Artificial Intelligence‐Based Delirium Prediction Model for Post‐Cardiac Surgery Patients: A Scoping Review

ABSTRACT

Background

Delirium is a common complication following cardiac surgery and significantly affects patient prognosis and quality of life. Recently, the application of artificial intelligence (AI) has gained prominence in predicting and assessing the risk of postoperative delirium, showing considerable potential in clinical settings.

Objective

This scoping review summarises existing research on AI‐based prediction models for post‐cardiac surgery delirium and provides insights and recommendations for clinical practice and future research.

Methods

Following the PRISMA‐ScR (Preferred Reporting Items for Systematic Reviews and Meta‐Analyses extension for Scoping Reviews) guidelines, eight databases were searched: China National Knowledge Infrastructure, Wanfang Database, China Biomedical Literature Database, Virtual Information Platform, PubMed, Web of Science, Medline, and Embase. Studies meeting the inclusion criteria were screened, and data were extracted on surgery type, delirium assessment tools, predictive factors, and AI‐based prediction models. The search covered database inception through January 12, 2025. Two researchers independently conducted the literature review and data analysis.

Results

Ten studies from China, Canada, and Germany involving 11,702 participants were included. The reported incidence of postoperative delirium ranged from 5.56% to 34%. The most commonly used assessment tools were Confusion Assessment Method for the Intensive Care Unit, Diagnostic and Statistical Manual of Mental Disorders‐5, and Intensive Care Delirium Screening Checklist. Key predictive factors included age, cardiopulmonary bypass time, cerebrovascular disease, and pain scores. AI‐based prediction models were primarily developed using R (6/10, 60%) and Python (4/10, 40%). Model performance, as measured by the area under the curve, ranged from 0.544 to 0.92. Among these models, Random Forest (RF) was the most effective (5/10, 50%), followed by XGBoost (3/10, 30%) and Artificial Neural Networks (2/10, 20%).

Conclusion

AI‐based models show promise for predicting postoperative delirium in cardiac surgery patients. Future studies should prioritise integrating these models into clinical workflows, conducting rigorous multicenter external validation, and incorporating dynamic, time‐varying perioperative variables to enhance generalizability and clinical utility.

Reporting Method

This review followed the PRISMA‐ScR (Preferred Reporting Items for Systematic Reviews and Meta‐Analyses extension for Scoping Reviews) guidelines.

Patient or Public Contribution

This study did not include patient or public involvement in its design, conduct, or reporting.

Implications for the Profession and/or Patient Care.

1. The scoping review identified ten peer‐reviewed studies conducted across six countries, detailing a variety of AI algorithms—including Random Forest, XGBoost, and deep learning models—used in prediction.

2. Key delirium predictors identified include age, cardiopulmonary bypass duration, cerebrovascular disease, and pain scores.

3. The review provides valuable context on delirium incidence, assessment tools, and risk factors while also critically appraising AI model performance.

1. Introduction

Due to population growth and aging, the overall prevalence and mortality rates of cardiovascular diseases continue to rise, with a trend toward younger patients (Conrad et al. 2024; Sun et al. 2023). Surgical procedures such as coronary artery bypass grafting and valve repair/replacement are important methods for treating heart diseases, improving heart function, and reducing associated health risks (Arora et al. 2019). Over 2 million cardiac surgeries are performed globally annually (Dimopoulos et al. 2023). However, factors such as the use of extracorporeal membrane oxygenation during surgery, delayed metabolism of anaesthetics, systemic immune response, increased inflammatory cytokines, postoperative hypoxia, hypoglycemia, and electrolyte imbalance can lead to postoperative delirium (Aldecoa et al. 2024; Ma et al. 2023; Madhavan et al. 2018). Delirium is an acute, transient disorder of consciousness characterised by inattention, disorientation, impaired memory, and hallucinations (Zarour et al. 2024). Patients often experience fluctuating mental states, including agitation or somnolence. If not managed promptly and effectively, delirium can lead to increased consumption of medical resources, such as prolonged mechanical ventilation, intensive Care Unit (ICU) and hospital stays, as well as impaired daily functioning, cognitive decline, and dementia (Faisal et al. 2023). In the United States, the annual medical costs associated with delirium exceed $33 billion, imposing a significant burden on families and society (Hshieh et al. 2023). Therefore, early identification of high‐risk factors and the implementation of preventive strategies can reduce the incidence of postoperative delirium in patients undergoing cardiac surgery, thereby improving outcomes and saving medical resources (Jiang et al. 2024).

Artificial intelligence (AI) has gradually been applied in the medical field for healthcare and nursing (Bindra and Jain 2024), encompassing two branches: virtual and physical. The core of the virtual branch is machine learning (ML), also known as deep learning, which is a mathematical modelling system that autonomously optimises performance through experiential data (Hamet and Tremblay 2017). Recently, research into the development and validation of risk prediction models for postoperative delirium in cardiac surgery patients has increased significantly worldwide (Cai et al. 2022). Studies have highlighted the use of AI algorithms such as Random Forest (RF), Support Vector Machine (SVM), Naive Bayes (NB), Artificial Neural Networks (ANN), and Extreme Gradient Boosting (XGBoost) for diagnosis, prediction, and patient management (Yasmin et al. 2021). AI‐based models can analyze clinical data to identify risk factors for delirium, supporting healthcare providers in designing personalised treatment and care plans (Zuhair et al. 2024).

