B110-10 Detecting Patient Identity Inconsistencies in Clinical Trials: AI-Driven Identification Using Respiratory Fingerprint Embeddings

Abstract Rationale Ensuring data integrity in clinical trials is critical. We propose an AI-based solution for mirror participants detection, instances where results representing multiple individuals come from a common source 1. Our approach can detect patient’s unique “respiratory fingerprint” using spirometry measurements and embedding approach. Methods We trained an AI model using more than 40,000 spirometry curves from an asthma clinical trial with 2,173 anonymized patients across 334 sites. The convolutional neural network trained on forced expiratory part and tidal breath allows to embed spirometry efforts into a latent vector space of 64 dimensions. By comparing these embedded values, our approach clusters patients with high similarities. Site level mirror patient detection is performed by: (1) computing embeddings for all randomization visit trials per patient at selected site, (2) selecting a representative trial for each patient, (3) calculating pairwise patient distances and applying a threshold based on intra-patient statistics, and (4) flagging pairs below this threshold as potential mirrors cases. Results To validate our approach, we used 10,000 spirometry curves from 582 anonymized patients with asthma. A mean average precision of 0.51 was reached when evaluated across the entire patient pool. This metric quantifies the model’s ability to rank trials from the same patient as more similar to each other than to those from different patients, indicating that intra-patient efforts are consistently positioned closer together in the latent space, even in the presence of similar patients from other sites. Figure 1 shows the 2D projection of embedded values for two sites, illustrating patient distances. In Site 1, patients are clearly separated, while in Site 2, Patients 3 and 6 appear mirrored, consistent with their similar flow-volume curves and suggesting possible identity duplication. Finally, in a preliminary retrospective analysis, we compared the sites flagged by the AI with those identified by human reviewers. The AI flagged 5 out of 9 sites, achieving a 77.8% alignment with the human assessments. Conclusions Embedding-based patient identity encoding provides a robust and scalable framework for detecting mirroring at both patient and site levels and assisting experts ensuring data integrity in clinical trials. Results are promising however additional validation steps are required. Future work will focus on testing its generalization to other conditions such as COPD, ILD, and healthy populations. The model could also be leveraged to track longitudinal changes in patient identity across visits and support broader clinical trial monitoring. References: 1. Cuyvers (2025). DOI: 10.1164/ajrccm.2025.211.Abstracts.A7946 This abstract is funded by: None