Systematic review and comparison of machine learning and conventional statistical models for predicting cardiovascular events in dialysis patients

Abstract

This systematic review aimed to evaluate the performance of machine learning (ML) models and conventional statistical models (CSMs) for predicting cardiovascular events in dialysis patients. Following PRISMA guidelines, eligible studies were searched through PubMed and Embase (January 2015–March 2025). Model performance (AUC/C-index) was compared using Mann–Whitney U test, while risk of bias was assessed via PROBAST. Furthermore, subgroup analyses stratified by algorithm type, validation method, and dataset size were conducted to explore heterogeneity. The review included 14 studies encompassing 29,310 patients and 34 models. Based on test/validation datasets only, ML models achieved comparable discrimination (mean AUC: 0.784 ± 0.112) than CSMs (0.772 ± 0.066), without statistical significance (p = 0.24). The PROBAST assessment indicated that 71.43% of studies had a low risk of bias. Subgroup analysis of performance revealed that deep learning models significantly outperformed both traditional ML and CSMs (p = 0.005), whereas traditional ML showed no advantage over CSMs (p = 0.727). Studies were predominantly originated from China (71.40%) and relied on internal validation (78.57%), limiting generalizability. Although deep learning algorithms show promises, ML models overall do not significantly outperform CSMs. CSMs remain viable, especially in resource-limited settings. Critical limitations include geographical bias, insufficient external validation, and tradeoffs between accuracy and interpretability. Future research should prioritize validation frameworks and clinical implementation over marginal accuracy improvements.

1. Introduction

End-stage kidney disease (ESKD) poses a formidable global health challenge, with dialysis serving as both a vital lifeline for patients and a significant strain on healthcare systems. As of 2022, United States Renal Data System (USRDS) reported that there were 642,353 dialysis patients in the United States, a number projected to rise by 3-4% [1] annually due to the increasing prevalence of diabetes and hypertension [2]. Among these patients, cardiovascular disease (CVD) emerges as the predominant cause of mortality in dialysis patients, accounting for approximately 40–50% of deaths [3–5]. Among these, heart failure, atrial fibrillation, acute coronary syndromes, and sudden cardiac death are the most prevalent. Cardiovascular events are markedly elevated in dialysis patients due to a complex interplay between traditional risk factors (e.g., dyslipidemia, smoking) and dialysis-specific pathophysiological mechanisms such as uremia, oxidative stress, and vascular calcification. This synergistic burden results in a risk of ischemic heart disease, heart failure, and sudden cardiac death up to 20 times higher than in the general population [6,7]. Given the distinct milieu of chronic inflammation, volume overload, and accelerated vascular calcification inherent to dialysis, this systematic review specifically targets this population [8].

Artificial intelligence (AI) has become increasingly prominent in healthcare field, offering the transformative tools to enhance patient care and clinical decision-making. Machine learning algorithms, including Decision Trees (DT), Random Forests (RF), Support Vector Machines (SVM), K-Nearest Neighbor (KNN), Naive Bayes (NB), and Extreme Gradient Boosting (XGBoost) algorithms, have been widely adopted to refine disease prediction models by analyzing large datasets to identify risk factors and predict patient outcomes [9,10]. In the domain of dialysis, machine learning (ML) application includes the predictive models for intradialytic hypotension [11], anemia control systems [12], and vascular access assessments [13], addressing critical challenges in renal care. Similarly, in the domain of cardiology, application encompasses the automated detection of arrhythmias from electrocardiogram, predictive models for heart failure readmissions, and assessments of coronary artery disease from cardiac imaging, which show the significant potential in cardiac care [14,15]. These advancements highlight AI’s potential to revolutionize healthcare by delivering personalized, data-driven solutions.

Recent advancements in ML hold promise in improving cardiovascular risk prediction by leveraging complex data patterns that traditional statistical models may overlook [16]. However, challenges such as the need for large, high-quality datasets and ensuring model interpretability remain significant barriers to clinical adoption, particularly in specialized populations like dialysis patients with unique clinical characteristics. Compared to conventional statistical models (CSMs), such as Logistic regression and Cox regression, ML models provide advantages for disease prediction, primarily through enhanced flexibility and data handling capabilities. ML algorithms can effectively analyze complex clinical datasets without rigid adherence to statistical assumptions often unmet in practice [17]. Furthermore, they leverage modern computational power to process large-scale electronic health record (EHR) data [18] and integrate diverse, multi-modal inputs (clinical, genomic, imaging) for more comprehensive predictions [19]. Unique challenges in dialysis patient data, including variable treatment modalities, heterogeneous protocols, and frequent missing laboratory values, make this population particularly suitable for ML approaches that excel at handling complex and incomplete datasets. Although these factors make ML attractive, selecting the optimal algorithm requires careful consideration. As highlighted by Uddin et al. [20] in a systematic review, although SVM is commonly used, RF frequently demonstrates superior accuracy, indicating that algorithm choice should be tailored to the specific disease and data properties.

Therefore, this systematic review aimed to compare the predictive performance of AI-based models versus traditional statistical models for forecasting cardiovascular events in dialysis patients. Through a synthesis of existing evidence, we analyzed key performance metrics, such as the Area Under the Curve (AUC), Concordance Index (C-Index), Accuracy, Sensitivity, and Specificity to delineate the strengths and limitations of each approach. Additionally, we assessed the interpretability, quality, and risk of bias in the development and validation processes of these models.

2. Methods

2.1. Search strategy

This systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) checklist [21] (Supplementary file S1). A comprehensive literature search was performed across PubMed and EMBASE to identify relevant articles published from 1 January 2015, to 20 March 2025. The search strategy was formulated by integrating Medical Subject Headings (MeSH) with free-text keywords. We constructed the Retrieval by using Boolean operators to comprise three core domains: artificial intelligence (encompassing terms such as ‘artificial intelligence,’ ‘machine learning,’ and ‘deep learning’), dialysis (incorporating terms such as ‘dialysis,’ ‘hemodialysis,’ and ‘peritoneal dialysis’), and cardiovascular events (including terms such as ‘heart failure,’ ‘coronary artery disease,’ ‘myocardial infarction,’ ‘myocardial ischemia,’ ‘atrial fibrillation,’ ‘sudden cardiac death,’ ‘atherosclerosis,’ and ‘stroke’). A detailed list of search terms and the full search strategy is provided in Supplemental file S2.

2.2. Inclusion criteria

Studies were eligible for inclusion if they met the following criteria: original research articles (e.g., cohort studies, case-control studies); ML models (e.g., random forests, decision trees, support vector machines, neural networks, or K-nearest neighbor) or Conventional Statistical Models (e.g., Logistic Regression, Cox Regression); predictive models for cardiovascular outcomes (e.g., major adverse cardiovascular events (MACE), sudden cardiac death (SCD), ischemic heart disease (IHD), heart failure (HF), or arrhythmias); patients undergoing dialysis (hemodialysis or peritoneal dialysis); human subjects; and journal articles published in English with full text or abstracts available.

