Drug-Induced Acute Kidney Injury Risk Prediction Models

Background: Acute kidney injury (AKI) risk prediction models can predict AKI with short lead times and excellent model performance. However, these prediction models have not ascertained the etiology of the AKI. Drugs are an important contributor to AKI, and it is difficult to distinguish drug causes from other etiologies. Summary: Clinical adjudication of AKI etiology can reduce misclassification associated with temporal relationships, since expert adjudicators are trained to assess an outcome in a consistent manner using standardized definitions. Drug-induced acute kidney injury (DI-AKI) varies by drug and is heterogeneous in onset and mechanisms, limiting a universal approach to risk prediction and necessitating expert review. DI-AKI models should be constructed using a high-quality prospective dataset to maximize model performance, internal and external validity. Predictor selection and engineering requires careful attention to unique issues arising from interactions such as drug dose and concentrations. Various statistical methods, such as least absolute shrinkage and selection operator regression or advanced machine learning techniques, may be applied to perform feature selection and capture trends and interactions between predictors. Finally, the model should be carefully evaluated by internal and external validation. Key Messages: The development of DI-AKI risk prediction models is needed to identify high-risk patients, identify new risk factors, formulate, and apply preventative strategies. DI-AKI risk prediction models require a well-defined dataset of clinically adjudicated cases to improve model performance, validity, and reduce the risk of misclassification.

© 2022 S. Karger AG, Basel

Introduction

Acute kidney injury (AKI) risk prediction models are evolving with the help of machine learning methods and neural networks enabling the use of electronic health record data to predict AKI with short lead times [1]. However, prediction models to date have not informed on the etiology of the AKI event. Drugs are an important contributor to AKI, and it is difficult to distinguish drug-induced acute kidney injury (DI-AKI) from other etiologies given the heterogeneity of pathophysiological mechanisms and the lack of drug-specific diagnostic biomarkers. Clinical adjudication is one approach to identifying the etiology of AKI through expert consensus. This enables more precise phenotyping of patient subpopulations based on etiology, risk factors, drug exposures, or serum creatinine (SCr) trajectories. The goals of risk prediction models are to enhance clinical care through earlier identification and causality assessment, risk stratify patients, and inform on targeted clinical interventions to improve patient outcomes.

Adjudication of Clinical Events

Adjudication of clinical outcomes and adverse events is employed to protect against detection bias and reduce systematic error, as well as to reduce overall bias and improve statistical power [2]. Consensus criteria, such as KDIGO, have been developed to define and stage AKI severity based on urine output and changes in SCr. Biomarker studies employing clinical adjudication of AKI events have demonstrated similar performance of clinical adjudication to KDIGO criteria for AKI cases with a high rate of agreement [3]. However, classification of AKI etiology or subtypes is complex and lacks standardized definitions or consensus criteria. Relying on clinical adjudication exclusively is not guaranteed to yield good discrimination. In the TRIBE-AKI study, Koyner and colleagues [4] found poor agreement between clinical adjudicators (Fleiss’ kappa 0.046) in differentiating acute tubular necrosis versus prerenal azotemia in patients undergoing cardiac surgery. To improve AKI subtypes identification in future studies, standardized phenotype criteria should be developed, and biomarkers may help better define AKI subtypes [2]. In the DIRECT study, we employed central clinical adjudication of DI-AKI events using a standardized consensus definition of DI-AKI subtypes to reduce AKI etiology misclassification and bias from a cohort study design [5]. We utilized causality assessment tools to categorize the causality relationship between drug and AKI event as definite, probable, possible, and unlikely. We developed an AKI-specific causality assessment process to address competing comorbidities, exposures, and patient risk factors [6]. Clinical adjudication and causality assessment enable more accurate phenotyping of DI-AKI events and the development of more precise risk prediction models. Despite the challenges of DI-AKI adjudication, inter-rater reliability can be improved by using standardized consensus definitions of DI-AKI, capturing the relevant clinical variables for these definitions in a consistent method across sites, and training clinical adjudicators to reduce variability in the adjudication process.

