A clinical, proteomics, and artificial intelligence-driven model to predict acute kidney injury in patients undergoing coronary angiography.

BACKGROUND
Standard measures of kidney function are only modestly useful for accurate prediction of risk for acute kidney injury (AKI).


HYPOTHESIS
Clinical and biomarker data can predict AKI more accurately.


METHODS
Using Luminex xMAP technology, we measured 109 biomarkers in blood from 889 patients prior to undergoing coronary angiography. Procedural AKI was defined as an absolute increase in serum creatinine of ≥0.3 mg/dL, a percentage increase in serum creatinine of ≥50%, or a reduction in urine output (documented oliguria of <0.5 mL/kg per hour for >6 hours) within 7 days after contrast exposure. Clinical and biomarker predictors of AKI were identified using machine learning and a final prognostic model was developed with least absolute shrinkage and selection operator (LASSO).


RESULTS
Forty-three (4.8%) patients developed procedural AKI. Six predictors were present in the final model: four (history of diabetes, blood urea nitrogen to creatinine ratio, C-reactive protein, and osteopontin) had a positive association with AKI risk, while two (CD5 antigen-like and Factor VII) had a negative association with AKI risk. The final model had a cross-validated area under the receiver operating characteristic curve (AUC) of 0.79 for predicting procedural AKI, and an in-sample AUC of 0.82 (P < 0.001). The optimal score cutoff had 77% sensitivity, 75% specificity, and a negative predictive value of 98% for procedural AKI. An elevated score was predictive of procedural AKI in all subjects (odds ratio = 9.87; P < 0.001).


CONCLUSIONS
We describe a clinical and proteomics-supported biomarker model with high accuracy for predicting procedural AKI in patients undergoing coronary angiography.

predict incident AKI and in some cases, earlier than when changes in creatinine or eGFR may occur. [4][5][6] In recent studies, machine learning was employed to develop models that predicted AKI in hospitalized patients with excellent accuracy; 7,8 and similarly, genomic and proteomic characterization of AKI has been undertaken with varying results. [9][10][11] To the best of our knowledge, machine learning for prediction of AKI in patients undergoing coronary angiography has not yet been studied. As such, we hypothesized that a proteomics-based and artificial intelligence-driven biomarker approach together with clinical risk factors would predict procedural AKI risk in patients enrolled in the Catheter Sampled Blood Archive in Cardiovascular Diseases (CASABLANCA) undergoing coronary angiographic procedures with or without interventions for various acute and non-acute indications.