While various studies have explored the use of AI in predicting postoperative delirium in patients undergoing cardiac surgery, a comprehensive summary of the technologies and model characteristics employed remains lacking. Therefore, this study summarised the effectiveness of AI models in predicting postoperative delirium and identified key predictive variables to support clinical decision‐making and inform future research.

2. Method

2.1. Design

This study was conducted as a scoping review following the methodological framework by Levac et al. (2010) to ensure rigour and transparency. We adhered to the Preferred Reporting Items for Systematic Reviews and Meta‐Analyses extension for Scoping Reviews checklist (Tricco et al. 2018).

2.2. Search Strategy

Relevant literature was retrieved from eight databases: China National Knowledge Infrastructure, Wanfang Database, China Biomedical Literature Database, Virtual Information Platform, PubMed, Web of Science, Medline (Ovid), and Embase. The main subject headings for the search include: “Cardiovascular Surgical Procedures, Cardiac Surgical Procedures, Cardiopulmonary Bypass (CPB), Extracorporeal Circulation, Heart Transplantation, Cardiac Catheterization, Delirium, Postoperative Delirium, Confusion, Artificial Intelligence, Machine Learning, Deep Learning, Support Vector Machine, Decision Tree, and Neural Network.” The literature search time frame is from the establishment of the database to January 12, 2025. Detailed examples of the search strings are provided in File S1.

2.3. Inclusion and Exclusion Criteria

The eligibility criteria were developed based on the PCC (Participants/Concept/Context) framework (Peters et al. 2020) covering the three core dimensions of study population, research topic, and research context to ensure a systematic and transparent screening of relevant studies.

2.3.1. Inclusion Criteria

1. Adult patients (≥ 18 years old) who have undergone cardiac surgery.

2. The study focused on the development, validation, or application of AI or ML models for predicting the onset of postoperative delirium.

3. The clinical setting includes the period following cardiac surgery, such as the ICU or cardiac surgery ward.

4. Studies must be original research published as full‐text articles in peer‐reviewed journals. Eligible study designs included randomised controlled trials, cohort studies, case–control studies, and cross‐sectional studies. Only Chinese‐ and English‐language studies were included due to the team's language proficiency

2.3.2. Exclusion Criteria

The exclusion criteria were applied to ensure the relevance, quality, and methodological transparency of the selected evidence, and to align with the practical scope of the review. Studies were excluded based on the following criteria:

1. Non‐peer‐reviewed publications, including grey literature (e.g., conference abstracts, preprints, theses, and technical reports).

2. Studies for which the full text was unavailable.

3. Studies that focus solely on the diagnosis, prediction, and treatment of delirium and do not involve AI‐based prediction.

4. Studies that failed to provide sufficient details about the AI model architecture, the features/variables used, or performance metrics.

5. Studies that used only traditional statistical methods without any AI or ML component.

2.4. Screening and Data Extraction

Two researchers independently conducted studies screening. Disagreements were resolved through discussion with a third, more experienced researcher. All retrieved studies were imported into EndNote (version 21 desktop), and duplicates were removed. Initial screening was based on titles and abstracts, followed by full‐text review to confirm eligibility. Data extraction was also independently completed by two researchers. The data extraction forms included the study title, year of publication, authors, study design, type of cardiac surgery, setting, model development methods, delirium diagnostic tools, evaluation time and incidence, the AI models used, validation methods, and model performance metrics (area under the curve [AUC], accuracy (ACC), sensitivity (SEN), and specificity (SPE)).

2.5. Risk of Bias and Applicability Assessment

To evaluate the potential bias of included studies, the Prediction model Risk Of Bias ASsessment Tool (PROBAST) tool was used (Moons et al. 2019). Two authors independently assessed bias risk and applicability. PROBAST's 20 questions across four domains (participants, predictors, outcomes, analysis) scrutinise studies on predictive models. Each signalling question can be answered as “yes,” “probably yes,” “no,” “probably no,” or “no information.” If at least one signalling question in a domain is answered as “no” or “probably no,” that domain should be considered at high risk of bias. Only when all domains are judged as low risk of bias, the overall bias can be considered low risk. The applicability assessment covers three areas: participants, predictors, and outcomes, following evaluation rules and procedures similar to those of bias risk assessment.

3. Results

3.1. Literature Screening Results

A total of 415 relevant records were retrieved. After screening, 10 studies met the inclusion criteria (Han et al. 2024; Huang et al. 2023; Li et al. 2024; Mufti et al. 2019; Nagata et al. 2023; Nowakowska et al. 2023; Sadlonova et al. 2024; Yang et al. 2024; Zhao et al. 2024; Zuo et al. 2023), including eight published in English and two in Chinese as shown in Figure 1.

FIGURE 1.

Flowchart of literature screening and analysis. AI, artificial intelligence; CBM, China Biomedical Literature Database; CNKI, China National Knowledge Infrastructure; ML, machine learning; VIP, Virtual Information Platform.

3.2. Characteristics of Included Studies

The included studies originated from six countries: China (5/10, 50%), and Canada, Germany, Japan, Poland, and South Korea (1/10, 10% each). All studies were published in peer‐reviewed journals, with five published in 2023 (5/10, 50%), four in 2024 (4/10, 40%), and one in 2019 (1/10, 10%). Surgical procedures studied included coronary artery bypass grafting, heart valve surgery, ascending aorta replacement, CPB, atrial septal defect repair, and ventricular septal defect repair. These procedures are widely recognised as significant risk factors for postoperative delirium due to features such as high complexity, substantial surgical trauma, and prolonged cardiopulmonary bypass duration (Huang et al. 2023). Of the 10 studies, five were prospective cohort studies, four were single‐center retrospective cohort studies, and one was a single‐center case–control study (Table 1).

