Methods
In this retrospective cohort study, we screened all patients admitted to the PICU at SLCH with sTBI, defined by a Glasgow Coma Scale (GCS) score of 8 or less after resuscitation in the Emergency Department [
16,
22]. Due to significant differences in pathophysiology and mortality, patients with GCS score of 3, bilateral fixed and dilated pupils on admission to the Emergency Department, cardiac arrest prior to admission to the pediatric intensive care unit (PICU), abusive head trauma, or gunshot wounds to the head were excluded from the analysis [
23,
24]. As the main purpose of the study was to evaluate the impact of the PNCP approach to sTBI care on hospital cost, patients with early death (<48 h) were excluded from the study [
16,
21]. After exclusions, we analyzed the cost data for 124 patients admitted to the SLCH PICU between July 15, 1999 and January 15, 2012. Of the 124 patients, SLCH discharged 64 patients before September 17, 2005 and these subjects were included in the pre-PNCP implementation group. SLCH discharged another 60 patients on or after September 17, 2005 and they were included in the post-implementation group.
Medical chart and demographic data were merged with the hospital’s cost of care data for the same patient hospitalizations. Inpatient costs were inflation-adjusted to 2012 dollars using the Medical Commodity Consumer Price Index (CPI) and Medical Services CPI [
25]. Upon examination of the data, we excluded a total of 7 subjects from the analysis due to either missing data (1 pre, 0 post), total costs more than 3 standard deviations from the mean (0 pre, 1 post (whose total cost exceeded10 standard deviations above the mean)) [
26‐
29], or who died within the first 48 h after admission (5 pre, 0 post). The final analytic sample contained 117 pediatric patients (58 pre and 59-post PNCP implementation).
A multilevel regression model adjusted for the effects of patients’ demographic and clinical characteristics on total cost, and random effects were incorporated to account for unobserved changes over time (covariates were nested within groupings of years: 1999–2002, 2003–2005, 2006–2008, and 2009–2012, with a similar number of subjects in each). To meet the assumption of normality, we log-transformed the total cost outcome variable [
30‐
32]. Demographic and clinical factors included patient sex, race, age at admission (years), mechanism of injury, lengths of stay (LOS) in days in the pediatric intensive care unit (PICU) and the hospital’s general ward (floor LOS). We represented the severity of TBI and risk of death for each patient with Glasgow Coma Score (GCS) and Pediatric Risk of Mortality (PRISM III) scores, respectively [
5,
33]. The GCS scale ranges from 3 to 15, and a score of 8 or less after resuscitation in the Emergency Department represents sTBI [
5]. The PRISM III score is measured in the first 12 h after admission to the PICU and reliably predicts risk of mortality in critically ill children, with higher scores representing higher risk of mortality [
33]. We merged clinical data with cost data from the hospital’s finance office. To assure patient and facility confidentiality (as approved by the IRB), we converted the cost per patient from actual dollars to the “percent of pre-PNCP patient mean total cost”. For example, a patient whose total cost was exactly equal to the mean of pre-PNCP patients’ costs is coded as 100.0 [
25]. A cost code of 105 signifies that a patient had total costs 5% higher than pre-PNCP average costs. We also analyzed subcategories of cost relevant to TBI care in the PICU for which financial data was available.
Descriptive analysis was performed for the pre-PNCP and post-PNCP groups separately, and statistics are presented in Table
1. Bivariate analysis included chi-squared tests for dichotomous variables, and for continuous variables, we used the Wilcoxon test for medians, and
t-test for means. As in prior cost studies, we used multi-level linear regression to model the effect of the PNCP program’s implementation on the total cost outcome variable (Table
