This study followed the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines and was prospectively registered in PROSPERO (registration ID: CRD42024617183) [23].

Detailed inclusion and exclusion criteria were formulated for this review. To improve the visualization effect, these standards were presented in tabular form (Textbox 1).

PubMed, Embase, Cochrane Library, and Web of Science were systematically searched up to November 12, 2025, using both Medical Subject Headings (MeSH) terms and free-text words. Search terms included combinations of “artificial intelligence,” “deep learning,” “machine learning,” “random forest,” “support vector machine,” and “left ventricular hypertrophy.” No filters were applied to constrain studies by publication year or geographic region. The details on the search strategy are available in Multimedia Appendix 1.

All identified records were uploaded to EndNote, and duplicates were filtered out. The titles and abstracts of the remaining studies were screened for potentially relevant studies. Full texts of the selected studies were obtained and evaluated for eligibility. An electronic extraction sheet was prepared in advance to record author information, publication year, country of the author, study design, patient sources, generation method of validation set, model types, and modeling variables. The entire screening process and data extraction were performed independently by 2 investigators, followed by cross-verification. Discrepancies were addressed through discussion with a third researcher.

The Prediction Model Risk of Bias Assessment Tool (PROBAST) was leveraged to appraise the risk of bias (RoB) in the included original studies [24]. The evaluation covers 4 domains: participants, predictor variables, outcomes, and statistical analysis, which collectively determine the overall RoB and applicability. Each domain contains specific signaling questions (2, 3, 6, and 9 questions, respectively), with 3 possible responses: “yes/probably yes” (low RoB), “no or probably no” (high RoB), and “no information” (unclear). Two researchers independently conducted the PROBAST assessments, followed by cross-checking. Any dissents were addressed through discussion with a third researcher.

Some studies had multiple validation sets. When both an internal and an external validation set were available, only the results from the external set were retained. If an external validation set is unavailable, the results from the internal validation set with optimal performance were selected.

A meta-analysis was conducted on the validation set, leveraging diagnostic 2×2 tables. For studies without 2×2 tables, sensitivity, specificity, positive predictive value, and accuracy, combined with case numbers, were used for estimation (formulas 1‐5). The bivariate mixed-effects model was used to pool diagnostic metrics (sensitivity, specificity, positive likelihood ratio [PLR], negative likelihood ratio [NLR], diagnostic odds ratio [DOR], and summary receiver operating characteristic [SROC]). Deek funnel plot was used for assessing publication bias [25], and Fagan nomogram was adopted for evaluating clinical utility [26]. Subgroup analyses were carried out based on modeling variables and model types. A significance threshold of P<.05 was established.

TP=Sensitivity×Events (1)

TN=Specificity×(sample size − Events) (2)

FN=Events - TP (3)

FP=samplesize - Events - TN (4)

OR

FP=TP/precision - TP (5)

A total of 1382 records were searched. After removing duplicates, 1086 records remained. Then, 1019 records were excluded through title and abstract screening, leaving 67 records for full-text review. Of those, 42 reports were removed due to incomplete data, incompatible outcomes, and incompatible interventions. Ultimately, 25 studies were selected for analysis (Figure 1).

Of the 25 case-control studies included (2018‐2025) [1,10,21,22,27-47], 19 used single-center designs, and 6 adopted multicenter approaches (including 4 studies based on databases that leveraged the UK Biobank resource). These studies were conducted across diverse regions: 10 studies in China, 6 in the United States, 2 in Japan, 2 in Korea, 1 in Mexico, 1 in the United Kingdom, 1 in Greece, 1 in Italy, and 1 in the United Arab Emirates. ECG-based models were included in 16 studies, clinical feature-based models were included in 5 studies, and ECHO-based models were included in 4 studies. Six studies performed external validation, including 5 ECG-based models and 1 clinical feature-based model. The main features of these studies are summarized in Table 1.

