Appendix E1 Data Retrieval Methods by Using Data Disovery and Query Builder and Life Sienes System All demographi and linial data were retrieved from our institutional eletroni medial reord databases by using the Data Disovery and Query Builder searh tool (IBM). Data Disovery and Query Builder is a query tool used to aess data ontained within the Life Sienes System database. Data Disovery and Query Builder and Life Sienes System were reated from a ollaboration pilot projet with IBM to provide researhers aess to linial data in a single entralized database. Mayo Clini Life Sienes System ontains patient demographis, diagnoses, hospital data, laboratory data, flowsheet data, linial notes, and pathology data obtained from multiple linial and hospital soure systems within our medial enter. Data in the Mayo Clini Life Sienes System were aessed with the Data Disovery and Query Builder toolset, onsisting of a Web-based graphial user interfae appliation and programmati appliation programming interfae. Data Variables Variables extrated from the Data Disovery and Query Builder inluded patient demographis (age, sex, rae, weight, seleted omorbidities as defined by International Classifiation of Diseases, Ninth Edition odes), linial status (inpatient or outpatient status), laboratory values (SCr laboratory values), and linial proedural odes defined by urrent proedural terminology oding. Time-assoiated variables (age, weight, omorbidities, linial status, SCr data, urrent proedural terminology oding) were referened to the time and date of CT sanning unless otherwise noted. Patients who underwent a ontrast-enhaned or unenhaned CT san at our institution were identified by using urrent proedural terminology odes of the following sans: unenhaned CT hest (71250), abdomen (74150), and/or pelvis (72192); ontrast-enhaned CT hest (71260, 71270), abdomen (74160, 74170), and/or pelvis (72193, 72194). Patients with preexisting dialysis requirements were exluded if they had a relevant urrent proedural terminology ode entered into their eletroni medial reords at or before the date of san (urrent proedural terminology odes: 99512, 90935, 90937, 90921, 90925, 90945, 90947, 99601, 99602). Patients who underwent a perutaneous angioardiographi study were identified and exluded by using urrent proedural terminology odes 92980 and 92982. Predisposing medial onditions reportedly assoiated with the development of AKI after ontrast material administration were identified by using relevant International Classifiation of Diseases, Ninth Edition odes as previously desribed (diabetes mellitus [250.X], diabeti nephropathy [250.4X], ongestive heart failure [428.X], hroni renal failure [see odes below], and aute renal failure [403.00, 403.01, 404.00, 404.01, 404.02, 404.03, 405.01, 453.3, 539.9, 580, 584, 590.1, 590.2, 590.3, 590.8, 593.81, 866]) (25 27). Two separate groups of International Classifiation of Diseases, Ninth Edition odes for hroni renal failure were used as desribed by Kern et al (25) to enompass both odes for hroni renal failure (hroni renal failure: 403.11, 403.91, 404.12, 404.13, 404.92, 404.93, 585, 586) and odes for hroni renal pathophysiology (pathophysiologi onditions, auses, or diseases assoiated with hroni renal Page 1 of 5
dysfuntion separate from hroni renal failure inluding hypertensive hroni kidney disease, gouty nephropathy, polyysti kidney, and other disorders of the kidneys or ureters) (274.1, 403.10, 403.90, 404.10, 404.11, 404.90, 404.91, 581, 582, 590.0, 593.6, 593.9, 753.12, 753.13, 753.14) (25). Speifi diagnoses of diabetes mellitus, diabeti nephropathy, ongestive heart failure, and hroni renal failure were assoiated with eah san reord if the diagnosis was present prior to sanning or added to the eletroni medial reord up to 30 days after the date of CT san. Speifi diagnoses of aute renal failure were assoiated with eah san reord if the diagnosis was present prior to sanning or on the day of san to avoid identifying patients who had AKI after the CT san. The Charlson omorbidity index was reated to predit 10-year mortality by examining 19 unique omorbidity groups (61). These omorbidities are weighted on the basis of their risk of mortality and summed to generate a sore that an be used as an overall assessment of a patient s health. A Charlson sore was generated from the eletroni medial reord for eah patient by identifying and inorporating the International Classifiation of Diseases, Ninth Edition odes of these 19 omorbidities as previously desribed (28). The omorbidities are myoardial infart (410.X, 411.X, 412), ongestive heart failure (428.X), peripheral vasular disease (443.9, 785.4), dementia (290.X), erebrovasular disease (430, 431, 432.X, 433.X, 434.XX, 435.X, 436, 437.X, 438.XX), hroni lung disease (490, 491, 492.X, 493.X, 494, 495, 496, 500, 501), onnetive tissue disease (710.0, 710.1, 710.4, 714.0, 714.2), pepti uler (531, 532, 533, 534), hroni liver