TABLE 1.

Characteristics of the included studies (N = 10).

Author	Year of Publication	Country	Study design	Setting	Type of surgery
Han et al. (2024)	2024	South Korea	Prospective cohort study	Yonsei University Health System	Undertaking cardiac surgery with or without CPB
Nagata et al. (2023)	2023	Japan	Single‐center prospective study	Osaka University Hospital	CABG, HVS, AAR, CPB
Zuo et al. (2023)	2023	China	Single‐center case–control study	The Second Affiliated Hospital of Army Medical University	Coronary atherosclerotic heart disease, VHD, CHD
Mufti et al. (2019)	2019	Canada	Single‐center retrospective cohort study	The QEII HSC in Halifax	CABG, HVS
Nowakowska et al. (2023)	2024	Poland	Single center prospective cohort study	The Central Clinical Hospital of the Medical University of Lodz	CABG
Sadlonova et al. (2024)	2023	Germany	Prospective, single‐center, observational	The University Medical Center Gottingen	CABG, HVS, CPB
Li et al. (2024)	2024	China	Single‐center retrospective cohort study	Guangdong Provincial People's Hospital	HVS, CPB
Huang et al. (2023)	2023	China	Prospective cohort study	China Medical University Northern Theatre Command General Hospital	CABG, HVS
Yang et al. (2024)	2023	China	Single‐center retrospective cohort study	The Affiliated Hospital of Traditional Chinese Medicine of Southwest Medical University	OPCAB, ONCAB, ASD, VSD repair, aortic replacement
Zhao et al. (2024)	2024	China	Single‐center retrospective cohort study	Nanjing First Hospital	Underwent cardiac surgery under CPB

Abbreviations: AAR, ascending aortic replacement; ASD, atrial septal defect; CABG, coronary artery bypass grafting; CHD, congenital heart disease; CPB, cardiopulmonary bypass; HVS, heart valve surgery; ONCAB, on‐pump coronary artery bypass; OPCAB, off‐pump coronary artery bypass; VHD, valvular heart disease; VSD, ventricular septal defect repair.

3.3. Incidence of Delirium

IN total 11,702 participants were included across the 10 studies. Reported delirium incidence ranged from 5.56% to 34%, with six studies (6/10, 60%) reporting an incidence rate > 20% (Table 2). Heterogeneity in patient populations and delirium assessment methods may have contributed to substantial variability in the reported incidence of postoperative delirium following cardiac surgery.

TABLE 2.

Basic characteristics of AI models for delirium after cardiac surgery (N = 10).

Author	Model	Development methodology	Diagnostic tool for delirium	Delirium assessment time	Sample size, n	Positive outcome, n	Verification/interpretation method	Best model
Nowakowska et al. (2023)	AdaBoost, XGBoost, RF, GBT	GridSearch	CAM‐ICU, MDAS	Assess once a day within 5 days after surgery	224	61, (34%)	10‐CV	RF
Huang et al. (2023)	GBDT, SVM, RF, LogR, KNN, DNN	Python 3.8	CAM‐ICU, 3D‐CAM	Assess every 24 h within 7 days after surgery	710	151, (21.3%)	Fivefold stratified cross‐validation	GBDT, RF
Sadlonova et al. (2024)	LASSO‐R, DT	R, 4.2.2	CAM‐ICU, I‐CAM	The first five postoperative days	504	106, (21.0%)	10‐CV	—
Nagata et al. (2023)	BNB, SVM, RF, ET, XGBoost	Python 3.8.3	DSM‐5	The first seven postoperative days	123	24, (27.6%)	Stratified hold‐out method	ET, XGBoost
Mufti et al. (2019)	LogR, ANN, SVM, NB, BBN, DT, RF	R, 3.1.0	CAM‐ICU, Mini‐Mental State Exam and CAM	—	5584	661, (11.4%)	10‐CV	RF, ANN
Li et al. (2024)	RF, LR, SVC, KNN, GNB, GBDT, Perceptron	Python, 3.12.0	CAM‐ICU, RASS	The first seven postoperative days (twice daily)	507	141, (28%)	SHAP	RF
Zhao et al. (2024)	Catboost, ANN, XGboost, RF	R, 3.1.0	CAM‐ICU, RASS	The first three postoperative days (every 12 h)	885	221, (25.0%)	SHAP	ANN
Yang et al. (2024)	RF, SVM, RBFNN, KNN, KRR	R 2021	DSM‐5, ICDSC	—	367	105, (28.6%)	10‐CV	RF
Zuo et al. (2023)	LR, XGBoost	R, 4.2.2	CAM‐ICU	The first three postoperative days (7:00/19:00)	684	38, (5.56%)	—	XGBoost
Han et al. (2024)	RF, XGBoost, ET, GBC, LGBM	R, 4.2.0 Python, 3.10.6	ICDSC	The first seven postoperative days	2114	260, (12.29%)	SHAP	XGBoost

