Machine learning integration in microRNA-based markers for cardiovascular diseases: A systematic review

Table 1 Characteristics of the selected studies that met the inclusion criteria

Ref.	Setting	CVD	Sample size	Identified miRNAs	Role	Main outcome
Kayvanpour et al[8], 2021	Germany	ACS	66 ACS patients and 68 healthy controls; 148 suspected ACS patients initially enrolled	Top 10 miRNAs selected via ANOVA F-value: MiR-142-5p, miR-151a-3p, miR-145-5p, miR-186-5p, miR-191-5p, miR-29c-5p, miR-30d-5p, miR-342-5p, miR-362-5p, and miR-589-5p	Diagnosis of ACS	Machine learning models, including neural networks, classified ACS with high diagnostic performance
Ren et al[13], 2024	United States	AMI/STEMI	24 screening samples; n = 6 each for no known CAD, known CAD, STEMI-pre, and STEMI-PCI; validation samples also used	Already identified: MiR-499, miR-1, miR-208b. Newly identified: MiR-331-3p, miR-142-5p, miR-200b-3p, miR-132-3p, miR-3605-5p, miR-18a-5p, miR-423-5p, miR-543, miR-301a-3p	Diagnosis and differentiation of AMI/STEMI from stable CAD	SCAD/LASSO regularized LR identified a 9-miRNA profile that differentiated no known CAD, known CAD, STEMI-pre, and STEMI-PCI, with ROC curves approaching 1 in selected comparisons; explored for rapid point-of-care diagnosis using MIX.miR ion-exchange membrane technology
Samadishadlou et al[14], 2023	Iran	AMI and stable CAD	Healthy (51), CAD (46), AMI (111)	Differentially expressed: Hsa-miR-21-3p, hsa-miR-32-3p, hsa-miR-186-5p. Additionally selected via AUC-ROC: Hsa-miR-197-5p, hsa-miR-29a-5p, hsa-miR-296-5p	Diagnosis of AMI; differentiating AMI from healthy samples and from CAD	Peripheral blood mononuclear cell-derived miRNA signatures were used to differentiate healthy controls, stable CAD, and MI samples
Samadishadlou et al[15], 2024	Iran	AMI	Training set: 62 MI and 94 healthy controls; independent test set: 8 MI and 6 healthy controls	Hsa-miR-375-3p, hsa-miR-601, hsa-miR-34a-5p, hsa-miR-29c-5p, hsa-miR-330-5p, hsa-miR-199b-5p, hsa-miR-142-3p, hsa-miR-200a-3p, hsa-miR-132-5p, hsa-miR-133a-3p	Diagnosis of early-stage AMI	ML model identified 10 miRNAs with accuracy of 0.86 and AUC of 0.83 for diagnosing AMI
Reel et al[16], 2025	United Kingdom	Essential HTN subtypes	Cushing’s syndrome (35), primary aldosteronism (109), paraganglioma/pheochromocytoma (75), primary HTN (111)	Hsa-miR-15a-5p, hsa-miR-32-5p, hsa-miR-485-3p, hsa-miR-495-3p, hsa-miR-1260a, hsa-miR-186-5p, hsa-miR-195-5p, hsa-miR-326, hsa-miR-139-5p, hsa-miR-133a-3p, hsa-miR-223-3p	Differentiation of endocrine HTN subtypes from primary HTN	Models trained with the miRNAs achieved balanced accuracy of 0.71-0.89 and AUCs of 0.8-0.9 in differentiating HTN subtypes and other conditions
Sajid et al[17], 2024	Pakistan	CAD	CAD cases (58), controls without CAD/stenosis < 50% (55)	MiR-21, miR-33a, miR-133a, miR-145, miR-146a	Diagnosis of CAD	ML models using miRNA biomarkers showed good diagnostic performance for angiography-defined CAD
Yerukala Sathipati et al[18], 2025	United States	Post-operative AF after CABG	Cases (7), controls (8)	Hsa-miR-19a-3p, hsa-miR-19b-3p, hsa-miR-184, hsa-let-7a-5p, hsa-miR-124-3p, hsa-miR-200a-3p, hsa-miR-423-5p, hsa-miR-96-5p, hsa-miR-100-5p, hsa-miR-17-5p	Prediction of post-operative AF after CABG	10 pre-operative circulating miRNA signatures were used to develop ML models for predicting POAF after CABG
Jusic et al[19], 2023	Luxembourg/Bosnia and Herzegovina	HTN	89 cases, 85 controls	MiR-361-3p and miR-501-5p	Diagnosis of HTN	SVM model using the two miRNAs plus clinical characteristics achieved accuracy of 0.87, specificity of 0.91, sensitivity of 0.83, and AUC of 0.90
Errington et al[20], 2021	United Kingdom	PAH	64 cases, 43 disease and healthy controls	MiR-636 and miR-187-5p	Diagnosis of PAH	Models using the two miRNAs showed high diagnostic accuracy in differentiating PAH patients from healthy controls

