance of the retrospective approach, also observed in diagnostic imaging research [99], raises concerns regarding potential biases and limited clinical generalizability. For instance, few studies have incorporated prospective validation or adhered to the robust design criteria essential for clinical implementation [97, 99]. These limitations can impact the reliability and clinical utility of the findings. To address these challenges, it is suggested that future research prioritize prospective designs and external validation to enhance the generalizability of findings. Furthermore, the development of design standards for the implementation of AI in clinical settings could help mitigate biases.

The predominance of machine learning (ML) over deep learning (DL) in predicting outcomes of cardiac surgery, as observed in our review of 64 studies, corroborates broader trends in clinical artificial intelligence research. With logistic regression (n = 41), random forest (n = 38), and XGBoost (n = 32) dominating algorithm use, MLs interpretability and computational efficiency [100, 101] make it more adaptable to clinical workflows, where transparency is critical [102]. For instance, gradient boosting and logistic regression models have consistently identified actionable predictors such as preoperative risk scores and biochemical markers [52, 103], enhancing their utility in surgical planning. Conversely, the limited adoption of deep learning aligns with its perceived drawbacks in clinical settings, including opacity in model outputs and the need for substantial training resources [102, 104]. as well as high computational resource demands that may exceed the infrastructure available in many clinical or research environments. Additionally, DL often requires large, high-quality datasets for effective training, and data availability in cardiac surgery—where patient cohorts can be heterogeneous and limited by privacy regulations or institutional silos—further contributes to its underrepresentation [105–107]. Clinician trust also plays a key role, as the “black box” nature of DL models can erode confidence in their outputs, particularly in high-stakes surgical decisions where explainability is paramount [108, 109]. Nevertheless, the superior performance of deep learning in processing complex and time-dependent data [110] suggests its potential for specialized applications, such as intraoperative monitoring, which may not yet be adequately explored in the scientific literature. Hybrid models, despite their less frequent representation in our review, constitute a promising approach towards balancing interpretability and predictive capacity, as evidenced by their success in other medical fields [110–112]; however, their underuse may stem from similar barriers, including the complexity of integrating multiple architectures, which demands advanced expertise and resources not always accessible in cardiac surgery research, alongside concerns over data scarcity and the need for rigorous validation to build clinician acceptance. Future research should prioritize the development of standardized frameworks for validating hybrid models in cardiac surgery, alongside efforts to improve the explainability of deep learning through techniques like attention mapping or surrogate models. Furthermore, collaboration between clinicians and data scientists is crucial for addressing the limitations of datasets and ensuring the alignment of artificial intelligence tools with real-world clinical needs [103, 113].

To demonstrate the practical implications of our findings, specific AI models from included studies directly influence clinical decision-making and workflows in cardiac surgery. For example, XGBoost-based models, which showed superior performance in 11 studies, predict acute kidney injury (AKI) post-cardiac surgery [43, 63]. These models enable early identification of high-risk patients, allowing clinicians to consider preventive interventions such as optimized perioperative fluid management, nephrotoxic agent avoidance, or enhanced postoperative monitoring, thereby supporting proactive patient management [40, 59]. Similarly, machine learning algorithms like random forest (used in 38 studies) facilitate mortality risk stratification, informing shared decision-making during preoperative consultations [43]. This can streamline resource allocation and improve informed consent processes, such as patient selection for elective procedures or escalation to multidisciplinary team reviews.

These applications align with broader AI integrations in other surgical specialties, offering transferable insights. For instance, in neurosurgery, AI-driven tools using convolutional neural networks enhance intraoperative precision through real-time tissue differentiation, reducing operative time and complications via improved surgical planning [114]. In orthopedics, predictive models analyze imaging to forecast postoperative complications or implant failure, guiding surgical approaches and rehabilitation plans, potentially reducing revision surgeries [115].

Consistent with AI’s role in improving diagnostic and predictive accuracy across healthcare [21], integrating similar AI-driven workflows into cardiac surgery, such as real-time predictive alerts in electronic health records, could enhance clinical practice once challenges like system interoperability and model validation are addressed. These examples highlight tangible ways AI can reshape perioperative care in cardiac surgery while reflecting transferable lessons from other surgical specialties.