Abbreviations: 10‐CV, 10‐fold cross‐validation; 3D‐CAM, 3D Cognitive Assessment Measure; AdaBoost, adaptive boosting; ANN, artificial neural network; BBN, Bayesian belief networks; BNB, Bernoulli Naive Bayes; CAM‐ICU, Confusion Assessment Method for the Intensive Care Unit; Catboost, Categorical Boosting; DNN, deep neural networks; DSM‐5, Diagnostic and Statistical Manual of Mental Disorders‐5; DT, Decision Tree; ET, Extra‐trees; GBC, Gradient boosting classifier; GBDT, Gradient Boosting Decision Tree; GBT, Gradient Boosting Tree; GNB, Gaussian Naive Bayes; I‐CAM, Information‐Communication Assessment Measure; ICDSC, Intensive Care Delirium Screening Checklist; KNN, K‐nearest Neighbours Classifier; KRR, Kernel Ridge Regression; LASSO‐R, Least Absolute Shrinkage and Selection Operator Regression; LGBM, light gradient boosting machine; LogR, logistic regression; MDAS, Memorial Delirium Assessment Scale; NB, Naïve Bayesian; RASS, Richmond Assessment Sedation Scale; RBFNN, Radial Basis Function Neural Network; RF, Random Forest; SHAP, Shapley Additive Explanations; SVM, Support Vector Machines; XGBoost, Extreme Gradient Boosting.

3.4. Delirium Assessment Tools and Timing

Delirium assessment scales are routinely used in clinical practice to help healthcare professionals quickly and accurately identify delirium in patients. We provided a detailed overview (Table 2) of the delirium diagnostic tools used across the 10 included studies and found that the Confusion Assessment Method for the ICU (CAM‐ICU) was the most commonly used tool in the ICU (7/10, 70%). In general wards, tools included the Diagnostic and Statistical Manual of Mental Disorders‐5 (DSM‐5), Intensive Care Delirium Screening Checklist (ICDSC), Memorial Delirium Assessment Scale, 3D Cognitive Assessment Measure, Information‐Communication Assessment Measure, and Mini‐Mental State Examination. Delirium assessments were primarily conducted within 7 days postoperatively (4/10, 40%), followed by 5 days (2/10, 20%) and 3 days (2/10, 20%).

3.5. AI Technical Characteristics of Included Studies

AI model development was performed using R (6/10, 60%) and Python (4/10, 40%), reflecting the prevalent use of these languages in statistical and machine learning applications due to their extensive libraries and community support. Ten‐fold cross‐validation (Motsinger and Ritchie 2006) and Shapley Additive Explanations (SHAP) (Hu et al. 2024) assess model reliability. They use internal validation and post hoc explanation techniques to reduce model opacity, revealing key individual risk factors, their directions, and contributions. To ensure robust model evaluation, four studies used 10‐fold cross‐validation, a method that reduces overfitting and provides a more reliable estimate of model performance. Three applied SHAP to enhance interpretability, allowing for clearer insight into feature contributions and model decision‐making—a critical aspect in clinical and applied settings. Two used 5‐fold stratified cross‐validation and the stratified hold‐out method, respectively. The most frequently used AI model was RF (5/10, 50%), followed by XGBoost (3/10, 30%) and Artificial Neural Networks (ANN, 2/10, 20%), as shown in Table 2.

3.6. AI Model Performance

AUC ranges from 0 to 1, with values closer to 1 indicating better model performance and greater discriminatory ability. In the 10 studies, the AUC values of 47 models ranged from 0.544 to 0.92, with a majority of AUC values falling within 0.80–0.92 (19/47, 40.42%). Model performance was also evaluated using ACC, SEN, and SPE. ACC represents the overall proportion of correct predictions, SEN measures the model's ability to correctly identify true positive cases, and SPE measures its ability to correctly identify true negatives. All studies reported ACC and SEN; 15 models (15/47, 31.91%) showed ACC between 0.80 and 0.92, whereas only five (5/47, 10.63%) had SEN in that range. Seven studies reported SPE, and most models (16/30, 53.33%) had values exceeding 0.80.

3.7. Predictive Factors of Delirium After Cardiac Surgery

We conducted a statistical analysis of predictors for delirium following cardiac surgery across 10 studies and synthesised them into four categories. First, demographic and preoperative comorbidities: the strongest predictor was advanced age (6/10, 60%), followed by a history of cerebrovascular diseases (2/10, 20%). Second, surgical and perioperative factors: these included cardiopulmonary bypass time (minutes) (2/10, 20%) and pain score (2/10, 20%). Third, clinical and laboratory markers: these encompassed serum creatinine, preoperative sRAGE, preoperative SpO₂/rSO₂, preoperative MCP‐1, ejection fraction, albumin, total bilirubin (TBIL), blood pH, and eGFR. Fourth, psychological and functional assessment scores: these included PHQ‐9, MoCA, Mini‐Cog < 4, Katz Index < 4, GAD‐7, BIS < 40 or PSI < 25, Barthel Index < 10, ASA grade, and 6‐CIT. Further details are presented in Table 3.

TABLE 3.

Performance and predictor of AI model for delirium after cardiac surgery (N = 10).

