Development of risk models for the prediction of new or worsening acute kidney injury on or during hospital admission: a cohort and nested study

Abstract

BackgroundAcute kidney injury (AKI) is a common clinical problem with significant morbidity and mortality. All hospitalised patients are at risk. AKI is often preventable and reversible; however, the 2009 National Confidential Enquiry into Patient Outcome and Death highlighted systematic failings of identification and management, and recommended risk assessment of all emergency admissions.ObjectivesTo develop three predictive models to stratify the risk of (1) AKI on arrival in hospital; (2) developing AKI during admission; and (3) worsening AKI if already present; and also to (4) develop a clinical algorithm for patients admitted to hospital and explore effective methods of delivery of this information at the point of care.Study designQuantitative methodology (1) to formulate predictive risk models and (2) to validate the models in both our population and a second population. Qualitative methodology to plan clinical decision support system (CDSS) development and effective integration into clinical care.Settings and participantsQuantitative analysis – the study population comprised hospital admissions to three acute hospitals of East Kent Hospitals University NHS Foundation Trust in 2011, excluding maternity and elective admissions. For validation in a second population the study included hospital admissions to Medway NHS Foundation Trust. Qualitative analysis – the sample consisted of six renal consultants (interviews) and six outreach nurses (focus group), with representation from all sites.Data collectionData (comprising age, sex, comorbidities, hospital admission and outpatient history, relevant pathology tests, drug history, baseline creatinine and chronic kidney disease stage, proteinuria, operative procedures and microbiology) were collected from the hospital data warehouse and the pathology and surgical procedure databases.Data analysisQuantitative – both traditional and Bayesian regression methods were used. Traditional methods were performed using ordinal logistic regression with univariable analyses to inform the development of multivariable analyses. Backwards selection was used to retain only statistically significant variables in the final models. The models were validated using actual and predicted probabilities, an area under the receiver operating characteristic (AUROC) curve analysis and the Hosmer–Lemeshow test. Qualitative – content analysis was employed.Main outcome measures(1) A clinical pratice algorithm to guide clinical alerting and risk modeling for AKI in emergency hospital admissions; (2) identification of the key variables that are associated with the risk of AKI; (3) validated risk models for AKI in acute hospital admissions; and (4) a qualitative analysis providing guidance as to the best approach to the implementation of clinical alerting to highlight patients at risk of AKI in hospitals.FindingsQuantitative – we have defined a clinical practice algorithm for risk assessment within the first 24 hours of hospital admission. Bayesian methodology enabled prediction of low risk but could not reliably identify high-risk patients. Traditional methods identified key variables, which predict AKI both on admission and at 72 hours post admission. Validation demonstrated an AUROC curve of 0.75 and 0.68, respectively. Predicting worsening AKI during admission was unsuccessful. Qualitative – analysis of AKI alerting gave valuable insights in terms of user friendliness, information availability, clinical communication and clinical responsibility, and has informed CDSS development.ConclusionsThis study provides valuable evidence of relationships between key variables and AKI. We have developed a clinical algorithm and risk models for risk assessment within the first 24 hours of hospital admission. However, the study has its limitations, and further analysis and testing, including continuous modelling, non-linear modelling and interaction exploration, may further refine the models. The qualitative study has highlighted the complexity regarding the implementation and delivery of alerting systems in clinical practice.FundingThe National Institute for Health Research Health Services and Delivery Research programme.

Development and validation of a prediction model for the probability of responding to placebo in antidepressant trials: a pooled analysis of individual patient data

“…This was an attempt to account for differences in baseline risks and in other unmeasured variables. Thus far, the performance of our models in the derivation and in the internal cross-validation was similar to many prediction models for various physical illnesses such as seizure recurrence,28 recurrence of venous thromboembolism29 and risk of acute kidney injury (Bedford et al , 2016), whose AUC ranged between 0.6528 and 0.75,30 and they often showed poorer external validation than internal validation 31. Therefore, further study is needed to assess the external validity of this model.…”

Does NGAL reduce costs? A cost analysis of urine NGAL (uNGAL) & serum creatinine (sCr) for acute kidney injury (AKI) diagnosis

“…This decision-tree was constructed based on previous literature [10, 11] and qualitative analysis with expert opinion. The decision-tree framework provided the clinical testing pathways.…”

Artificial Intelligence in Acute Kidney Injury Risk Prediction

“…Risk prediction scores for AKI have been reported in several clinical settings, mostly in critical care, surgery, and contrast-induced nephropathy [71][72][73][74][75][76][77]. Still, most AKI cases are reported in general hospital wards, and risk scores in this setting are scarce [78][79][80][81].…”