Prediction of major postoperative events after non-cardiac surgery for people with kidney failure: derivation and internal validation of risk models

Tyrone G Harrison, Brenda R Hemmelgarn, Matthew T James, Simon Sawhney, Braden J Manns, Marcello Tonelli, Shannon M Ruzycki, Kelly B Zarnke, Todd A Wilson, Deirdre McCaughey, Paul E Ronksley* (Corresponding Author)

*Corresponding author for this work

Research output: Contribution to journalArticlepeer-review

3 Downloads (Pure)


BACKGROUND: People with kidney failure often require surgery and experience worse postoperative outcomes compared to the general population, but existing risk prediction tools have excluded those with kidney failure during development or exhibit poor performance. Our objective was to derive, internally validate, and estimate the clinical utility of risk prediction models for people with kidney failure undergoing non-cardiac surgery.

DESIGN, SETTING, PARTICIPANTS, AND MEASURES: This study involved derivation and internal validation of prognostic risk prediction models using a retrospective, population-based cohort. We identified adults from Alberta, Canada with pre-existing kidney failure (estimated glomerular filtration rate [eGFR] < 15 mL/min/1.73m2 or receipt of maintenance dialysis) undergoing non-cardiac surgery between 2005-2019. Three nested prognostic risk prediction models were assembled using clinical and logistical rationale. Model 1 included age, sex, dialysis modality, surgery type and setting. Model 2 added comorbidities, and Model 3 added preoperative hemoglobin and albumin. Death or major cardiac events (acute myocardial infarction or nonfatal ventricular arrhythmia) within 30 days after surgery were modelled using logistic regression models.

RESULTS: The development cohort included 38,541 surgeries, with 1,204 outcomes (after 3.1% of surgeries); 61% were performed in males, the median age was 64 years (interquartile range [IQR]: 53, 73), and 61% were receiving hemodialysis at the time of surgery. All three internally validated models performed well, with c-statistics ranging from 0.783 (95% Confidence Interval [CI]: 0.770, 0.797) for Model 1 to 0.818 (95%CI: 0.803, 0.826) for Model 3. Calibration slopes and intercepts were excellent for all models, though Models 2 and 3 demonstrated improvement in net reclassification. Decision curve analysis estimated that use of any model to guide perioperative interventions such as cardiac monitoring would result in potential net benefit over default strategies.

CONCLUSIONS: We developed and internally validated three novel models to predict major clinical events for people with kidney failure having surgery. Models including comorbidities and laboratory variables showed improved accuracy of risk stratification and provided the greatest potential net benefit for guiding perioperative decisions. Once externally validated, these models may inform perioperative shared decision making and risk-guided strategies for this population.

Original languageEnglish
Article number49
Number of pages11
JournalBMC Nephrology
Issue number1
Early online date10 Mar 2023
Publication statusPublished - 10 Mar 2023

Bibliographical note

TGH was supported by a Kidney Research Scientist Core Education and National Training Program postdoctoral fellowship (co-sponsored by the Kidney Foundation of Canada and Canadian Institutes of Health Research) and the Clinician Investigator Program at the University of Calgary. BJM is supported by the Svare Chair in Health Economics. MT is supported by the David Freeze Chair in Health Services Research. MTJ was the principal investigator of an investigator-initiated research grant from Amgen, Canada, which is not related to this work. These funding sources had no role in study design, data collection, analysis, reporting, or the decision to submit for publication.


  • Kidney disease
  • Perioperative
  • Surgery
  • Risk prediction
  • outcomes


Dive into the research topics of 'Prediction of major postoperative events after non-cardiac surgery for people with kidney failure: derivation and internal validation of risk models'. Together they form a unique fingerprint.

Cite this