Comparative Diagnostic Accuracy of Artificial-intelligence derived risk stratification versus conventional risk stratification methods in Pulmonary Hypertension patients: A systematic review and meta-analysis

Faizan Ahmed¹Faseeh Haider, MD²Muhammad Arham³Allah Dad³Kinza Bakht³Muhammad Moseeb Ali Hashim⁴Paweł Łajczak⁵Muhammad Hassan²Fatima Binte Athar⁶Muhammad Adnan Haider⁷Muhammad Usman⁸Najam Gohar⁸Tehmasp Mirza⁹Mushood Ahmed¹⁰Mark Moshiyakhov¹Brett Sealove¹Swapnil Patel¹Jesus Almendral¹Mohamed Bakr¹Yasar Sattar^11*Fawaz Alenezi¹²

• ¹Hackensack Meridian Jersey Shore University Medical Center, Neptune City, United States
• ²Allama Iqbal Medical College, Lahore, Pakistan
• ³Sheikh Zayed Medical College, Rahim Yar Khan, Pakistan
• ⁴University of Missouri, Columbia, United States
• ⁵Slaski Uniwersytet Medyczny w Katowicach, Katowice, Poland
• ⁶Karachi Medical and Dental College, Karachi, Pakistan
• ⁷Hospitalist Physician, Mission Hospital, Asheville, North Carolina, Asheville, United States
• ⁸Post Graduate Medical Institute Ameer-ud-Din Medical College, Lahore, Pakistan
• ⁹Shalimar Medical and Dental College, Lahore, Pakistan
• ¹⁰Rawalpindi Medical University, Rawalpindi, Pakistan
• ¹¹Department of Interventional Cardiology, Tidal Health, Salisbury, Maryland, Salsbury, United States
• ¹²Duke University Division of Cardiology, Durham, United States

Background: Accurate risk stratification in pulmonary hypertension (PH) is integral for optimizing therapeutic strategies and improving patient outcomes. Recent artificial intelligence (AI) models have demonstrated notable efficacy in risk stratification of PH, achieving Area Under Curve (AUC) values of 0.94 and 0.81 in internal and external validation cohorts, respectively. This meta-analysis aims to demonstrate the effectiveness of AI models in the risk stratification of PH by comparing their performance to conventional risk stratification methods. Methods: A systematic search of five databases (PubMed, Embase, ScienceDirect, Scopus, and the Cochrane Library) was conducted from inception to March 2025. Statistical analysis was performed in R (version 2024.12.1+563) using 2×2 contingency data. Sensitivity, specificity, and diagnostic odds ratio (DOR) were pooled using a bivariate random-effects model (reitsma from the mada package), while the AUC was meta-analyzed using logit-transformed values via the metagen() function from the meta package. Results: Six studies were included in the final synthesis, comprising 14,095 patients: 4,481 in internal test datasets and 4,948 in external datasets. AI risk stratification models showed significant performance with a logit mean difference of 0.26 (95% CI 0.09 - 0.43; p=0.31) having low heterogeneity (I2=14.3%) as compared to conventional methods. Furthermore, pooled sensitivity and specificity were 0.77 (95% CI 0.74 - 0.79) and 0.72 (95% CI 0.70 - 0.75) in favour of AI methods respectively. The heterogeneities for pooled sensitivity and specificity were 57.1% (p=0.04) and 91.8% (p<0.0001), underscoring high variability across all studies. Lastly, DOR was substantially high, 8.53 (6.59 - 11.04) in favour of AI models with a high heterogeneity of 73.6% (p=0.002). Heterogeneity (I2) for pooled sensitivity went to 25.9% after excluding major outlier but it remained high for pooled specificity and DOR upon leave-one-out sensitivity analysis. Conclusions: Artificial intelligence-based risk stratification demonstrates significantly higher diagnostic performance compared to conventional methods in pulmonary hypertension. The higher pooled AUC, sensitivity, specificity, and DOR highlight AI's potential to enhance predictive accuracy, guiding better treatment strategies. Nonetheless, more superior quality studies are needed to validate AI models for clinical integration.