2) and address the random effect of yearly variation [
34]. The analysis of service type use and cost is presented in Table
3. SAS v.9.4 was used for all analyses, with significance level set at 0.05.
Table 1
Descriptive statistics by PNCP implementation status (N = 117)
Sex (male) | 32 (55%) | 33 (56%) | 0.934 |
Race (non-white) | 14 (24%) | 17 (29%) | 0.567 |
Injury mechanism (motor vehicle) | 36 (62.1%) | 41 (69.5%) | 0.397 |
Age at admission (years) |
Mean (SD) | 11.0 (4.7) | 11.9 (4.5) | 0.271 |
Median (IQR) | 12.1 (8.4–14.6) | 12.4 (8.7–16.2) | 0.250 |
--- 0 to 5 | 9 (15.5%) | 8 (13.6%) | 0.601 |
--- 6 to 11 | 19 (32.8%) | 19 (32.2%) | |
--- 12 to 15 | 20 (34.5%) | 16 (27.1%) | |
--- 16 to 18 | 10 (17.2%) | 16 (27.1%) | |
Glasgow Coma Score (GCS)a |
Mean (SD) | 5.1 (1.8) | 4.5 (1.6) | 0.041 |
Median (IQR) | 5.0 (3.0–7.0) | 4.0 (3.0–6.0) | 0.047 |
GCS 3–5 N (%) | 23 (39.7%) | 33 (55.9%) | 0.078 |
GCS >5 N (%) | 35 (60.3%) | 26 (44.1%) | |
Pediatric Risk of Mortality Score (PRISM III) |
Mean (SD) | 5.4 (4.3) | 5.9 (5.8) | 0.557 |
Median (IQR) | 4.5 (2.0–9.0) | 5.0 (2.0–8.0) | 0.928 |
1st Quartile N (%) | 6 (10.3%) | 9 (15.3%) | 0.877 |
2nd Quartile N (%) | 23 (39.7%) | 16 (27.1%) | |
3rd Quartile N (%) | 14 (24.1%) | 22 (37.3%) | |
4th Quartile N (%) | 15 (25.9%) | 12 (20.3%) | |
PICU length of stay (days) |
Mean (SD) | 12.0 (7.9) | 14.7 (9.0) | 0.090 |
Median (IQR) | 12.0 (6.0–15.0) | 14.0 (7.0–22.0) | 0.097 |
1st Quartile N (%) | 10 (17.2%) | 8 (13.6%) | 0.144 |
2nd Quartile N (%) | 18 (31.0%) | 15 (25.4%) | |
3rd Quartile N (%) | 19 (32.8%) | 16 (27.1%) | |
4th Quartile N (%) | 11 (19.0%) | 20 (33.9%) | |
General/Floor length of stay (days) |
Mean (SD) | 33.7 (33.4) | 30.3 (28.7) | 0.549 |
Median (IQR) | 26.0 (7.0–42.0) | 22.0 (8.0–38.0) | 0.664 |
1st Quartile N (%) | 14 (24.1%) | 11 (18.6%) | 0.700 |
2nd Quartile N (%) | 11 (19.0%) | 19 (32.2%) | |
3rd Quartile N (%) | 16 (27.6%) | 15 (25.4%) | |
4th Quartile N (%) | 17 (29.3%) | 14 (23.7%) | |
Overall hospital length of stay (days) |
Mean (SD) | 45.8 (38.3) | 45.0 (34.1) | 0.908 |
Median (IQR) | 34.0 (17.0–56.0) | 37.0 (17.0–58.0) | 0.946 |
Total inpatient cost percentageb | 100 (78.0) | 107 (72.4) | 0.607 |
Mean (SD) | 100.0 (78.0) | 107.2 (72.4) | 0.607 |
Median (IQR) | 67.6 (46.9–119.6) | 88.7 (57.5–132.2) | 0.377 |
Table 2
Total cost regression modela
Post-PNCP Implementation | 0.933 (0.722, 1.205) | 0.594 |
Sex (male) | 0.992 (0.903, 1.090) | 0.869 |
Race (non-white) | 1.007 (0.904, 1.121) | 0.902 |
Age at admission (years) |
Age 0–5 years | (reference) |
Age 6–11 years | 1.061 (0.920, 1.224) | 0.416 |
Age 12–15 years | 1.062 (0.921, 1.223) | 0.407 |
Age 16–18 years | 1.104 (0.949, 1.284) | 0.199 |
Injury mechanism (non-motor vehicle) | 0.963 (0.872, 1.063) | 0.452 |
Glasgow Coma Score (GCS > 5)a | 0.936 (0.853, 1.028) | 0.165 |
Pediatric Risk of Mortality Score (PRISM III) |
2nd Quartile | 1.030 (0.891, 1.190) | 0.691 |
3rd Quartile | 1.016 (0.879, 1.173) | 0.832 |
4th Quartile | 1.026 (0.878, 1.200) | 0.746 |
PICU LOS |
1st Quartile (vs. cost for 2nd quartile LOS) | 0.543 (0.471, 0.627) | <0.001 |
3rd Quartile (vs. cost for 2nd quartile LOS) | 1.413 (1.251, 1.596) | <0.001 |
4th Quartile (vs. cost for 2nd quartile LOS) | 2.091 (1.827, 2.392) | <0.001 |
General/Floor LOS |