| METHODS
All study procedures were approved by the Partners Healthcare Institutional Review Board and carried out in accordance with the Declaration of Helsinki.
The design of the CASABLANCA (NCT NCT00842868) study has been detailed previously. 12 Briefly, 1251 patients undergoing coronary and/or peripheral angiography with or without intervention between 2008 and 2011 were prospectively enrolled at the Massachusetts General Hospital in Boston, Massachusetts. Patients were referred for angiography for various acute and non-acute indications.
Of the 1251 patients enrolled, we excluded patients who did not undergo a coronary angiogram, patients who had a history of renal replacement therapy, those with missing blood urea nitrogen or creatinine values, and those with an insufficient quantity of sample. This left us with 889 patients undergoing coronary angiography with available blood samples.
After informed consent was obtained, detailed clinical and historical variables were recorded using a standardized case report form at the time of the angiographic procedure. This case report form included more than 100 clinical variables acquired at the time of study entry as well as results of coronary angiography. Angiographic results were based on visual interpretation by the operator, verified through the catheterization report.
Median follow-up was 4 years, with a maximum follow-up of 6 years. Follow-up was complete for all patients. Processes for identification and adjudication of clinical endpoints were as previously described 12 and included review of medical records, as well as phone follow-up with patients and/or managing physicians and was performed by physicians blinded to biomarker concentrations. The Social Security Death Index and/or postings of death announcements were used to confirm vital status. A detailed definition of endpoints for CASABLANCA was previously published. 12 Specific to this analysis, procedural AKI was defined as an abrupt reduction in kidney function with an absolute increase in serum creatinine of more than or equal to 0.3 mg/dL, a percentage increase in serum creatinine of ≥50%, or a reduction in urine output (documented oliguria of <0.5 mL/kg per hour for >6 hours), within 7 days after contrast exposure.
Baseline characteristics between those who developed procedural AKI and those who did not were compared. Dichotomous variables were compared using Fishers exact test, while continuous variables were compared using t test or Wilcox Rank sum test.
A total of 15 mL of blood was obtained immediately before the angiographic procedure through a centrally-placed vascular access sheath. The blood was immediately centrifuged for 15 minutes, serum and plasma aliquoted on ice, and frozen at −80 C until biomarker measurement. The samples for the present study were analyzed after the first freeze-thaw cycle for baseline biomarker values only. Luminex xMAP technology, is a bead-based multiplexed immunoassay system in a microplate format. The multiplexed assays were developed by Myriad RBM at their Austin, Texas facility. Each analyte assay was individually designed in a single assay format. The individual assays were validated according to CLSI Standards and thoroughly tested at the simplex stage before multiplexing. After multiplexing key performance parameters, such as LLOQ, LDD, and precision were established prior to every kit release. During the assay runs, laboratory information management system (LIMS) provided chain of custody and data logging information for samples throughout the testing process. Sample plating was verified by two technicians and run in temperature-controlled lab. Native controls were run in duplicate alongside samples. Standard curves were at the front and back of each plate to minimize between and within run impression. All samples and reagent handling were automated. A minimum of 50 beads were analyzed per analyte and a 8-point standard curve fitting with advanced algorithms ensured accuracy for sample concentrations. Controls followed Westgard rules to monitor unwanted trending. Sample results were manually reviewed before release. The data was backed up on site with long-term off-site storage. We measured 109 biomarkers in blood (Supporting Information, Table S1) from 889 patients undergoing coronary angiographic procedures for various indications.
A complete case analysis was performed; blood urea nitrogen, or creatinine values were missing with some patients (n = 167), so these patients were removed from the analysis. One other patient was removed from the analysis for having an insufficient quantity of sample, leaving 889 samples available for analysis. For any biomarker result that was below the limit of detection, we utilized a standard approach of imputing concentrations 50% below the limit of detection.
To facilitate the machine learning analysis, the concentrations for all proteins underwent the following transformations: (a) they were log-transformed to achieve a normal distribution, (b) outliers were clipped at the value of three times the median absolute deviation, and not statistically significant, it was removed from the panel and the analysis repeated until the predictive contribution of all variables was statistically significant. With our final panel, we evaluated its performance using the MCCV process described above, and we also determined its in-sample performance using a final prognostic model developed on all of the available data with LASSO with logistic regression. A cutoff was determined using the optimal Youdens index.
In all statistical analyses, a two-tailed P-value of <0.05 was considered statistically significant. All analyses were performed using the R statistical computing platform, version 3.4.4.
As expected, those who developed procedural AKI had higher Those who developed procedural AKI had numerically lower concentrations of CD5 antigen-like (3600 vs 3755 pg/mL) compared to those with did not develop procedural AKI (Table 1).
