Date Published: July 15, 2019
Publisher: Public Library of Science
Author(s): Michael Simonov, Ugochukwu Ugwuowo, Erica Moreira, Yu Yamamoto, Aditya Biswas, Melissa Martin, Jeffrey Testani, F. Perry Wilson, Maarten W. Taal
Abstract: BackgroundAcute kidney injury (AKI) is an adverse event that carries significant morbidity. Given that interventions after AKI occurrence have poor performance, there is substantial interest in prediction of AKI prior to its diagnosis. However, integration of real-time prognostic modeling into the electronic health record (EHR) has been challenging, as complex models increase the risk of error and complicate deployment. Our goal in this study was to create an implementable predictive model to accurately predict AKI in hospitalized patients and could be easily integrated within an existing EHR system.Methods and findingsWe performed a retrospective analysis looking at data of 169,859 hospitalized adults admitted to one of three study hospitals in the United States (in New Haven and Bridgeport, Connecticut) from December 2012 to February 2016. Demographics, medical comorbidities, hospital procedures, medications, and laboratory data were used to develop a model to predict AKI within 24 hours of a given observation. Outcomes of AKI severity, requirement for renal replacement therapy, and mortality were also measured and predicted. Models were trained using discrete-time logistic regression in a subset of Hospital 1, internally validated in the remainder of Hospital 1, and externally validated in Hospital 2 and Hospital 3. Model performance was assessed via the area under the receiver-operator characteristic (ROC) curve (AUC). The training set cohort contained 60,701 patients, and the internal validation set contained 30,599 patients. External validation data sets contained 43,534 and 35,025 patients. Patients in the overall cohort were generally older (median age ranging from 61 to 68 across hospitals); 44%–49% were male, 16%–20% were black, and 23%–29% were admitted to surgical wards. In the training set and external validation set, 19.1% and 18.9% of patients, respectively, developed AKI. The full model, including all covariates, had good ability to predict imminent AKI for the validation set, sustained AKI, dialysis, and death with AUCs of 0.74 (95% CI 0.73–0.74), 0.77 (95% CI 0.76–0.78), 0.79 (95% CI 0.73–0.85), and 0.69 (95% CI 0.67–0.72), respectively. A simple model using only readily available, time-updated laboratory values had very similar predictive performance to the complete model. The main limitation of this study is that it is observational in nature; thus, we are unable to conclude a causal relationship between covariates and AKI and do not provide an optimal treatment strategy for those predicted to develop AKI.ConclusionsIn this study, we observed that a simple model using readily available laboratory data could be developed to predict imminent AKI with good discrimination. This model may lend itself well to integration into the EHR without sacrificing the performance seen in more complex models.
Partial Text: Among hospitalized patients, acute kidney injury (AKI) is strongly associated with increased costs, length of stay, and mortality [1, 2]. As such, hospital-acquired AKI is being evaluated as a potential quality measure by the Centers for Medicare and Medicaid Services . AKI is diagnosed in relation to a rise in creatinine, but this marker rises late in the course of the syndrome [4, 5]. Real-time prediction of AKI prior to a creatinine increase holds promise to preempt such events through medication adjustment, avoiding nephrotoxins, optimizing hemodynamics, or engaging in other diagnostic or therapeutic procedures, including biomarker measurement .
After exclusion criteria were applied, the cohort contained 60,701 patients in the training set. The validation sets at YNHH, SRH, and BH contained 30,599 patients, 43,534 patients, and 35,025 patients, respectively. Table 1 displays baseline characteristics of patients included in the analysis. While significant differences existed across the hospitals, the cohort was characteristic of a hospitalized population with a median age ranging from 61 to 68 years. Statistically significant testing differences were expected between hospitals given very large sample sizes. Across hospitals, a range of 44%–49% of patients were male, 16%–20% were black, and 23%–29% were admitted to surgical wards.
In this study, we assessed the performance of a predictive model built from EHR data from three US hospitals to predict the onset of AKI within 24 hours. Our complete model, which utilized all potential covariates, displayed moderately good performance for predicting 24-hour AKI (average AUC across hospitals of 0.73) as well as the clinically pertinent outcomes of requirement for renal replacement therapy and mortality. A simpler model utilizing only time-updated laboratory data performed nearly as well as the complete model and maintained its performance across the three hospitals and across the outcomes of sustained AKI and requirement for renal replacement therapy.