What was the biggest change in the structure of icd-10 that did not exist in icd-9 codes?

The analysis used Kentucky statewide inpatient hospital discharge administrative billing data for Kentucky residents treated in Kentucky acute care hospitals between January 1, 2012 and December 31, 2017. Data were provided by the Kentucky Office of Health Policy, Cabinet for Health and Family Services. State data policy requires the removal of personal identifiers from state administrative billing data sets (see Kentucky Revised Statutes 216.2927). Thus, transfers from one hospital to another or readmissions for active treatment of the same injury could not be identified. Data presented here reflect instances of inpatient hospitalizations, rather than distinct injured patients or injuries. A hospital injury data set was extracted based on a principal diagnosis of injury. The selection of injury diagnoses was based on the CDC State Injury Indicators Report: Instructions for Preparing 2015 Data, for ICD-9-CM-coded data (Thomas and Johnson 2017); and CDC’s Proposed Framework for Presenting Injury Data using ICD-10-CM External Cause of Injury Codes, for ICD-10-CM-coded data (Annest et al. 2014).

Measures

This study defined three injury hospitalization measures used as model outcome variables.

1. 1)

   The ECOI completeness was defined as the percentage of records with a principal diagnosis of injury that were supplemented with ECOI codes specifying the mechanism/cause and intent of injury. The ECOI completeness measure was calculated for each month during the study period, forming a time series of 72 observations.

2. 2)

   Injury intent was identified for each injury hospitalization record with an ECOI code (CDC 2011; Annest et al. 2014; CDC/NCHS 2016). Monthly measures “percentage of injury hospitalizations by intent” (unintentional, intentional self-harm, assault, and undermined intent) were calculated.

3. 3)

   Injury mechanism (cause of injury) was identified for each injury hospitalization record with a valid ECOI (CDC 2011; Annest et al. 2014). Monthly crude rate (per 100,000 population) was calculated for each injury mechanism.

Analysis

The transition to ICD-10-CM on October 1, 2015, was a significant policy change (intervention) that could have affected (interrupted) the previously established trend of ECOI completeness as well as the injury hospitalization measures of cause and intent. Segmented regression analysis is one method for statistical modeling of an interrupted time series data and it was used in this study to evaluate the impact of the transition to ICD-10-CM coding on the monitored trend of a specific injury hospitalization measure (Wagner et al. 2002; Bernal et al. 2017). We viewed the ICD-10-CM transition as an intervention that split the time series of an injury hospitalization measure. The trend within each segment can be approximated by a linear regression line, defined by two parameters: level and slope. The level is the value (intercept) of the time series at the beginning of a time interval (segment). The slope is the trend, i.e. the average rate of monthly change during the time segment. Each segment has its own intercept and slope. Segmented regression analysis is used to evaluate simultaneously both the significance of the immediate change in the injury hospitalization measure after the coding transition (immediate intervention effect), and the increase or decrease in the slope of the regression line after the transition in comparison to the trend/slope of the regression line before the transition (change in trend). We used the following model:

$$ {Y}_{\mathrm{t}}={\beta}_0+{\beta_1}^{\ast }{time}_t+{\beta_2}^{\ast } ICD10{CM}_t+{\beta_3}^{\ast } time- after- ICD10{CM}_t+{\varepsilon}_t, $$

where Yt is the injury hospitalization measure in month t, with t taking values from 1 (for January 2012) to 72 (for December 2017).The first segment, before the transition, is represented by the intercept β0 (baseline level of the injury hospitalization measure at time t = 0, December 2011) and the slope β1 (average monthly change in the injury hospitalization measure before the transition to ICD-10-CM). The variable ICD10CM is equal to 0 before October 1, 2015 and equal to 1 for observations on or after October 2015. The variable time-after-ICD10CM measures the time units after the intervention (0 for January 2012 to September 2015), and takes values from 1 to 27 for October 2015 to December 2017). The coefficient β2 represents the immediate effect of the transition, i.e., the change in level (drop or jump) of the injury hospitalization measure immediately after the transition. The coefficient β3 represents the change in the slope (increase or decrease) of the monthly injury hospitalization measure after the transition to ICD-10-CM, compared to the slope β1 for the segment before the transition. Thus, β1 + β3 represents the post-intervention slope of the segment after the transition, i.e., the average percentage increase in the injury hospitalization measure from 1 month to the next after the transition to the ICD-10-CM coding. The error term ɛt represents the random variability not explained by the variables in model (e.g., unaccounted seasonal variations, data quality anomalies).

One assumption of the ordinary least squares regression analysis is that error terms are uncorrelated. When this assumption is violated, statistical tests for the significance of the parameters are not correct. The autoregressive error model can correct for autocorrelation. The general form of the autoregressive error model of order k would be

where ɛt are independently normally distributed with mean zero and variance σ2.