Following our machine learning-driven approach to panel development, six predictors were present in the final model: four (history of diabetes, BUN/creatinine ratio, CRP, and osteopontin) had a positive association with AKI risk; while two (CD5 antigen-like and Factor VII) had a negative association with AKI risk. Using the model-building procedure described above for subsets of variables, the addition of each biomarker provided a statistically significant improvement in the AUC and the likelihood ratio, while decreasing the AIC and the BIC ( Table 2).
The final model had a cross-validated area under the receiver operating characteristic curve (AUC) of 0.79 for predicting procedural AKI, and an in-sample AUC of 0.82 (P < 0.001). The optimal score cutoff had 77% sensitivity, 75% specificity, and a negative predictive value of 98% for procedural AKI (Figure 1). An elevated score was predictive of procedural AKI in all subjects (odds ratio = 9.87; P < 0.001).
In addition, we tested our model in several subgroups and found that in women (n = 358) the AUC = 0.76; in those whose age ≥ 75 years The rationale for our study is based on the fact that AKI following coronary angiographic procedures is associated with significant morbidity and mortality that has potential to alter patient management if predicted early. 13,14 Ability to predict onset of AKI earlier might alter management in efforts toward its prevention, such as alteration of angiography plans (ie, minimizing dye exposure and employing biplane angiography, for example), avoidance of nephrotoxins, or preprocedure hydration. In those at risk for CKD progression because the presence of comorbidities, such as diabetes and HF, interventions might be considered to reduce its incidence including lifestyle changes, better control of such comorbidities, avoidance of nephrotoxins, and consideration of delaying elective angiography plans until such comorbidities are better managed.
Prior work has examined this question, mostly based on clinical variables. Among patients in the Minnesota Registry of Interventional Cardiac Procedures, diabetes, increased age, higher dose and route of contrast administration, HF, hypertension, peri-procedural shock, baseline anemia, post-procedural drop in hematocrit, use of nephrotoxins, volume depletion, increased creatinine kinase-muscle/brain enzyme, and need for cardiac surgery after contrast exposure were associated with increased risk of procedural AKI. 15 Mehran et al developed a simple risk score that included pre-and periprocedural risk factors including hypotension, intra-aortic balloon pump, HF, CKD, diabetes, age > 75 years, anemia, and volume of contrast with good discriminative power (c-statistic 0.67). 4 In another AKI risk prediction model developed by Brown et al, pre-procedural serum creatinine, HF, and diabetes accounted for >75% of the predictive model. 16,17 While BUN and serum creatinine are most often used to predict procedural AKI, they are not very sensitive or specific for the diagnosis of AKI because they are affected by many renal and non-renal factors that are independent of kidney injury or kidney function. 18 As such, several biomarkers and biomarker panels with and without clinical risk factors have been examined to more accurately predict AKI. Our risk prediction model included the BUN/creatinine ratio in addition to clinical and biomarker risk factors to better predict procedural AKI. Given the proximity of collection of pre-and post-procedure samples and the slower rise in creatinine, than BUN, it is understandable why the BUN and ratio of BUN/creatinine was predictive of renal dysfunction than creatinine alone.
Inflammation may play an important role in presence and severity of AKI. CRP is an acute-phase protein of hepatic origin that is a marker of inflammation synthesized in response to factors released by macrophages and adipocytes. 19 CRP has been associated with cardiovascular risk 20 and has also been associated with renal dysfunction. 21 Tang et al demonstrated that elevated serum CRP concentrations were associated with increased serum creatinine and urea concentrations (P < 0.01) in patients with AKI; CRP concentrations subsequently fell after recovery from AKI. 22 In older patients with AKI, CRP was an independent risk factor for mortality. 23 CRP has also been studied for its ability to predict risk for AKI. In a study of 1656 patients undergoing coronary artery bypass grafting, pre-operative CRP concentrations predicted post-operative AKI and mortality; the addition of CRP to an existing risk model improved net reclassification and discrimination. 24 That we found concentrations of CRP as a predictor of procedural AKI is consistent with this body of evidence.
Osteopontin is an extracellular matrix protein and proinflammatory cytokine thought to facilitate the recruitment of monocytes/macrophages and to mediate cytokine secretion in leukocytes. It plays a role in many physiological and pathological processes, including biomineralization, tissue remodeling, and inflammation. 25 It is found mainly in the loop of Henle and distal nephrons in normal kidneys and can be upregulated in all tubular and glomerular segments following kidney damage, and may also have a role in renal repair. 26 In the last several years, the role of osteopontin in the pathogenesis of diabetic nephropathy has been explored. 25 Osteopontin has been reported to be highly expressed in the tubular epithelium of the renal cortex and in glomeruli in rat and mouse models of diabetic nephropathy 27

| CONCLUSIONS
In a typical at-risk population undergoing coronary angiography for various acute and non-acute indications, we describe a clinical and proteomics-supported biomarker model with high accuracy for predicting procedural AKI in patients undergoing coronary angiography.
The ability to predict AKI may allow for earlier interventions in at-risk patients to reduce future AKI risk. We plan to test our risk prediction model in an external validation cohort in the future.

ACKNOWLEDGMENTS
This study was sponsored by a grant from Prevencio, Inc. Dr. Nasrien

CONFLICTS OF INTEREST
Dr. Nasrien E. Ibrahim has received presentation fees from Novartis.