Author	Model	AUC	ACC	SEN	SPE	F	PPV	NPV	Predictors
Nowakowska et al. (2023)	GBT	0.781	0.722	0.719	0.723	—	—	—	Age (years), Preop‐MCP‐1 (ng/mL), Preop‐sRAGE (ng/mL)
	RF	0.787	0.709	0.757	0.657	—	—	—
	XGBoost	0.744	0.669	0.67	0.664	—	—	—
	AdaBoost	0.727	0.583	0.92	0.245	—	—	—
Huang et al. (2023)	GBDT	0.86	0.77	0.76	—	0.75	—	—	GAD‐7, PHQ‐9, Chill, VAS pain score > 3 points, Mitral valve surgery, BMI
	SVM	0.79	0.71	0.72	—	0.69	—	—
	RF	0.85	0.76	0.76	—	0.74	—	—
	LR	0.67	0.63	0.7	—	0.62	—	—
	KNN	0.67	0.63	0.64	—	0.61	—	—
	DNN	0.78	0.75	0.61	—	0.67	—	—
Sadlonova et al. (2024)	LASSO	0.64	0.802	0.12	0.937	—	0.272	0.844	MoCA, 6‐CIT, TMTA, TMTB
	DT	0.621	0.816	0.166	0.982	—	0.714	0.821
Nagata et al. (2023)	BNB	0.76	0.73	0.55	0.94	0.6	0.77	—	Age (years), use of psychotropic drugs, Mini‐Cog < 4, Barthel Index < 100, history of stroke or cerebral haemorrhage, and eGFR < 60
	SVM	0.74	0.74	0.63	0.89	0.6	0.68	—
	RF	0.70	0.68	0.55	0.89	0.55	0.56	—
	ET	0.76	0.7	0.63	0.78	0.56	0.52	—
	XGBoost	0.75	0.74	0.67	0.79	0.59	0.55	—
Mufti et al. (2019)	ANN	0.804	0.717	0.718	0.716	0.717	0.717	0.717	Age (years), gender, ejection fraction, cerebrovascular disease, intra‐operative TEE, blood product transfusion within the first 48 h, MV than 24 h, length of stay in the ICU
	BBN	0.774	0.713	0.722	0.712	0.701	0.699	0.713
	DT	0.772	0.701	0.681	0.729	0.709	0.729	0.726
	LR	0.814	0.733	0.698	0.767	0.723	0.75	0.756
	NB	0.799	0.73	0.648	0.795	0.727	0.744	0.795
	RF	0.813	0.725	0.743	0.717	0.741	0.721	0.728
	SVM	0.811	0.713	0.602	0.838	0.672	0.778	0.831
Li et al. (2024)	RF	0.92	0.83	0.86	—	0.81	—	—	Pain score, anaesthesia duration, blood PH, TBLL, serum creatinine
	GBDT	0.9	0.84	0.8	—	0.8	—	—
	SVM	0.83	0.83	0.81	—	0.8	—	—
	LR	0.8	0.76	0.74	—	0.72	—	—
	GNB	0.79	0.76	0.76	—	0.73	—	—
	KNC	0.78	0.76	0.6	—	0.61	—	—
	Perceptron	0.77	0.78	0.73	—	0.73	—	—
Zhao et al. (2024)	ANN	0.684	0.316	0.545	—	0.4	—	—	Preop:ASA grade, degree of education, albumin. Post: APACHE II, Pre‐op MMSE score, CPB duration
	Xgboost	0.632	0.545	0.273	—	0.364	—	—
	Catboost	0.655	0.302	0.455	—	0.369	—	—
	RF	0.544	0.25	0.045	—	0.077	—	—
Yang et al. (2024)	RF	0.92	0.879	0.692	0.953	—	—	—	Age (years), education level, visual impairment, ASA grade, albumin, CPB time, aortic clamping
	SVM	0.893	0.842	0.645	0.919	—	—	—
	RBFNN	0.896	0.858	0.674	0.931	—	—	—
	KNN	0.733	0.752	0.35	0.912	—	—	—
	KRR	0.9	0.841	0.664	0.912	—	—	—
Zuo et al. (2023)	LR	0.732	0.947	0.5	1	0.49	0	0.947	Age (years), preoperative SpO2/rSO2
	XGBoost	0.659	0.951	0.67	0.985	0.71	0.571	0.965
Han et al. (2024)	XGBoost	0.851	0.851	0.621	0.89	0.545	0.486	0.933	eGFR, age (years), duration of BIS < 40 or PSI < 25, Katz grade 4
	ET	0.874	0.812	0.69	0.832	0.513	0.408	0.941
	LGBM	0.877	0.871	0.345	0.96	0.435	0.588	0.897
	RF	0.856	0.663	0.931	0.618	0.443	0.29	0.982
	GBC	0.871	0.827	0.448	0.89	0.426	0.406	0.906

Abbreviations: 6‐CIT, Six‐Item Cognitive Impairment Test; APACHE II, Acute Physiology and Chronic Health Evaluation II; ASA grade, American Society of Anesthesiologists Physical Status Classification System; BIS, bispectral index; BMI, body mass index; Chill, postoperative shivering; eGFR, estimated glomerular filtration rate; GAD‐7, Generalised Anxiety Disorder‐7; Mini‐Cog, Mini‐Cognitive Assessment; MMSE, Mini‐Mental State Examination; MoCA, Montreal Cognitive Assessment; MV, mitral valve surgery; PHQ‐9, Patient Health Questionnaire‐9; Preop‐MCP‐1, Preoperative Monocyte Chemoattractant Protein‐1; Preop‐sRAGE, Preoperative Soluble Receptor for Advanced Glycation End‐products; PSI, patient state index; TBLL, total bilirubin; TEE, transesophageal echocardiography; TMTA, Trail Making Test A; TMTB, Trail Making Test B; VAS, Visual Analog Scale.