2.3. Exclusion criteria

Studies were excluded if they met any of the following conditions: non-original research (e.g., reviews, systematic reviews, conference abstracts, letters, guidelines, meta-analyses, or case reports); lack of prediction or risk assessment of cardiovascular complications in dialysis patients (e.g. studies focusing solely on incidence, treatment, or chronic kidney disease populations without specifying dialysis patients); did not utilize the predictive models or were purely descriptive without predictive analysis (e.g., reporting only statistical associations); non-human subjects (e.g., animal studies, in vitro studies, or pure cell experiments); study populations not explicitly limited to dialysis patients; not published in English, or full text was unavailable. Studies with only abstracts were excluded due to insufficient detail for comprehensive data extraction and quality assessment required for this systematic review; studies with missing data on key performance metrics were excluded, while studies with partial missing data were included if the available metrics were sufficient for analysis.

2.4. Data extraction

Data extraction was conducted independently by two researchers (YFL and CYC), with discrepancies resolved through consensus or by a third researcher (HLX). The following data were extracted from each included article: (i) first author, publication year, country; (ii) study characteristics (dataset, study design, population, outcome, sample size); (iii) model specifications (model type, algorithms, predictor variables); (iv) model metrics performance. For studies reporting multiple models or outcomes, we prioritized data most directly applicable to predicting major cardiovascular events in dialysis patients. Performance metrics were extracted from validation or test datasets to ensure the unbiased estimates. When studies reported incomplete performance metrics or missing data, we contacted corresponding authors for additional information. If data remained unavailable, we documented this as a limitation and proceeded with available data.

2.5. Quality assessment

The PROBAST was employed to evaluate the transparency and completeness of reporting in model development and validation processes [22]. Additionally, this tool was applied to assess the risk of bias and applicability across four key domains: participants, predictors, outcomes, and analysis (Supplementary file S3). Each study was classified as having low, moderate, or high risk of bias and applicability concerns. The assessment was conducted independently by two reviewers (YFL and JXQ), with any discrepancies resolved through discussion or consultation with a third reviewer (HLX) to ensure objectivity and reliability.

2.6. Statistical analysis

To evaluate model performance, we extracted metrics such as AUC or C-Index from each study. AUC/C-index was selected as primary metrics given their established role as gold standards for discriminative ability assessment and consistent reporting across clinical studies [23]. Due to substantial heterogeneity in dialysis modalities, follow-up periods, and outcome definitions, formal meta-analysis was deemed inappropriate. We therefore applied descriptive statistics: means ± standard deviations (SD) for normally distributed data (mean AUC ± SD), and medians with interquartile ranges (IQR) for skewed data (median AUC [IQR]).

Models were categorized as machine learning or conventional statistical. Performance was compared using AUC values, with an independent samples t-test for normally distributed AUCs (normality assessed via Shapiro-Wilk test and histograms) or the Mann–Whitney U test otherwise. To address heterogeneity concerns and explore performance variations, we conducted subgroup analyses stratified by [1]: algorithm type (conventional statistical models, traditional machine learning, and deep learning models) [2]; validation method (internal validation vs. internal and external combined validation) [3]; dataset size (<1,000, 1,000–3,000, and >3,000 patients). Kruskal-Wallis tests were used for multi-group comparisons with post hoc pairwise comparisons using Bonferroni correction.

Effect sizes were calculated to quantify the practical significance of observed differences. Cohen’s d was computed for parametric comparisons and Cliff’s Delta for non-parametric assessments, with interpretation following established conventions (negligible: <0.2, small: 0.2–0.5, medium: 0.5–0.8, large: >0.8). Sensitivity analyses were performed to test the robustness of findings [1]: excluding extreme AUC values (<0.6 or >0.9) [2]; restricting analysis to test and validation datasets only; and [3] selecting only the best-performing model from each study to address data dependency concerns. Statistical analyses were conducted with Python 3.10.7, with significance set at p < 0.05.

3. Results

3.1. Study characteristics

The initial literature search yielded 1,006 records from PubMed (n = 464) and Embase (n = 542). After removing 252 duplicate records, 754 unique records remained for screening. Two independent reviewers evaluated the titles and abstracts, excluding 400 records that did not meet the predefined inclusion criteria. The full texts of the remaining 354 articles were retrieved and independently assessed. Studies were excluded if they did not exclusively involve dialysis patients, did not predict or assess cardiovascular risks, or did not utilize predictive models. Following this assessment, 14 studies were eligible and included in the review. The complete study selection process was illustrated in the flowchart (Figure 1).

Figure 1.

PRISMA flow diagram of study selection.

Of the 14 publications included in this review (2015–2025), geographical representation was primarily from East Asia (China: n = 10, 71.40%; Japan: n = 2, 14.20%), with limited European contribution (Germany: n = 1, 7.10%; Italy: n = 1, 7.10%). Geographic distribution analysis revealed pronounced clustering in East Asian countries, particularly China and Japan, with minimal Western population representation. This geographic concentration may introduce population-specific factors that may affect the generalizability of predictive relationships and model performance across diverse healthcare settings and patient populations.

Study designs comprised retrospective cohort studies (n = 12, 85.70%), prospective cohort study (n = 1, 7.10%), and longitudinal observational study (n = 1, 7.10%). Data sources were categorized as national registries (n = 7, 50.00%), multi-center databases (n = 3, 21.40%), and single-center records (n = 4, 28.60%). The study populations encompassed various dialysis modalities: hemodialysis/maintenance hemodialysis (n = 7, 50.00%), peritoneal dialysis/maintenance peritoneal dialysis/continuous ambulatory peritoneal dialysis (n = 5, 35.70%), and unspecified dialysis (n = 2, 14.30%). The aggregate sample size was 29,310 patients, demonstrating marked heterogeneity in sample distribution (range: 314–7,539; mean: 2,093.6 ± 2,188.5; median:1020.5, IQR: 2934.5). Cardiovascular outcome definitions were classified into isolated clinical endpoints (n = 8, 57.10%) and composite endpoints (n = 6, 43.90%). Studies employed both CSMs (n = 12) and ML models (n = 15) to predict the cardiovascular events. The detailed information is depicted in Table 1.

Table 1.

Extracts of selected studies.

