Artificial intelligence–based cuffless blood pressure systems have not yet demonstrated sufficiently consistent, generalizable performance for routine clinical decision-making, according to a systematic review and meta-analysis published in npj Digital Medicine. 

“Our analysis shows that inconsistent cohort characterization, heterogeneous performance reporting, and insufficient validation protocols remain major barriers to clinical translation,” wrote first author Ana Cisnal of the Spinal Cord Artificial Intelligence Lab in Zurich, Switzerland, and colleagues.  

The researchers performed a systematic review and meta-analysis of 86 published studies evaluating machine learning-based cuffless blood pressure monitoring approaches published from January 2017 through March 2025. 

The review found that most published studies (73%) evaluated data-driven machine learning regression models, with fewer examining hemodynamics-derived approaches (16%) or arterial waveform reconstruction (11%). Photoplethysmography was the predominant sensing modality, appearing in 70 of the 86 studies, reflecting its central role in current cuffless blood pressure monitoring research. 

Validation methods 

Validation methods varied widely across studies, and the investigators found that the differences in strategy substantially influenced reported performance. Among the reviewed studies: 

• 34 (39%) used subject-specific validation 

• 48 (55%) used holdout validation 

• 38 (44%) used cross-validation approaches. 

Although cross-validation became more common after 2023, the researchers found that many studies failed to maintain subject-wise separation between training and testing datasets, increasing the risk of information leakage and inflated performance estimates. 

The review also identified important differences between calibration-based and calibration-free algorithms. Calibration-based approaches require periodic reference measurements using a cuff device, whereas calibration-free models attempt to estimate blood pressure without individualized calibration. The researchers noted that these two paradigms require distinct validation strategies and should not be evaluated using identical reporting frameworks. 

The investigators also performed a random-effects meta-analysis of reported systolic blood pressure mean error to assess consistency across published studies. Across all validation strategies, pooled systolic blood pressure mean error measured 1.12 mmHg, indicating only a slight average overestimation. However, heterogeneity was extremely high, indicating that reported performance varied substantially across studies and should not be interpreted as evidence of consistently reliable clinical performance, according to the researchers. 

Subject-wise validation produced lower pooled mean error than record-wise validation, providing empirical support for subject-level evaluation as the preferred validation approach. Studies using subject-wise validation demonstrated a pooled systolic blood pressure mean error of 0.81 mmHg, compared with 1.50 mmHg for studies using record-wise validation. The researchers interpreted this pattern as consistent with the possibility that participant overlap between training and testing produced overly favorable estimates of generalizability. 

Risk of bias 

No study achieved an overall low risk of bias. Validation independence was the most problematic domain, with 63% of studies classified as having unclear risk and 37% classified as high risk. Studies with high overall risk of bias demonstrated larger pooled systolic blood pressure mean errors (1.68 mm Hg vs 0.39 mm Hg among studies with unclear overall risk). 

Publication bias analyses found no strong evidence that selective publication substantially distorted the literature. Instead, the investigators concluded that inconsistent methodologies, small sample sizes, and heterogeneous study designs represented the primary obstacles to reliable interpretation. 

The review found considerable inconsistency in how investigators reported model performance. Although many studies reported mean deviation, mean absolute difference, or both, few adhered completely to existing reporting recommendations. None simultaneously reported all recommended evaluation metrics, including mean deviation with standard deviation, mean absolute difference, cumulative percentage errors at 5, 10, and 15 mmHg, mean absolute percentage difference, and correlation coefficients. 

The researchers also noted that several studies incorrectly reported compliance with American Association for the Advancement of Medical Instrumentation standards or used nonrecommended performance metrics, limiting meaningful comparisons across studies. They further observed that many publications included Bland-Altman plots without verifying the assumption that measurement differences followed an approximately normal distribution, an important requirement for valid agreement analysis. 

Reporting of demographic characteristics was limited. Among the 86 reviewed studies: 

• one reported skin tone 

• two reported ethnicity 

• three stratified results by sex 

• two stratified performance by age 

• five reported performance according to hypertension stage 

• one described antihypertensive medication use. 

The researchers noted several limitations. Across published studies, inconsistent cohort characterization, heterogeneous validation protocols, limited ambulatory testing, inadequate baseline model comparisons, and nonstandardized reporting made it difficult to determine whether reported performance reflected true predictive capability or favorable study conditions. They also noted that nearly all published studies were conducted in controlled or clinic-based environments, whereas evidence supporting reliable performance during real-world ambulatory monitoring remains limited. 

The investigators recommended that future studies should include subject-independent validation, external testing across diverse populations and blood pressure ranges, dynamic testing that captures physiologic variation during daily activities, and randomized clinical trials demonstrating improved patient outcomes. They also highlighted the need to evaluate performance in older adults, patients with hypertension, patients receiving antihypertensive medications, and other populations that remain underrepresented in existing studies. 

"We believe that the adoption of harmonized validation standards and transparent reporting practices is essential to improve reproducibility, enable fair comparison across studies, and accelerate the development of clinically meaningful cuffless BP monitoring technologies suitable for continuous real-world and ambulatory use,” wrote Cisnal and colleagues. 

The researchers declared no conflicts of interest.  

(Editor’s Note: The researchers noted that the study manuscript will undergo further editing before final publication, and there may be errors present that affect the content. All legal disclaimers apply.)