ICD to CCSR

Overview & background

The icd_to_ccsr function returns the CCSR (Clinical Classifications Software Refined) categories for a user-specified set of diagnosis codes.

CCSR provides a grouping of individual ICD-10 diagnosis codes into broader, clinically meaningful disease categories. The original CCSR grouping was developed based on US ICD-10-CM codes. GEMINI developed an algorithm mapping Canadian diagnosis codes (ICD-10-CA) to CCSR categories. The full mapping procedure is described here.

The purpose of this vignette is to provide a practical guide on how to apply the icd_to_ccsr function in different research contexts, and how to further analyze the outputs provided by the function. For more details on the inner workings on the function itself, please refer to the function documentation (?icd_to_ccsr).

When and why to use CCSR

There are currently more than 70,000 unique ICD-10 codes, with additional country-specific variations. ICD-10 codes contain fine-grained diagnosis details that may not be relevant for research purposes. CCSR aggregates individual diagnosis codes into ~540 clinically relevant categories across 22 body systems. A full list of current CCSR categories can be found here.

CCSR facilitates disease-specific healthcare utilization and outcome as well as dimensionality reduction in statistical modeling. CCSR categories are particularly useful when analyzing large, heterogeneous cohorts (e.g., GIM patients). In those cases, grouping patients into broad disease categories can provide important insights into the patient case mix and allows for a meaningful comparison of baseline characteristics and clinical outcomes across different patient sub-populations (see Example applications of CCSR) .

When not to use CCSR:

For research studies that focus on a single, well-defined condition (e.g., COVID-19), it is often better to simply filter the cohort by the specific ICD-10-CA code(s) of interest (e.g., U071 = confirmed COVID-19) instead of using broader disease categories that might result in a loss of information and specificity.
Additionally, some conditions do not have their own CCSR category. For example, there is no CCSR category for delirium. Instead, any diagnosis codes associated with delirium are grouped into CCSR category NVS011 = “Neurocognitive disorders”, together with a large range of other neurocognitive disorders.

The decision on whether to use ICD-10-CA codes vs. CCSR categories should be made on a project-by-project basis and researchers need to carefully consider what approach is most appropriate for their analyses.

Running `icd_to_ccsr()` with default arguments

To run the icd_to_ccsr function, users need to provide 1) a valid database connection and 2) a dxtable input containing all diagnosis codes of interest. Typically, dxtable is created based on the ipdiagnosis table (or a subset thereof) in the GEMINI database, which contains all diagnosis codes assigned to each genc_id.

Here is an example of how to load the relevant data and run the icd_to_ccsr function with default settings:

# Load required libraries
library(RPostgreSQL)
library(DBI)
library(getPass)

# Establish database connection
db <- DBI::dbConnect(drv,
                     dbname = "DB_name",
                     host = "172.XX.XX.XXX",
                     port = 1234,
                     user = getPass("Enter user:"),
                     password = getPass("Enter Password:"))

# query relevant columns of ipdiagnosis table
ipdiagnosis <- dbGetQuery(db, "SELECT genc_id, diagnosis_code, diagnosis_type FROM ipdiagnosis;") 

# Run default version of icd_to_ccsr         
ipdiagnosis_ccsr <- icd_to_ccsr(db, dxtable = ipdiagnosis) 

head(ipdiagnosis_ccsr, 10)

Mock output table of icd_to_ccsr():

genc_id	diagnosis_type	diagnosis_code	diagnosis_code_desc	ccsr_default	ccsr_default_desc	ccsr_1	ccsr_2
1	M	I500	Congestive heart failure	CIR019	Heart failure	CIR019
2	M	J159	Bacterial pneumonia, unspecified	RSP002	Pneumonia (except that caused by tuberculosis)	INF003	RSP002
3	6	F009	Dementia in Alzheimer’s disease, unspecified	NVS011	Neurocognitive disorders	NVS011
4	M	N390	Urinary tract infection, site not specified	GEN004	Urinary tract infections	GEN004
5	M	I214	Acute subendocardial myocardial infarction	CIR009	Acute myocardial infarction	CIR009
6	M	J440	Chronic obstructive pulmonary disease with acute lower respiratory infection	RSP008	Chronic obstructive pulmonary disease and bronchiectasis	RSP008
7	M	I634	Cerebral infarction due to embolism of cerebral arteries	CIR020	Cerebral infarction	CIR020
8	M	N179	Acute renal failure, unspecified	GEN002	Acute and unspecified renal failure	GEN002
9	M	I2510	Atherosclerotic heart disease of native coronary artery	CIR011	Coronary atherosclerosis and other heart disease	CIR011
10	M	A4150	Sepsis due to Escherichia coli [E.coli]	INF002	Septicemia	INF002	INF003

