Abstracts

Rationale: While genetic testing can readily be performed in a large number of patients, phenotyping still remains a non-scalable manual task that represents a bottle neck in precision medicine studies. Electronic medical records (EMR) provide a way to obtain systematic longitudinal data over the entire disease course with typically hundreds of data points per individual. We assessed similarities between longitudinal phenotypes as documented in the EMR of patients with known or presumed genetic epilepsies. Methods: We extracted EMR data using the PCORnet protocol on pediatric epilepsy patients with a known or presumed genetic cause recruited within the Genomics Research and Innovation Network, a collaboration between the Children’s Hospital of Philadelphia, Boston Children’s Hospital, and Cincinnati Children's Hospital Medical Center. We transformed the time course of EMR entries per patient into density curves as a proxy for disease activity over time. We then assessed pairwise correlations and manually reviewed phenotypes in patients with high pair-wise correlations. Also, we assessed correlations between patient with known genetic etiologies.  Results: We identified 24,536 EMR entries in 128 patients with a median of 117 encounters per patient (range 2-1493 encounters). The extracted data spanned a total of 594 patient years with a median of 2.54 years per patient. The highest correlation was found in three separate patient pairs, including two patients with infantile spasms starting at the age of six months that subsequently resolved, two patients with explosive seizure onset at the age of 2 years, and two patients with a bimodal peak of seizure activity at the age of 12 months and 24 months. None of these three patient pairs had a common genetic cause. Patients with mutations in SCN1A (n=3), SCN2A (n=3), or STXBP1 (n=3) did not show significant pair-wise correlations and were not represented in patient clusters with a high degrees of correlation.  Conclusions: Comparison of extracted EMR data allows for the identification of patients with strikingly similar epilepsy phenotypes. Particularly, groups of patients with a sudden onset of seizures can readily be identified by our data-driven approach that is agnostic to provider-specific information on phenotypic features or treatment. Self-resolving infantile spasms, characterized by a sudden spike followed by significant drop of entries, are the prime example of clinically relevant phenotypes identified with this strategy. However, the longitudinal phenotype of patients with known genetic epilepsies due to mutations SCN1A, SCN2A, or STXBP1 is too diverse to be captured. Further development of extraction techniques including problem lists, medications, and full-text patients notes have the potential to refine methods to extract systematic, longitudinal phenotypes from electronic medical records.  Funding: This study was in part funded by the Genomics Research and Innovation Network  (GRIN), a collaboration between the Children’s Hospital of Philadelphia, Boston Children’s Hospital, and Cincinnati Children's Hospital Medical Center.