Proc AUTOREG with a BACKSTEP option was used to select the correct order of the autoregressive error model. Since seasonality produces autocorrelation at the seasonal lag, it is recommended that the initial full model has an order equal to or larger than the order of any potential seasonality (SAS/ETS(R) 9.3 User’s Guide 2018; Penfold and Zhang 2013). In our monthly models, we specified NLAG = 12 in the stepwise autoregression. The AUTOREG procedure automatically tests for correlations and estimates the autoregressive parameters. Backward elimination sequentially removes autoregressive parameters not significant at the 0.05 level. Sensitivity analyses were performed using a less restrictive threshold at the 0.1 level (SLSTAY = 0.1). The stepwise autoregressive process in the AUTOREG procedure is performed using the Yule-Walker method. The maximum likelihood estimates are produced after the order of the model is determined from the significance test of the preliminary Yule-Walker estimates (SAS/ETS(R) 9.3 User’s Guide 2018). The Fit Diagnostic panel (proc AUTOREG PLOTS = all) for each model outcome (injury hospitalization measure) was reviewed to evaluate model assumptions and fit, inspecting the residuals plot, the white noise probability plot, the autocorrelation function (ACF) plot, and the partial autocorrelation function (PACF) plot (Chvosta 2009).

Initial inspection of the quality and completeness of the Kentucky injury hospitalization billing records exposed a significant drop (about 10%) in the percentage of the ECOI completeness in April, May, and June of 2013. During this period, ECOI completeness for the hospital with the largest volume of injury cases in Kentucky fell from above 95% to about 50%. The decline in this hospital’s ECOI completeness alone accounted for a drop of over 5% in the total state ECOI completeness. Three other relatively large volume hospitals had between 10% to 20% declines in ECOI completeness during the same period, adding a cumulative 3% to the statewide drop. So, the 10% drop in the statewide ECOI completeness can be explained as a data quality/reporting issue, and the ECOI measurements during April–June, 2013 are outliers. ECOI reporting is neither required for billing reimbursement nor mandated in Kentucky. Therefore, these are plausible observations reflecting the reality of billing data collection, particularly during changes in coder staffing and other transitions. In order to isolate the effect of this quarter’s data quality on the parameter estimates in our models, we included an indicator variable that had a value of 1 for the months April – June, 2013, and values of 0 for all other months. If the parameter estimate for the data quality issue was not significantly different from zero, the indicator variable was excluded from the final model.

For each modeled injury hospitalization measure of interest, we reported the maximum likelihood regression coefficient estimates from the final model. When the backward elimination of autoregressive terms resulted in a final model that contained autoregressive terms, we reported the parameters of interest from the SAS AUTOREG procedure’s final model parameter estimates with “autoregressive parameters assumed given” (see Additional file 1 for an example). We reported only the parameters of interest for time (an important control for overall secular trend in the modeled outcome), ICD10CM (level change in modeled outcome after the transition to ICD-10-CM coding), and time-after-ICD10CM (slope change in trend line after the transition to ICD-10-CM compared to slope of the trend line before the transition).

The estimated trends for each injury hospitalization measure were described visually (Figs. 1, 2 and 3) by the observed values of the modeled measure, estimated trend, and 95% confidence limits for estimated trend (UCLM = and LCLM = options in the OUTPUT statement).

a Percentage of External-Cause-of-Injury (ECOI) completeness in injury hospitalization data, Kentucky resident inpatient hospitalizations, January 2012 – December 2017 (72 monthly observations). b Percentage of ECOI completeness in injury hospitalization data, Kentucky resident inpatient hospitalizations, January 2012– December 2017, (October 2015 observation removed; 71 monthly observations)

Kentucky Resident Injury Hospitalizations by Intent of Injury, January 2012 – December 2017

Trends in injury hospitalization rates per 100,000 population, by pattern of observed change after the transition to ICD-10-CM coding, and by injury mechanism, Kentucky resident inpatient hospitalizations, January 2012 – December 2017: a Significant level change without significant slope change. b Significant slope change without significant level change. c Significant level and slope changes. d No significant changes in level or slope

Formal trend analysis was not performed for “drowning” and “pedestrian, other” mechanisms due to very small counts. A different rationale led to the omission of formal trend analysis for injuries due to the “overexertion” mechanism. Overexertion codes were omitted in the ICD-10-CM version 2015 that came into effect on October 1, 2015, so there was no overexertion ICD-10-CM code between October 1, 2015 and September 30, 2016. The code X50 (overexertion and strenuous or repetitive movements) became available for use on October 1, 2016. The Proposed ICD-10-CM External Cause of Injury Matrix includes injuries coded as W18.4, “slipping, tripping and stumbling without falling”, in the mechanism “overexertion”. Inclusion of these injuries gave the impression of a significant temporary drop in overexertion injuries, whereas in reality, overexertion injuries were not captured with a dedicated code between October 1, 2015 and September 30, 2016.

While immediate changes in the injury trends after the transition to ICD-10-CM were expected and could be explained (in most cases) with the new ICD-10-CM coding guidelines, consistent increases/decreases in long-term injury trends after the transition could indicate true change in the injury incidence. Examining such trends, as a sensitivity analysis, models with different join points for the segmented regression were compared to identify statistically if the initial point for sustained changes in injury hospitalization trends preceded October 2015. Models with different join points were compared based on the maximum likelihood estimates for the Akaike’s Information Criterion (AIC) and the corrected Akaike’s information criterion (AICC) (the model with the lowest index was preferred).

The study was approved by the University of Kentucky Institutional Review Board as part of the Kentucky Injury Surveillance Quality Improvement program.

Page 2

1. Notes: Each model had a significant intercept
2. *p < 0.05
3. **p < 0.001
4. aThe final model included autoregressive parameter(s) significant at the 0.05 level