Trustworthy AI for Biomedical Networks: Safety, Robustness, and Clinical Reliability

Artificial intelligence applied to patient-level and clinical networks, including patient similarity graphs, multimodal clinical networks, electronic health record (EHR)–derived graphs, disease and comorbidity networks, and clinical knowledge graphs, is creating new opportunities for disease diagnosis, prognosis, treatment planning, and clinical decision support. However, the reliability of these systems in real-world settings remains a major concern. Clinical networks are often incomplete, noisy, and heterogeneous, and may be affected by missing records, inconsistent coding, measurement errors, and differences between patient populations and healthcare systems. Network-based models may also be vulnerable to adversarial perturbations, data leakage, distribution shifts, and poorly calibrated predictions. These challenges raise important questions about whether model outputs can be reproduced, interpreted, and used safely in clinical practice.Recent studies have shown the potential of graph neural networks and other network-based learning methods across a range of clinical applications. Yet their performance can decline when tested on independent datasets, at different institutions, or under changing clinical conditions. Current efforts address these challenges through uncertainty estimation, robust training, explainable modelling, external validation, and systematic benchmarking. Still, widely accepted standards for assessing the safety, robustness, transparency, and clinical reliability of patient-level network AI are lacking.This Research Topic aims to advance the development and evaluation of trustworthy AI methods for patient-level and clinical network analysis. It will examine how models built from patient similarity graphs, EHR data, comorbidity structures, and clinical knowledge graphs can remain reliable when faced with noisy or missing information, adversarial perturbations, heterogeneous populations, and shifts between research and clinical environments. Contributions may investigate ways to improve model robustness, quantify and communicate predictive uncertainty, identify sources of bias, and make model decisions understandable and auditable. The Topic also seeks to clarify how computational performance should connect to clinical utility, patient safety, reproducibility, and regulatory expectations. Relevant questions include how clinical network models generalise across sites and populations, how interpretability relates to clinical validity, and what is needed for responsible integration into healthcare workflows.We welcome contributions on methods, applications, evaluation strategies, and translation across patient similarity, EHR-derived, comorbidity, and clinical knowledge network settings. Topics include, but are not limited to:- Robustness of graph neural networks and other AI methods for patient similarity graphs, EHR-derived graphs, and clinical knowledge graphs- Safety evaluation and risk assessment of network-based clinical models- Uncertainty quantification, calibration, and confidence estimation- Adversarial attacks, defence mechanisms, and stress testing- Learning from noisy, incomplete, dynamic, or missing clinical data- Detection and mitigation of dataset shift, institutional effects, and population bias- Explainable and interpretable AI for patient, disease, comorbidity, and clinical knowledge networks- Clinical validation of network-based model explanations- Benchmarking, reproducibility, and external validation across datasets and institutions- Fairness, privacy, security, and responsible use of patient network data- Clinical decision support, real-world utility, and integration into healthcare workflows- Regulatory, ethical, and governance frameworks for trustworthy clinical AI Topic editors: Daqian Shi, Marco Fernandes Source: https://www.frontiersin.org/research-topics/85797/trustworthy-ai-for-biomedical-networks-safety-robustness-and-clinical-reliability