Our findings align with previous research on ML applications in cardiac surgery outcomes, while also revealing important nuances. Consistent with earlier studies [113, 116], LR remains the most frequently employed algorithm due to its interpretability and ease of implementation. However, its predictive performance is consistently outperformed by ensemble methods such as XGBoost and random forest (RF). For instance, XGBoost demonstrated superior discrimination, with AUC values ranging from 0.83 to 0.98 in prior studies [62, 117], emerging as the top-performing algorithm in 34.4% of the studies in which it was used. This suggests its robustness in modeling complex, non-linear relationships inherent in surgical datasets. XGBoost’s consistent outperformance in surgical datasets can be attributed to its built-in regularization techniques that prevent overfitting, efficient handling of sparse and missing data common in electronic health records, and ability to capture intricate feature interactions without requiring extensive preprocessing [118–120]. Furthermore, dataset characteristics play a pivotal role: surgical outcomes often involve class imbalance (e.g., rare events like mortality or complications), where XGBoost excels through mechanisms like scale_pos_weight to adjust for minority classes, reducing bias toward majority outcomes [41, 121, 122]. The tabular structure of most cardiac surgery datasets—comprising structured variables from patient demographics, labs, and procedures—aligns well with XGBoost’s tree-based architecture, which outperforms simpler models like LR on non-linear patterns and deep learning on smaller, heterogeneous datasets where interpretability is valued [123–125]. The strong performance of RF in our review is supported by the findings of Elmahrouk et al. [126], who reported 98% accuracy, particularly in studies requiring feature interaction analysis. Gradient boosting variants, such as LightGBM [73] and CatBoost [19, 76], showed promising results, though their adoption is less widespread than XGBoost, likely due to computational complexity. While AdaBoost was less frequently utilized in our review, its reported efficacy in predicting long-term mortality (AUC 0.801) by Yu et al. [52] indicates untapped potential for adaptive boosting in specific surgical contexts.

The discrepancies in algorithm performance may stem from variability in dataset size, feature engineering practices, or outcome definitions across different studies. For example, the relatively modest performance of SVM in both our review and existing literature [127, 128] may be attributed to its sensitivity to imbalanced data and challenges in hyperparameter tuning. Future research should prioritize external validation of these models in diverse populations and the development of standardized reporting of performance metrics (e.g., calibration metrics such as the Brier score). Moreover, hybrid models—combining the interpretability of LR with the predictive power of ensemble methods—could bridge the gap between clinical utility and algorithmic transparency. Addressing the “black box” nature of advanced ML algorithms through explainable AI techniques may also enhance clinical adoption [62, 117].

This scoping review contributes to the existing body of knowledge by offering several new insights that distinguish it from previous reviews. First, by synthesizing 64 studies, we identified clear trends in algorithm use and performance. Our analysis reveals that although logistic regression remains the most common method (n = 41), XGBoost has emerged as the most consistently high-performing model, achieving the best performance in 34.4% of the studies in which it was implemented (11 of 32 studies)—a finding not quantitatively emphasized in earlier reviews [19].

Second, we highlight the significant underutilization of deep learning and hybrid approaches in cardiac surgery despite their potential advantages. This trend is consistent with findings across broader surgical fields; a systematic review on surgical outcomes prediction noted that while ensemble machine learning models are prevalent, the application of more complex deep learning architectures remains limited, primarily due to challenges with interpretability and data requirements [31].

Third, our review underscores critical systemic gaps that impede clinical translation, such as an overwhelming reliance on retrospective data (89.06% of studies), limited inclusion of diverse patient populations (e.g., pediatrics and elderly cohorts), and inconsistent model validation practices. A systematic review by Kenig et al. [129] emphasizes the need for standardized validation methods and highlights the limited diversity in existing datasets, which underscores the importance of including diverse patient demographics to enhance the applicability of AI models in surgical settings [129].

By mapping these evolving trends and pinpointing specific methodological shortcomings, our synthesis provides a more detailed and actionable roadmap for future research, aiming to foster the development of more equitable and clinically applicable AI tools in cardiac surgery.

Beyond the mere assessment of algorithmic performance, the successful clinical translation of AI models in cardiac surgery necessitates a critical examination of associated ethical considerations. A central concern is algorithmic bias and the requisite for fair deployment of these models, as models trained on datasets lacking representative demographic characteristics have the potential to perpetuate or exacerbate existing health disparities by exhibiting poor performance in specific underrepresented groups [130, 131]. Ensuring equity demands proactive measures to include diverse populations in training data and to conduct rigorous fairness audits prior to clinical implementation [132]. Furthermore, the “black box nature” of many advanced models (such as XGBoost and deep learning networks) challenges transparency and can lead to the erosion of clinical trust. As a result, the development and application of Explainable AI (XAI) techniques, such as SHAP and LIME, are considered not merely technical exercises but ethical imperatives to ensure that model predictions are understandable and justifiable to clinical specialists [133, 134]. Finally, achieving the equitable deployment of AI tools is essential; this implies that these models must be validated and accessible across diverse healthcare settings, including resource-limited hospitals, to prevent the emergence of an “AI divide” that benefits only well-resourced centers [135]. Accordingly, future research must prioritize the development of fair, transparent, and accountable AI systems and integrate ethical audits as a standard step in the model development lifecycle to bolster trust and ensure equal outcomes for all patients.

