A convenience sample of 7 distinct population-based state hospital discharge databases, representing diverse geographic areas, were used for this study. Hospital discharge data from Arizona, Colorado, Michigan, New Jersey, New York, Utah, and Washington State were obtained from the State Inpatient Databases (SID), Healthcare Cost and Utilization Project (HCUP).¹⁷ The SID contain nearly all community hospital discharges for the respective states. SID data for 2015 were available to us, having been purchased for another study (data re-use was approved by HCUP). The ICD-10-CM lexicon took effect on October 1, 2015.¹⁸ The SID diagnosis fields contained diagnoses based on the ICD-9-CM lexicon for the first 3 calendar quarters of 2015; the ICD-10-CM lexicon was used for the fourth calendar quarter of 2015. Inpatients under 18 years old were excluded because the FCI was developed for adult populations. All available diagnosis fields for each hospitalization were used to identify comorbidities—the most sensitive approach. The number of diagnosis fields did not vary across the 4 quarters within each state, but did vary across states (Arizona, New Jersey, New York, and Washington State had 25; Colorado and Michigan had 30; Utah had 21). Although this project did not involve individually identifiable human subjects, it was conducted under the parent study approved by the University of Washington Institutional Review Board.

The FCI was originally developed as an unweighted index of 18 self-reported chronic conditions.¹⁰ It was developed to predict functional outcomes in community-based adult populations, rather than the outcomes for which comorbidity indices are typically developed: mortality, hospital length of stay, charges, or costs among inpatient populations.^{1, 2, 19, 20} Many of the chronic conditions in the FCI are not included in other comorbidity indices; they were selected based on their theorized association with functional status, through a process involving literature review and a series of focus groups conducted among patients and health care professionals.¹⁰ The FCI was further developed and tested using a randomly sampled Canadian survey of noninstitutionalized adults, the Canadian Multi Centre Osteoporosis Study (N=9,423) and a U.S. database of patients with spine conditions from the National Spine Network (N=28,348).¹⁰ It is commonly used as an additive index, but a weighted index or a set of indicators for individual comorbidities may perform better in some circumstances.^{10, 21, 22}

Our goal during the development process was to optimize concordance between the ICD-9-CM and ICD-10-CM code lists—minimizing discrepancies in capture of individual comorbidities and FCI counts across the two lexicons. The FCI code lists presented herein were developed by the co-authors. JMS is a registered nurse with a doctorate in health services research and over 20 years of experience with ICD coding for research and clinical purposes; her expertise includes surveillance methodology and use of administrative billing data to develop various prevalence and trend metrics based on ICD and other coding lexicons. SDR is a licensed physical therapist and epidemiologist; he has led several research projects assessing/validating the FCI and comparing the FCI to other risk adjustment tools.

We started with published FCI code lists, including the ICD-9-CM code lists developed by Kumar et al (2016),¹¹ and the ICD-10-AM code list developed by Gabbe et al (2013).¹⁶ We also obtained unpublished ICD-9-CM and ICD-10-CM code lists used by Marcum et al (2018).¹⁴ Where there were discrepancies between these existing lists, we used Groll et al (2005)¹⁰ to guide decisions based on original intent, scope, and FCI category descriptions. We also reviewed the FCI-related literature (including later articles co-authored by Groll^23–25) to check for conceptual/definitional drift or refinement. For the stroke category, we relied heavily on the American Heart Association/American Stroke Association expert consensus document.²⁶ We reviewed Charlson and Elixhauser ICD-9-CM and ICD-10 comorbidity code lists published by Quan et al (2005),⁹ and added codes as indicated. However, the FCI focus on chronic conditions was intentionally maintained; i.e., acute comorbidity codes on the Charlson and Elixhauser lists were not added. During this process, we avoided expanding or modifying category definitions where feasible. We aimed to optimize concordance between codes included on the ICD-9-CM and ICD-10-CM lists, in order to facilitate (1) measurement continuity across the lexicon change and (2) use of data sets involving both lexicons. Using ICD coding and mapping manuals,^27–29 as well as the American Academy of Professional Coders (AAPC) online code converter,³⁰ we translated and back-translated all codes between ICD-9-CM and ICD-10-CM to identify potential discrepancies, and made adjustments to ensure concordance across lexicons.

