Patient representation from clinical narratives for early diagnosis of rare diseases: a comparative study on Activated PI3K-δ Syndrome

Abstract Background: Activated phosphoinositide 3-kinase delta syndrome (APDS), first described in 2013, is a rare primary immunodeficiency with an estimated prevalence of 1–2 cases per million. Caused by mutations in PIK3CD or PIK3R1 , it leads to immune dysregulation, lymphoproliferation, recurrent infections, and malignancy. Early identification is critical to enable timely initiation of personalized targeted therapies. This study aimed to evaluate whether machine learning applied to clinical narratives could support earlier identification of APDS patients using information extracted from electronic clinical notes. Results: Two approaches derived from clinical notes were evaluated: transformer-based embeddings and phenotype vectors based on UMLS concepts. Machine learning models were trained on 222 patients, including 29 APDS patients. Both approaches achieved strong performance on the full dataset (F1 = 0.82 ± 0.13 and 0.80 ± 0.12, respectively). Phenotype-based analyses included statistical testing, feature importance, ablation studies, and expert review. These analyses identified clinically relevant features including immunodeficiency, adenopathy, lymphopenia, and splenomegaly. Performance remained robust when evaluated on a temporally constrained subset using only pre-diagnosis data (F1 = 0.79 ± 0.14 and 0.74 ± 0.19), supporting the feasibility of earlier detection. Post hoc validation on an independent cohort of 100 patients further supported real-world applicability. Conclusions: Machine learning approaches based on clinical narratives can reliably identify APDS patients, including prior to diagnosis. These findings suggest that such methods could support earlier detection of undiagnosed APDS patients, reducing diagnostic odyssey and improving patient care.