AI predictions in reviewed studies mostly focused on postoperative outcomes, such as mortality (26 studies), acute kidney injury (AKI; 10 studies), and complications (6 studies), while preoperative focus was limited (4 studies) and intraoperative phases were absent. This highlights the need to target high-impact post-surgical risks to improve prognostic accuracy in open-heart surgery, aligning with the primary objectives of this review. The distribution matches previous research prioritizing mortality and survival, where AI models outperformed traditional tools for 30-day mortality, prolonged ICU/hospital stays, and complications like septic shock or thrombocytopenia. AI models also addressed readmissions (AUC 0.75) and specialized risks like liver dysfunction [43, 100, 136]. The imbalance likely results from postoperative biological vulnerabilities and the methodological reliance on retrospective datasets that favor quantifiable endpoints. Clinically, these findings facilitate risk stratification for interventions that reduce adverse events, readmissions, and costs, thereby improving patient outcomes and resource utilization. However, a potential limitation is the selection bias toward dominant outcomes, leading to an oversight of intraoperative dynamics. Therefore, future studies should focus on integrating real-time data to develop intraoperative AI models, enabling comprehensive perioperative strategies.

In summary, this scoping review directly addressed our three initial research questions. First, most studies applied machine learning techniques, particularly logistic regression, random forest, and XGBoost, with deep learning and hybrid approaches being less common. Second, XGBoost consistently demonstrated superior predictive accuracy across multiple studies, with evaluation metrics such as AUC, C-index, and Brier score often reported. Third, significant limitations were identified, including predominant reliance on retrospective data, inconsistent validation, and limited population diversity, which restrict clinical applicability. Collectively, these findings reaffirm the alignment of our review with its stated objectives and underscore clear directions for advancing AI toward more robust and equitable applications in cardiac surgery.

Strengths and limitations

This scoping review possesses several strengths. First, a comprehensive and systematic search strategy across four major databases (PubMed, Web of Science, IEEE, and Scopus) ensured broad coverage of relevant literature. Second, adherence to the PRISMA-ScR guidelines provided a robust methodological framework, bolstering transparency and reproducibility. Third, the inclusion of 64 studies facilitated a detailed synthesis of AI techniques, performance metrics, and limitations, offering a comprehensive overview of the field. These strengths support the reliability and relevance of our findings to advancing AI in open-heart surgery.

Nevertheless, the review also has limitations. First, the predominance of retrospective studies (89.06%) restricts generalizability, as prospective validation is crucial for clinical applicability. Second, the potential for publication bias must be acknowledged, as studies with positive results are more likely to be published, which may lead to an overestimation of AI model performance in the literature. Third, significant heterogeneity in outcome definitions (e.g., for mortality, AKI, delirium) and study designs across the included studies complicates direct comparison and synthesis of findings, potentially affecting the consistency and generalizability of the reported results. This heterogeneity in AI algorithms also precluded a meta-analysis, limiting quantitative synthesis. Furthermore, the underrepresentation of diverse patient populations, such as pediatric and elderly groups, may hinder the applicability of findings to these cohorts. Lastly, the review’s focus on English-language articles may have excluded relevant studies in other languages. These limitations underscore the need for future research to address these gaps through prospective designs, standardized reporting, and inclusive datasets.

Conclusion

This scoping review emphasizes the transformative potential of artificial intelligence (AI) in advancing risk prediction and clinical decision-making for cardiac surgery outcomes. While machine learning (ML) algorithms, particularly XGBoost and random forest, demonstrate superior predictive accuracy compared to traditional statistical methods, their clinical adoption remains hindered by methodological limitations, including the predominance of retrospective designs and insufficient validation in diverse patient populations. The field’s reliance on interpretable ML models reflects a pragmatic balance between algorithmic transparency and performance, though the underutilization of hybrid and deep learning approaches suggests untapped opportunities for innovation in complex surgical scenarios.

To translate AI’s promise into clinical practice, future research must prioritize prospectively validated frameworks, standardized reporting of calibration metrics (e.g., Brier score), and the integration of explainable AI techniques to demystify “black-box” models. Addressing the ethical and practical challenges of dataset bias, algorithmic generalizability, and equitable representation of vulnerable populations (e.g., pediatric and elderly patients) is equally critical. Collaborative efforts between clinicians, data scientists, and policymakers will be essential to develop robust, patient-centered AI tools that align with real-world surgical workflows. By bridging technical innovation with clinical rigor, AI can redefine precision medicine in cardiac surgery, optimizing outcomes while ensuring trust, equity, and actionable utility at the bedside.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

SS, TKH and NA developed the concept for the study. SS and NA, PA, FK carried out the data extraction, analysis and interpretation under the supervision of TKH. Finally, SS and TKH drafted the manuscript. All authors reviewed the content and approved it.

Funding

This work was supported by a grant from Hamadan University of Medical Sciences Research Council.

Data availability

All data generated or analyzed during this study are included within this article.

Declarations

Ethics approval and consent to participate

The study was conducted in accordance with the Declaration of Helsinki and approved by a local ethics committee in Iran, namely Ethics Committee of the Hamadan University of Medical Sciences (ethical code: IR.UMSHA.REC.1403.399).

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.