Author (Year)	Country	Dataset	Study design	Population	Outcome	Sample Size	Model type
Yu et al. (2018) [24]	China	HPDR	Retrospective cohort study	PD	CM	3054	LR
Mezzatesta et al. (2019) [25]	Italy	IFC-CNR HEMO clinical trial	Retrospective cohort study	Dialysis	CVD CHF IHD ARR CCVD	1383	LR, KNN, CART, NB, SVCL, SVCR, SVCP
Wang, YF. et al. (2022) [26]	China	The Third Affiliated Hospital of Southern Medical University The Third Affiliated Hospital of Sun Yat-sen University	Retrospective cohort study	HD	HF	353	XGBoost, LR
You, X. et al. (2021) [27]	China	The Third Affiliated Hospital of Southern Medical University The Third Affiliated Hospital of Sun Yat-sen University	Retrospective cohort study	HD	Cardiovascular Events	370	Cox Regression
Xu, Z. et al. (2024) [28]	China	PDTAP	Retrospective cohort study	PD	MACE	7,539	LR, CVEformer, Hi-BEHRT, SAFARI, A-LSTM, Deepr, RF, SVM
Zhang et al. (2022) [29]	China	Dialysis Facility of the General Hospital of Northern Theater Command	Retrospective cohort study	HD	5-year risk of CV Events All-Cause Mortality	398	LR
Matsubara et al. (2017) [30]	Japan	J-DOPPS	Longitudinal observational study	HD	MACE All-Cause Death	3,601	LR
de Gonzalo-Calvo et al. (2020) [31]	German	AURORA	Retrospective cohort study	HD	MACE	810	CART
Inaguma et al. (2019) [32]	Japan	AICOPP	Prospective cohort study	consecutive dialysis	First CV Events All-Cause Death	1,520	Cox Regression
Xu, X. et al. (2024) [33]	China	PDTAP	Prospective cohort study	PD	All-Cause Death Cardiovascular Death	7,539	CVDformer, LR, LSTM, ConCare
Wang, YX. et al. (2023) [34]	China	Kidney Disease Center of the First Affiliated Hospital, Zhejiang University School of Medicine	Retrospective cohort study	MPD	CVC	1,035	LR
You, X. et al. (2022) [35]	China	The Third Affiliated Hospital of Southern Medical University The Third Affiliated Hospital of Sun Yat-sen University	Prospective cohort study	HD	First CVD Event	388	Cox Regression, LR
Xu, L. et al. (2024) [36]	China	Department of Nephrology, Fujian Provincial Hospital	Retrospective cohort study	CAPD	HF Hospitalization All-Cause Mortality	1,006	XGBoost, RF, AdaBoost
Tong et al. (2024) [37]	China	The First Affiliated Hospital of Hainan Medical College	Retrospective cohort study	MHD	AIS	314	LR

Note: PD: Peritoneal Dialysis; HD: Hemodialysis; MPD: Maintenance Peritoneal Dialysis; CAPD: Continuous Ambulatory Peritoneal Dialysis; MHD: Maintenance Hemodialysis; CM: Cardiovascular Mortality; CVD: Cardiovascular Disease; CHF: Congestive Heart Failure; IHD: Ischemic Heart Disease; ARR: Arrhythmia; CCVD: Coronary Cardiovascular Disease; HF: Heart Failure; CV: Cardiovascular; MACE: Major Adverse Cardiovascular Event; CVC: Cardiac Valve Calcification; AIS: Acute Ischemic Stroke; LR: Logistic Regression; KNN: K-Nearest Neighbors; CART: Classification and Regression Trees; NB: Naive Bayes; SVCL: Support Vector Classifier with Linear kernel; SVCR: Support Vector Classifier with Radial Basis Function kernel; SVCP: Support Vector Classifier with Polynomial kernel; LSTM: Long Short-Term Memory; RF: Random Forest; Hi-BEHRT: Hierarchical Bidirectional Encoder Representations from Transformers; SAFARI: Sequential Attention-based Feature and Representation Integrator; A-LSTM: Attention-based Long Short-Term Memory; Deepr: Deep Learning for Patient Representation; SVM: Support Vector Machine; XGBoost: Extreme Gradient Boosting.

3.2. Model characteristic

3.2.1. Model development and validation

A total of 34 prediction models were identified and classified into two distinct categories based on their methodological type. These categories include: Conventional Statistical Models, such as Logistic Regression (29.41%, 10/34) and Cox Regression (8.82%, 3/34); Machine Learning Models (61.76%, 21/34), such as KNN (K-Nearest Neighbors), CART (Classification and Regression Trees), NB (Naive Bayes), SVCL (Support Vector Classifier Linear), SVCR (Support Vector Classifier Radial), SVCP (Support Vector Classifier Polynomial), XGBoost (Extreme Gradient Boosting), CVEformer (Cardiovascular Events Transformer), Hi-BEHRT (Hierarchical Bidirectional Encoder Representations from Transformers), SAFARI (Self-Attention based Framework for Adverse event Recognition and Interpretation), A-LSTM (Attention Long Short-Term Memory), Deepr (Deep Residual Network), RF (Random Forest), SVM (Support Vector Machine), CVDformer (Cardiovascular Disease Transformer), LSTM (Long Short-Term Memory), ConCare (Contextual Care Model), and AdaBoost (Adaptive Boosting) (Table 2).

Table 2.

Model classification Count and percentage.

Category	Model	Count	Percentage
Conventional Statistical Model	LR	10	29.41%
	Cox Regression	3	8.82%
Machine Learning Model	KNN	1	2.94%
	CART	2	5.88%
	NB	1	2.94%
	SVCL	1	2.94%
	SVCR	1	2.94%
	SVCP	1	2.94%
	XGBoost	2	5.88%
	CVEformer	1	2.94%
	Hi-BEHRT	1	2.94%
	SAFARI	1	2.94%
	A-LSTM	1	2.94%
	Deepr	1	2.94%
	RF	2	5.88%
	SVM	1	2.94%
	CVDformer	1	2.94%
	LSTM	1	2.94%
	ConCare	1	2.94%
	AdaBoost	1	2.94%

For model validation, 11 studies conducted internal validation only, none focused exclusively on external validation, 3 studies employed both internal and external validation (Table 3). Validation methodology analysis revealed a striking paucity of external validation across included studies. The vast majority of studies relied exclusively on internal validation techniques, with external validation using independent datasets or different institutions being notably rare. This validation pattern was consistently observed across both ML and conventional statistical approaches, indicating that validation limitations were not specific to particular methodologies but rather reflected broader trends in the field.

Table 3.

Validation method.

Validation method	Number of studies	Percentage
Internal validation only	11	78.57%
External validation only	0	0.00%
Both internal and external validations	3	21.43%

3.2.2. Model predictors

Our systematic review encompassing 14 studies identified a total of 460 predictors incorporated into the final prediction models. After eliminating duplications, 144 unique predictor variables were selected. These predictors could be categorized into five distinct domains: Demographics, Clinical Characteristics, Dialysis-related Variables, Laboratory Variables, and Novel Biomarkers.

Demographics include factors such as age, sex, and socioeconomic status. Clinical Characteristics encompass medical history and symptoms, such as diabetes, hypertension, and heart failure. Dialysis-related Variables cover treatment-specific aspects, such as dialysis modality and duration. Laboratory Variables consist of routine test results, including hemoglobin, phosphorus, and albumin levels. Novel Biomarkers represent innovative indicators, such as specific microRNAs. The complete list of predictors within each category can be found in the Supplementary file S4.