3.8. Results of Risk Bias Assessment

Two researchers independently assessed the risk of bias and applicability of the included prediction models using the PROBAST tool. In the participants domain, three studies were judged to be at high risk of bias, mainly due to inappropriate data sources. In the analysis domain, all 10 studies were rated as high risk of bias, primarily because of insufficient sample sizes—failing to meet the recommended events per variable > 20 in eight studies—and reliance on univariable predictor selection in seven studies. Overall, all 10 studies exhibited a high risk of bias, whereas concerns regarding applicability were low, suggesting methodological shortcomings during model development or validation; applicability was generally good (Table 4).

TABLE 4.

Evaluation results of bias risk in literature prediction models (N = 10).

Study	ROB				Applicability			Overall
	Participants	Predictors	Outcome	Analysis	Participants	Predictors	Outcome	ROB	Applicability
Nowakowska et al. (2023)	+	+	+	−	+	+	+	−	+
Huang et al. (2023)	+	+	+	−	+	+	+	−	+
Sadlonova et al. (2024)	+	+	+	−	+	+	+	−	+
Nagata et al. (2023)	+	+	+	−	+	+	+	−	+
Mufti et al. (2019)	+	+	+	−	+	+	+	−	+
Li et al. (2024)	+	+	+	−	+	+	+	−	+
Zhao et al. (2024)	−	+	+	−	+	+	+	−	+
Yang et al. (2024)	−	+	+	−	+	+	+	−	+
Zuo et al. (2023)	−	+	+	−	+	+	+	−	+
Han et al. (2024)	+	+	+	−	+	+	+	−	+

Note: “+” indicates low risk of bias (ROB)/low concern regarding applicability, “−” indicates high ROB/high concern regarding applicability, and “?” indicates unclear ROB/unclear concern regarding applicability.

4. Discussion

This study investigates the application of artificial intelligence in predicting postoperative delirium among cardiac surgery patients, providing a basis for early prevention and the development of personalised care plans. We found that the incidence of postoperative delirium ranges from 5.56% to 34%, and commonly used assessment tools include CAM‐ICU, DSM‐5 criteria, and ICDSC. Advanced age, prolonged CPB, a history of cerebrovascular disease, and pain are core risk factors. AI prediction models (RF/XGBoost/ANN) performed well (AUC 0.80–0.92), but SEN was generally inadequate (models with high sensitivity accounted for only 10.63%).

4.1. Assessment and Prediction of Delirium Risk After Cardiac Surgery

Delirium can be categorised into three clinical phenotypes: hyperactive, hypoactive, and mixed (Krewulak et al. 2018). Differences among these phenotypes complicate assessment and pose significant challenges for the selection of assessment tools, the timing of assessments, and clinical management. In the 10 included studies, we found substantial variation in delirium assessment instruments and inconsistency in assessment timing; these differences may be the primary drivers of the heterogeneity in reported incidence. Despite methodological variations, current evidence consistently shows that postoperative delirium after cardiac surgery is highly prevalent and closely associated with prolonged hospital stay, increased complications, and delayed functional recovery (de la Varga‐Martínez et al. 2023), underscoring the clinical need for accurate and standardised assessment. The CAM‐ICU (Chen et al. 2021) was designed mainly for intensive care settings where hyperactive and mixed subtypes are more common, and it may be less sensitive in detecting the hypoactive subtype. To maximise detection accuracy, it is advisable to combine the CAM‐ICU with other assessment tools (Miranda et al. 2023), such as the ICDSC or DSM‐5–based clinical evaluation. The core objective of the 3D‐CAM is to enable a rapid, accurate, and standardised diagnosis of delirium in general inpatients (Marcantonio et al. 2014). The DSM‐5 provides not only diagnostic criteria but also guidance on differential diagnosis, subtype classification, severity assessment, and etiological considerations (Widiger and Hines 2022). Therefore, for patients on general wards, we recommend a stepwise, collaborative assessment pathway that uses the 3D‐CAM as a rapid screening tool and applies DSM‐5 criteria as the gold standard for definitive diagnosis and subtyping.

4.2. AI‐Based Predictive Models Effectively Forecast Postoperative Delirium Following Cardiac Surgery

AI‐based predictive models effectively assessed the risk of postoperative delirium in patients undergoing cardiac surgery by capturing both linear and nonlinear relationships and complex variable interactions within clinical datasets. This ability significantly enhances their predictive performance. Among the 10 studies reviewed, the most effective AI models included RF, XGBoost, and ANN. RF excels at managing multivariable interactions while providing interpretable risk prediction outputs (Hu and Szymczak 2023). Compared to traditional statistical techniques, XGBoost captures complex variable relationships, overcomes linear assumptions, and ranks risk factor importance. ANN is well‐suited to modelling nonlinear patterns, requires fewer parameters, and reduces computational demands. It performs well in training and accurately predicts unseen data, effectively uncovering latent patterns (Ranade and Ranade 2023). Therefore, AI‐based predictive models can effectively assess the risk of postoperative delirium in patients undergoing cardiac surgery, and future research should expand the application of AI in this field.

