Variation in NICU Head CT Utilization Among U.S. Children’s Hospitals

We performed a multicenter, retrospective cohort study using the Pediatric Health Information System (PHIS) database, which contains administrative billing data from 47 tertiary care children’s hospitals in the United States.²⁰  Data are housed and maintained by the Children’s Hospital Association (CHA; Lenexa, KS), which assures data quality in conjunction with participating hospitals. For each hospital encounter, PHIS captures patient demographic variables, International Classification of Diseases, 9th and 10th Revision, Clinical Modification (ICD-9 and 10-CM) diagnosis and procedure codes, and date-stamped billing data. For this study, we included 31 hospitals in the final cohort after excluding 16 hospitals that did not have data available for the entire 12-year study period. The Institutional Review Board of Boston Children’s Hospital determined the study was not human subjects research and was therefore exempt from review and informed consent.

We included patients admitted to the NICU and discharged between January 1, 2010 and December 31, 2021. We excluded patients admitted from the emergency department.

The primary outcome variable was head CT utilization. We used PHIS imaging billing codes to identify encounters billed for head CT, including both contrast and noncontrast head CTs. Only 0.6% of admitted infants had more than 1 head CT within the same encounter, therefore we defined head CT rate as the proportion of admissions that included at least 1 head CT. A sensitivity analysis in which only hospitals with <10% inborn patients (ie, ≥90% outborn) in 2021 was conducted to address differences in proportion of transferred patients across hospitals. Outborn and inborn designations were made using the source of admission indicator in PHIS. Five hospitals were excluded because ≥10% of patients were inborn. Further, to address the possibility that infants may have received a head CT at an outside hospital before transfer to a PHIS-participating hospital, we performed an additional sensitivity analysis examining head CT rates restricting to infants admitted at ≤7 days of life and to infants >7 days of life. Secondary outcomes included the utilization of head MRI (including both contrast and noncontrast) and head ultrasound, also identified through billing codes.

For the analysis of hospital-level variation in 2021, we adjusted for patient-level demographic and clinical characteristics potentially associated with head CT utilization: age at admission (<1 month, 1–2 months, 3–5 months, 6–11 months, 12+ months), sex, race and ethnicity (Non-Hispanic white, Non-Hispanic Black, Hispanic, Asian, multiple races or ethnicities, other, and unknown),²¹  insurance type (public, private, other, or unknown), indicator variables for each of 12 complex chronic condition (CCC) categories,²²  and diagnoses and procedures associated with head CT imaging.

Diagnoses and procedures associated with head CT were determined by examining the ICD-9 and 10-CM diagnosis codes for each patient who underwent head CT within the 2010 to 2021 cohort. Each ICD-9 and 10-CM code was categorized as possible diagnoses and procedures for head CT (yes or no), with codes subsequently divided into 1 of 9 possible imaging diagnoses and procedure categories: trauma; stroke; VP shunt; seizure and encephalopathy; neoplasm; infection; abnormal neurologic finding; other ear, nose, and throat (ENT) or cranial or facial diagnosis; and other disorders of the nervous system. Associated diagnostic and procedure categories (potential indications) for head CT were identified by study investigators after examining frequency of ICD-9 and 10-CM codes associated with patients undergoing head CT.

The large sample size of the study provided high statistical power to detect small differences, so we evaluated associations between head CT utilization and patient characteristics with standardized differences (SD), using thresholds of 0.40 < SD ≤ 0.80 and SD > 0.8 to indicate moderate and strong associations, respectively. The standardized difference has the benefit of comparing the difference in means in units of the pooled SD, while not being affected by sample size.²³  Standardized differences are particularly helpful when sample sizes are so large that small differences may result in very small P values but do not indicate meaningful differences between groups.

We used mixed effects logistic regression to estimate the overall head CT utilization rate for each of the 12 years in the study period. The model included random intercepts and random coefficients for each year by hospital to account for hospital-level effects. To assess trends over time, we used segmented linear regression to model the estimated annual head CT rate versus continuous year.²⁴  Head MRI utilization and head ultrasound utilization were modeled similarly. Additionally, to further assess for trends in individual hospital imaging rates over the study period, individual hospital rates of head CT, head MRI, and head ultrasound (US) imaging (as a proportion of hospital NICU admissions) were determined for each year within the study period. Further, we conducted a sensitivity analysis in which only hospitals with <10% of inborn patients in 2021 (ie, ≥90% outborn patients) were included to address differences in proportion of transferred patients across hospitals. Within this sensitivity analysis, we used segmented linear regression to model the estimated annual head CT rate versus continuous year and tested for a change in intercept and slope in 2015, the midpoint in our study period.

