The increasing reliance on digital health platforms and online medical communities has raised critical concerns regarding the presence of fake or malicious user profiles that can compromise the integrity of shared medical information. Ensuring trustworthiness in such environments is essential for reliable biomedical data analysis and decision-support systems.
In this work, we propose a graph-based machine learning framework for detecting fake and anomalous user profiles in digital health networks. The method models user interactions, behavioral patterns, and content similarities as a heterogeneous graph, enabling the capture of both structural relationships and semantic dependencies. Graph representation learning techniques are then applied to extract meaningful embeddings that reflect user behavior in the network.
To enhance detection performance, we integrate temporal activity features with graph embeddings and evaluate several machine learning models, including Graph Neural Networks and ensemble-based classifiers. Experiments conducted on simulated and real-world inspired datasets demonstrate that incorporating relational and temporal information significantly improves detection accuracy compared to traditional feature-based approaches.
The proposed approach highlights the effectiveness of graph-based artificial intelligence for improving data integrity in digital health ecosystems and contributes to the development of robust tools for trustworthy biomedical data analysis.