Systematic Review and Meta–analysis of Algorithms Used to Identify Drug–induced Liver Injury (DILI) in Health Record Databases

Eng Hooi Tan; En Xian Sarah Low; Yock Young Dan; Bee Choo Tai

Disclosures

Liver International. 2018;38(4):742-753. 

In This Article

Methods

Search Strategy

The systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta–Analyses (PRISMA) guidelines.[24] The review protocol is registered in the PROSPERO International Prospective Register of systematic reviews, registration number CRD42016036652. We obtained relevant citations from PubMed, EMBASE and Scopus, using a combination of medical subject heading terms and keywords pertaining to DILI in medical record databases (Data S1). Searches were limited to human studies published in English or Chinese. The search was performed since database inception to 17 October 2016. We also hand–searched the references of included studies for additional relevant articles not identified in the database search.

Study Selection

The titles and abstracts retrieved from the search were independently evaluated by two reviewers (a pharmacist and a physician in the department of gastroenterology and hepatology). Any disagreement was resolved and eligible studies were then included in the full text review. A study was included in the systematic review and/or meta–analysis if it met the following criteria: (i) specified laboratory criteria used in DILI detection, (ii) data sources were from computerized/electronic databases, (iii) validated acute liver injury (ALI) with analysis on drug exposures and exclusion of alternative causes of liver injury, (iv) validated cases detected in the algorithm against medical record review (applicable to the meta–analysis only). For this review, we limit the definition of DILI to that caused by synthetic drugs because consumption of herbal and dietary supplements may be inconsistently recorded and more prone to missing information in the EMR systems.[25] Studies which focused on liver injury resulting from herbs, poisoning or overdose were excluded. Animal studies, nonclinical studies, case reports, case series, trial safety analyses and review articles were also excluded.

Data Extraction

Study data were abstracted using standardized forms that recorded the first author, publication year, country, study design, setting, data source, age and gender of study population, number of cases compared against reference standard, algorithm used, and drug(s) of interest. Study quality was assessed using the modified Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS–2) tool,[26,27] excluding the question on generalizability to the UK population because international studies were included in this review. An overall quality score ranging from 0 to 13 was derived by combining scores for reporting quality, generalizability and low risk of bias.

Statistical Analysis

Positive predictive value (PPV) was calculated for all studies, defined as the percentage of confirmed DILI cases from the number of medical records reviewed for cases detected by the algorithm. The pooled estimates of PPV and 95% confidence intervals were generated via a random–effects model using Metaprop,[28] a statistical program implemented to perform meta–analyses of proportions. Summary estimates were generated for prespecified subgroups: (i) algorithms which specified laboratory and/or diagnosis code criteria, (ii) algorithms which specified study drug and (iii) algorithms which validated ALI vs DILI cases. The random–effects model was chosen to account for interstudy heterogeneity. The model was based on the DerSimonian and Laird method,[29] with the estimate of heterogeneity taken from the inverse–variance method. Heterogeneity is quantified using the I[2] statistic, which describes the percentage of total variation across studies because of heterogeneity rather than chance alone.[30] Publication bias was assessed via visual inspection of a funnel plot of standard error of PPV against PPV using the trim and fill method.[31] All statistical analyses were performed using STATA 14.1.

processing....