1068 — Improving Signal-to-noise Ratio in Big Data to Identify Multifactorial Risks of Drug-induced Liver injury (DILI)
Lead/Presenter: Ayako Suzuki, COIN - Durham
All Authors: Suzuki A (Durham VA Medical Center HSR&D Center of Innovation)
Tillmann HL (Durham VA Medical Center)
Williams JS (Central Arkansas Veterans Healthcare System)
Austen MA (Central Arkansas Veterans Healthcare System)
Hunt CM (Cooperative Studies Program Epidemiology Center – Durham)
The VHA DILI database includes comprehensive clinical and administrative data from 8.7 million veterans who received 124 drugs exhibiting hepatotoxicity (PPO 15-155). Using big data analytic techniques, we can identify multifactorial DILI risks, health disparities, and"unrealized" associations for even modest effect sizes. However, separating the signal from the noise is challenging. In this pilot, we implemented an electronic algorithm for DILI case/control identification for big data generation while enhancing the signal-to-noise ratio.
We identified over 1,800,000 exposures to amoxicillin/clavulanate (2004-2014) with available laboratory data. The high-risk period for liver injury events was defined as at most the first 90 days after drug initiation or at least drug initiation to 30 days following drug discontinuation, depending on the total length of the exposure. Liver events were defined during high-risk periods by: when median baseline = < upper limit normal (ULN), ALT > = 5xULN or ALP > = 2xULN, otherwise, ALT > = 5x or ALP > = 2x median baseline values. Non-DILI events were excluded using ICD-9 codes or relevant labs. Liver injury frequencies at baseline and the high-risk period assessed the signal-to-noise ratio.
Use of lab LOINC codes yielded an unexpectedly high rate of liver events or "noise" identified in the pre-exposure period. By data-mapping and validating lab and ICD-9 codes, and stratifying by comorbidities, we significantly reduced and characterized baseline noise and increased the signal-to-noise ratio in identifying acute liver injury. The computed DILI frequency revealed significant age-, gender-, race/ethnicity-disparities in the DILI risks.
Exploring the pre-drug-exposure period is a crucial step in assessing noise for DILI-assessment. Rigorous data validation of lab and ICD-9 codes successfully improved the signal to noise ratio in the generated big data and identified novel disparities in DILI risks.
The VA's high rate of polypharmacy is concerning as it increases DILI risk due to drug: drug interactions. There are no strategies currently available to prevent DILI. The implemented algorithm combined with the generated big data will enable us to investigate complex drug: host and drug: drug interactions in DILI. These discoveries will inform DILI preventive strategies and enhance VA drug safety.