disease (571.0, 571.2, 571.4, 571.5, 571.6, 571.9), hemiplegia (342.X), moderate or severe kidney disease (583.0, 583.1, 583.2, 583.4, 583.6, 583.7), diabetes (250.0, 250.1, 250.2, 250.3, 250.7), diabetes with ompliation (250.4, 250.5, 250.6), solid tumor (140.X-195.X), leukemia (204, 205, 206, 207), lymphoma (200, 202, 203), moderate or severe liver disease (572.2, 572.3, 572.4, 572.8), malignant tumor (172.X), metastati solid tumor (196.X, 197.X, 198.X, 199.X), and HIV (042, 043, 044). All SCr data assoiated with eah reord within a 28-day window of CT sanning (14 days before, 14 days after) were extrated from the eletroni medial reord and sorted by time with respet to the time and date of eah unique CT san reord by using R software (R Foundation for Statistial Computing, Vienna, Austria) on a shared-resoure multinode Beowulfstyle superomputing luster (46). In ases where multiple SCr values were olleted within a 1- day interval, the mean and maximum for eah day were alulated. Baseline renal funtion for eah patient was determined from the maximum SCr level reorded 24 hours prior to eah san event. San reipients were stratified with respet to their presumptive risk for AKI by using baseline SCr level as follows: low risk, SCr level < 1.5 mg/dl; medium risk, SCr level of 1.5 2.0 mg/dl; high risk, SCr level > 2.0 mg/dl (3,29). Patients were also stratified on the basis of baseline estimated glomerular filtration rate, derived from the Modifiation of Diet in Renal Disease equation, by using Kidney Disease Outomes Quality Initiative defined utoffs for hroni kidney disease (51). Propensity Sore Analysis All propensity sore estimates were performed by using the R software pakage. Propensity sores were estimated for eah study partiipant meeting inlusion and exlusion riteria from a logisti regression model of the likelihood of ontrast material exposure on the basis of demographi (age, sex, rae), International Classifiation of Diseases, Ninth Edition defined linial odes (108 odes defining ongestive heart failure, hroni renal failure Page 2 of 5
pathophysiology, diabetes mellitus, diabeti nephropathy, aute renal failure, and additional unique odes used in the Charlson omorbidity sore) and additional linial (baseline SCr level, inpatient status) ovariates. Propensity sore estimates of the full data set, representing one or more sans per patient, were performed by using stratifiation and 1:1 mathing methods found in the MathIt software pakage (R Foundation for Statistial Computing) (47). Propensity sore estimates of the single-san per patient data set were performed by using 1:1 mathing methods, inverse propensity weighting, and weighting by the odds by using the MathIt (R Foundation for Statistial Computing) and Twang (R Foundation for Statistial Computing) software pakages for R (47,48). Stratifiation Stratifiation of the entire data set was performed as an initial means to minimize the bias from ovariate imbalane vis-à-vis omparing subsets of patients with similar propensity sores for more aurate estimates of outomes (37,38). Data from the high-, medium-, and low-risk subgroups of the full data set were separately stratified by propensity sore estimates into deiles by using the sublassifiation routine of MathIt. The improvement in ovariate balane after stratifiation was alulated by using separate onditional logisti regression, onditioned on the sublass deile. The risk of AKI in eah risk subgroup was determined by using logisti regression analysis omparing the inidene of AKI with ontrast material exposure. The overall risk of eah subgroup was alulated by using Cohran-Mantel-Haenszel estimates of risk with onditional logisti regression, onditioned to the strata identity. 1-to-1 Mathing One-to-one mathing of the full and most reent san data sets was performed on the logit of the propensity sore by using the nearest neighbor mathing (greedy-type mathing) routine of the software pakage MathIt (62 64). This tehnique mathes a treated and untreated pair of patients on the basis of propensity sore similarity. Mathes were rejeted if there was no mathing ontrol group within 0.15 standard deviations of a treated patient (aliper width) (62). Nonmathed individuals were disarded from the final data set. Eah mathed pair was assigned a unique pair identifiation reated by using R ode. The improvement in ovariate balane after 1:1 mathing was performed by using onditional logisti regression, onditioned on the pair identifiation. The risk of AKI in eah risk subgroup was determined by using logisti regression analysis omparing the inidene of AKI with ontrast material exposure among mathed individuals. Weighting Methods In addition to 1:1 mathing, propensity sore weighting methods were performed by using Twang and IPTW on the smaller single-san per patient data set to estimate the effet of ontrast material exposure on the inidene of AKI (48,65). Inverse probability of treatment weighting and weighting by the odds were performed on this smaller data set as follows. In the ase of inverse probability of treatment weighting, the weight (w) for patient i was defined Ti 1 Ti as wi, where T i represents the treated subgroup (1 = treated, 0 = ontrol) and P i P i 1 p i represents the estimated propensity sore for eah individual (38,66,67). These weights were stabilized and trimmed to remove onfounding effets from rare patients with very large or Page 3 of 5