The frequency analysis revealed that age was the most common predictor (92.86%, 13/14), followed by diabetes (71.43%, 10/14), phosphorus (42.86%, 6/14), CRP (42.86%, 6/14), albumin (35.71%, 5/14), and sex (35.71%, 5/14) as the six most frequently utilized predictors. Other important predictors included systolic blood pressure (28.57%, 4/14), WBC (28.57%, 4/14), hypertension (28.57%, 4/14), hemoglobin (28.57%, 4/14), BMI (21.43%, 3/14), smoking status (21.43%, 3/14), dialysis time per session (21.43%, 3/14), urine volume (21.43%, 3/14), hemoglobin (21.43%, 3/14), serum creatinine (21.43%, 3/14), LDL (21.43%, 3/14), and sodium (21.43%, 3/14). The frequency of predictors in included studies is shown in Figure 2.

Figure 2.

Frequency of predictors in included studies.

Conventional statistical models and machine learning models demonstrated distinct approaches to predictor selection. CSMs utilized fewer variables (mean: 9.33 ± 6.44, median:7, IQR:5) selected primarily through clinical relevance and univariate significance testing. In contrast, machine learning models incorporated substantially more predictors (mean:62.67 ± 94.70, median:15.50, IQR:86), with some models utilizing extensive variable arrays. The expanded capacity for predictor inclusion in machine learning approaches allowed more comprehensive representation of the complex pathophysiology underlying cardiovascular risk in dialysis patients. Average predictors count by model type and predictor count distribution by study is shown in Figure 3.

Figure 3.

Average predictor count and predictor count distribution.

Note: Left panel: Mean predictor count comparison (Statistical Models: 9.33 ± 6.44; Machine Learning Models: 62.67 ± 94.70). Right panel: Predictor distribution by study, with study IDs corresponding to Table 1. Blue: Statistical Models, Red: Machine Learning Models. The horizontal line within each category denotes the mean number of predictors.

3.2.3. Model performance

Across the 14 reviewed studies, 16 unique performance metrics were identified. These metrics comprised four categories: Discrimination metrics were most prevalent, with AUC/AUROC/AUPRC appearing in 14 studies (100.00%) and C-index/C-statistic in 4 studies (28.57%); classification metrics included accuracy (28.57%), sensitivity/recall (28.57%), precision/PPV (21.43%), F1-score (21.43%), specificity (7.14%), and confusion matrices (7.14%); reclassification metrics were represented by Net Reclassification Improvement Index (NRI) and Integrated Discrimination Improvement (IDI), each appearing in 2 studies (14.29%); less commonly utilized were calibration metrics (calibration slope, 7.14%), clinical utility metrics (Decision Curve Analysis, 7.14%), and explainability metrics (SHAP, 7.14%). Additional metrics included model validation metrics (test error with cross-validation, 7.14%) (Table 4).

Table 4.

List of metrics and their occurrence in number and percentages.

Metric	Studies (n)	Percentage (%)
Discrimination metrics
AUC/AUROC/AUPRC	14	100.00
C-index/C-statistic	4	28.57
Classification metrics
Accuracy	4	28.57
Sensitivity/recall	4	28.57
Precision/PPV	3	21.43
F1-score	3	21.43
Specificity	1	7.14
Confusion matrices	1	7.14
Reclassification metrics
NRI	2	14.29
IDI	2	14.29
Other metrics
Calibration slope	1	7.1
Decision curve analysis	1	7.1
SHAP	1	7.1
Test error (CV)	1	7.1

The performance of predictive models varied across the 14 included studies, with AUC/C-index ranging from 0.500 to 0.943 when considering test and validation set results across all models. Conventional statistical models demonstrated AUC/C-index values ranging from 0.684 to 0.915, while machine learning models exhibited a broader range of 0.500 to 0.943. ML models consistently demonstrated superior discriminative performance compared to traditional statistical approaches. Advanced ML models, such as Random Forest (0.943), CVDformer (0.903) and CVEformer (0.880), achieved the highest performance metrics in test sets. A detailed summary of model types, prediction targets, dataset types, and performance metrics for CSMs and ML models is provided in Tables 5 and 6, respectively.

Table 5.

Characteristics of conventional statistical models.

ID	Model type	Prediction target	Dataset type	AUC/C-index (95% CI)	Other performance metrics
1	LR	CM	Test set	0.732 (0.699–0.765)	Calibration slope: 1.096 (0.575–1.557)
1	LR	CM	Validation set	0.721 (0.680–0.761)	Calibration slope: 1.028 (0.733–1.322)
3	LR	HF	Test set	0.722	Accuracy: 0.748, sensitivity: 0.756, specificity: 0.744
4	Cox regression	3/5/10-year CV event risk	Training set	0.808 (0.773–0.844)	AUROC: 0.836/0.845/0.869
4	Cox regression	3/5/10-year CV event risk	10-fold cross- validation	0.794	AUROC: 0.832/0.832/0.845
4	Cox regression	3/5/10-year CV event risk	Hold-out validation	0.798	AUROC: 0.822/0.832/0.860
5	LR	MACE	Test set	0.630	Accuracy: 0.6272, Recall: 0.5468, Precision: 0.7096 F1: 0.6176
6	LR	Composite endpoint	Training set	0.824 (0.783–0.866)	All-cause mortality AUC: 0.836 (0.794–0.879)
6	LR	Composite endpoint	Internal validation	0.843 (0.794–0.876)	Superior to reference model (C-statistic: 0.767), NRI: 2.51%, IDI: 8.45%
7	LR	MACE	Male cohort	0.760	Superior to FRS model (C-statistic: 0.64)
7	LR	MACE	Female cohort	0.770	Superior to FRS model (C-statistic: 0.60)
7	LR	MACE	Bootstrap	0.750 (0.74–0.76)	Good calibration curve
9	Cox regression	CV events	Simple model	0.737 (0.712–0.761)	NRI: 0.44 (0.34–0.53), IDI: 0.02 (0.02–0.03)
9	Cox regression	CV events	Complex model	0.765 (0.741–0.788)	NRI: 0.44 (0.34–0.53), IDI: 0.02 (0.02–0.03)
10	LR	All-cause Mortality	Test set	0.684 (0.637–0.736)	Accuracy: 0.688, sensitivity: 0.646, PPV:0.721, F1: 0.681, AUPRC: 0.687
10	LR	CM	Test set	0.694 (0.645–0.743)	Accuracy: 0.695, sensitivity: 0.641, PPV:0.747, F1: 0.690, AUPRC: 0.678
11	LR	CVC	Derivation cohort	0.845 (0.803–0.886)	NR
11	LR	CVC	Validation cohort	0.845 (0.781–0.909)	NR
12	LR	First CVD event	Training set	0.703 (0.652–0.753)	NR
12	Cox regression	First CVD event	Validation set	0.688 (0.655–0.720)	NR
14	LR	AIS	Training set	C-index: 0.877, AUC: 0.857	DCA: Positive net benefit within threshold risk range of 2-96%
14	LR	AIS	Validation set	C-index: 0.915, AUC: 0.905	NR

Note: LR: Logistic regression; CM: cardiovascular mortality; HF: heart failure; CVD: cardiovascular diseases; MACE: major adverse cardiovascular events; CVC: cardiac valve calcification; AIS: acute ischemic stroke; NR: not reported.