For the analysis of hospital-level variation in infant inpatient head CT utilization in 2021, we adjusted for patient demographic and clinical characteristics potentially associated with CT utilization: age, sex, race and ethnicity,²¹  insurance type, separate indicator variables for each of 12 complex chronic condition categories,²²  and separate indicator variables for each of the 9 diagnoses and procedures associated with head CT. We used logistic regression to estimate the multivariable associations between these variables and head CT utilization. We calculated the expected CT utilization rate at each hospital by taking the average of the predicted probabilities across its cases and then divided the observed rate at each hospital by the expected rate to generate O:E (observed:expected) ratios.²⁵  We calculated adjusted head CT rates by multiplying each hospital’s O:E ratio by the overall head CT rate in the entire 2021 cohort.

To understand how each diagnosis and/or procedure associated with head imaging may have contributed to variation in utilization, we calculated head CT rate for each diagnosis and procedure (proportion of admissions with given imaging diagnoses and procedure that used head CT) overall and on an individual hospital basis. Patients with more than 1 potential diagnosis and/or procedure for head CT were included in multiple imaging diagnoses and procedure categories. Interquartile ranges (IQR) were used to assess degree of hospital-level variation. To limit effect of random chance, only hospitals with 20 or greater admissions within a given imaging diagnoses or procedure category were included for analysis. We defined high priority potential diagnoses and procedures for practice standardization as those with the greatest contributions to head CT utilization (ie, top quartile of diagnoses and procedures) accompanied by wide variation in head CT utilization (ie, IQR > 10%). All analyses were performed with SAS System for Windows v9.4 (Cary, NC) and statistical significance was achieved with a 2-sided P value < .05.

Between 2010 and 2021, there were a total of 338 644 infants admitted to the NICU within 31 US children’s hospitals, with 10 052 admissions utilizing 1 or more head CT (3.0%). Patient characteristics are shown in Table 1 and Supplemental Table 2. Characteristics strongly associated with head CT utilization (ie, standardized difference (SD) > 0.8) included neurologic or neuromuscular complex chronic condition (CCC) and technology dependent CCC (Table 1). Premature or neonatal CCC was moderately associated with head CT utilization (SD = 0.45).

The overall head CT utilization rate decreased from 4.9% of total admissions in 2010 to 2.6% in 2021, an average decrease of 0.21% per year (Fig 1A). Rate of change between the beginning and end of our study period (ie, between 2010 and 2015 and between 2016 and 2021) was then examined. This analysis demonstrated that the decrease in head CT utilization was most significant within the first half of our study period (2010–2015 slope −0.52%; 95% CI: [–0.59% to −0.46%]; P < .0001). Head CT utilization then stabilized for the remainder of the study period (2016–2021 slope 0.03%; 95% CI: [−0.04%, 0.09%]; P = .36; Fig 1A). A sensitivity analysis in which only hospitals with <10% inborn patients in 2021 (ie, ≥90% outborn patients) was subsequently conducted and was consistent with the overall cohort, demonstrating a decrease in head CT utilization rate from 4.8% of total admissions in 2010 to 2.4% in 2021, an average decrease of 0.22% per year (Supplemental Fig 4A). Rate of change between the beginning and end of our study period (ie, between 2010–2015 and between 2016–2021) was also examined within our sensitivity analysis and demonstrated that the decrease in head CT utilization was most significant within the first half of our study period (2010–2015 slope −0.48%; 95% CI: [−0.58% to −0.38%]; P < .0001; Supplemental Fig 4A). Head CT utilization then stabilized for the remainder of the study period (2016–2021 slope −0.03%; 95% CI: [−0.13% to 0.07%]; P = .55; Supplemental Fig 4A).

We also performed a sensitivity analysis examining head CT rates restricting to infants admitted at ≤7 days of life and to infants >7 days of life, given the high number of outborn infants, to address concern that patient may have received head CT imaging before transfer to a PHIS-participating hospital. Sensitivity analysis demonstrated the 85% of patients within our study were admitted within ≤7 days of life and that both subgroups demonstrated imaging trends consistent with our overall cohort (Supplemental Fig 5).