The function returns the GEMINI-derived ICD-to-CCSR mappings. Each row in the output corresponds to a single ICD-10-CA code (diagnosis_code) of a given encounter (genc_id). Each diagnosis code is returned with its description (diagnosis_code_desc) and one or multiple (up to 6) CCSR categories (ccsr_1 to ccsr_6). One CCSR category has been assigned as the default category (ccsr_default), which corresponds to the main disease group and is typically the category of interest for further analyses (see Example applications of CCSR). The output also contains the description of the CCSR default category (ccsr_default_desc).

Note: Although the GEMINI-derived CCSR mappings have been validated by a medical expert, users are advised to spot-check the output to verify that the ICD-10-CA descriptions broadly match the CCSR default descriptions. Please inform the GEMINI team in case you identify any inconsistencies. Moreover, please note that the mappings may be subject to change as we continuously update our database to ensure consistency with the official CCSR tool.

Optional input arguments

`type_mrdx`

default: TRUE

As illustrated in the example output table above, the function by default only returns the most responsible diagnosis (MRDx) code per genc_id. That is, only a single row per encounter is returned, which contains the type-M (or type-6 if present) diagnosis (in line with the CIHI definition of MRDx; see page 13). If users want to obtain CCSR categories for any diagnosis types, the input argument type_mrdx can be set to FALSE, as follows:

ipdiagnosis_ccsr <- icd_to_ccsr(db, ipdiagnosis, `type_mrdx` = FALSE)

In that case, the function will not apply any filtering by diagnosis type, and instead, will return all rows from the original input table. Note that this output contains multiple rows per genc_id, corresponding to the different diagnosis types. For more information on diagnosis type coding, see here.

Mock output table of icd_to_ccsr() with type_mrdx = FALSE:

genc_id	diagnosis_type	diagnosis_code	diagnosis_code_desc	ccsr_default	ccsr_default_desc	ccsr_1	ccsr_2	ccsr_3
1	1	F067	Mild cognitive disorder	MBD013	Miscellaneous mental and behavioral disorders/conditions	MBD013
1	3	I4890	Atrial fibrillation, unspecified	CIR017	Cardiac dysrhythmias	CIR017
1	M	I500	Congestive heart failure	CIR019	Heart failure	CIR019
2	1	C793	Secondary malignant neoplasm of brain and cerebral meninges	NEO070	Secondary malignancies	NEO070
2	3	G4731	Sleep apnoea, central	NVS016	Sleep wake disorders	NVS016
2	M	J159	Bacterial pneumonia, unspecified	RSP002	Pneumonia (except that caused by tuberculosis)	INF003	RSP002
3	1	Z951	Presence of aortocoronary bypass graft	CIR011	Coronary atherosclerosis and other heart disease	CIR011	FAC009
3	6	F009	Dementia in Alzheimer’s disease, unspecified	NVS011	Neurocognitive disorders	NVS011
4	1	E1042	Type 1 diabetes mellitus with autonomic neuropathy	END003	Diabetes mellitus with complication	END003	END004	NVS015
4	M	N390	Urinary tract infection, site not specified	GEN004	Urinary tract infections	GEN004

`unique_mrdx`

default: FALSE

Typically, each encounter should have exactly one MRDx code. However, in certain circumstances, the function may identify encounters with multiple MRDx codes (e.g., two type-M diagnosis codes). This could be due to data quality issues, or it can happen if users combine in-patient and ED diagnoses or create a long-table format when pre-processing their data. By default, the function will show a warning if any encounters with multiple MRDx codes are found, but it will run as usual and return multiple rows for any encounters with more than 1 MRDx. If users want the requirement for a unique MRDx to be strict, they can specify unique_mrdx = TRUE, which will cause the function to terminate if any encounters with multiple MRDx codes are found. In this case, users are advised to carefully check their input table and remove any entries causing the error message. Note that this input argument is only relevant if type_mrdx is set to TRUE, otherwise, it will be ignored.

`replace_invalidpdx`

default = TRUE

Some diagnosis codes will be returned with ccsr_default = "XXX000", which indicates that the diagnosis code “is unacceptable as a principal diagnosis” (invalid PDX) according to US diagnostic coding following Medicare Code Edits guidelines (see here for more details). Since the CCSR default category is meant to reflect the main disease category (~MRDx), the official CCSR tool sets ccsr_default to "XXX000" for any codes that are not valid as a PDX/MRDx.