1st Quartile N (vs. cost for 2nd quartile LOS) | 0.714 (0.622, 0.820) | <0.001 |
3rd Quartile N (vs. cost for 2nd quartile LOS) | 1.311 (1.157, 1.486) | <0.001 |
4th Quartile N (vs. cost for 2nd quartile LOS) | 2.143 (1.872, 2.454) | <0.001 |
Table 3
Descriptive statistics of services provided before and after PNCP implementation status (N = 117)
Admission CT‡ scans per patient | 0.95/0.22 (55) | 0.97/0.32 (55) | 0.727 | −41% | 0.001* | −40% | 0.001* |
Subsequent CT scans per patient | 2.62/2.43 (51) | 2.24/1.76 (48) | 0.331 | −36% | 0.001* | −46% | 0.018* |
End Tidal CO2 (days monitored) | 7.98/6.65 (52) | 11.37/7.82 (57) | 0.013* | −6% | 0.001* | +47.8% | 0.005* |
ICP Monitoring (days monitored) | 2.59/3.02 (35) | 4.44/5.49 (42) | 0.026* | −8% | 0.002* | +57.4% | 0.059 |
External Ventricular Drains (EVD) (ave.# procedures per patient) | 0.21/0.41 (12) | 0.05/0.22 (3) | 0.013* | (not available) | | (not available) | |
Hyperosmolar Therapy: 3% Hypertonic Saline or Mannitol (doses per patient) | 4.83/5.70 (42) | 8.46/8.26 (52) | 0.007* | +22% | 0.027* | +77.8% | 0.056 |
Pentobarbital (doses per patient) | 0.43/0.80 (17) | 1.66/3.24 (19) | 0.006* | +2325% | 0.001* | +13,638% | 0.005* |
Epinephrine (infusions per patient) | 0.09/0.28 (5) | 0.10/0.36 (5) | 0.796 | +30% | 0.430 | +155.9% | 0.291 |
Norepinephrine (infusions per patient) | (no utilization) | 0.53/1.44 (9) | n/a | (no pre-PNCC utilization) | n/a | (no pre-PNCC utilization) | n/a |
Dopamine (infusions per patient) | 1.48/3.09 (19) | 1.63/2.89 (23) | 0.795 | −14% | 0.567 | +18.8% | 0.810 |
Fentanyl (doses per patient) | 8.17/6.94 (56) | 8.86/5.31 (59) | 0.546 | +129% | 0.001* | +361.7% | <0.001* |
Midazolam (doses per patient) | 7.64/5.85 (57) | 9.95/6.07 (56) | 0.038* | −73% | 0.001* | −34.8% | 0.262 |
Morphine (doses per patient) | 2.55/2.99 (34) | 12.24/12.62 (49) | 0.001* | −28% | 0.129 | +363.5% | 0.020* |
Dexmedetomidine (infusions per patient) | (no utilization) | 2.14/3.46 (29) | n/a | (no pre-PNCC utilization) | n/a | (no pre-PNCC utilization) | n/a |
sTBI-care labs (lab tests per patient)† | 56.83/39.28 (58) | 69.88/42.37 (59) | 0.087 | −6% | 0.023* | +16.0% | 0.272 |
Sodium monitoring (ave.# tests per patient) | 0.45/0.50 (26) | 0.54/0.50 (32) | N/S | (not available) | | (not available) | |
Osmolality monitoring (ave.# tests per patient) | 0.72/0.45 (42) | 0.78/0.42 (46) | N/S | (not available) | | (not available) | |
EEG monitoring (ave.# tests per patient) | 0.22/0.42 (13) | 0.31/0.46 (18) | N/S | (not available) | | (not available) | |
Discussion
Although longer LOS in the PICU and on the floor are associated with increased costs, holding all else constant, the PNCP program did not result in increased costs. Previous findings attributed this increase in LOS to increased survival rates - lives previously lost to sTBI were more likely to be saved, but to require more time in the PICU [
21]. We surmise that the magnitude of the post-implementation decrease in LOS on the floor was enough to counter the post-implementation increased LOS in the PICU (though neither had statistically significant effects).
As expected, supporting care based on the BTF guidelines increased utilization of products and services related to TBI care. As noted in Table
