Artificial intelligence for patent ductus arteriosus-a systematic review

Optimal management of patent ductus arteriosus (PDA) remains controversial. Complexity in severity appraisal, high-dimensional data, and the need for longitudinal, individualized assessment make PDA a compelling candidate for Artificial Intelligence (AI)-driven approaches. This systematic review evaluates AI research in the context of PDA, identifying strengths, limitations, and future directions.

Following PRISMA 2020, databases were searched for peer-reviewed articles from January 1, 2010, to May 31, 2025. Eleven studies met inclusion criteria. Data on design, population, sources, AI methods, performance, validation, limitations, and explainability were extracted. Risk of bias was assessed using the Prediction model Risk of Bias Assessment Tool and Joanna Briggs Institute Critical Appraisal Checklist; reporting quality using the Minimum Information about Clinical AI Modeling checklist. Heterogeneity precluded meta-analysis; therefore findings were synthesized narratively.

Eleven studies addressed diagnosis/screening (n = 5), treatment-response prediction (n = 2), risk-factor identification (n = 2), treatment-complication prediction (n = 1), and subphenotype analysis (n = 1). Ten were retrospective; nine single-center, one multi-center, and one used a national registry. Sample sizes were mostly <500 (range: 66-8,369). Definitions of PDA subgroups-symptomatic and hemodynamically significant PDA-varied significantly. Populations included preterm, neonatal and pediatric cohorts, often excluding other congenital heart disease, pulmonary hypertension, or early mortality. Input data ranged from multimodal parameters to high-dimensional unimodal sources. Ten studies used supervised learning; nine traditional machine learning; five deep learning. No study performed adequate external validation. Diagnostic models achieved AUCs of 0.74-0.93, however risk of bias was high, particularly in analysis, suggesting overfitting. Models addressing other aspects showed modest performance. None of the included studies exhibited low risk of bias. Most studies addressed explainability to some degree; only one addressed clinical utility; none evaluated fairness. Reproducibility was hindered by manual preprocessing and limited sharing of data, models, or code.

Artificial intelligence shows feasibility for supporting PDA risk stratification, diagnosis, severity assessment, and prediction of treatment-related outcomes. However, current applications remain in early, pilot-stage development and are not yet suitable for clinical implementation. Future work should prioritize clinically meaningful tasks, scientifically rigorous and bias-aware methodologies, larger and more representative cohorts, and systematic external validation. Fairness, explainability, and reproducibility must be addressed to support translation. Continued methodological refinement and clinical grounding will be key to unlocking the potential of these technologies for this highly vulnerable patient population in the future.