xdm20.train {BART}

A real data example for recur.bart.

Description

A matrix containing a 20% random sample of the training set for a real data example of recurrent events survival analysis. There are 100 patients in the cohort: 50 in the training set and 50 in the testing set. The full data set, xdm.train, can be obtained online at https://www.mcw.edu/-/media/MCW/Departments/Biostatistics/tr064zip.zip There are 488 patients in the full cohort: 235 in the training set and 253 in the testing set. See the Reference below (and the References therein) for more detailed information; a brief synopsis follows.

We explored the hospital admissions for a cohort of patients with diabetes cared for by the Froedtert and Medical College of Wisconsin health network. These patients were identified via their Electronic Health Records (EHR) which include vital signs, diagnoses, procedures, laboratory values, pharmacy orders and billing data. This human subjects research and de-identified data release was approved by the Medical College of Wisconsin and Froedtert Hospital joint Institutional Review Board. To maintain patient privacy, roughly one fourth of patients were randomly sampled for inclusion as well as other de-identification procedures.

We identified likely incident diabetes mellitus type 2 patients by tabulating their first diagnosis code of primary diabetes (ICD-9 codes 250.x0 and 250.x2) in 2006 or 2007, i.e., no such codes were found for these patients prior to 2006 for as far back as each patient's records go which is variable. We restricted the population to adults aged 21 to 90 by 01/01/2008. Among the patients treated in this health system, the vast majority were racially self-identified as either white or black so our inclusion criteria is restricted to these groups. Since our interest is in patients with primary diabetes, we excluded those patients who were diagnosed with either secondary diabetes or gestational diabetes.

For this cohort, we identified every hospital admission between 01/01/2008 and 12/31/2012. For convenience, follow-up begins on 01/01/2008, rather than from each patient's actual incident diagnosis date which varied over the preceding 2 years. Following all patients concurrently allows us to temporally adapt, via our model, for seasonal/epidemic hospital admissions such as the H1N1 influenza outbreak in the US from April to June 2009.

We investigated the following risk factors: gender, race, age, insurance status (commercial, government or other), diabetes therapy (insulin, metformin and/or sulfonylurea), health care charges, relative value units (RVU), vital signs, laboratory values, comorbidity/complication diagnoses and procedures/surgeries (we will refer to vital signs and laboratory values collectively as signs; and comorbidity/complication diagnoses and procedures/surgeries collectively as conditions). In total, we considered 85 covariates of which 82 are external factors as described above and three are temporal factors: time, t, the counting process, N_i(t-), and the sojourn time, v_i(t). Among these potential predictors only gender, race and age are time-independent. The rest are defined as last value carried forward.

For insulin, metformin and sulfonylurea, we only had access to prescription orders (rather than prescription fills) and self-reported current status of prescription therapy during clinic office visits. Since, generally, orders are only required after every three fills, and each fill can be for up to 90 days, we define insulin, metformin and sulfonylurea as binary indicators which are one if there exists an order or current status indication within the prior 270 days; otherwise zero.

Health care charges and relative value units (RVU) are measures related to the services and procedures delivered. However, they are so closely related that recent charges/RVUs are of no practical value in this analysis. For example, just prior to a patient's hospital admission on a non-emergent basis, they often have a series of diagnostic tests and imaging. Similarly, for an emergent admission, the patient is often seen in the emergency department just prior to admission where similar services are conducted. We do not consider these charges/RVUs predictive of an admission because we are interested in identifying preventive opportunities. Therefore, we investigate charges/RVUs that are the sum total of the following moving windows of days prior to any given date: 31 to 90, 91 to 180, 181 to 300.

For many patients, some signs were not available for a given date so they were imputed; similarly, if a sign was not observed within the last 180 days, then it was imputed (except for height which never expires, weight extended to 365 days and body mass index which is a deterministic function of the two). We utilized the Sequential BART missing imputation method. However, instead of creating several imputed data sets, we imputed a new sign at each date when it was missing, i.e., in order to properly address uncertainty within one data set, a new value was imputed for each date that it was missing and never carried forward.

Conditions are binary indicators which are zero until the date of the first coding and then they are one from then on. Based on clinical rationale, we identified 26 conditions (23 comorbidities and 3 procedures/surgeries) which are potential risk factors for a hospital admission many of which are possible complications of diabetes; besides clinical merit, these conditions were chosen since they are present in more than just a few subjects so that they may be informative. Similarly, we employed 15 general conditions which are the Charlson diagnoses and 18 general conditions from the RxRisk adult diagnoses which are defined by prescription orders. Seven conditions are a composite of diagnosis codes and prescription orders.

Usage

data(xdm20.train)

References

Sparapani, Rein, Tarima, Jackson, Meurer (2020). Non-parametric recurrent events analysis with BART and an application to the hospital admissions of patients with diabetes. Biostatistics doi:10.1093/biostatistics/kxy032

See Also

xdm20.test

Examples

data(xdm20.train)
head(xdm20.train[ , 1:10])