All 10 included studies reported AUC, ACC, and SEN, whereas seven studies reported SPE. A considerable proportion of models achieved AUC values between 0.80 and 0.92 (19/47, 40.42%) and ACC values within the same range (15/47, 31.91%), indicating strong discriminative ability and prediction ACC. However, only a few models achieved similarly high sensitivity (5/47, 10.63%), reflecting limitations in correctly identifying patients who developed delirium and the potential for missed diagnoses. Among the seven studies that reported specificity, more than half of the models (16/30, 53.33%) had SPE values exceeding 0.80, indicating effective identification of patients who did not develop delirium. Besides, when high SPE is accompanied by low SEN, common contributors include pronounced class imbalance in the training data (far fewer delirium than non‐delirium cases), overly conservative decision thresholds, or overly aggressive feature selection, which can bias the model toward the negative class (Granata et al. 2024). We noted that Davoudi et al. (2017) applied data‐level techniques—oversampling and undersampling—but observed no significant improvement (AUROC, 0.79–0.86). Mufti et al. (2019) undersampled the majority class to achieve a 1:1 class ratio, thereby producing a more balanced dataset and narrowing inter‐class disparities. Provided that class imbalance and related issues are appropriately addressed, AI models such as RF, XGBoost, and ANN show promise in analysing clinical data from cardiac surgery patients, enhancing risk prediction and enabling early identification of individuals at high risk for postoperative delirium.

4.3. Translating Predictors Into Practice: Explainable AI to Enable Early Identification

Bridging the gap between model development and clinical implementation requires addressing performance limitations and enhancing interpretability. Emerging AI technologies are expanding the boundaries of early delirium identification by leveraging data sources beyond structured clinical variables. For instance, Fu et al. (2022) developed a natural language processing (NLP) algorithm based on the CAM framework to efficiently and accurately identify delirium events from complex texts in electronic health records, offering substantial support for early detection. Similarly, Wang et al. (2022) applied sentiment‐based NLP methods, demonstrating that extracting emotional cues from clinical text data can improve identification accuracy and highlighting the potential value of emotional information in diagnosis. Beyond textual data, Sun et al. (2019) employed a deep learning model combining convolutional and recurrent neural networks to monitor consciousness levels and detect delirium via frontal lobe EEG signals in the ICU. Unlike behaviour‐based assessments, this model directly analyzes physiological signals, eliminating subjective interpretation and providing a more timely and accurate assessment of patient condition. These approaches exemplify the potential of AI to enable earlier, more objective identification of delirium. However, one of the key challenges in translating AI models from research into clinical practice is striking a balance between model performance and interpretability. Owing to the models' inherent black‐box nature and limited generalizability, studies often lack sufficient validation, making it difficult to build trust among physician and nursing teams and to ensure predictive accuracy, thereby impeding the development of concrete, actionable interventions (Mohammad‐Rahimi et al. 2024; Muhammad and Bendechache 2024). The papers we included primarily employed 10‐fold cross‐validation and SHAP to assess model reliability; these internal validation techniques and post hoc explanation methods help reduce model opacity: they not only provide risk scores but also reveal, at the individual level, the main risk drivers along with their direction and relative contribution. For example, results from 10‐fold cross‐validation and SHAP indicate that delirium risk is commonly associated with advanced age, prolonged cardiopulmonary bypass time, and cerebrovascular disease factors (Mufti et al. 2019; Yang et al. 2024). This not only further strengthens confidence in the predictive model but also enables physicians and nurses to devise proactive, personalised prevention and management plans tailored to each patient's specific risk profile. Therefore, we advocate that interpretability be treated as a core requirement, rather than an optional add‐on, in future model development and validation.

Another priority issue is the suboptimal sensitivity observed in some models, which may hinder early identification of high‐risk patients. To mitigate patient safety risks from missed cases, we recommend using AI‐identified key predictors as cues and implementing nurse‐led, concrete, and actionable bedside strategies: prioritise comprehensive preoperative assessment for high‐risk patients (de la Varga‐Martínez et al. 2021), intensify continuous postoperative monitoring, and lead nonpharmacologic interventions such as reorientation, encouraging family involvement and presence, providing hearing aids and eyeglasses to reduce sensory deprivation, and implementing sleep‐promotion measures (Olotu et al. 2022). These initiatives can enable early prevention and mitigation of delirium, reduce the potential risks associated with suboptimal model sensitivity, and thereby decrease the occurrence of patient safety incidents.

