HSR&D Home » Research » CDP 09-387 – HSR&D Study
Development of an Automated Nephrotoxicity Pharmacosurveillance System (CDA 08-020)
Michael E Matheny, MD MS MPH
Tennessee Valley Healthcare System Nashville Campus, Nashville, TN
Funding Period: September 2009 - August 2015
The mission of the Veterans Health Administration to provide high-quality care to the nation's Veterans is challenged by the medical complexity of the population; however, routinely collected data from the electronic health record provides unique opportunities for knowledge discovery and continuous quality improvement initiatives. For example, acute kidney injury (AKI) among hospitalized patients is associated with significantly elevated inpatient mortality risk and post-discharge morbidity. The incidence of AKI among the general U.S. population ranges from 1-5% in all hospital admissions and up to 20% in intensive care patients. In addition, increasing rates of AKI have been noted in patients with additional comorbidities and particularly those with polypharmacy. The veteran population is more complex than the general population and is at increased risk for AKI.
The objective of this project is to establish the necessary informatics infrastructure and clinical expertise needed to develop prospective surveillance of medication-related AKI in VA. This will require a literature review of the evidence on risk factors for AKI and development of automated data extraction protocols for VA data. The retrospective database will then be evaluated to establish risk adjustment models and expected event rates for AKI. Finally, the release of a new nephrotoxic medication will be simulated in these data and evaluated with my automated surveillance application.
This project represents an opportunity to quantify the magnitude of risk for known factors and identify new risk factors for AKI among the Veteran patient cohort, as well as advance patient safety by allowing identification of high risk inpatients with a risk prediction model. In future work, an automated surveillance system using the products of this research could provide complementary post-marketing medication surveillance to the FDA's adverse event reporting system.
The project is comprised of five studies: (1) Clinical Domain Synthesis, (2) Data
Processing and Validation, (3) Retrospective Evaluation, (4) Risk Prediction Modeling, and (5) Simulated Prospective Monitoring.
This grant will execute a series of objectives in order to pilot a prospective medication surveillance system using the VA informatics infrastructure. Data during 2002-2009 from the VISN 9 regional VA network will be used in this proposal. First, I must address data requirements from routine clinical data including parsing free text, integrating a large variety of data sources and identifying combinations of data elements indicative of clinical diagnoses. Next, I propose a series of hypothesis-driven and knowledge discovery retrospective evaluations of the resulting organized data, including nephrotoxicity synergy from multiple concurrent medications that affect the kidney through a pre-renal mechanism, evaluating the effect of hyperglycemia on AKI when patients are administered nephrotoxic agents and generating a risk prediction model for the development of AKI. Last, I propose to adapt state-of-the-art surveillance methods and an automated application for VA surveillance of AKI using the data infrastructure developed in this application and performing risk adjustment using the developed risk models. Lastly, I will pilot the system with a simulated prospective evaluation in which AKI events related to a medication class are inserted into the retrospective data in order to determine the sensitivity and specificity of the system as a proof of concept.
Specific Aim 1 is a systematic review of risk factors for acute kidney injury, and most of this work has been completed. As of FY2013, we performed a literature update resulting in a total of 3204 documents that met the pubmed search criteria, of which 1514 met the abstract inclusion criteria and underwent full text review. Two independent reviewers performed both the abstract and full text reviews for each record. Because this volume is too large to be published in a single manuscript, we published a sub-analysis AKI following PCI/PTCA. The AKI following PCI review article was completed in Year 3.
Specific Aim 2 is in the NLP processing part of the aim using yTEX (Cindy Brandt Yale Collaboration) i , and we hope to complete the analysis in the coming months. As the processed documents come online, we will then incorporate the NLP products into the association and risk modeling studies.
Specific Aim 3 is the evaluation of acute kidney injury among a retrospective cohort of hospitalized veterans. We have completed the structured data cleaning and analysis in advance of the unstructured data products of Aim 2 for the VISN9 data. Preliminary analysis was presented at the 2012 HSR&D conference as a poster. A total of 42,391 patients among 4 of the 6 sites, of which 8% experienced AKIN Stage 1, 1% Stage 2, and 1% stage III acute kidney injury. Among the preliminary multivariable risk association analysis, key risk factors among inpatient data were low albumin, low hemoglobin, elevated calcium, and preexisting elevated creatinine, and elevated troponin I. Among medication exposures, we found that ACE inhibitors, aminoglycosides, loop diuretics, thiazide diuretics, vancomycin, and anhydrase inhibitors were all associated with acute kidney injury.
Specific Aim 4 was the evaluation of acute kidney injury among a retrospective cohort of hospitalized veterans. A cohort of 1,620,898 patient hospitalizations from 116 Veterans Affairs hospitals between 2003 and 2012 was used. Predictor variables including medications, diagnoses, laboratory values, vital signs, and radiologic studies, with the modeling techniques of logistic regression, lasso regression, and random forests were used to predict hospital acquired AKI stage 1+, stage 2+ and acute dialysis. The models were evaluated with AUCs, IDIs, NRIs and Brier scores using 50 bootstrap iterations of training and test sets. In final models, we report odds ratios, 95% confidence intervals and importance rankings for predictor variables to evaluate their significance. The best performing model was logistic regression with a median AUC of 0.712 (95% CI: 0.712 - 0.712) for stage 1+, 0.705 (95% CI: 0.702 - 0.709) for stage 2+ and 0.819 (95% CI: 0.812 - 0.825) for dialysis. Multiple risk factors were important across all models including pre-admission glomerular filtration rate, sodium, chloride, bicarbonate, and blood urea nitrogen lab values. The manuscript is ready to be submitted.
Specific Aim 5 will be concluded in the following months.
Each phase of the proposal can provide benefits to the VA population. Development of the data processing infrastructure necessary for research can facilitate a wide variety of research endeavors beyond the scope of this application. Retrospective evaluation of the data, in addition to being preparatory to the prospective system, can identify new risk factors for AKI in the VA population as well as characterize the magnitude of risk for known nephrotoxic medications. Such information can be used to inform therapies for patients. The risk prediction models can be used to predict hospitalized patients at high risk for AKI. Establishing a medication surveillance system in VA could identify a number of unknown risks for AKI among VA patients over time, leading to improved overall patient safety. Impact remains to be seen.
External Links for this Project
Dimensions for VADimensions for VA is a web-based tool available to VA staff that enables detailed searches of published research and research projects.
If you have VA-Intranet access, click here for more information vaww.hsrd.research.va.gov/dimensions/
VA staff not currently on the VA network can access Dimensions by registering for an account using their VA email address. Search Dimensions for this project
DRA: Kidney Disorders
MeSH Terms: none