3, there was an increase in intracranial pressure (ICP) monitoring post-PNCP. We acknowledge that length of ICP monitoring is influenced by increased survival, but importantly, it still represents increased cost to the healthcare system. There was also increased utilization of medications used to treat intracranial hypertension (hyperosmolar therapies, pentobarbital, midazolam and morphine). Fentanyl is also used to treat intracranial hypertension, but utilization did not change. Dexmedetomidine is a sedative used to reduce agitation that may lead to intracranial hypertension, but was not commercially available pre-PNCP. The guidelines also recommend optimizing blood pressure to maintain adequate cerebral perfusion pressure, but do not specify the type of medication used to achieve this goal. While Norepinephrine was introduced post-PNCP, likely reflecting practice preferences, overall there was no change in utilization of medications used to maintain adequate blood pressure (Epinephrine and Dopamine).
The guidelines also recommend avoiding hyperventilation, which requires continuous monitoring of exhaled CO2 (End Tidal CO2 monitoring). We found that End Tidal CO2 monitoring utilization increased post-PNCP, suggesting increased compliance with the guidelines. However, retrospectively, we cannot rule out that this increase was simply a consequence of increased length of mechanical ventilation (data on cost of mechanical ventilation was unavailable for too many pre-PNCP patients, precluding inclusion in our analysis). As is the case of length of ICP monitoring, length of End Tidal CO2 monitoring in the ICU may also be influenced by survival.
Finally, there was an increase in the utilization of laboratory studies related to sTBI care (for a complete list, see online supplemental Table
4). This finding, however, was not statistically significant. Table
3 also includes data on computed tomography (CT) imaging. While there was no change in utilization of CT scans, reductions in cost per scan resulted in overall lower costs of CT scans per patient.
Table 4
Descriptions of 21 sTBI-care Labs
Basic Metabolic Panel | |
CBC Auto with Auto Diff | |
CBC Auto without Auto Diff | |
Gases Blood Any Combination | |
Osmolality | |
Gases Blood with O2 Sat | |
Thromboplastin Time Portal | |
Prothrombin Time | |
Sodium Body Fluid | |
Sodium | |
Comp Metabolic Panel | |
Blood Count Hemoglobin | |
Sodium Urine | |
Osmolality Urine | |
Electrolyte Panel | |
Glucose | |
CBC with Manual Diff | |
Blood Gas | |
Cortisol Total | |
Renal Function Panel | |
Hematocrit | |
Our results are encouraging for provider organizations that seek to improve outcomes in the wake of economic constraints. Together with the findings of improved patient outcomes by Pineda et al. and others [
7,
11,
21], this analysis supports implementation of the BTF guidelines in clinical practice. Importantly, the PNCP supported this approach to care without requiring creation of a separate service or intensive care unit, making such approach to care easier to reproduce [
35]. As the global impact of neurological diagnoses in critically ill children and the approaches to pediatric neurocritical care continue to be better defined [
36‐
39] cost of implementation will become easier to assess and compare.