A high RoB was identified in 21 case-control studies (single- and multi-center) due to potential selection bias. An unclear RoB was assigned to 4 studies based on public databases. The 21 case-control studies were rated as having a high RoB due to selection bias. Due to inherent selection biases, such as the “healthy volunteer” effect commonly found in biobanks, the 4 studies based on public databases were rated as having an unclear RoB. Five case-control studies utilized clinical features for modeling and were rated as having a high RoB due to case-control influence during data collection. The remaining studies, which used image-based ML models, were assessed as having a low RoB in terms of predictors. All studies used valid outcome definitions, and the avoidance of predictor variables in model construction ensured a low RoB. For the imaging-based validation set, the accurate calculation of events per variable (EPV) proved challenging, leading to an uncertain RoB. Developers typically rely on EPV, especially EPV of 10 higher, to determine the minimum sample size required and the maximum number of candidate predictions that can be tested [48]. However, when building models based on images, the basic unit of validation data is the image, not the traditional “patient case.” Therefore, we cannot calculate the sample size that meets the EPV requirements based on the number of patient cases, making it difficult to directly apply this standard in such studies. In contrast, studies using clinical features for modeling consistently met the EPV>10 criterion and thus were classified as low RoB. In 2 studies, data were segmented according to images, and patients’ images were shared between the training set and the validation set, which may lead to data leakage, and thus were classified as high RoB (Figure 2).

This study examined the efficacy of 3 types of ML models for LVH detection. In the validation cohort, ECG-based models had a sensitivity of 0.76 (95% CI 0.66‐0.84) and a specificity of 0.84 (95% CI 0.78‐0.89). The clinical feature-based models had a sensitivity of 0.78 (95% CI 0.69‐0.85) and a specificity of 0.71 (95% CI 0.65‐0.76). The ECHO-based models had a sensitivity ranging from 0.71 to 0.94 and a specificity ranging from 0.67 to 0.96. These findings confirm the significant efficacy of ML in LVH detection.

ML is increasingly accurate in LVH detection, which enables the identification and treatment of earlier subclinical cardiovascular diseases [49]. In clinical practice, ECG is now widely used as a screening method for LVH due to its affordability and high accessibility, and is recommended as part of standard clinical evaluation for patients with hypertension in the Chinese Hypertension Clinical Practice Guidelines [50]. ECG-based LVH detection has been a major research priority in hypertensive cardiac target organ damage studies [1]. The clinical diagnosis of LVH via 12-lead ECG relies principally on the quantitative assessment of QRS complex (Q-wave, R-wave, S-wave) voltage parameters [51,52]. Despite recent innovations in diagnostic criteria, the overall sensitivity remains suboptimal, particularly in Chinese and other East Asian populations [1,53,54]. The low sensitivity of ECG in LVH detection limits its clinical application [32]. Consistent with existing research, our results showed that ECG-based models produced a sensitivity of 0.76 (95% CI 0.66‐0.84) and a specificity of 0.84 (95% CI 0.78‐0.89). Although limited by its sensitivity, studies have confirmed that ECG-detected LVH is associated with adverse cardiovascular outcomes [32,55]. Given its widespread availability and low cost, ECG remains an ideal initial assessment tool for LVH screening in patients with hypertension [56]. Therefore, some researchers recommend using ECG primarily for the risk stratification and prognosis prediction of LVH [43,57].

It has also been noted that different ML methods show significant differences in their ability to identify positive or outcome events. With its powerful automatic feature extraction capabilities, DL models can directly identify complex and subtle patterns from imaging data, integrate multiple imaging variables, and improve diagnostic accuracy [58,59]. Its automated segmentation reduces the subjective bias of traditional manual segmentation in radiomics and can integrate segmentation with diagnosis, offering significant advantages [60,61]. Therefore, it is imperative to prioritize the development of automated methods based on DL. Decision tree models are highly interpretable models that generate rules by recursively splitting variables in a dataset [62]. Decision tree models add a penalty term proportional to the absolute value of the feature coefficient for feature selection and model building [63]. Multiple studies have confirmed that RF models typically exhibit higher predictive performance than other ML methods because they can model nonlinear relationships and are insensitive to overfitting [64]. RF models demonstrate significant advantages in handling imbalanced datasets [65]. RF models excel at handling nonlinear and missing data, are conducive to high-dimensional data analysis, and demonstrate strong predictive performance [66]. Support vector machine (SVM) models have high predictive accuracy [67]. For structured and well-defined data, SVM models can skillfully capture the association between biomarkers and diseases through mathematical and statistical principles, thereby achieving high diagnostic accuracy [68,69]. SVM models are known for their ability to find the optimal decision boundary in high-dimensional feature spaces, resulting in effective classification based on extracted features [70,71].

ML models that incorporated clinical features demonstrated a sensitivity of 0.78 (95% CI 0.69‐0.85) and a specificity of 0.71 (95% CI 0.65‐0.76) in the validation set. Unlike traditional ECG criteria, the clinical features used to construct these models encompassed demographic characteristics (eg, sex, age) and clinical factors (eg, systolic blood pressure, duration of hypertension, and abdominal obesity). The possible reasons for using these factors are as follows.

