Earlier this week, I met with my external advisors at Creighton to regroup and come up with a more concrete plan. Having gone through the data myself, I have a better idea of what kind of project we can actually do. Since the initial plan was to model the risk of contracting C. diff given patient-to-patient or patient-to-healthcare provider contact in hospitals and nursing homes, taking antimicrobial stewardship programs into account.
However, the H-CUP NIS database, while rich with data, is limited. Nursing homes are not included, and there is no way to even identify a hospital after 2011, let alone determine if it has a stewardship program.
After some discussion, we came to a consensus on assessing the use of Fecal Microbiota Transplants (FMT) and the effect on readmission rates. This requires inclusion of another database, the Nationwide Readmission Database (NRD).
Note: Aetna provides some good background on C. diff and FMTs here: http://www.aetna.com/cpb/medical/data/800_899/0844.html
The disease C. diff is coded in the DXx columns (DX1, DX2, …, DX25) as 00845, which corresponds to ICD-9-CM code 008.45. Using this code, we can filter our database on any patient that has been diagnosed with C. diff.
Of interest in those results, is what procedure was used. Procedure codes are listed in the PRx columns (PR1, PR2, …, PR15). These correspond to ICD-9-CM, however, there are no ICD-9-CM codes for FMTs. There are Current Procedural Terminology (CPT) and Healthcare Common Procedure Coding System (HCPCS) codes for it though. The HCPCS codes is G0455, “Preparation with instillation of fecal microbiota by any method, including assessment of donor specimen” and the CPT Code is 44705, “Preparation of fecal microbiota for instillation, including assessment of donor specimen”. These are used for billing and insurance purposes. What ARHQ have done is created their own codes that abstract the CPT/HCPCS codes, using their own Clinical Classifications Software for Services and Procedures (CCS-Services and Procedures). This is a lossy encoding, meaning we cannot always get back the actual CPT/HCPCS code given a CCS code.
These are stored in the PRCCSx columns (PRCCS1, PRCCS2, …, PRCCS15). The encoding is given in ranges, so some programming was required to break the ranges into discrete values in order to do the many-to-one mappings.
Some codes begin with a letter, such as our G0455, and some end with a letter. Many have leading zeros that are significant. This must all be taken into account as three separate cases.
ccs.codes <- read.csv('data/formats/2017_ccs_services_procedures.csv', stringsAsFactors=FALSE)
ccs.codes$Code.Range <- gsub("'", "", ccs.codes$Code.Range)
final.df <- data.frame(Code=c(""), CCS=c(""), CCS.Label=c(""))
for (i in 1:dim(ccs.codes)[1]) {
print(paste0("row: ", i))
# Get the current code range in row i
code.range <- ccs.codes$Code.Range[i]
# Split it into two codes
code.range <- unlist(strsplit(code.range, "-"))
tmp.df <- c()
# Look for codes with a letter at the beginning of the code
if (length(grep("^([A-Z]).*", code.range, ignore.case = FALSE, perl = TRUE, value = FALSE)) == 2) {
gsub("^([A-Z]).*", "\1", code.range)
letters <- gsub("^([A-Z]).*", "\\1", code.range)
numbers <- gsub("^[A-Z](.*)", "\\1", code.range)
leading.zeros <- gsub("^(0*).*", "\\1", numbers)
# If there is no range only use the existing code, otherwise use the range
if (numbers[1] == numbers[2]) {
tmp.df <- data.frame(Code=as.character(paste0(letters[1], leading.zeros[1], as.integer(numbers[1]))))
} else {
tmp.df <- data.frame(Code=as.character(paste0(letters, leading.zeros, seq(as.integer(numbers[1])), as.integer(numbers[2]), by=1)))
}
}
# Look for codes with a letter at the end of the code
if (length(grep(".*([A-Z])$", code.range, ignore.case = FALSE, perl = TRUE, value = FALSE)) == 2) {
letters <- gsub(".*([A-Z])$", "\\1", code.range)
numbers <- gsub("(.*)[A-Z]$", "\\1", code.range)
leading.zeros <- gsub("^(0*).*", "\\1", numbers)
# If there is no range only use the existing code, otherwise use the range
if (numbers[1] == numbers[2]) {
tmp.df <- data.frame(Code=as.character(paste0(leading.zeros[1], as.integer(numbers[1]), letters[1])))
} else {
tmp.df <- data.frame(Code=as.character(paste0(leading.zeros, seq(as.integer(numbers[1]), as.integer(numbers[2]), by=1), letters)))
}
}
# Look for codes with no letters
if (length(grep("^([0-9])+$", code.range, ignore.case = FALSE, perl = TRUE, value = FALSE)) == 2) {
numbers <- gsub("^(.*)$", "\\1", code.range)
leading.zeros <- gsub("^(0*).*", "\\1", numbers)
# If there is no range only use the existing code, otherwise use the range
if (numbers[1] == numbers[2]) {
tmp.df <- data.frame(Code=as.character(paste0(leading.zeros[1], as.integer(numbers[1]))))
} else {
tmp.df <- data.frame(Code=as.character(paste0(leading.zeros, seq(as.integer(numbers[1]), as.integer(numbers[2]), by=1))))
}
}
tmp.df$CCS <- as.character(ccs.codes$CCS[i])
tmp.df$CCS.Label <- as.character(ccs.codes$CCS.Label[i])
final.df <- rbind(final.df, tmp.df)
}
final.df <- final.df[-1, ]
write.csv(final.df, "nis-ccs-codes.csv", row.names=FALSE)Finally, we can do a lookup for our codes and we find they are coded as CCS code 95, Other non-OR lower GI therapeutic procedures.
We can now use this in conjunction with our C. diff diagnosis code to make an assumption that the patient was treated with an FMT.
The next major milestone is in obtaining, cleaning, and importing the NRD dataset. There will be some addition overhead there, similar to the NIS dataset. I anticipate that will take about another week. Once the NRD database is up, I can assess the feasibility of this new project goal using the available data.
I also need to keep in mind, the assumption that CCS code 95 corresponds to an FMT is not precise. This will require a little more digging into to validate the assumption.