Table 6.

Characteristics of machine learning models.

ID	Model type	Prediction target	Dataset type	AUC/C-index	Other performance metrics
2	SVCR	Ischemia	Test Set 1	0.500	Accuracy: 0.94, recall: 0.94, precision: 0.89, F1: 0.91, test error with CV:0.94
2	SVCR	Ischemia	Test Set 2	0.740	Accuracy: 0.75, recall: 0.75, precision: 0.77, F1: 0.75, test error with CV:0.73
2	SVCR	Ischemia	Test Set 3	0.670	Accuracy: 0.73, recall: 0.73, precision: 0.72, F1: 0.72, test error with CV:0.74
2	SVCR	Ischemia	Test Set 4	0.730	Accuracy: 0.58, recall: 0.58, precision: 0.94, F1: 0.69, test error with CV:0.95
2	SVCR	Ischemia	Test Set 5	0.500	Accuracy: 0.48, recall: 0.48, precision: 0.23, F1: 0.31, test error with CV:0.73
3	XGBoost	HF	Test set	0.814	Accuracy: 0.785, sensitivity: 0.796, specificity: 0.781
5	CVEformer	MACE	Time series data	0.880	Accuracy: 0.804, recall: 0.762, precision: 0.819, F1: 0.790
5	Hi-BEHRT	MACE	Test set	0.830	Accuracy: 0.811, recall: 0.772, precision: 0.838, F1: 0.803
5	SAFARI	MACE	Test set	0.800	Accuracy: 0.802, recall: 0.756, precision: 0.812, F1: 0.783
5	A-LSTM	MACE	Test set	0.780	Accuracy: 0.784, recall: 0.658, precision: 0.800, F1: 0.721
5	Deepr	MACE	Test set	0.770	Accuracy: 0.759, recall: 0.606, precision: 0.793, F1: 0.687
5	RF	MACE	Time series data	0.720	Accuracy: 0.709, recall: 0.619, precision: 0.733, F1: 0.671
5	SVM	MACE	Time series data	0.680	Accuracy: 0.745, recall: 0.585, precision: 0.661, F1: 0.621
8	CART	24-month CVD Risk	Test set	iAUC: 0.71	Significant improvement compared to model without miRNAs
10	CVDformer	All-cause Mortality	Test set	0.877 (0.813–0.905)	Accuracy: 0.861, sensitivity: 0.835, precision: 0.880, F1: 0.857, AUPRC: 0.934
10	CVDformer	CM	Test set	0.903 (0.840–0.935)	Accuracy: 0.878, sensitivity: 0.825, precision: 0.888, F1: 0.855, AUPRC: 0.907
10	ConCare	All-cause Mortality	Test set	0.825 (0.745–0.864)	Accuracy: 0.809, sensitivity: 0.770, precision: 0.820, F1: 0.794, AUPRC: 0.845
10	ConCare	CM	Test set	0.829 (0.772–0.869)	Accuracy: 0.820, sensitivity: 0.775, precision: 0.813, F1: 0.794, AUPRC: 0.802
10	LSTM	All-cause Mortality	Test set	0.788 (0.713–0.835)	Accuracy: 0.765, sensitivity: 0.743, precision: 0.788, F1: 0.765, AUPRC: 0.807
10	LSTM	CM	Test set	0.814 (0.751–0.877)	Accuracy: 0.791, sensitivity: 0.781, precision: 0.805, F1: 0.793, AUPRC: 0.785
13	XGBoost	HF Hospitalization	Training set	0.844	NR
13	XGBoost	HF Hospitalization	Validation set	0.793	5-fold cross-validation
13	XGBoost	1-year HF Hospitalization	Training set	0.874	NR
13	XGBoost	1-year HF Hospitalization	Validation set	0.691	5-fold cross-validation
13	XGBoost	5-year HF Hospitalization	Test set	0.698	Accuracy: 0.639, sensitivity: 0.655, specificity: 0.738, PPV: 0.692, NPV: 0.578, F1: 0.673
13	XGBoost	1-year Mortality	Test set	0.669	Accuracy: 0.765, sensitivity: 0.857, specificity: 0.582, PPV: 0.056, NPV: 0.925, F1: 0.104
13	RF	HF Hospitalization	Training set	0.931	NR
13	RF	HF Hospitalization	Validation set	0.794	5-fold cross-validation
13	RF	HF Hospitalization	Test set	0.853	Accuracy: 0.746, sensitivity: 0.710, specificity: 0.849, PPV: 0.788, NPV: 0.696, F1: 0.747
13	RF	1-year HF Hospitalization	Training set	0.943	NR
13	RF	1-year HF Hospitalization	Validation set	0.728	5-fold cross-validation
13	RF	1-year HF Hospitalization	Test set	0.729	Accuracy: 0.739, sensitivity: 0.923, specificity: 0.575, PPV: 0.179, NPV: 0.912, F1: 0.299
13	RF	All-cause Mortality	Test set	0.871	Accuracy: 0.821, sensitivity: 0.780, specificity: 0.892, PPV: 0.806, NPV: 0.829, F1: 0.793
13	RF	5-year Mortality	Test set	0.829	Accuracy: 0.743, sensitivity: 0.765, specificity: 0.806, PPV: 0.722, NPV: 0.765, F1: 0.743
13	AdaBoost	HF Hospitalization	Training set	0.821	NR
13	AdaBoost	HF Hospitalization	Validation set	0.794	5-fold cross-validation
13	AdaBoost	1-year HF Hospitalization	Training set	0.840	NR
13	AdaBoost	1-year HF Hospitalization	Validation set	0.704	5-fold cross-validation

Note: HF: heart failure; MACE: major adverse cardiovascular events; CVD: cardiovascular diseases; CM: cardiovascular mortality; NR: not reported.

To quantify overall performance, we calculated summary statistics based on test and validation set AUC/C-index values. Among all 50 test/validation set values, 4.00% (2/50) achieved excellent discrimination (AUC ≥0.9), 40.00% (20/50) demonstrated very good performance (AUC 0.8–0.89), and 38.00% (19/50) showed good performance (AUC 0.7–0.79), with 18.00% (9/50) falling below 0.70. Analysis showed that ML models had a mean AUC of 0.784 ± 0.112 (median: 0.794, IQR: 0.112), compared to 0.772 ± 0.066 (median: 0.765, IQR: 0.102) for CSMs. Non-normal distributions (Shapiro-Wilk: CSM: p = 0.0001; ML: p = 0.0003) prompted the use of the Mann–Whitney U test, which revealed no significant difference in AUC between ML and CSM (U = 320, p = 0.24). However, substantial heterogeneity existed within model categories, necessitating the detailed subgroup analyses to identify factors influencing performance.

