From: Validated risk prediction models for outcomes of acute kidney injury: a systematic review
Chawla et al. [25] | Itenov et al. [38] | James et al. [39] | Lee et al. [40] | |
---|---|---|---|---|
Model development | ||||
Sample of patients | Patients who survive AKI | Patients admitted to the ICU for at least 24 h and with AKI | patients with a prehospitalization eGFR of more than 45 mL/min/1.73m2 and who had survived hospitalization with AKI | Adult (age > 18 years) who developed dialysis-requiring AKI (AKI-D) |
Study design | Prospective cohort study | Prospective cohort study | Prospective cohort study | Retrospective cohort study |
Number of centers | 1 center | 9 academic ICUs | Multicenter (population-based repository) | 21 hospitals |
AKI definition | RIFLE | KDIGO | KDIGO | RRT + SCr > 50% rise |
Derivation cohort sample size | 5351 | 568 | 9973 | 2214 |
Derivation time period | October 1999 - December 2005 | 2006–2010 | April 2004 - March 2014, with follow-up to March 2015 | January 2009 - September 2015 |
The outcome of interest | Risk for progression to CKD stage 4 | Recovery after AKI within 28 days | Progression of AKI to advanced CKD | Recovery after dialysis-requiring AKI within 90 days |
Number of prediction models | Three logistic regression models | Two cause-specific Cox regression models: one for the hazard of recovery and one for death without recovery | Five multivariate logistic regression | Two models: Logistic regression and classification and regression tree (CART) |
Predictor selection method (e.g.full model approach, backward elimination) | Model1: stepwise logistic regression, Model2: based on the most heavily weighted factors from model1, Model3: based on sentinel clinical events | Model1: most likely predictors, Model2: full model | Stepwise backward logistic regression at P < 0.05 with bootstrap selection (1000 samples) | Stepwise logistic regression with bootstrap selection (1000 samples) |
Incidence of outcome | 13.6% entered CKD4 | 15.1% risk of not recovering | 2.7% developed advanced CKD | 59.1% not recovered after AKI-D |
Validation method | ||||
Validation cohort sample (e.g. split sample, bootstrap) | Separate cohort | Separate cohort | Internal (one-third of derivation cohort) and separate cohort | Internal validation (10-fold cross-validation) |
Validation cohort sample size | 11,589 | 766 | 2761 (external cohort) | - |
Validation time period | October 1999 - December 2005 | 1 January 2012–31 December 2013 | June 2004 - March 2012, with a follow-up to March 2013 | January 2009 - September 2015 |
Incidence of outcome | 8.5% entered CKD4 | 10% risk of not recovering | 2.2% developed advanced CKD | 59.1% not recovered after AKI-D |
Performance statistics | c − statistics = 0.81–0.82 | AUROC = 73.1% for predicting recovery | c − statistic = 0.87 | Logistic regression: c − index = 0.645, CART: c − index = 0.61 |
Model performance statistics: calibration | Not reported | The calibration plot used, noted as nicely calibrated | P (slope) = 0.92, 0.88, 0.8, 0.89, 0.67 | The calibration plot used, noted as excellent calibration |
Chen et al. [ 41 ] | He et al. [ 42 ] | Pike et al. [ 44 ] | Huang et al. [ 43 ] | |
Model development | ||||
Sample of patients | Patients diagnosed with cardiac surgery-associated AKI (CSA-AKI) | Patients with sepsis-associated AKI | Critically ill patients receiving RRT with AKI | ICU patients with AKI-3 |
Study design | Prospective cohort study | Prospective cohort study | Prospective cohort study | Prospective cohort study |
Number of centers | 1 center | 1 center | Multicenter | Multicenter (seven ICUs) |
AKI definition | Not mentioned | KDIGO | Not mentioned | KDIGO |
Derivation cohort sample size | 196 | 209 | 1124 | 229 |
Derivation time period | not mentioned | January 2015 - December 2020 | November 2003 - July 2007 | August 2007 - November 2010 |
The outcome of interest | Postoperative AKI requiring RRT or in-hospital death | Predict the occurrence of acute kidney disease (AKD) in patients with sepsis-associated AKI | Renal recovery and mortality for ill patients with AKI requiring RRT at day 60 | Two outcomes: 1) complete recovery and 2) complete or partial recovery at hospital discharge |
Number of prediction models | Five logistic regression models with different combinations of the 3 selected predictors | Three models: Recurrent Neural Network-Long Short-Term Memory (RNN-LSTM), decision trees, and logistic regression | Four logistic regression models (ATN clinical model, reduced ATN model, LASSO model, stepwise-selected model, and parsimonious model) | Multiple Least absolute shrinkage and selection operator (LASSO) models |
Predictor selection method (e.g. full model approach, backward elimination) | LASSO logistic regression and random forests | LASSO | Model1: reduced ATN model, Model2: LASSO, Model3: stepwise logistic regression, Model4: routinely available predictors | Correlation-based feature selection (n = 4) and one feature added based on the literature |
Incidence of outcome | 16.3% | 55.5% | 36.5% | 37.55% (complete recovery) |
Validation method | ||||
Validation cohort sample (e.g. split sample, bootstrap) | Internal validation (bootstrap) and separate cohort | Separate cohort (MIMIC III database) | Internal validation (2-fold split) | Internal validation (stratified 10-fold cross-validation) and a separate cohort |
Validation cohort sample size | 52 | 509 | 562 | 244 |
Validation time period | Not mentioned | 2008–2014 | November 2003 - July 2007 | August 2007 - November 2010 |
Incidence of outcome | 21.1% | 46.4% | - | 33.20% (complete recovery) |
Performance statistics | ROC-AUC = 97.1% | AUROC for LSTM = 1.00 AUROC for decision trees = 0.872 AUROC for logistic regression = 0.717 | Renal recovery using model 4: AUROC = 0.76% | Complete recovery: AUROC = 0.53%, complete or partial recovery: AUROC = 0.61% |
Model performance statistics: calibration | Calibration score assessed by Brier score and HL test and noted as good | The calibration plot used, noted as nicely calibrated | HL: P = 0.08–0.45 | Calibration plot used |