Inborn errors of immunity (IEIs) often remain undiagnosed due to many factors including the fragmentation of care across subspecialties. The increasing complexity of these disorders poses a diagnostic challenge, even for immunologists. Artificial intelligence (AI) algorithms trained on electronic health records (EHRs) and expert knowledge can identify a broad range of IEI phenotypes. Clinical validation of these tools is needed to identify undiagnosed patients and impact care.
We sought to describe patients with putative IEIs identified by an algorithm trained on EHR signatures.
We published an algorithm PheNet trained on EHR data of patients with IEIs [1]. PheNet identified suspected IEI patients across the University of California (UC) medical database. Algorithmically identified high-risk individuals were reviewed by clinicians and bioinformaticians; those of interest following chart review were recruited, consented, and underwent genetic and immune phenotyping.
The top 500 cases identified via PheNet were subsequently reviewed by an immunologist. Each center found ∼15 subjects that were considered suspicious for IEIs. Currently, there are 25 subjects who consented who have undergone sequencing and full immune workup. Immune phenotyping revealed T and B cell defects, while genetic testing has identified several actionable variants.
AI algorithms are useful in the identification of patients with IEI phenotypes, with meaningful clinical implications. By combining AI to identify subjects with concerning clinical phenotypes with in-person clinical examination and sequencing, we identified ∼20 subjects who had remained undiagnosed despite receiving specialty care. One outcome of this effort will be to expand surveillance beyond simple “warning signs” to more sophisticated EHR signatures of IEIs.
Reference