3.2.4. Subgroup and sensitivity analysis

To address heterogeneity and explore factors influencing model performance, we conducted comprehensive subgroup analyses using exclusively test and validation set results (n = 50, 18 CSMs, 32 ML models). The results of subgroup analyses are presented in Figure 4.

Figure 4.

Subgroup performance analysis of ML models and CSMs.

Note: Panel A shows significant performance differences by algorithm type (Kruskal-Wallis H = 11.405, p = 0.003), with deep learning models outperforming both CSMs (p = 0.031) and traditional ML models (p = 0.005), while no difference was observed between CSMs and traditional ML models (p = 0.727). Panel B demonstrates no significant association between dataset size and performance (H = 0.023, p = 0.989). Panel C shows no difference between validation methods (U = 246.5, p = 0.683). Box plots display median (center line), interquartile range (box), and red dashed lines indicate group means. *p < 0.05, **p < 0.01 (Bonferroni corrected) and ns means no significance.

Stratification by algorithm type revealed significant performance differences (H = 11.405, p = 0.003). Deep learning models (n = 11), including transformer-based architectures (CVEformer, Hi-BEHRT, SAFARI, CVDformer) and LSTM variants (A-LSTM, Deepr, LSTM, ConCare), achieved mean AUC of 0.827 ± 0.042 (median: 0.825). Traditional machine learning models (n = 21), comprising SVCR, RF, XGBoost, SVM, CART, and AdaBoost, demonstrated mean AUC of 0.725 ± 0.093 (median: 0.728). Conventional statistical models (n = 18) showed mean AUC of 0.760 ± 0.067 (median: 0.755). Post hoc pairwise comparisons with Bonferroni correction revealed significant differences between deep learning and traditional ML models (adjusted p = 0.005) as well as deep learning models and CSMs (adjusted p = 0.031). Notably, no significant difference was observed between traditional ML models and CSMs (adjusted p = 0.727), suggesting that advanced deep learning architectures may offer performance improvements.

Unexpectedly, analysis by dataset size showed no significant differences (H = 0.023, p = 0.989). Models from studies with <1,000 patients (n = 15) achieved mean AUC of 0.763 ± 0.068, those with 1,000–3,000 patients (n = 16) reached 0.745 ± 0.110, and studies >3000 patients (n = 19) attained 0.770 ± 0.069. Given the non-significant overall test, post hoc pairwise comparisons were not conducted as no meaningful differences existed between dataset size categories. This lack of association contradicts theoretical expectations and may reflect the confounding effect of algorithm choice, as all deep learning models originated from large-scale studies.

Validation methodology showed no significant impact on performance. Models from studies employing both internal and external validation (n = 12) demonstrated mean AUC of 0.777 ± 0.072, compared to 0.754 ± 0.088 for internal validation only (n = 38), with no statistical difference (U = 246.5, p = 0.683). This finding raises concerns about the adequacy of current validation practices in the field.

Effect size analysis demonstrated negligible practical differences between ML and CSMs. Cohen’s d of −0.005 indicated virtually no effect, while Cliff’s Delta of 0.054 confirmed minimal practical distinction. These near-zero effect sizes demonstrate that ML models, when considered collectively, provide no meaningful performance advantage over conventional statistical approaches.

Sensitivity analyses confirmed the robustness of our primary findings. After excluding extreme AUC values (<0.6 or >0.9), ML models (n = 29) demonstrated a mean AUC of 0.773 compared to 0.752 for CSMs (n = 17), with no statistically significant difference (U = 286.5, p = 0.369). These consistent results across multiple analytical approaches suggest that deep learning algorithms may demonstrate potential superiority over conventional statistical methodologies.

3.3. Quality assessment

According to the quality assessment conducted using the PROBAST tool, 28.57% (n = 4) of the included studies demonstrated high methodological quality, while the remaining 71.43% (n = 10) exhibited medium quality. Quality evaluation was performed across four critical domains: Participants and Data Sources, Predictors, Outcomes, and Analyses, with performance in each domain classified as High (H), Medium (M), or Low (L). Notably, the majority of studies achieved high-quality ratings specifically in the Predictors and Outcomes domains. Regarding applicability, all studies (100.00%, n = 14) consistently displayed high applicability across most assessment domains, indicating strong external validity. In the assessment of bias risk, 71.43% (n = 10) of the included studies exhibited an overall low risk of bias, with particularly robust performance in the Predictors and Outcomes domains, where all studies uniformly demonstrated low bias risk (Table 7). Details of quality assessment can be found in the Supplemental file S5.

Table 7.

PROBAST assessment.

	Model development assessment				Model validation assessment
Study	Participants and data sources	Predictors	Outcomes	Analyze	Participants and data sources	Predictors	Outcomes	Analyze
	Quality/applicability				Risk of bias/applicability
Yu et al. (2018)	M/M	H/H	H/H	H/H	L/H	L/H	L/H	M/H
Mezzatesta et al. (2019)	M/M	M/H	H/H	M/ M	M/M	L/H	L/H	M/M
Wang, Y. et al. (2022)	M/H	H/H	H/H	M/H	M/H	L/H	L/H	M/H
You, X. et al. (2021)	M/H	H/H	H/H	H/H	M/H	L/H	L/H	L/H
Xu, Z. et al. (2024)	H/H	H/H	H/H	H/H	L/H	L/H	L/H	L/H
Zhang et al. (2022)	M/H	H/H	H/H	H/H	M/H	L/H	L/H	L/H
Matsubara et al. (2017)	M/H	H/H	H/H	H/H	M/H	L/H	L/H	M/H
de Gonzalo-Calvo et al. (2020)	H/H	H/M	H/H	H/M	L/H	L/M	L/H	L/M
Inaguma et al. (2019)	H/H	H/H	H/H	M/H	L/H	L/H	L/H	L/H
Xu, X. et al. (2024)	M/H	H/H	H/H	M/H	M/H	L/H	L/H	M/H
Wang, Y. et al. (2023)	M/H	H/H	H/H	M/H	L/H	L/H	L/H	L/H
You, X. et al. (2022)	M/H	H/H	H/H	M/H	M/H	L/H	L/H	M/H
Xu, L. et al. (2024)	M/H	H/H	H/H	H/H	M/H	L/H	L/H	L/H
Tong et al. (2024)	M/H	H/H	H/H	M/H	M/H	L/H	L/H	M/H

Note: H (high); M (medium); L (low).

3.4. Performance variability and algorithm Tradeoffs

To investigate performance heterogeneity, we analyzed the performance distribution of CSMs and ML models. As shown in Figure 5A, box plots with overlaid data points illustrate the median, interquartile range, and mean AUC values, highlighting comparable median AUCs (CSMs: 0.765, ML: 0.794) with no significant difference (Mann–Whitney U test, p = 0.24). Figure 5B further presents violin plots to depict the distribution density, revealing greater variability in ML models (coefficient of variation: 14.3%) compared to CSMs (8.6%). These findings informed our multi-dimensional comparative analysis of accuracy, interpretability, data requirements, and clinical feasibility (Figure 6).

