SCALING RARE DISEASE INSIGHTS THROUGH AI-DRIVEN PHENOTYPE SEARCH ON FEDERATED NETWORKS
Author(s)
Gabriel de Maeztu, MD1, Elizabeth Romelia Viera Ramírez, MD2, Leonor Fayos de Arizon, MD2, Melissa Pilco Teran, MD2, Maria Quijada Lopez, Master3, Roser Torra Balcells, PhD2.
1President, IOMED, Barcelona, Spain, 2FUNDACIO PUIGVERT, Barcelona, Spain, 3IOMED, Barcelona, Spain.
1President, IOMED, Barcelona, Spain, 2FUNDACIO PUIGVERT, Barcelona, Spain, 3IOMED, Barcelona, Spain.
OBJECTIVES: Rare hereditary diseases often face a prolonged "diagnostic odyssey" and significant data scarcity.
This study aims to validate a Longitudinal Patient Similarity (LPS) framework deployed across a federated network of hospitals. The primary objective is to demonstrate that advanced AI can identify high-fidelity patient cohorts suitable for accelerating diagnosis, constructing synthetic clinical trial arms, and enabling federated HEOR analyses.
METHODS: Leveraging the EHDEN network infrastructure, we implemented a federated, privacy-preserving AI framework across multiple hospital sites. Using the OMOP CDM to harmonize electronic health records (EHR), pharmacy, and laboratory data, we applied a vector-similarity AI model. This model was trained on small "seed cohorts" (n=10-22 confirmed patients) to identify undiagnosed individuals by matching their unique clinical trajectories, the chronological sequence and timing of health events, against known disease phenotypes. Validation was conducted retrospectively across 1.2 million patient records and prospectively in routine clinical workflows for five distinct pathologies: Alport Syndrome, Fabry disease, Primary Hyperoxaluria, and Tuberous Sclerosis Complex.
RESULTS: The LPS framework demonstrated robust generalizability across multiple independent hospital sites. In a preselected cohort of 72,611 patients with Chronic Kidney Disease, the model achieved a sensitivity >84% for all disease cohorts. The effectiveness of the screening was confirmed by a clinically manageable Number Needed to Screen (NNS), with a median of 30 patient reviews required to identify one confirmed case. High-similarity cohorts were validated as representative of the target rare disease phenotypes, confirming their suitability for secondary research use cases, such as the generation of large-scale, real-world evidence pools and the construction of synthetic control arms for clinical trial augmentation.
CONCLUSIONS: This study demonstrates that federated RWD networks augmented by AI-driven longitudinal similarity, provide a scalable "safety net" for rare disease diagnosis. Beyond diagnostic utility, these high-fidelity cohorts bridge the gap between rare disease populations and rigorous RWE generation.
This study aims to validate a Longitudinal Patient Similarity (LPS) framework deployed across a federated network of hospitals. The primary objective is to demonstrate that advanced AI can identify high-fidelity patient cohorts suitable for accelerating diagnosis, constructing synthetic clinical trial arms, and enabling federated HEOR analyses.
METHODS: Leveraging the EHDEN network infrastructure, we implemented a federated, privacy-preserving AI framework across multiple hospital sites. Using the OMOP CDM to harmonize electronic health records (EHR), pharmacy, and laboratory data, we applied a vector-similarity AI model. This model was trained on small "seed cohorts" (n=10-22 confirmed patients) to identify undiagnosed individuals by matching their unique clinical trajectories, the chronological sequence and timing of health events, against known disease phenotypes. Validation was conducted retrospectively across 1.2 million patient records and prospectively in routine clinical workflows for five distinct pathologies: Alport Syndrome, Fabry disease, Primary Hyperoxaluria, and Tuberous Sclerosis Complex.
RESULTS: The LPS framework demonstrated robust generalizability across multiple independent hospital sites. In a preselected cohort of 72,611 patients with Chronic Kidney Disease, the model achieved a sensitivity >84% for all disease cohorts. The effectiveness of the screening was confirmed by a clinically manageable Number Needed to Screen (NNS), with a median of 30 patient reviews required to identify one confirmed case. High-similarity cohorts were validated as representative of the target rare disease phenotypes, confirming their suitability for secondary research use cases, such as the generation of large-scale, real-world evidence pools and the construction of synthetic control arms for clinical trial augmentation.
CONCLUSIONS: This study demonstrates that federated RWD networks augmented by AI-driven longitudinal similarity, provide a scalable "safety net" for rare disease diagnosis. Beyond diagnostic utility, these high-fidelity cohorts bridge the gap between rare disease populations and rigorous RWE generation.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
P35
Topic
Methodological & Statistical Research, Real World Data & Information Systems, Study Approaches
Topic Subcategory
Distributed Data & Research Networks
Disease
Rare & Orphan Diseases, Urinary/Kidney Disorders