Validation of Electronic Health Record Phenotyping of Bipolar Disorder Cases and Controls

Abstract

Objective To validate the use of electronic health records (EHRs) for the diagnosis of bipolar disorder (BD) and controls. Methods EHR data were obtained from a healthcare system of more than 4.2 million patients spanning more than 20 years. Chart review by experienced clinicians was used to identify text features and coded data consistent or inconsistent with a diagnosis of BD. Natural language processing (NLP) was used to train a diagnostic algorithm with 95% specificity for classifying BD. Filtered coded data were used to derive three additional classification rules for cases and one for controls. The positive predictive value (PPV) of EHR-based BD and subphenotype diagnoses was calculated against direct semi-structured interview diagnoses by trained clinicians blind to EHR diagnosis in a sample of 190 patients. Results The PPV of NLP-defined BD was 0.85. A coded classification based on strict filtering achieved a PPV of 0.79, but BD classifications based on less stringent criteria performed less well. None of the EHR-classified controls was given a diagnosis of BD on direct interview (PPV = 1.0). For most subphenotypes, PPVs exceeded 0.80. The EHR-based classifications were used to accrue 4500 BD cases and 5000 controls for genetic analyses. Conclusions Semi-automated mining of EHRs can be used to ascertain BD cases and controls with high specificity and predictive value compared to a gold-standard diagnostic interview. EHRs provide a powerful resource for high-throughput phenotyping for genetic and clinical research.

Phenome-wide Heritability Analysis of the UK Biobank

“…However, our head-to-head comparison of the heritability estimates between self-reported illness and ICD-10 codes showed largely consistent results, indicating that both phenotypic approaches at least captured comparable variations in these phenotypes. Prior research evaluating phenotypes derived from electronic health records (EHR) indicate that greater phenotypic validity can be achieved when diagnostic codes are supplemented with text mining methods [55][56][57][58]. The specificity of the disease codes might also be improved by leveraging the medication records in the UK Biobank.…”

Efficient Genome-wide Association in Biobanks Using Topic Modeling Identifies Multiple Novel Disease Loci

“…Specifically, we apply latent Dirichilet allocation (LDA), a means of identifying commonly cooccurring features, to derive a set of 50 disease topics. Then we test those topics for association with common genetic variation and compare this approach to capture a given diagnosis varies widely, even when diagnosis-specific classifiers are applied to augment single codes (2,3). As such, approaches that focus on individual diagnostic codes are limited by inaccurate, missing or heterogeneous diagnoses; eg, where individuals with cystic fibrosis might be represented by male infertility, diabetes and chronic rhinosinusitis even in the absence of a diagnostic code for cystic fibrosis (4).…”

Real world big data for clinical research and drug development

“…Specifically, given the lack of treatments in neurological disorders, the use of algorithms in the identification of new patents is a pressing need for the biopharmaceutical sector. One study outlined the effective application of semiautomated mining of EHRs to ascertain bipolar disorder patients and control subjects with high specificity and predictive value when compared with diagnostic interviews [28]. This technique could have broad applicability across many research areas in neurology.…”