Many of the external cause diagnosis codes (diagnosis_type = 9) fall into that category as they are not meant to be coded as MRDx, and thus, will be returned with ccsr_default = "XXX000". Note that these codes still have valid ccsr_1 (to ccsr_6) categories:

Example codes with ccsr_default = "XXX000":

genc_id	diagnosis_type	diagnosis_code	diagnosis_code_desc	ccsr_default	ccsr_default_desc	ccsr_1	ccsr_2
11	9	V180	Pedal cyclist injured in noncollision transport accident, driver, nontraffic accident	XXX000	Code is unacceptable as a principal diagnosis PDX (only used for the inpatient default CCSR)	EXT008	EXT020
12	9	W01	Fall on same level from slipping, tripping and stumbling	XXX000	Code is unacceptable as a principal diagnosis PDX (only used for the inpatient default CCSR)	EXT003	EXT020
13	9	Y04	Assault by bodily force	XXX000	Code is unacceptable as a principal diagnosis PDX (only used for the inpatient default CCSR)	EXT016	EXT022
14	9	Y832	Surgical operation with anastomosis, bypass or graft as the cause of abnormal reaction or later complication, without mention of misadventure at the time of the procedure	XXX000	Code is unacceptable as a principal diagnosis PDX (only used for the inpatient default CCSR)	EXT025
15	9	Y848	Other medical procedures as the cause of abnormal reaction or later complication, without mention of misadventure at the time of the procedure	XXX000	Code is unacceptable as a principal diagnosis PDX (only used for the inpatient default CCSR)	EXT025

When type_mrdx = TRUE, invalid PDX categories are typically rare. However, due to country-specific differences in coding and mapping ambiguity for some ICD-10-CA codes, users may find that even some type M/6 (MRDx) codes are returned with CCSR default category "XXX000". This usually only affects a few injury/poisoning codes or codes starting with “Z”, including pregnancy-related, procedural, or health status/personal history codes:

Example MRDx codes with ccsr_default = "XXX000":

genc_id	diagnosis_type	diagnosis_code	diagnosis_code_desc	ccsr_default	ccsr_default_desc	ccsr_1	ccsr_2	ccsr_3	ccsr_4	ccsr_5	ccsr_6
16	M	Z37000	Single live birth, pregnancy resulting from both spontaneous ovulation and conception	XXX000	Code is unacceptable as a principal diagnosis PDX (only used for the inpatient default CCSR)	PRG030
17	M	Z588	Other problems related to physical environment	XXX000	Code is unacceptable as a principal diagnosis PDX (only used for the inpatient default CCSR)	EXT011

The invalid PDX category is in line with coding for insurance purposes, however, it typically is not meaningful in research contexts and can result in a loss of information. For example, in the scenario shown above, users may still want to group the patient with diagnosis_code = "Z37000" into a relevant, pregnancy-related CCSR category. In fact, the code has been assigned a CCSR 1 category of "PRG030" (“Maternal outcome of delivery”), which contains useful information about the encounter.

The function therefore provides an optional input argument replace_invalidpdx to overwrite any "XXX000" values with one of the (valid) CCSR 1-6 categories. In fact, replace_invalidpdx is set to TRUE by default given that for most research projects, the distinction between valid vs. invalid principle diagnosis coding may not be relevant (or if it is, should be determined based on diagnosis_type instead of mapped CCSR category). Additionally, users may specifically be interested in CCSR groups for any diagnosis types (type_mrdx = FALSE), in which case the definition of CCSR default categories as valid vs. invalid PDX categories is not meaningful.

As an example, when replace_invalidpdx is set to TRUE (default), "XXX000" is replaced with "PRG030" for diagnosis code "Z37000":

Example output illustrating replace_invalidpdx = TRUE:

genc_id	diagnosis_type	diagnosis_code	diagnosis_code_desc	ccsr_default	ccsr_default_desc	ccsr_1	ccsr_2	ccsr_3	ccsr_4	ccsr_5	ccsr_6
16	M	Z37000	Single live birth, pregnancy resulting from both spontaneous ovulation and conception	PRG030	Maternal outcome of delivery	PRG030

Note: Invalid PDX CCSR categories are replaced as follows: For any codes that have only been mapped to a single CCSR category, "XXX000" will be replace with the ccsr_1 category. For any codes that have been mapped to multiple CCSR categories, the function chooses the CCSR 1-6 category that is the most common default category among diagnosis codes of the same ICD-10 chapter (same first 3 characters). Otherwise, if this would still result in mapping to "XXX000" (e.g., if all codes of the same chapter have been mapped to "XXX000"), the function defaults to ccsr_1.

