OBJECTIVES : Pyruvate kinase deficiency (PKD) is a rare, inherited (autosomal recessive) red blood cell enzyme disorder that causes lifelong hemolytic anemia. Because there is no ICD9/10 code available to identify PKD in real-world database study, this study aims to evaluate the application of natural language processing (NLP) on physician notes to identify or confirm a diagnosis of PKD.

METHODS : We used the Veterans health Information Systems and Technology Architecture (VistA) which is a rich, automated environment that supports day-to-day operations at local Department of Veterans Affairs (VA) health care facilities across the United States. In Vista, the Text Integration Utility (TIU) is a set of files that stores progress notes, consult reports, discharge summaries and other free text documents. Natural language processing (NLP) enables to extract data from the unstructured free text TIU notes as part of the US Veterans Affairs electronic medical records (10/1/1999 onward). The keywords: “Pyruvate”, “kinase” and “deficiency” were used to search TIU notes. The results of NLP were compared with those searched by SQL-like operator with wildcard “%”. The efficiency of NLP was further manually examined with TIU note review among the results of NLP.

RESULTS : Among the 3811 patients possibly associated with PKD diagnosis from January 1, 1999 onwards were selected, only 46 patients were extracted using NLP, while 450 possible patients were identified using SQL-like operator, including all of 46 patients from NLP. The manual search only found 18 patients (mean age 56.8±13.6, female 5.6%, White 83.3%) with confirmed PKD including 9 death. The median follow-up time was 6.0 years. The annualized emergency room visits, out-patient visits and hospitalization admissions were 5.5±13.9, 118.0±251.6 and 2.8±5.2, respectively.

CONCLUSIONS : NLP appears to be more accurate than SQL like operator in identifying PKD cases NLP application may have great potential in rare disease real world research.