We used two internal data sets for development purposes, to assess concordance and identify inconsistencies in coding and code translation: (1) adult inpatients in Washington State, dual-coded by hospital coders in 2014 (N=2,351), and (2) a set of Washington State workers’ compensation data for allowed work-related conditions (N=6,528). The latter data set was not technically dual-coded, but included ICD diagnoses using both ICD lexicons, because this subset of workers’ compensation claims had remained open across the lexicon change, and codes in the alternate lexicon were either autofilled using computer algorithms or coded by hand. Neither data set allowed for assessment of the full range of FCI comorbidities. Hearing impairment was not fully coded in the former data set. The latter data set consisted of conditions accepted for workers’ compensation coverage as a direct result of—or related to recovery from—the work injury, and had very low prevalence of FCI comorbidities other than back disease, arthritis, anxiety, and depression. However, among those individual comorbidities having adequate prevalence, kappa values for most individual comorbidities were well above 0.81 (considered almost perfect agreement³¹), and kappa values for the FCI counts were 0.83 (SE 0.01; FCI ranged 0 to 10) for the inpatient data set and 0.85 (SE 0.01; FCI ranged 0 to 3) for the workers’ compensation data set. One notable exception was the angina category, which had a kappa value of only 0.30 (considered fair agreement³¹) in the inpatient data set (there too few cases to test angina using the workers’ compensation data set). This was primarily due to cases that were coded as angina using ICD-9-CM but as congestive heart failure or heart disease using ICD-10-CM (because the ICD-10-CM codes I25.11× and I25.7× included angina together with other heart disease diagnoses). We did not make further code list changes to improve cross-lexicon concordance for angina, because there was no apparent way to improve concordance for the angina category without degrading concordance for the heart disease category, or without violating our decision rules, which included not listing the same diagnosis code in more than one FCI category. Kappa and related statistics are provided for these analyses in Supplemental Table S1 (Supplemental Digital Content 1, http://links.lww.com/MLR/C109).

We have provided our updated ICD diagnosis code lists for the FCI herein (Table 1), which can be modified as needed. Further details, including decision rules and rationale guiding code inclusion/exclusion, are provided for transparency (see Supplemental Table S2, Supplemental Digital Content 1, http://links.lww.com/MLR/C109). Stata code for calculating FCI from ICD diagnoses is available from the first author on request.

The SID do not contain the type of long-term functional outcomes that the FCI was developed to predict. However, the FCI has already been well-validated in that respect.^{3, 10, 15, 16, 23} For purposes of assessing criterion validity spanning the ICD lexicon change, length of stay and discharge disposition were used as proxies for short-term functional status. Though limited, there is some evidence of association between functional status and these short-term outcomes. Need for functional assistance is associated with prolonged length of stay (measured as >7 days in one study,³² and ≥90 days in another³³). Lower inpatient mobility was associated with three-fold higher odds of a non-routine (vs. routine) discharge in a single-hospital study (adjusted OR: 3.1; 95% CI: 2.8 to 3.6),³⁴ and greater loss of function (based on severity of illness subclasses) was strongly and monotonically associated with the same outcome in a national study.³⁵

HCUP calculates and cleans the length of stay uniform variable (LOS), which is generally based on subtracting the admission date from the discharge date. Length of stay was missing for 0.004% of hospital discharges in our sample.

HCUP also calculates and cleans the discharge disposition uniform variable (DISPUNIFORM). We converted this categorical variable to a binary indicator for routine discharge (discharged to home, self-care, or court/law enforcement), vs all other discharges (including transfers to short-term hospitals, skilled nursing facilities, home health care, hospice, discharges against medical advice, and deaths). The DISPUNIFORM categories for missing, invalid, or destination unknown were converted to missing values. The resulting routine discharge indicator was missing for 0.104% of hospital discharges in our sample.

FCI frequency distributions were compared descriptively, by ICD lexicon, calendar quarter (ICD-9-CM during quarters 1–3, and ICD-10-CM during quarter 4), gender, age, and state. Frequencies for individual comorbidities were also compared descriptively by calendar quarter and ICD lexicon. Criterion validity across the two ICD lexicons was assessed using regression models with robust variance estimates—gamma-log regression (generalized linear models with gamma family and log link to account for right skew³⁶) for the length of stay outcome regressed on FCI, and logistic regression for the routine discharge outcome regressed on FCI. Covariates for each regression model included age category, gender, state fixed effects, an indicator for ICD lexicon, and a term for the interaction between ICD lexicon and FCI. We used a standard two-sided alpha of 0.05; however, our main focus was on effect size, since the large sample size (over 5 million hospitalizations) would likely assure statistical significance of unimportant differences. All analyses were conducted using Stata/MP 15.1 for Windows (College Station, Texas).³⁷

During code list development, certain inclusion/exclusion decisions were arbitrary. We have provided our ICD code lists and decision rules for transparency, which can be modified as needed. These code lists were not designed for surveillance of individual comorbidities across the lexicon change; further adjustments to the code lists might be needed for that purpose, particularly for the categories of arthritis, chronic respiratory disease, angina, depression, and back disease. The SID did not contain longer-term functional measures with which to assess criterion validity; however, the FCI has been previously validated in that regard.^{3, 10, 15, 16, 23} Given the absence of better alternatives, length of stay and discharge disposition were used as proxies for short-term functional status. Using the binary version of the discharge disposition variable may potentially have masked important variation. The primary goal of this study was to develop diagnosis code lists that could be used across lexicons, and the available shorter-term outcomes served that purpose, despite low explained variance. We encourage further research to validate these FCI diagnosis code lists using long-term functional outcomes. The FCI has been based variously on patient interviews,²⁵ patient questionnaires,²⁴ chart review,²³ and diagnoses in administrative data.^{3, 11, 16} In future research, it would be useful to compare FCI performance across these varying sources.

The FCI ICD-9-CM and ICD-10-CM code lists provided herein can be used for studies spanning the ICD lexicon change in 2015, or with data sets that include diagnosis codes from both lexicons. Using these code lists, the FCI demonstrated general concordance and similar distribution across lexicons, though there were small discrepancies in prevalence of a few individual comorbidities.