Premium
Systematic review and meta‐analysis of algorithms used to identify drug‐induced liver injury (DILI) in health record databases
Author(s) -
Tan Eng Hooi,
Low En Xian Sarah,
Dan Yock Young,
Tai Bee Choo
Publication year - 2018
Publication title -
liver international
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.873
H-Index - 110
eISSN - 1478-3231
pISSN - 1478-3223
DOI - 10.1111/liv.13646
Subject(s) - false positive paradox , algorithm , meta analysis , medicine , inclusion and exclusion criteria , systematic review , medical diagnosis , scopus , confidence interval , medline , database , data mining , computer science , machine learning , pathology , alternative medicine , political science , law
Abstract Background & Aims Drug induced liver injury (DILI) is largely underreported, leading to underestimation of its burden. Electronic detection of DILI in healthcare databases shows promise to overcome the issues of spontaneous reporting. The performance of detection algorithms may vary because of inconsistent DILI definition and detection criteria. We performed a systematic review and meta‐analysis to identify the DILI detection criteria used in health record databases and determine the performance characteristics of the detection algorithms. Methods We searched PubMed, EMBASE and Scopus for studies that utilized laboratory threshold criteria to identify DILI cases. Validation studies were included in the meta‐analysis. Data were abstracted using standardized forms and quality was assessed using modified Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS‐2) criteria. We evaluate the performance characteristics of the detection algorithm by obtaining the pooled estimate of the positive predictive value (PPV) assuming a random effects model. Results A total of 29 studies met the inclusion criteria for the systematic review; 25 of these studies (n = 35 948) had PPV estimates for performing the meta‐analysis. The PPV of DILI detection algorithms was low, ranging from 1.0% to 40.2%, with a pooled estimate of 14.6% (95% CI 10.7‐18.9). Algorithms that performed better had prespecified exclusion diagnoses as well as drugs of interest to minimize false positives. Conclusion Algorithm performance varied with different case definitions of DILI attributed to different laboratory threshold criteria, diagnosis codes, and study drugs.