Developing Risk Prediction Model

The development of DI-AKI risk prediction models can help healthcare providers identify high-risk patients and aid them in making preemptive clinical decisions (e.g., discontinuing nephrotoxic drugs, optimizing volume status, monitoring SCr and urine output) to prevent this adverse event. Additionally, these models may identify new DI-AKI risk factors, helping to formulate prevention strategies [7]. DI-AKI varies by drug and is heterogeneous in onset and mechanisms, affected by patient-specific, disease-specific, and process of care factors; thus, a universal DI-AKI risk prediction model is challenging to develop. For these reasons, a serious attempt should be made to construct DI-AKI models using a high-quality prospective dataset to maximize model performance, internal and external validity (Fig. 1). When examining vancomycin-associated AKI models, some predictors (e.g., trough concentrations and concomitant piperacillin/tazobactam) are consistent across all models, while other predictors such as residence in the intensive care unit and surgery are unique to each model, which can be attributed to the quality of the dataset or population-specific factors [8, 9]. Thus, if feasible, we recommend DI-AKI risk prediction models to be drug- and population-specific, for example, a risk prediction of vancomycin-induced AKI in hospitalized adult patients [8].

Fig. 1.

Key elements of drug-induced acute kidney injury risk prediction model development and validation.

/WebMaterial/ShowPic/1455016

Identifying candidate predictors for a DI-AKI risk prediction model is a meticulous process that requires a careful evaluation of relevant clinical studies and guidelines. Predictor selection is usually followed by data preprocessing to enhance the quality of the data. Infrequent categorical variables can be merged to avoid overfitting and increase statistical power. For example, it is reasonable to combine coronary artery bypass surgery or valve replacement surgery – known AKI risk factors – under a categorical variable of cardiac surgery [10]. Additionally, the Agency for Healthcare Research and Quality has developed resources to guide the grouping of clinical variables into clinically meaningful categories. Continuous variables should be evaluated for outliers, multicollinearity, and transformation. A hybrid approach of statistical and clinical evaluation should be adopted to identify and resolve such observations.

Various statistical strategies have been utilized to reduce data dimensionality and avoid overfitting. A priori definition and inclusion of clinically important predictors in univariate selection methods are limited since the risk factors are not well defined and vary between drugs, sub-phenotypes, and sub-populations. Stepwise selection methods are limited when outcomes, such as DI-AKI, have low event rates, since the variable selection is unstable, and the estimated regression coefficients are biased toward overestimating the model performance. The least absolute shrinkage and selection operator (LASSO) limits model overfitting but is not robust to collinearity in the data. For example, drug-specific risk factors (e.g., total daily dose, trough concentration) can be correlated but may be important independent predictors of DI-AKI. LASSO will often select one variable and shrink the coefficients of other variables to zero – this selection can vary even with minor changes in the data [11]. Advanced machine learning techniques such as recurrent neural networks may better capture nonlinear relationships, trends, and interactions between predictors. Using a deep-learning approach, Tomašev and colleagues [1] were able to predict 55.8% of all inpatient episodes of AKI with excellent model performance (area under the receiver-operating characteristic curve 0.921). Using a gradient boosting model, Koyner and colleagues [12] could predict stage 2 AKI within 24–48 h with excellent model performance. The SCr trajectory is a significant covariate in many AKI prediction models. In DI-AKI risk prediction, SCr trajectory and its temporal relationship to the candidate drug exposure play a significant role in adjudicating DI-AKI. Our analysis of the DIRECT study revealed that SCr trends between predefined time points such as hospital admission, start of drug exposure, and AKI onset were important predictors of highly probable DI-AKI cases [6]. Additionally, nephrotoxic drugs-specific factors are also shown to be a significant predictor of AKI [12]. While the risk of nephrotoxicity has been established to vary by pharmacological class, the number of nephrotoxic drugs, treatment length, and intensity have been associated with an increased risk of nephrotoxicity. Accounting for these potential predictors may potentially enhance the performance of DI-AKI risk prediction [13].