Sex constitutes a significant determinant of LVH [1,37]. Substantial evidence has established the female sex as an independent predictor. In previous predictive models, women scored 70, while men scored 0, possibly due to the older study population, where most women had reached menopause [1]. Women exhibit higher salt sensitivity than men due to female hormonal influences, which amplifies adverse renin-angiotensin-aldosterone system effects [72,73]. Regardless of race or menopausal status, women with salt-sensitive hypertension face a higher LVH risk because this sex-specific mechanism promotes aldosterone-driven endothelial dysfunction [72,73]. This association is further exacerbated by diminished cardiovascular protection due to decreased estrogen levels in postmenopausal women [37].

Age plays a pivotal role in forecasting LVH in young patients with salt-sensitive hypertension. The evidence reveals a dual pattern: (1) patients 40 years or older demonstrate a significantly higher risk than younger populations, consistent with age-related left ventricular mass increase [1,37,74], while (2) patients with early-onset hypertension exhibit more severe cardiac damage and elevated cardiovascular mortality [74,75]. Such observations suggest that the interaction between age and the timing of hypertension requires special attention.

The development of LVH is closely associated with systolic blood pressure level and duration of hypertension. Prospective studies demonstrate that each 19 mm Hg increase in systolic blood pressure elevates the LVH risk by 49% [6], while effective systolic blood pressure control markedly reduces the disease risk in old patients [76]. These findings provide crucial evidence for clinical treatment decisions. Hypertension duration, as a cumulative exposure metric, has been consistently validated in multiple studies to be significantly associated with LVH. Shortening the course of hypertension can effectively decrease the risk of myocardial remodeling, which shows the importance of early intervention in delaying LVH progression [77].

Obesity constitutes an independent risk factor for LVH [78,79]. Waist circumference, the principal diagnostic metric for abdominal obesity, is linked with both cardiovascular and metabolic disorders [80]. The pathophysiological linkage between abdominal obesity and LVH can be attributed to the robust correlation between waist circumference and excessive visceral adipose tissue accumulation [80,81]. As a metabolically active depot, visceral adipose tissue exhibits substantially higher energy demands compared to subcutaneous fat [81]. This augmented metabolic requirement exacerbates cardiac workload in patients with hypertension, ultimately resulting in LVH [78,81]. Studies have identified waist circumference as a superior predictor of cardiovascular risk and mortality compared to BMI [80,82]. Notably, elevated systolic blood pressure and prolonged hypertension duration may synergistically accelerate concentric cardiac remodeling, while abdominal obesity plus aging compounds metabolic risk.

The complex interactions among clinical features were quantified and integrated into ML models. While previous studies have confirmed the underlying mechanism of individual factors, models combining features from diverse dimensions provide more complete predictive insights. Specifically, sex-specific risk mechanisms dynamically interact with age-related cardiovascular changes, while blood pressure’s temporal cumulative effects and the spatial distribution of metabolic factors form a 3D LVH prediction network. This finding provides a theoretical foundation for developing clinical feature–based ML tools for precise prediction.

ECHO serves as the primary imaging tool for LVH identification. However, its clinical application faces multiple technical challenges, including a relatively low signal-to-noise ratio, inconsistent image quality, and suboptimal reproducibility [28,35]. These limitations are particularly pronounced in patients with obesity due to enhanced chest wall and subcutaneous tissue thickness, which significantly impede image acquisition [28,35]. ML technology offers a novel approach to overcoming these technical barriers. ML models incorporating ECHO demonstrated a sensitivity range of 0.71 to 0.94 and a specificity range of 0.67 to 0.96.

Existing hospital ECHO systems are integrated with this technology to enable automated image analysis and real-time LVH diagnosis. Without replacing existing manual verification procedures, it markedly increases diagnostic efficiency and proves particularly beneficial as a clinical decision aid for less experienced physicians [28]. In contemporary practice, LVH detection chiefly relies on ECHO, where ML proves to be powerful in image processing stages [35,83]. Conventional approaches that depend on manual segmentation and feature extraction are prone to information loss, whereas DL models incorporate image processing, texture extraction, and feature selection throughout the modeling process, which better preserves original image information [35,84,85]. All ML models based on ECHO in this study were DL models. Beyond improved diagnostic efficiency, ML models can automatically identify high-risk cases and enable cardiologists to prioritize urgent interventions, thus shortening care pathways and improving outcomes [49]. Additionally, this technique can enhance diagnostic capabilities at primary care levels, optimize the allocation of cardiovascular imaging resources, reduce unnecessary tests, and be significantly cost-effective in population screening and subclinical disease management [28].