CVD: Cardiovascular disease, CAD: Coronary artery disease, HTN: Hypertension; AMI: Acute myocardial infarction; ACS: Acute coronary syndrome, AF: Atrial fibrillation, PAH: Pulmonary arterial hypertension, CABG: Coronary artery bypass grafting; miRNA: MicroRNA; STEMI: ST elevation myocardial infarction; PCI: Percutaneous coronary intervention; SCAD: Smoothly clipped absolute deviation; LASSO: Least absolute shrinkage and selection operator; LR: Logistic regression; ROC: Receiver operating characteristic; MI: Myocardial infarction; AUC: Area under the curve; POAF: Post-operative atrial fibrillation; SVM: Support vector machines.

Table 2 Characteristics of the model used in the studies

Ref.	Models evaluated	Internal validation strategy	External validation	Training dataset	Test/validation dataset	Performance metrics
Kayvanpour et al[8], 2021	LR, kNN, LDA, NB, RF, CT, SVM, XGB, and ANN	The subjects were divided into training and test sets in the ratio of 9:1, respectively. This was repeated 10 times to enable ten-fold cross-validation	None	90% of subjects; 121 samples per split	10% of subjects; 13 samples per split	Accuracy, sensitivity, specificity, and ROC-AUC
Ren et al[13], 2024	Regularized LR using either SCAD or LASSO	Leave-one-out cross-validation	None for the ML model; selected miRNAs were biologically evaluated in matched clinical samples using an ion-exchange membrane sensor platform	24 subjects; 800-miRNA screening library (100%)	None; leave-one-out cross-validation was used because of small sample size	ROC curves and AUC (used to evaluate the selected miRNA combinations)
Samadishadlou et al[14], 2023	A 2-layer architecture utilizing SVM (with linear, polynomial, and RBF kernels), LR, RF, kNN, GB, XGB, and DT models (layer 1 isolated healthy vs not-healthy; layer 2 separated MI vs CAD)	The data was split in a 7:3 ratio into the training and test sets, respectively. A ten-fold cross-validation followed this	None	70% of all the samples	30% of all the samples	AUC-ROC, accuracy, sensitivity, specificity, and confusion matrix
Samadishadlou et al[15], 2024	SVM, GB, XGB and hard voting ensemble model	Done in 2 phases. In miRNA selection: The LASSO method was cross-validated using the dataset 10-fold to select the best miRNA to be used in model development. In model selection, the training dataset was split in a 7:3 ratio into training and validation datasets. The models were then cross-validated 5-fold on the datasets. The best-performing models were then tested on the independent dataset	Performed using an independent dataset (GSE29532)	GSE61741 (62 MI samples and 94 healthy samples)	GSE29532 (8 MI samples and 6 healthy samples)	Accuracy, AUC-ROC, sensitivity, and specificity
Reel et al[16], 2025	J48, NB, IBk, RF, LB, LMT, SL, and SMO	The data was randomly split into training and testing sets in an 8:2 ratio for model development and validation	None	80% of all the samples	20% of all the samples	Balanced accuracy is the primary metric. Other metrics include sensitivity, specificity, AUC-ROC, F1 score and Kappa score
Sajid et al[17], 2024	LR, SVM, nonlinear kNN, tree-based (DT, RF), GB, XGBM, CBoost, ABoost, and ensemble voting	The data subset was first split in an 8:2 ratio into a CV subset and a hold-out subset for final evaluation. The CV subset was then divided into 10 folds. Nine folds were used for training and one-fold for testing, and the process was repeated 10 times. The best models were then tested on the hold-out dataset	None	80% of the 113 subjects (cohort: 58 CAD cases, 55 healthy controls)	20% of the 113 subjects (hold-out subset)	Accuracy, sensitivity, specificity, AUC-ROC, performance evaluation measure, F-statistic, and P values
Yerukala Sathipati et al[18], 2025	kNN, XGB, SVM, and RF	The data was split in a 8:2 ratio into training and validation datasets, respectively	Performed using an independent GEO dataset (GSE222739)	80% of the dataset (n = 12)	20% of the dataset (n = 3)	AUC-ROC, accuracy, specificity, and sensitivity
Jusic et al[19], 2023	RF, SVM, MLP, XGB, kNN, Logit	Hyperparameter tuning used two repeated 10-fold CV. The final model was also evaluated using leave-one-out cross-validation	None	147 subjects (89 from the validation cohort + 58 from the sequenced discovery cohort)	23 subjects (randomly extracted as 20% of the 112-subject validation cohort)	AUC-ROC, balanced accuracy, F1 score, precision, sensitivity, and specificity
Errington et al[20], 2021	RF, Rpart, LASSO, XGB, and Ensemble	The data was split into training and validation data sets. The models were then CV 10-fold in the training dataset	The models were externally validated using publicly available datasets	Two-thirds of the samples	One-third of the samples (validation set)	Sensitivity, specificity, AUC, correct classification rate (accuracy), positive predictive value, and negative predictive value

