Text Mining of Electronic Health Records Can Accurately Identify and Characterize Patients With Systemic Lupus Erythematosus

Abstract

Objective Electronic health records (EHR) are increasingly being recognized as a major source of data reusable for medical research and quality monitoring, although patient identification and assessment of symptoms (characterization) remain challenging, especially in complex diseases such as systemic lupus erythematosus (SLE). Current coding systems are unable to assess information recorded in the physician’s free‐text notes. This study shows that text mining can be used as a reliable alternative. Methods In a multidisciplinary research team of data scientists and medical experts, a text mining algorithm on 4607 patient records was developed to assess the diagnosis of 14 different immune‐mediated inflammatory diseases and the presence of 18 different symptoms in the EHR. The text mining algorithm included key words in the EHR, while mining the context for exclusion phrases. The accuracy of the text mining algorithm was assessed by manually checking the EHR of 100 random patients suspected of having SLE for diagnoses and symptoms and comparing the outcome with the outcome of the text mining algorithm. Results After evaluation of 100 patient records, the text mining algorithm had a sensitivity of 96.4% and a specificity of 93.3% in assessing the presence of SLE. The algorithm detected potentially life‐threatening symptoms (nephritis, pleuritis) with good sensitivity (80%‐82%) and high specificity (97%‐97%). Conclusion We present a text mining algorithm that can accurately identify and characterize patients with SLE using routinely collected data from the EHR. Our study shows that using text mining, data from the EHR can be reused in research and quality control.

Use of electronic health record data mining for heart failure subtyping

“…In these studies, Labrosse [ 9 ], Brunekreef [ 11 ] and Mull [ 12 ] used a string character or rule-based text mining algorithm that was very similar to ours and resulted in analogous results. The accuracy of SLE detection was very similar to ours: 71% had a complete agreement in diagnosis in a validation sample of 100 randomly selected patients [ 11 ]. Labrosse et al [ 9 ] manually reviewed all records, and their algorithm was superior to detecting pregnancy (35 of 36) compared to manual EHR assessment (30 of 36).…”

“…In contrast, text mining of several dichotomous disease states has been previously attempted, such as for pregnancy status in a sample of 344 patients [ 9 ], the presence of colorectal cancer in a sample of 1,262,671 patient reports and pathology notes [ 10 ], systematic lupus erythematosus (SLE) in a sample of 4,607 patients [ 11 ], and cardiac implantable device infections in a sample of 19,212 implant procedure patients records [ 12 ]. In these studies, Labrosse [ 9 ], Brunekreef [ 11 ] and Mull [ 12 ] used a string character or rule-based text mining algorithm that was very similar to ours and resulted in analogous results. The accuracy of SLE detection was very similar to ours: 71% had a complete agreement in diagnosis in a validation sample of 100 randomly selected patients [ 11 ].…”

Microarray testing in patients with systemic lupus erythematosus identifies a high prevalence of CpG DNA-binding antibodies

“…However, our text mining algorithm performs very well with a very high sensitivity and specificity in detecting patients with SLE (>90%), which should be adequate in a large cohort like this. 10 Ideally, the results of this study should be validated in patients meeting the classification criteria for SLE, rather than patients with a clinical diagnosis of SLE. Our study did not include the use of immunosuppressive medication in the analysis, although some types of medication, such as belimumab and rituximab, are known to be able to influence levels of autoantibodies.…”

“…Details of the text mining algorithm used in this study have been described in detail. 10 In short, the text mining algorithm searches for pre-defined key words indicating the presence of a certain diagnosis or symptom in the written documents. Multiple diagnoses could be assigned to a single patient at the same time.…”

Development and evaluation of a scalable alternative to chart review for phenotype case adjudication using standardized structured data from electronic health records

“…On the other hand, most of the content in charts is highly redundant and useful information can be buried under duplicated notes [19]. With the advances in data extraction and mining [20][21][22], a growing body of literature uses various natural language processing techniques to extract diagnostic information [23][24][25][26][27][28]. While these models show high flexibility and adaptability, they tend to be disease-specific, which limits their scalability.…”