Missing vs. unmapped diagnosis codes

Some entries in the output table might have missing CCSR default categories (i.e., is.na(ccsr_default)). This can be due to either 1) missing/empty diagnosis codes or 2) diagnosis codes that have not been mapped to any CCSR categories. Users can distinguish between the two scenarios based on ccsr_default_descr, which will be either 1) "Missing diagnosis code" or 2) "Unmapped".

Missing diagnosis codes

Diagnosis codes that are returned with ccsr_default_desc = "Missing diagnosis code" refer to entries in the dxtable input where the diagnosis_code was NA or NULL, for example:

Mock output table showing missing diagnosis codes:

genc_id	diagnosis_type	diagnosis_code	diagnosis_code_desc	ccsr_default	ccsr_default_desc
18	M	NA	NA	NA	Missing diagnosis code
19	3	NA	NA	NA	Missing diagnosis code
20	1	NA	NA	NA	Missing diagnosis code

This is typically due to data quality issues and should only affect a very small number of encounters.

If type_mrdx is set to TRUE, encounters that do not have any rows containing a type-6 or type-M diagnosis will also be returned with ccsr_default = "Missing diagnosis code" (and diagnosis_code = NA). This is to ensure that the number of rows in the output table corresponds to the number of unique encounters in the provided dxtable input. The function will additionally show a warning message to inform users if any encounters with missing MRDx codes have been found.

Unmapped diagnosis codes

For rows where ccsr_default = "Unmapped", a valid ICD-10-CA diagnosis code exists, however, that code has not been mapped to any CCSR categories (yet). For example:

Mock output table showing unmapped diagnosis codes:

genc_id	diagnosis_type	diagnosis_code	diagnosis_code_desc	ccsr_default_desc
21	4	98633		Unmapped
22	1	U85	Resistance to antineoplastic drugs	Unmapped
23	9	Y466	Other and unspecified antiepileptics causing adverse effect in therapeutic use	Unmapped
24	1	T062	Injuries of nerves involving multiple body regions	Unmapped
25	1	D302	Benign neoplasm of ureter	Unmapped
26	3	U075	Personal history of COVID-19	Unmapped
27	3	E1330	Other specified diabetes mellitus with background retinopathy	Unmapped

Type-4 morphology codes (e.g., row 1 in example output above) will always be returned as "Unmapped" since they are numeric codes based on ICD-0 (oncology) diagnoses, which are not meant to be mapped to CCSR categories.

For any other diagnosis types, codes that are returned as "Unmapped" have simply not been mapped by the GEMINI team yet. We constantly update our database to add new ICD-10-CA codes and strive to provide mappings for as many codes as possible. However, our team prioritizes mapping MRDx diagnosis codes. Therefore, unmapped codes are more likely to occur when users run the icd_to_ccsr() function with type_mrdx = FALSE. If users plan to analyse CCSR categories for non-MRDx codes and find that a high percentage of codes is unmapped, we recommend using our publicly available GEMINI CCSR mapping algorithm, which enables (semi-)automatic ICD-to-CCSR mapping. For more details on how to use the algorithm, please get in touch with the GEMINI team.

Example applications of CCSR

The examples below illustrate a few common applications of CCSR in clinical research.

Patient case mix

Analyzing the frequency of each CCSR default category in large, heterogeneous cohorts can provide important insights into the patient case mix. For example, when running the icd_to_ccsr() function on MRDx codes of a typical GIM cohort, the 5 most common CCSR default categories among GIM patients can be obtained as follows:

library(data.table)

ccsr_summary <- ipdiagnosis_ccsr[, .(N = .N, `% patients` = .N/nrow(ipdiagnosis_ccsr) * 100), 
                                 by = c("ccsr_default","ccsr_default_desc")]

head(ccsr_summary[order(N, decreasing = TRUE)],5)

Mock output table showing top 5 most frequent CCSR default categories in a GIM cohort:

ccsr_default	ccsr_default_desc	N	% patients
CIR019	Heart failure	4775	4.98
RSP002	Pneumonia (except that caused by tuberculosis)	4007	4.23
RSP008	Chronic obstructive pulmonary disease and bronchiectasis	3760	3.85
GEN004	Urinary tract infections	3561	3.51
NVS011	Neurocognitive disorders	2981	3.23

Users could additionally analyse the cumulative percentage of patients across CCSR categories to determine the number of unique disease categories accounting for 50% (or 75%/90%…) of patients. This provides a useful quantitative measure of the breadth of conditions managed at a given hospital/medical ward.