Figure 5.

Performance distribution of CSMs and ML models.

Note: Panel A shows box plots with overlaid individual data points for CSMs (n = 18) and ML models (n = 32). Panel B displays violin plots showing distribution density. Box plots indicate median (center line), interquartile range (box), mean (red diamond), and 1.5 × IQR whiskers. Despite ML models achieving higher median AUC (0.794 vs. 0.765), Mann–Whitney U test showed no significant difference (p = 0.24). Coefficient of variation: CSMs 8.6%, ML 14.3%. All metrics derived from test/validation datasets only.

Figure 6.

Radar plot for multi-dimensional algorithm tradeoff analysis.

Note: Four axes represent evaluation dimensions scaled 0–10: Accuracy (based on AUC), Interpretability (model transparency), Data Requirements (sample size), and Clinical Feasibility (bias risk score). Blue Quadrilateral: deep learning; Red Quadrilateral: conventional statistical models; Green Quadrilateral: traditional machine learning. Shaded areas show overall performance profiles.

Each dimension was normalized to a 0–10 scale for quantitative comparison. Accuracy scores were derived through linear transformation (Score = AUC × 10) of mean test/validation performance, yielding scores of 8.3/10 for deep learning (AUC: 0.827 ± 0.042), 7.7/10 for CSMs (AUC: 0.772 ± 0.066), and 7.3/10 for traditional ML (AUC: 0.725 ± 0.093).

Data requirements were inversely scored using the formula: Score = 10 × [1 – log(n_min)/log(n_max)], where n represents sample size. CSMs demonstrated optimal data efficiency (8.5/10), operating effectively with 314 patients and 9.33 ± 6.44 predictors. Traditional ML models showed intermediate requirements (5.0/10, –1,000 patients), while deep learning approaches demanded substantial resources (3.0/10, >3,000 patients).

Clinical feasibility integrated PROBAST assessments through weighted scoring: Score = 0.7 × (proportion with low bias) + 0.3 × (applicability). CSM-based studies achieved 8.3/10 with 83.3% (10/12) demonstrating low bias risk. ML implementations scored 6.0/10 for traditional algorithms and 7.0/10 for deep learning, reflecting their respective methodological variability.

Interpretability assessment followed the three-tier taxonomy established by Rudin [38] and Hassija [39]: inherently interpretable (score: 8–10), post hoc interpretable (score: 4–7), and black-box models (score: 1–3). All CSMs provided direct clinical interpretation through regression coefficients, scoring 9.0/10. Traditional ML algorithms, despite theoretical capability for feature importance analysis via SHapley Additive exPlanations (SHAP) [40], demonstrated practical implementation in only 14.3% of studies, yielding 5.0/10. Deep learning architectures provided no interpretability measures across all implementations, resulting in 2.5/10.

The radar plot synthesis demonstrates that deep learning’s modest accuracy gain (7.3% over CSMs) is counterbalanced by substantial deficits in interpretability (72% reduction) and data efficiency (65% higher requirements). This multi-dimensional tradeoff analysis indicates that algorithm selection must prioritize context-specific clinical needs over pure discriminative performance, as no approach achieves comprehensive superiority.

4. Discussion

This systematic review investigated the comparison between CSMs and ML models for predicting cardiovascular events in dialysis patients which focused on performance metrics, algorithm variations, advantages, challenges, and clinical implications.

It is the first study to systematically compare ML and CSMs specifically for cardiovascular event prediction in dialysis patients, a high-risk population with unique clinical needs.

4.1. Principal findings and comparative evidence

The absence of statistically significant performance difference between ML models and CSMs in this systematic review contrasts with the prevailing narrative of ML superiority in healthcare prediction. This finding gains particular significance when contextualized within broader landscape of clinical prediction modeling. Song et al. [41] reported similar performance parity in acute kidney injury prediction, with conventional logistic regression paradoxically outperforming ML approaches despite theoretical disadvantages. Shin et al. [42] further demonstrated that while ML models achieved numerically superior discrimination in heart failure prediction, this apparent advantage was undermined by critical methodological limitations, with 90% of studies exhibiting high risk of bias and lacking adequate external validation.

The negligible overall effect sizes observed between ML and CSMs, coupled with substantial within-category heterogeneity, suggest that implementation quality and data characteristics may contribute more to predictive success than algorithm selection alone. However, the superior performance of deep learning architectures indicates that meaningful improvements remain achievable with sufficient resources. Therefore, the critical question is not whether ML can outperform CSMs but whether the marginal gains justify the substantially increased implementation complexity.

4.2. Stratified performance patterns and multi-dimensional tradeoffs

The stratified analysis revealed performance heterogeneity that challenges simplistic comparisons between modeling approaches. Algorithm type significantly influenced performance (H = 11.405, p = 0.003), with deep learning models achieving meaningful improvements over both CSMs and traditional ML models. However, traditional ML algorithms showed no advantage over CSMs despite intermediate complexity (adjusted p = 0.727), suggesting a threshold effect where incremental sophistication provides no benefit without fundamental architectural advances.

The multi-dimensional tradeoff analysis exposed a fundamental paradox: technological sophistication inversely correlates with clinical implementability. Deep learning’s 7.3% accuracy advantage was counterbalanced by 72% reduced interpretability and 65% increased data requirements. This inverse relationship has profound implications for global health equity, as the minimum threshold of 3,000 patients for deep learning deployment excludes large parts of the dialysis facilities worldwide. The radar plot synthesis demonstrates that no single approach achieves Pareto optimality across evaluation dimensions, necessitating context-specific model selection based on available resources and clinical priorities.

Unexpectedly, dataset size showed no association with performance (H = 0.023, p = 0.989), contradicting the established machine learning theory. The absence of sample size effects (p = 0.989) warrants careful interpretation. Despite theoretical expectations of performance improvements with larger datasets, the analysis revealed nearly identical performance across sample size categories. This paradox likely reflects confounding factors: deep learning concentration in large-scale studies, data quality potentially outweighing quantity, and possible ceiling effects in cardiovascular prediction for dialysis patients.

4.3. Clinical implementation implications and enduring value of conventional approaches

The negligible overall effect sizes (Cohen’s d=-0.005, Cliff’s Delta = 0.054) between ML models and CSMs have critical implications for resource allocation. Watson et al. [43] identified five key implementation barriers, with financing representing the predominant concern—a challenge amplified by our finding that CSMs achieve comparable performance with minimal infrastructure. The complete absence of interpretability mechanisms in deep learning implementations compounds these barriers, as nephrology practice requires transparent reasoning for critical decisions.

Conventional statistical models demonstrated unexpected resilience in our analysis, maintaining competitive performance while offering substantial practical advantages. The higher methodological rigor observed in CSM studies (83.3% low bias risk versus 60-70% for ML) reflects decades of refinement in statistical frameworks and validation protocols. CSMs’ ability to function effectively with approximately 300 patients and fewer than 10 predictors by average which makes them uniquely suitable in high-burden and resource-limited regions. Furthermore, their inherent interpretability (scoring 9.0/10) facilitates clinical acceptance and regulatory compliance in ways that opaque ML models cannot achieve.