4.4. From Identification to Intervention: Nurse‐Led Risk Mitigation Strategies

Delirium prediction is intrinsically linked to the management of risk factors, and the key predictors identified by AI models inform highly targeted nursing interventions. Among the 10 studies included in this review, age emerged as the most prominent factor, primarily due to age‐related declines in neuroplasticity, neuroendocrine dysregulation, and reduced functional reserve (Liu et al. 2022). For older patients, nurses should adopt a team‐based collaborative model involving clinicians and family members to ensure comprehensive support across the preoperative, intraoperative, and postoperative phases. Such collaboration enables effective identification and management of potential delirium risks as well as helps alleviate patient anxiety and improve postoperative recovery quality through individualised care plans and continuous monitoring. Prolonged CPB is another significant predictor because it can disrupt cerebral blood flow, lead to hypoxia, and cause metabolic imbalance (Jufar et al. 2021). Extended CPB duration is significantly associated with delirium risk (Chen et al. 2024). CPB is a known trigger of systemic inflammatory responses; proinflammatory cytokines released during CPB can disrupt the blood–brain barrier and induce neuroinflammation, leading to neuronal dysfunction (Gilbey et al. 2023). Therefore, nurses need to closely monitor hemodynamic stability and cerebral oxygenation, promptly communicate findings to the medical team, and implement goal‐directed hemodynamic management protocols. A history of cerebrovascular disease has also emerged as a key predictor. Its associations with systemic inflammation and inadequate oxygen delivery increase cerebral vulnerability during and after surgery (Johansen and Gottesman 2021; Kumar et al. 2023). To address this, nurses should focus on interventions such as optimising mechanical ventilation, using vasoactive agents, timely transfusion, and reducing ICU length of stay, which may improve cerebral oxygenation and reduce the incidence of delirium (Tian et al. 2021). Pain, considered the fifth vital sign, can trigger physiological stress, sympathetic activation, inflammation, and impaired cerebral perfusion, all of which exacerbate cognitive dysfunction and delirium (O'Gara et al. 2021). Pain is a key factor that nurses can independently assess and initially intervene on. Effective pain assessment tools (e.g., NRS, CPOT) should be used, and the principles of scheduled and multimodal analgesia should be followed to optimise analgesia and sedation regimens, avoid over‐ or under‐sedation, promote early mobilisation, and encourage rehabilitation exercises. These interventions improve circulation, support physical function, reduce pain, and help prevent delirium (Hume et al. 2024).

Beyond static risk factors, multiple studies have shown differences in dynamically monitored indicators between delirium and non‐delirium groups. Two studies observed that the mean regional cerebral oxygen saturation was lower in the delirium group than in the non‐delirium group; one reported a statistically significant difference (55.9% [49.4, 60.4] vs. 58.7% [54.2, 62.9], p < 0.001) (Han et al. 2024), whereas the other showed a downward trend that did not reach statistical significance (58% [53, 62] vs. 61% [56.5, 65.5], p = 0.067) (Zuo et al. 2023). In addition, the delirium group had lower preoperative systolic blood pressure (119 [110, 135] mmHg vs. 133 [116, 148] mmHg, p = 0.002) and greater variability in mean arterial pressure on continuous monitoring (Zuo et al. 2023). Another study reported slightly higher postoperative blood glucose levels in the delirium group (9.1 mmol/L vs. 8.9 mmol/L) (Zhao et al. 2024). As frontline perioperative practitioners and care coordinators, nurses should deliver interdisciplinary, patient‐tailored interventions based on comprehensive assessments of both static and dynamic risk factors, thereby reducing the incidence of delirium and enhancing perioperative recovery.

4.5. Limitations

This review ensured research quality and relevance through a rigorous screening process. However, excluding non‐English and non‐Chinese literature may result in some studies on using AI to predict delirium in post‐cardiac surgery patients being left out. Since most studies (8 out of 10, or 80%) are single‐center investigations, their results are limited in applicability to other populations and healthcare systems. Therefore, it is crucial to conduct multicenter studies that account for geographical and clinical diversity, as well as to perform external validation. Although RF and XGBoost demonstrated strong performance in these studies, notable variability was observed in key performance indicators (e.g., AUC, ACC, SEN, SPE). Two main factors contributed to this variation. At the data level, differences in dataset size, quality, and distribution affected generalizability. Small or imbalanced datasets weakened model performance, especially on minority class predictions. At the modelling level, discrepancies in feature selection methods and hyperparameter tuning also significantly impacted results (Gui et al. 2024; Welvaars et al. 2023; Xie et al. 2020). Currently, most delirium prediction models are built primarily on static clinical variables (such as preoperative comorbidities, demographic characteristics, and surgery‐related timing variables), while the use of dynamic high‐frequency time‐series data remains limited and has been reported in only a few studies. Although static variables have some predictive value, they struggle to reflect the continuous changes in patients' postoperative physiological status. Future research should prioritise integrating dynamic, high‐frequency time‐series data—such as continuous noninvasive physiological monitoring data (heart rate, blood pressure, oxygen saturation, etc.) (Park et al. 2025), wearable device data (e.g., sleep–wake rhythms) (Angelucci et al. 2025), and EEG signals (Mulkey et al. 2023). Combining these sources to develop delirium prediction tools with higher sensitivity and more timely responsiveness. At the same time, model explainability and cross‐center external validation should be strengthened to promote a paradigm shift in AI‐based tools from “static risk prediction” to “dynamic risk monitoring.”

5. Conclusion

This review evaluated the application of artificial intelligence in predicting postoperative delirium following cardiac surgery. Although accuracy is high, most studies remain at the stages of model development and internal validation, with limited clinical integration or real‐world external validation. Bridging this gap will require multicenter external validation, workflow‐centered algorithm optimization, and incorporation of dynamic perioperative variables. We also identified consistent predictors; these findings provide a foundational evidence base for designing targeted screening protocols and proactive interventions for high‐risk patients, ultimately aiming to improve postoperative outcomes and patient safety.

Author Contributions

Centao Qin: visualisation, writing – original draft, writing – review and editing. Lu Zeng: writing – review and editing. Jinbo Zhang: writing – review and editing. Juan Zhang: writing – review and editing. Ming Tao: writing – review and editing. Jiamei Zhou: writing – review and editing.

Funding

The authors have nothing to report.

Consent

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Data S1: jan70456‐sup‐0001‐supinfo.docx.

Data Availability Statement

Available from the first author upon request.