Despite the superior image-processing capabilities of DL models, their clinical integration with ultrasound systems must overcome real-world challenges, including real-time performance, device compatibility, and patient data privacy. Furthermore, the stability of ML models across demographics needs further validation in multicenter research. Future research should prioritize enlarging and diversifying ECHO image databases and validating the actual performance of ML in the detection of LVH in prospective trials.

Meta-analysis of ECG-based ML models for detecting LVH revealed substantial heterogeneity. This restricts the statistical reliability of pooled performance metrics, preventing their interpretation as definitive generalization conclusions. Potential sources of such heterogeneity are manifold. Methodologically, substantial variation in ECG signal preprocessing, feature extraction, and used algorithms introduces fundamental differences. Clinically, diverse underlying disease profiles and severity among studied populations affect model consistency. Regional demographic variations further increase result dispersion. Consequently, pooled estimates from our research require cautious interpretation. Their value lies more in demonstrating the broad performance range and limited confirmatory nature of current evidence than in providing a stable performance expectation. Future efforts should prioritize methodological standardization and prospective multicenter validation to enhance evidence quality in this domain.

When the PROBAST was used to assess model quality, we found that the results raised some concerns, particularly in terms of its strict evaluation of the study population. We believe that some of its criteria may be overly stringent. Although the PROBAST is suitable for evaluating multivariate models for diagnosis and prediction, it is primarily based on retrospective case-control studies, with few prospective studies included [86]. This design preference often leads to a high risk of bias in assessment results, which poses a significant challenge to diagnostic model research. In the studies we included, most involved diagnostic models, and the case selection process itself might introduce bias. The PROBAST uses the EPV principle for sample size estimation, but image-based ML models face challenges in applying EPV rules for bias risk assessment. In future studies, we may explore prior sample size estimation during the research design phase, leveraging ML-based expected performance metrics, such as area under the curve and its 95% CI. Moreover, learning curve analysis serves as a powerful empirical tool. By plotting how a model’s performance on both the training and validation sets changes as the sample size increases, we can intuitively determine whether the model’s performance has reached a plateau, thereby establishing the optimal number of cases. In the field of DL, there exist several widely accepted rules. For instance, in complex image classification tasks, it is generally recommended that each category requires at least several thousand or even tens of thousands of training images to achieve good generalization performance. Although these guidelines are somewhat rough, they provide a practical benchmark for assessing sample sizes in large-scale imaging studies. In addition, the PROBAST requires the verification of the consistency between model statistical weights and actual reported results, which is particularly difficult for models with low interpretability (such as SVM), as they typically do not provide the weight information required in the original validation set [24]. These issues indicate that while the PROBAST is widely used for bias risk assessment in ML research, the observed limitations arise more fundamentally from the inherent biases of the retrospective case-control design—which is applied by most of the current ML diagnostic studies—than from an inherent excessive stringency in the PROBAST criteria. Advancing this field requires constructing and validating models through prospective, multicenter, double-blinded cohort studies, thereby improving evidence quality at its source. Specifically, future work should emphasize prospective sample size estimation and learning curve analysis during development, enhance model interpretability, conduct independent external validation, and ultimately assess the real-world impact of a model on clinical workflows and patient outcomes via pragmatic effectiveness studies. Additionally, we recommend using QUADAS-AI (Quality Assessment of Diagnostic Accuracy Studies Using AI) to assess the risk of bias in included original studies for diagnostic ML tasks [87]. QUADAS-AI focuses on data sources, sample size, eligibility criteria, the rationale for splitting training, validation, test sets, imaging protocols, and preprocessing methods—rather than simply labeling all case-control studies as “high risk.” In the studies we included, a substantial number of studies clearly described the source, size, and eligibility criteria, the principles of splitting into training, validation, and test sets, as well as the imaging protocols and preprocessing methods, thereby mitigating the high risk of bias associated with the study design.