ML: Machine learning; CV: Cross validation; SVM: Support vector machines; CT: Classification tree; NB: Naïve bayes; LR: Regularized logistic regression; SCAD: Smoothly clipped absolute deviation; LASSO: Least absolute shrinkage and selection operator; RF: Random forests; XGB: XGBoost; GB: Gradient boosting; kNN: K-nearest neighbors; DT: Decision tree models; LB: Logitboost; SL: Simple logistic; LMT: Logistic model tree; SMO: Sequential minimal optimization; CBoost: CatBoost; ABoost: AdaBoost; ANN: Artificial neural networks; Rpart: Regression partition tree; LDA: Linear discriminant analysis; miRNA: MicroRNA; ROC: Receiver operating characteristic; AUC: Area under the curve; RBF: Radial basis function; CAD: Coronary artery disease; MI: Myocardial infarction; XGBM: Extreme gradient boosting machine; MLP: Multilayer perceptron.

Table 3 Key diagnostic performance metrics of machine learning models integrating microRNAs for cardiovascular disease diagnosis

Ref.	Best model(s)	AUC-ROC (range or best)	Accuracy (best reported)	Sensitivity (best)	Specificity (best)	Notes on interpretation
Kayvanpour et al[8], 2021	ANN was the best-performing model (SVM, kNN, LDA, and RF also performed highly)	0.87-0.99	0.87-0.96	0.87-0.95	0.87-1.00	Good internal discriminative performance, but no external validation; risk of optimistic bias
Ren et al[13], 2024	Regularized LR (LASSO/SCAD)	0.5 to approximately 1.0	NR	NR	NR	Focus on miRNA identification; no full diagnostic model metrics
Samadishadlou et al[14], 2023	Two-layer architecture utilizing SVM (RBF)	0.96 (layer 2) to 1.0 (layer 1)	0.96 (overall two-layer architecture)	0.97 (layer 2) to 1.0 (layer 1)	0.86 (layer 2) to 1.0 (layer 1)	Good internal performance (two-layer approach isolated healthy samples perfectly), but no external validation cohort utilized
Samadishadlou et al[15], 2024	HVE (aggregating SVM, GB, and XGB)	0.83 (HVE on test set)	0.86	1.00	0.67	Very small test set (14 samples total: 8 MI, 6 healthy) limits reliability; platform differences between training and test sets impacted individual model performance
Reel et al[16], 2025	LMT/LogitBoost (along with SL and SMO)	0.80-0.90	0.71-0.89 (balanced accuracy)	0.43-0.95	0.83-1.00	Moderate-large sample; balanced accuracy used
Sajid et al[17], 2024	AdaBoost (for miRNA biomarkers) and GB (for atherosclerosis inflammatory biomarkers)	0.88-0.95 (CV)/0.76-0.93 (hold-out)	0.87-0.90 (CV)/0.78-0.96 (hold-out)	0.88-0.92 (CV)/0.71-0.86 (hold-out)	0.96-1.00 (CV)/0.81-1.00 (hold-out)	Moderate sample; strong internal metrics but no external validation
Yerukala Sathipati et al[18], 2025	RF/XGB	0.76-0.83	0.73-0.80	0.75-0.87	0.71	Very small sample, high risk of overfitting, though external validation was performed
Jusic et al[19], 2023	SVM	0.90	0.87	0.83	0.91	Moderate sample; internal only
Errington et al[20], 2021	RF, XGB and Ensemble model	0.82-0.85	0.81-0.83	0.86-0.91	0.64-0.71	Study with external validation, more reliable estimates