Once a model is generated, a careful evaluation must be performed to assess accuracy, discrimination, and calibration. Internal validation is commonly performed by randomly dividing the dataset into a training cohort and an internal validation cohort. The low event rates of DI-AKI and small cohorts limit this approach, and we recommend using more efficient methods such as cross-validation. In 10-fold cross-validation, the model is developed using 90% of the dataset, validated in the remaining 10%, and the process is repeated such that all cases have been included in the train and validation groups at least once. External validation evaluates the model’s reproducibility and generalizability using a new cohort. Despite the importance of external validation, it is estimated that only 5% of all risk prediction models undergo external validation given the resource intensity needed to generate high quality, clinically adjudicated cohorts [14]. After completing model construction, a reporting guideline, such as TRIPOD, should be used to communicate fundamental aspects of risk prediction model development and validation. This enables readers, care providers, and policymakers to judge which models are useful in which situations clearly [15].

Despite improvements in the performance of DI-AKI prediction models, it is important to address existing limitations to enhance future work. First, current DI-AKI prediction models rely on SCr trajectories following treatment with nephrotoxic drugs to identify cases and controls. Currently, differentiating AKI subphenotypes is challenging and could be enhanced by clinically adjudicating cases. Second, routine drug concentration monitoring is not available for all drugs making the incorporation of alternative predictors such as drug treatment duration and drug dosing critical to model performance. Often, such variables are not included due to data availability or technical requirements for data processing. Third, detailed accounting for concomitant AKI risk factors is often lacking from datasets or due to study design limitations.

Conclusions

Advanced machine learning methods have improved risk prediction models for AKI, but these models are limited in ascertaining the AKI etiology. Clinical adjudication offers a standardized approach to reducing bias in outcome ascertainment, but poor agreement among expert adjudicators limits our ascertainment of DI-AKI. Development of risk prediction models for DI-AKI requires a combined approach of expert adjudication and advanced statistical methods to ensure the specificity of cases and optimize therapeutic interventions.

Statement of Ethics

The Rationale and Design of the Genetic Contribution to Drug-Induced Renal Injury (DIRECT) Study was approved by the Institutional Review Board of the University of California San Diego (121651) and registered at clinicaltrials.gov (NCT02159209). The study was conducted in accordance with the Helsinki Declaration, and all participants provided their written informed consent.

Conflict of Interest Statement

The authors have no sponsorship or funding arrangements to disclose relating to this manuscript.

Funding Sources

No funding was received for this study.

Author Contributions

Zaid Yousif and Linda Awdishu contributed to conceptualizing, writing, editing, and revising this paper. The final version of the manuscript was approved by Zaid Yousif and Linda Awdishu.

Data Availability Statement

No public data repository exists for the data of the DIRECT Study. Further inquiries can be directed to the corresponding author.

Copyright: All rights reserved. No part of this publication may be translated into other languages, reproduced or utilized in any form or by any means, electronic or mechanical, including photocopying, recording, microcopying, or by any information storage and retrieval system, without permission in writing from the publisher.
Drug Dosage: The authors and the publisher have exerted every effort to ensure that drug selection and dosage set forth in this text are in accord with current recommendations and practice at the time of publication. However, in view of ongoing research, changes in government regulations, and the constant flow of information relating to drug therapy and drug reactions, the reader is urged to check the package insert for each drug for any changes in indications and dosage and for added warnings and precautions. This is particularly important when the recommended agent is a new and/or infrequently employed drug.
Disclaimer: The statements, opinions and data contained in this publication are solely those of the individual authors and contributors and not of the publishers and the editor(s). The appearance of advertisements or/and product references in the publication is not a warranty, endorsement, or approval of the products or services advertised or of their effectiveness, quality or safety. The publisher and the editor(s) disclaim responsibility for any injury to persons or property resulting from any ideas, methods, instructions or products referred to in the content or advertisements.

留言 (0)

沒有登入
gif