This study has high statistical heterogeneity. Even when we conducted subgroup analyses using various ML approaches and validation methods to further elucidate the sources of heterogeneity, our explanatory power remains limited. This may be due to underlying heterogeneity present during model training. We consider that the following ML processes may introduce potential heterogeneity. First, regarding image preprocessing, the included studies exhibited certain variations in their approaches. These differences may introduce significant heterogeneity. For different images, there is considerable variation in image parameters, yet there is a lack of effective discussion on how this impacts modeling. The images in the original studies exhibit significant variations across imaging protocols. Although we performed subgroup analyses based on different models, even within the same model, differences in model iterations and parameter adjustment rules persist, potentially contributing to heterogeneity.

Our systematic review is the first to quantitatively examine ML performance in LVH detection. While informative, several limitations should be considered when interpreting these results. First, the systematic search, while rigorous, identified solely small case-control studies, thereby narrowing the spectrum of studies for subgroup comparisons. Second, the limited number of included studies precluded meaningful comparisons between different ML methodologies. Third, while most studies completed external validation, some relied solely on random sampling for internal validation. Owing to the limited number of publications, a systematic comparison between internal and external validation approaches is infeasible. Fourth, the RoB in the included studies is high, which is inevitable given the stringent criteria for evaluating ML-based research. This observation corresponds with methodological appraisals indicating that 87% of ML models for medical prediction carry a high RoB [88]. Future studies should optimize design, implementation, reporting, and validation approaches to facilitate clinical translation of ML-based prediction models. Fifth, some studies divide data according to images, and patients’ images are shared between the training set and validation sets, which may result in data leakage. Sixth, there are different reference standards for different models. This discrepancy in reference standards indicates that any comparison of diagnostic accuracy between different model types is inherently confounded. The presented pooled estimates for each subgroup should be interpreted independently and not compared directly to each other. Seventh, the findings of this study may have been influenced to some extent by the fact that 3 papers used the 2×2 formula for estimation. The results should be interpreted with caution. Eighth, several included studies used “clinical consensus” as the reference standard for diagnosing LVH. While this consensus typically adhered to established guidelines (defining LVH as left ventricular mass index >115 g/m² in men and >95 g/m² in women [89]), original studies generally did not explicitly describe whether the consensus process was entirely independent of the outputs of evaluated ML models. Theoretically, “incorporation bias” could be introduced if the predictions of a model were known or incorporated during the adjudication of the reference standard, potentially leading to an overestimation of its diagnostic accuracy. This constitutes a key limitation for evidence interpretation in this systematic review and underscores the necessity for future original research to explicitly report reference standard independence in methodological descriptions.

ML is widely used in LVH research, but its clinical application still faces multiple challenges. Many models are developed based on limited samples and lack standardized processes for key steps, such as image selection, segmentation, and feature extraction, resulting in high reliance of models on the experience of developers [62,90]. This raises concerns about the robustness of the models. Significant differences in image quality across different institutions further weaken the universality of the models. Existing models often focus on accuracy and neglect population risk stratification and regional differences in medical resources, limiting their clinical utility. Considering interpretability and a lack of trust among physicians, it remains difficult to integrate ML into clinical practice [91,92]. Although some models perform well in retrospective studies, their effectiveness and applicability in real-world clinical settings still need to be verified through rigorous prospective trials. Before clinical application, it is necessary to fully address the issues of patient data privacy protection, lack of trust among doctors due to insufficient model interpretability, and ethical considerations associated with new systems [93-95]. Additionally, we must be wary of overreliance on technological means at the expense of humanistic care. The need for systematic management and individualized treatment should not be overlooked. Technology should serve, rather than replace, patient-centered medical practices.

To promote the safe, effective, and responsible application of ML in LVH clinical practice, it is necessary to standardize the model development process, strengthen research on model generalizability and interpretability, conduct rigorous prospective validation, and focus on risk stratification and resource differences. Furthermore, it is necessary to establish a comprehensive ethical and regulatory framework covering data privacy, patient rights, algorithm transparency, and clinical workflow integration. Clear implementation guidelines are needed to ensure that technological development truly benefits patients and improves the quality of medical care.

This study suggests that ML demonstrates diagnostic potential in detecting LVH. However, the overall certainty of the evidence is low, primarily constrained by the limited number of included studies and substantial heterogeneity. Consequently, current conclusions about model performance should be cautiously interpreted. Future investigations should adopt prospective designs and implement standardized data acquisition and model validation protocols to develop and evaluate ML models with enhanced robustness and clinical interpretability. This will advance this technology and make it a reliable diagnostic tool.