unstable inverse propensity sores (68,69). In the ase of weighting by the odds method, the weight, wi Ti 1 Ti, was applied to eah individual. Individuals in the treated group (T i 1 = 1) reeived a weighting of 1, whereas individuals in the ontrol group (T i = 0) were weighted aording to the term (40). The effet of weighting on ovariate balane was assessed by 1 using Twang. The effet of ontrast material exposure on AKI after both weighting methods was assessed by using logisti regression of the weighted outome data. Counterfatual Analysis By using the R programming language, a parsing sript was developed to identify patients who underwent both a ontrast-enhaned and unenhaned CT san (14 days or greater apart) over the 11-year study period. Nearest neighbor 1:1 mathing was performed on patients with two or more sans after parsing, by using the MathIt pakage, to math pairs of ontrast-enhaned and unenhaned CT sans from the same patient that took plae in lose temporal proximity to eah other. Covariate balane of the resulting paired data set was assessed by using onditional logisti regression. Similarly, the risk of ontrast material exposure on the development of AKI was determined by using onditional logisti regression, aounting for these mathed pairs. As the propensity sore of a given patient may be expeted to flutuate in the interval between CT sans beause of hanges in age and/or linial harateristis, adjusted odds ratios were generated by using inverse probability weighted logisti regression to aount for these potential hanges over time. The ausal relation between intravenous ontrast material administration and the development of AKI was assessed by using the MNemar test (49). The paired response for eah patient in the ounterfatual data set (AKI response after ontrast material exposure and AKI response after unenhaned san) was displayed as a 2 2 ontingeny table. The MNemar test interrogates the equivalene, or marginal homogeneity, of the off-diagonal marginal frequenies b and (Table E5). Cases of paired positive outomes (a in Table E5) and paired negative outomes (d in Table E5) have no effet on the result (50). 2 In the ase of the MNemar test, the 2 2 value (df = 1) is derived as follows: 2 0.5, is often applied to the MNemar test to redue the likelihood of overestimation of signifiane when the frequeny of off-diagonal terms are low (70). 2 (49). Yates ontinuity orretion, If marginal homogeneity is present in a ase ontrol type experiment (eg, the null hypothesis is not rejeted), it implies that the outome of interest is unrelated to the treatment or exposure of interest. This finding suggests that the ourrene of the outome of interest is just as likely to our after the ontrol experiment (represented as paired response in Table E5 where the outome ours in the unexposed experiment but not the exposed experiment) as it is Page 4 of 5
to our following exposure to ontrast material (represented as paired response b in Table E5 where the outome ours in the exposed experiment but not the unexposed experiment) (50). Referenes 61. Charlson ME, Pompei P, Ales KL, MaKenzie CR. A new method of lassifying prognosti omorbidity in longitudinal studies: development and validation. J Chroni Dis 1987;40(5):373 383. 62. Rosenbaum PR, Rubin DB. The bias due to inomplete mathing. Biometris 1985;41(1):103 116. 63. Rubin DB. Mathing to remove bias in observational studies. Biometris 1973;29(1):159 183. 64. Rubin DB. The use of mathed sampling and regression adjustment to remove bias in observational studies. Biometris 1973;29(1):185 203. 65. van der Wal WM, Geskus RB. ipw: an R pakage for inverse probability weighting. J Stat Softw 2011;42(13):1 23. 66. Czajka JC, Hirabayashi S, Little R, Rubin DB. Projeting from advane data using propensity modeling. J Bus Eon Stat 1992;10(2):117 131. 67. Robins JM, Hernán MA, Brumbak B. Marginal strutural models and ausal inferene in epidemiology. Epidemiology 2000;11(5):550 560. 68. Potter FJ. The effet of weight trimming on nonlinear survey estimates. Proeedings of the Setion on Survey Researh Methods of Amerian Statistial Assoiation. San Franiso, Calif: Amerian Statistial Assoiation, 1993. 69. Sharfstein DO, Rotnitzky A, Robins JM. Adjusting for non-ignorable drop-out using semiparametri nonresponse models. J Am Stat Asso 1999;94(448):1096 1120. 70. Yates F. Contingeny tables involving small numbers and the χ2 test. Suppl J R Stat So 1934;1(2):217 235. Page 5 of 5