Based on our comprehensive analysis, we propose a stratified implementation framework for optimal model selection in diverse clinical contexts: CSMs with strong interpretability should be prioritized in the following scenarios. First, resource-constrained settings include community dialysis units with fewer than 1,000 patients, and facilities with limited IT infrastructure and medical institutes where computational resources and technical expertise are scarce. Secondly, routine clinical decision-making in dialysis care where physicians need to understand and trust the prediction logic, as most clinicians lack specialized training in machine learning and require transparent models to integrate predictions into their clinical judgment. The inherent interpretability of CSMs allows physicians to verify that model predictions align with established medical knowledge and pathophysiological principles, ensuring safer and more accountable clinical decisions [44]. Thirdly, regulatory compliance scenarios where healthcare systems must justify predictions to external stakeholders, including quality metrics reporting to government agencies and clinical audit processes that mandate algorithmic transparency and accountability [45]. These contexts require accurate predictions as well as clear documentation of decision rationale.

Deep learning models should be considered in specific high-resource environments where their advanced pattern recognition capabilities provide distinct advantages. First, large-scale integrated delivery networks benefit from deep learning’s capacity to detect subtle phenotypic clusters and nonlinear interactions among clinical variables, particularly in identifying subpopulations with differential cardiovascular risk trajectories that traditional regression-based approaches fail to capture [46]. Secondly, academic research settings within medical centers leverage deep learning for exploratory data analysis, including unsupervised feature learning for novel biomarker identification, temporal pattern mining in longitudinal cohorts, and discovery of latent cardiovascular risk phenotypes from high-dimensional omics data. Thirdly, multi-modal data integration scenarios in advanced health systems with comprehensive electronic health records combine structured clinical data, continuous monitoring streams, laboratory time series, and imaging results, where deep learning’s ability to simultaneously process diverse data types offers unique value. Nevertheless, successful implementation requires establishing robust governance frameworks to address the ‘black box’ nature of these models, including the development of complementary interpretability tools such as SHAP values or attention visualization and continuous performance monitoring systems to ensure clinical safety and effectiveness.

4.4. Challenges and limitations

This systematic review has few limitations that may affect the generalizability of the findings. The geographical concentration (71.40% from China) indicates a significant geographical bias that may limit the results’ extrapolation to other regions with distinct dialysis patient populations and healthcare practices. The predominance of internal validation (78.57%) likely overestimates model performance, potentially masking true generalizability limitations when applied to diverse populations. This validation gap reflects a broader issue in contemporary predictive analytics research, which has disproportionately concentrated on model development rather than comprehensive investigation of clinical integration and longitudinal impact assessment [47].

Inconsistent reporting standards across studies, particularly regarding missing data handling, sample size justification, and methods to address class imbalance, introduced potential bias and limited comparability of results. Several studies failed to report how missing predictors were managed during the model development, while others did not justify their sample size calculations or address the inherent class imbalance in cardiovascular event prediction. These methodological inconsistencies may partially explain the substantial performance heterogeneity observed within model categories.

The heterogeneity in study designs, sample sizes, dialysis modalities, outcome definitions, and follow-up periods precluded formal meta-analysis, limiting the ability to provide precise pooled estimates with confidence intervals. This heterogeneity reflects the diverse nature of dialysis populations and healthcare systems but constrains the strength of our conclusions. Additionally, the absence of calibration assessment in most studies represents a critical gap, as discrimination alone provides incomplete information about clinical utility.

4.5. Future research directions

Future research should prioritize several strategic initiatives to advance clinical implementation. Rigorous multi-center external validation frameworks should be established across geographically diverse institutions to assess true model generalizability [48]. Such validation can span different healthcare systems, dialysis modalities, and ethnic populations to ensure global applicability.

To address the geographical bias while preserving data privacy, federated learning approaches offer transformative potential, enabling collaborative model development across continents without sharing patient data [49]. This technology could democratize access to large-scale predictive modeling while maintaining local data governance, potentially bridging the gap between resource-rich and resource-limited settings.

The complementary strengths revealed by our multi-dimensional analysis suggest that hybrid approaches combining the interpretability of traditional statistical methods with the pattern recognition capabilities of ML warrant systematic investigation. Recent developments in interpretable machine learning indicate that future algorithms may transcend current tradeoffs between accuracy and interpretability [50]. Such hybrid models could leverage ensemble methods or hierarchical structures to maintain clinical transparency while capturing complex nonlinear relationships.

Additionally, standardized reporting frameworks should mandate comprehensive documentation of missing data handling, sample size justification, class imbalance methods, calibration metrics, and interpretability assessments. Prospective implementation studies comparing patient-centered outcomes between centers using different prediction approaches are essential, as discrimination metrics alone cannot determine real-world clinical impact. The field must shift focus from marginal accuracy improvements to developing robust, equitable solutions that can meaningfully improve cardiovascular risk across diverse dialysis populations globally.

5. Conclusion

This systematic review demonstrates that machine learning models do not exhibit overall superiority over conventional statistical models for predicting cardiovascular events in dialysis patients. While deep learning architectures show promising discriminative capabilities, traditional machine learning algorithms offer no meaningful advantage despite increased complexity. Conventional statistical models remain competitive, providing robust performance with distinct practical advantages including inherent interpretability, minimal computational requirements, and established regulatory frameworks that facilitate clinical implementation.

However, few limitations currently hinder the clinical translation of all predictive approaches. The geographical concentration of studies, predominance of internal validation, and limited model interpretability restrict the generalizability and real-world applicability across diverse healthcare settings. The heterogeneity in study designs and outcome definitions further complicates direct comparisons and evidence synthesis, highlighting critical gaps in current research methodologies.

Future priorities should accomplish transition from algorithm-centric development to implementation-focused research that prioritizes clinical utility, equity, and real-world impact. Rather than pursuing incremental accuracy gains, future efforts should address fundamental barriers to clinical translation while ensuring that advances benefit diverse global dialysis populations, not just well-resourced centers.

Funding Statement

This study was supported by the Three-Year Action Plan for Strengthening the Construction of Public Health System in Shanghai (2023–2025) of Key discipline construction project (Grant No. GWVI-11.1-49), Three-Year Action Program of the Shanghai Municipality for Strengthening the Construction of the Public Health System (Grant No. GWVI-6), Research on Key Technology of Typical Chronic Disease Management Application Based on Machine Learning Methods under the Perspective of Data Elements – Taking Chongming District as an Example (Grant No. CKY2024-65), Project of the National Health Commission Institute of Hospital Management (Grant No. SZ2024HL023), and Shanghai Science and Technology Innovation Action Plan Project (Grant No. 23640770100).

Disclosures statement

No potential conflict of interest was reported by the author(s).