Which Rare Diseases Are Best Suited to Diagnostic AI? A Systematic Selection Framework

Abstract

Background Patients with rare diseases are fundamentally challenging to diagnose. Despite the fact that some estimates place rare diseases as collectively as prevalent as one in ten, diagnosis often takes >4 years and oftentimes much, much longer, and can lead to lifelong negative health consequences or early death. Expert diagnosticians and clinical geneticists are a scarce commodity, and wait times are usually measured in months to years. Therefore, there is a great need to assist healthcare systems and service providers of all kinds in improving diagnostic capabilities. Artificial intelligence and machine learning (AI/ML) make assisting primary care and non-expert providers in advancing diagnostics for undiagnosed patients much more feasible; however, there are no comprehensive, longitudinal, multimodal datasets available to train AI models that could be subsequently deployed in healthcare. Methodology We developed a rigorous prioritization framework to identify rare diseases that would be well-suited for the development of AI diagnostic models. The prioritization framework was designed to guide the selection of conditions for inclusion in the collection of multi-site, multi-modal benchmarking datasets. We prioritized diseases with low diagnostic rates relative to their prevalence and progressive and multi-system diseases where pleiotropic manifestations across specialties may be more amenable to signal detection. We considered the impact of diagnostic delay where earlier detection could meaningfully improve outcomes or where diagnosis would lead to actionable management changes. Finally, we considered the feasibility of confirming a suspected diagnosis and prioritized conditions with inexpensive or widely available tests. Results Our informatics approach leverages the Mondo disease classification, ontology-based characterization of phenotypic breadth and depth, real-world and published prevalence information, availability of therapeutics, and clinical expertise. We approximated which of the 16,403 rare disease classes in Mondo had similarly labeled ICD 10-CM or ICD-11 codes, finding a maximum similarity match of 2.7% and 15%, respectively. Using our prioritization framework, we selected 3,079 diseases(1,2) across all anatomical systems, including rare diseases with higher and lower prevalence. Preliminarily, approximately 9% of rare diseases had an approved or research grade drug associated. Conclusions Artificial intelligence advances are expected to greatly improve diagnostic processes and efficacy, but require quality data and benchmarking before they can be deployed for clinical use. The outcome of the preliminary analysis presented here provisions a robust, targetable set of rare diseases that would most likely benefit from the development of diagnostic AI models.

Ähnliche Arbeiten