Assessing the precision of machine learning for diagnosing pulmonary arterial hypertension: a systematic review and meta-analysis of diagnostic accuracy studies

Abstract

Introduction: Pulmonary arterial hypertension (PAH) is a severe cardiovascular condition characterized by pulmonary vascular remodeling, increased resistance to blood flow, and eventual right heart failure. Right heart catheterization (RHC) is the gold standard diagnostic technique, but due to its invasiveness, it poses risks such as vessel and valve injury. In recent years, machine learning (ML) technologies have offered non-invasive alternatives combined with ML for improving the diagnosis of PAH.

Objectives: The study aimed to evaluate the diagnostic performance of various methods, such as electrocardiography (ECG), echocardiography, blood biomarkers, microRNA, chest x-ray, clinical codes, computed tomography (CT) scan, and magnetic resonance imaging (MRI), combined with ML in diagnosing PAH.

Methods: The outcomes of interest included sensitivity, specificity, area under the curve (AUC), positive likelihood ratio (PLR), negative likelihood ratio (NLR), and diagnostic odds ratio (DOR). This study employed the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool for quality appraisal and STATA V.12.0 for the meta-analysis.

Results: A comprehensive search across six databases resulted in 26 articles for examination. Twelve articles were categorized as low-risk, nine as moderate-risk, and five as high-risk. The overall diagnostic performance analysis demonstrated significant findings, with sensitivity at 81% (95% CI = 0.76-0.85, p < 0.001), specificity at 84% (95% CI = 0.77-0.88, p < 0.001), and an AUC of 89% (95% CI = 0.85-0.91). In the subgroup analysis, echocardiography displayed outstanding results, with a sensitivity value of 83% (95% CI = 0.72-0.91), specificity value of 93% (95% CI = 0.89-0.96), PLR value of 12.4 (95% CI = 6.8-22.9), and DOR value of 70 (95% CI = 23-231). ECG demonstrated excellent accuracy performance, with a sensitivity of 82% (95% CI = 0.80-0.84) and a specificity of 82% (95% CI = 0.78-0.84). Moreover, blood biomarkers exhibited the highest NLR value of 0.50 (95% CI = 0.42-0.59).

Conclusion: The implementation of echocardiography and ECG with ML for diagnosing PAH presents a promising alternative to RHC. This approach shows potential, as it achieves excellent diagnostic parameters, offering hope for more accessible and less invasive diagnostic methods.

Systematic review registration: PROSPERO (CRD42024496569).

Keywords: area under receiving operator curve; area under the curve; diagnostic method; machine learning; pulmonary arterial hypertension.

Figure 1

Preferred reporting items for systematic reviews and meta-analyses (PRISMA) flowchart for study identification and selection. The original database search resulted in 1,809 studies from six databases searched, namely PubMed, ScienceDirect, ProQuest, Taylor & Francis, Embase, and EBSCO. Through title and abstract screening, 1,762 articles were removed, and 47 articles were screened for duplication. Duplicate screening resulted in 11 removed articles. Thirty-six articles were further assessed for eligibility and ten articles were removed due to irrelevant data, evaluation, or outcomes. The final step resulted in 26 clinical trials included in the qualitative synthesis.

Figure 2

Risk of bias summary using the QUADAS-2 tool for diagnostic studies. The green region represents studies with a low risk of bias, the yellow region shows studies with a moderate risk of bias, and the red region shows studies with a high risk of bias.

Figure 3

Forest plot showing overall sensitivity (left forest plot) and specificity (right forest plot) with corresponding heterogeneity statistics. The gray square and solid lines represent the odds ratio with 95% confidence intervals. The rhombus indicates the pooled estimate with 95% confidence intervals.

Figure 4

Summary ROC curve with confidence and prediction regions around mean operating sensitivity and specificity point.

Figure 5

Bivariate boxplot with most studies clustering within the median distribution with three outliers suggesting indirectly a lower degree of heterogeneity.

Figure 6

Funnel plot with superimposed regression line. The vertical axis displays the inverse of the square root of the effective sample size [1/root(ESS)]. The horizontal axis displays the diagnostic odds ratio (DOR). This Deek's funnel plot asymmetry test is a useful tool for assessing the potential publication bias in studies.

Figure 7

Forest plot showing echocardiography subgroup mean sensitivity (left forest plot) and specificity (right forest plot) with corresponding heterogeneity statistics. The gray square and solid lines represent the odds ratio with 95% confidence intervals. The rhombus indicates the pooled estimate with 95% confidence intervals.

Figure 8

Forest plot showing blood biomarkers subgroup mean sensitivity (left forest plot) and specificity (right forest plot) with corresponding heterogeneity statistics. The gray square and solid lines represent the odds ratio with 95% confidence intervals. The rhombus indicates the pooled estimate with 95% confidence intervals.

Figure 9

Forest plot showing microRNA subgroup mean sensitivity (left forest plot) and specificity (right forest plot) with corresponding heterogeneity statistics. The gray square and solid lines represent the odds ratio with 95% confidence intervals. The rhombus indicates the pooled estimate with 95% confidence intervals.

Figure 10

Forest plot showing other subgroups’ mean sensitivity (left forest plot) and specificity (right forest plot) with corresponding heterogeneity statistics. The gray square and solid lines represent the odds ratio with 95% confidence intervals. The rhombus indicates the pooled estimate with 95% confidence intervals.

Figure 11

Forest plot showing ECG subgroup mean sensitivity (left forest plot) and specificity (right forest plot) with corresponding heterogeneity statistics. The gray square and solid lines represent the odds ratio with 95% confidence intervals. The rhombus indicates the pooled estimate with 95% confidence intervals.