A central strength of our study was its pre/post design and use of inflation-adjusted hospital cost data rather than hospital charges. This addressed a key limitation of previous studies that used only hospital charges to approximate the economic effects of BTF guideline implementation [
7,
11,
15‐
17]. In addition, this is the first research into the utilization changes for specific products and services related to sTBI care after guideline implementation.
This research has limitations. The data were restricted to only the 117 sTBI patients treated at one major academic medical center in the Midwest over 12 years. Actual costs of care also vary greatly between and within different regions of the country, and globally.
To provide some insights into the nature of the cost variations, Table
3 presents the changes in costs per item within the institution over time. It is reasonable to accept these variations in unit cost as reflecting common hospital accounting practice, introduction of new drugs, established drugs coming off-patent, and periodic changes in vendor pricing. However, in combination with utilization changes, unit cost changes can amplify or diminish reductions in total cost per patient or total cost increases (such as increase in both utilization and unit cost for pentobarbital, a medication commonly used to treat patients with sTBI).
Multilevel regression with random effects modeling was adopted to address these variations based on the assumption that the effect of periodic variation is random. Variations in cost are determined by multiple factors including accounting practices for capital equipment and contracts with vendors. Additionally, we assumed that if a service is performed, the hospital always accounts for the cost of that item. However, in our retrospective analysis, it is not possible to identify cases in which zero cost represents lack of utilization versus omission of cost data entry.
Retrospective cohort studies are also limited by changes in clinical practice over time, and it is not possible to control for all factors that may affect both clinical practice and cost. We attempted to mitigate this limitation by incorporating random effects (based on year) into our statistical model, and exploring within-center practice variability and secular trends in this cohort of patients [
21]. It is also important to highlight the large amount of between-center variability in severe TBI care –multicenter studies will be needed to broadly address the economic impact of clinical practice [
40‐
42].
Our analysis excluded patients with early deaths and a patient who was an outlier (cost over 10 standard deviations from the mean). While this approach facilitates comparison with previous studies and may more accurately reflect adherence to the BTF guidelines [
17], it could also introduce selection bias. Some of these limitations may be addressed in larger, prospective studies; although, even with a larger sample, cautious interpretation is advised.
Finally, the PNCP operating costs (i.e., physician, clinical nurse coordinator, data management and overhead or fixed cost allocations) were not included in this analysis; PNCP resources are not exclusively dedicated to TBI patients, making this cost assignment difficult to quantify or reasonably allocate.
There are a number of opportunities for future research in this area. The apparent trade-off between increases in costly PICU days and reductions in less costly general inpatient floor days is ripe for future research. In addition, it would behoove researchers to explore in more detail the directions and magnitudes of cost fluctuations by category of sTBI product and service. Further, a longitudinal study of the cost of care for sTBI patients across the care continuum is needed. Such a study would be able to consider the costs to survivors, families, and society of TBI rehabilitation and disability services. Although longitudinal costs across multiple care providers were beyond the scope of the current study, it is reassuring to see that care based on recommendations from the BTF did not increase the cost of the acute initial hospitalization for these patients. To better understand the extent to which the investment in PNCP care yields cost savings to society, future work is planned to analyze total patient costs over a longer time span, including care after hospital discharge. Incorporating all trauma centers in the state will also enable the comparison of BTF guideline adopters vs. non-adopters.