HVE: Hard voting ensemble; CV: Cross validation; SVM: Support vector machines; ROC: Receiver operating characteristic; AUC: Area under the curve; ANN: Artificial neural networks; kNN: K-nearest neighbors; RF: Random forests; LR: Regularized logistic regression; LASSO: Least absolute shrinkage and selection operator; SCAD: Smoothly clipped absolute deviation; miRNA: MicroRNA; RBF: Radial basis function; XGB: XGBoost; GB: Gradient boosting; LMT: Logistic model tree; SL: Simple logistic; SMO: Sequential minimal optimization; MI: Myocardial infarction; LDA: Linear discriminant analysis; NR: Not reported.

Table 4 Risk-of-bias assessment using QUADAS-2 domains

Ref.	Patient selection (risk of bias/applicability)	Index test (including feature selection and overfitting risk)	Reference standard	Flow and timing (including external validation)	Overall ML-specific concerns
Kayvanpour et al[8], 2021	Low/Low	Low (10-fold CV was reported, but feature-selection timing was not fully clear)	Low (diagnosis strictly adjudicated by a board of 3 expert cardiologists using ESC guidelines)	Low (internal CV only; no external)	Moderate overfitting risk due to no external validation
Ren et al[13], 2024	Unclear/Low	Low (LASSO/SCAD used for selection; validation on matched sample)	Unclear	High (no formal validation split reported)	Limited validation; ML mainly for feature identification
Samadishadlou et al[14], 2023	Low/Low	Low (7:3 split and 10-fold CV)	Unclear	Low (internal only)	Small test set; class imbalance addressed using sample weighting
Samadishadlou et al[15], 2024	Low/Low	Low (LASSO and 5/10-fold CV)	Unclear	Low (internal and independent test set)	Improved validation compared to prior work
Reel et al[16], 2025	Low/Low	Low (8:2 split and balanced accuracy metric)	Unclear	Low (internal only)	Use of balanced accuracy helps with potential imbalance
Sajid et al[17], 2024	Low/Low	Low (8:2 and 10-fold CV and hold-out)	Unclear	Low (internal only)	Ensemble methods; good internal practices
Yerukala Sathipati et al[18], 2025	Unclear/Low (small n = 15)	Low (8:2 split)	Unclear	Low (external validation performed via independent GEO dataset)	Very small sample; high risk of overfitting
Jusic et al[19], 2023	Low/Low	Low (8:2 and 10-fold CV)	Unclear	Low (internal only)	Small cohort
Errington et al[20], 2021	Low/Low	Low (10-fold CV and external datasets)	Unclear	Low (external validation performed)	Strongest validation approach among included studies

CV: Cross validation; ML: Machine learning; ESC: European society of cardiology; LASSO: Least absolute shrinkage and selection operator; SCAD: Smoothly clipped absolute deviation; GEO: Gene Expression Omnibus.