Identification of Persons with Acquired Hemophilia in a Large Electronic Health Record Database

Background: Electronic health records (EHR) capture rich clinical information on a patient's encounter with their health care system. Studying national EHRs has the potential to provide epidemiological and granular clinical insights of the patient's diagnosis, treatment patterns and clinical outcomes. Acquired hemophilia (AH) occurs in 1:1-1.5 million people and is characterized by anti-factor VIII autoantibodies, and bleeding in elderly and post-partum patients often in the absence of prior bleeding history. The AH literature includes case reports and registries of known patients; no study to date has attempted to extract population-level information using a large EHR.

Objective: This study aimed to identify patients with AH using a large EHR for the purposes of studying treatment patterns.

Methods: This retrospective study attempts to identify AH patients by combining standard identification methods used in secondary database analysis along with clinical information from the EHR, and a text search of physician notes using Natural Language Processing (NLP) output. Records were accessed from a large, national, de-identified, longitudinal EHR database (Humedica) between January 1, 2007 and July 31, 2013. Given the rare nature of the disease, broader selection criteria were initially applied by using an ICD-9 code for hemorrhagic disorder due to intrinsic circulating anticoagulants (286.5 and all sub-codes), and confirmation of records in the EHR, 6 months before and 12 months after first (index) diagnosis. Additional selection criteria were: no anticoagulant use, not having a diagnosis of systemic lupus erythematosus, accessible physician notes, membership in an integrated delivery network (IDN), mention of "bleeding" within physician notes, and a normal prothrombin test (PT) but an elevated activated partial thromboplastin time (aPTT).

Results: From 6,348 patients with a diagnosis code of 286.5 or any sub-codes, there were 20 males and 18 females who met selection criteria. The median age was 78.5 years (age top coded at 89 years).

All 38 had a note for "pain" (most commonly chest [66%] and back [50%]), 21 (55%) had a note for "bruising", while none of the females had an ICD-9 code for abnormal menstruation or pregnancy.

All 286* diagnoses for the 38 patients included:

No patients received typical AH treatment for bleeding (factor VIII, bypassing agent or DDAVP). The most common prescriptions were azithromycin (34%) and prednisone (32%).

Data review: Individual Signs, Diseases, and Symptoms (SDS) associated with the patient via NLP key word extracts were randomly reviewed to validate if the identified cohort indeed had AH. (1) >89-year-old male; ~6 months of bruising and ecchymoses, "suspect" "clotting factor deficiency". (2) 70-year-old female breast cancer patient; isolated mention of hematemasis from Mallory-Weiss tear, multiple mentions of hematuria, hemorrhagic cystitis, and paradoxically a renal pelvis blood clot in the SDS on the same date. (3) A 32-year-old pregnant female reports bruising and irregular menses, and conflicting mention of venous thrombosis and blot clots. NLP output included FVIII deficiency, and there was mention of congenital and FVIII inhibitor. All 3 had bleeding symptoms, normal PT and elevated aPTT, but none had AH diagnosis or hemostatic treatment.

Conclusions: NLP approaches to analysis of EHRs hold promise and have demonstrated utility in population-based studies for some disorders. However, in this study, ICD-9 codes, lab results, NLP output, and treatments were not consistently aligned. This study highlights that in ultra-rare disorders, ICD-9 coding alone may not be sufficient to identify cohorts, and multimodal analysis combined with in-depth reviews of physician notes might be more effective.