Risk adjustment using CCSR

When comparing clinical outcomes (e.g., mortality or readmission rates) across patients/physicians/hospitals, it is important to consider patients’ baseline characteristics, such as diagnosis group. Risk adjustment could be performed by including CCSR categories as predictors in regression models, or splitting cohorts into different subgroups based on patients’ CCSR default category. Risk adjustment by CCSR could also be used when analyzing quality of care, resource usage, healthcare costs or other indicators.

As an example, see Roberts et al. (2023) performing risk adjustment using CCSR in GEMINI data.

Filtering cohorts by CCSR category

If users want to filter their cohort by a certain diagnosis group, they could choose the relevant CCSR category(s) from this list and then filter by the corresponding CCSR category number(s).

For example, to filter for any genc_ids with Septicemia, the relevant CCSR category is INF002. Users have 2 options:

Filter encounters where INF002 is the CCSR default category:

# option 1: filter by CCSR default category only
ipdiagnosis_ccsr <- ipdiagnosis_ccsr[ccsr_default=="INF002",]

Filter encounters where INF002 is among any of the CCSR 1-6 categories (recommended):

# option 2: filter by CCSR 1-6 (includes default category)
ipdiagnosis_ccsr <- ipdiagnosis_ccsr[
  rowSums(ipdiagnosis_ccsr[,c("ccsr_1","ccsr_2","ccsr_3","ccsr_4","ccsr_5","ccsr_6")] == "INF002", 
          na.rm = TRUE) > 0,]

The difference between the two approaches is that the CCSR default category is meant as the main disease category. Focusing on that single category is usually only relevant if users need to ensure that each diagnosis code has exactly one corresponding CCSR category. For example, in the analyses in the previous section, patients are grouped according to their default CCSR category because we don’t want to double-count patients whose MRDx code has been mapped to more than 1 CCSR category. By contrast, when filtering for patients who fit a certain CCSR criterion, a 1:1 mapping between diagnosis code and CCSR categories is usually not required. Instead, we likely want to capture all patients who have the CCSR category of interest among any of the mapped CCSR 1-6 categories.

In addition to that, the official CCSR tool assigns default categories based on hierarchical rules, i.e., some CCSR categories are less likely to be a default category than others. For example, for diagnosis codes related to diabetes, the distinction between diabetes with vs. without complications (CCSR END003 vs. END002) takes priority over the distinction between type 1 vs. type 2 diabetes (CCSR END004 vs. END005). As a result, if researchers want to filter for encounters with type-1 diabetes, searching by CCSR default category alone would result in many relevant patients being missed because most diabetes patients have CCSR default categories END002 or END003 (instead of END004/END005). Therefore, the general recommendation when filtering cohorts by CCSR category is to use all CCSR 1-6 categories (option 2 shown above), instead of filtering by CCSR default alone.

Additional notes on filtering by CCSR category:

When investigating specific, well-defined diseases, it may be more appropriate to filter by individual ICD-10-CA codes instead of relying on broad CCSR categories.
If filtering by CCSR category is appropriate, users should still verify that the individual ICD-10-CA codes that are included are relevant for the research question.
It is recommended to always filter by CCSR category number ("INF002") instead of the description ("Septicemia") as the category descriptions are subject to small changes with newer releases of the official CCSR tool.

Additional considerations

Initial vs. subsequent encounters

For some diagnoses (e.g., injuries or poisoning), the US ICD-10 codes contain a suffix that differentiates between initial vs. subsequent encounters. As a result, some CCSR categories make that same distinction. For example: INJ018 = “Crushing injury, initial encounter”, whereas INJ055 = “Crushing injury, subsequent encounter”. However, Canadian ICD-10 codes do not differentiate between initial vs. subsequent encounters. Therefore, ICD-10-CA codes were mapped to “initial encounter” by default. If researchers are interested in analyzing patients’ medical history, they should not rely on the CCSR description to differentiate between initial vs. subsequent encounters, and instead, should look at the encounter history of a given patient within the database. More generally, it means that any CCSR description of “initial” or “subsequent” encounter in the GEMINI database should be ignored (i.e., INJ018 = INJ055 = “Crushing injury”).