Trends in alternative imaging modalities are demonstrated in Fig 1B. During the 12-year study, there was an overall rise in mean head MRI utilization, from 12.1% to 18.7% of encounters between 2010 and 2021, respectively, with an average annual increase of 0.60% throughout the 12-year study period (Fig 1B). This increase was most notable in the second half of our study period (2016–2021 slope 0.98%; 95% CI: [0.63% to 1.33%]; P = .0002) compared with the first half (2010–2015 slope 0.39%; 95% CI: [0.05% to 0.74%]; P = .03). These results are consistent with our sensitivity analysis that demonstrated an increase in MRI utilization from 11.6% of admissions in 2010 to 19.0% in 2021, an increase of 0.67% per year (Supplemental Fig 4B). Sensitivity analysis also demonstrated that MRI utilization increased most notably in the second half of our study period (2016–2021 slope 0.98% per year; 95% CI: [0.60 to 1.36]; P = .0004) compared with the first half of the study period (2010–2015 slope 0.44%; 95% CI: [0.06 to 0.82]; P = .03). Head ultrasound imaging rate increased slightly during this study period, rising from 41.3% to 43.4% of encounters in 2010 vs 2021, reflecting an average annual increase of 0.19% (Fig 1B). This increase was most notable in the second half of our study period (2016–2021 slope 1.53% per year; 95% CI: [0.80 to 2.27]; P = .001) compared with the first half of the study period (2010–2015 slope −1.25%; 95% CI: [−1.98 to −0.51]; P = .005). Again, this was consistent with our sensitivity analysis, which showed an increase in head ultrasound imaging rate from 41.9% in 2010 to 43.9% in 2021, an annual increase of 0.18% (Supplemental Fig 4B), and that head US utilization increased most notable in the second half of our study period (2016-2021 slope 1.53% per year; 95% CI: [0.80 to 2.27]; P = .001) compared with the first half of the study period (2010–2015 slope −1.25%; 95% CI: [−1.98 to −0.51]; P < .0001).

Of note, although there was an overall decrease in hospital head CT utilization and an increase in head MRI utilization during the study period, this pattern was not consistent across all hospitals (Supplemental Fig 6). For example, hospital 23 demonstrated a decrease in head CT utilization, with a concomitant increase in head MRI utilization, whereas hospital 26 demonstrated a relatively stable rate of both head CT and head MRI utilization across the study period (Supplemental Fig 6). Further, Supplemental Fig 6 appears to illustrate that some hospitals are demonstrating substitution patterns for MRI in place of head CT (see hospitals 11, 12, 16, 23), whereas other institutions are simply exhibiting growth of any advanced imaging (hospitals 27, 29, 31).

Infant head CT utilization varied substantially across hospitals during the entire study period (Fig 1A). In 2021, there were a total of 28 067 admissions to the NICU, of which 727 used head CT (2.6% of admissions). Observed and case-mix adjusted head CT rates for individual hospitals are illustrated in Fig 2. Unadjusted hospital-specific rates ranged from 0% to 10% of admissions (Fig 2). After risk-adjustment, hospital head CT utilization ranged from 0% to 6.2% of admissions, reflecting persistent variation in hospital-level rates of head CT utilization.

Given findings of hospital variation in head CT utilization following risk-adjustment, we sought to determine whether specific diagnoses might be associated with this variation. Within the 2021 cohort, diagnoses and procedures associated with the highest rate of head CT utilization included (in decreasing order): VP shunt (19.3% underwent head CT), neoplasm (12.3%), seizure or encephalopathy (12.0%) (Supplemental Table 3). There was also significant hospital-level variation in head CT utilization among these diagnoses and procedures, as demonstrated in Fig 3. The largest degree of interhospital imaging rate variation included, in decreasing order: seizure and encephalopathy (hospital-level head CT rate IQR = 11.6%; 50th percentile 12.0%; range 0% to 28.9%), VP shunt (IQR = 10.8%; 50th percentile 15.4%; range 0% to 54.0%), and infection (IQR = 10.1%; 50th percentile 7.5%; range 0% to 27.1%) (Supplemental Table 3). Sensitivity analysis in which hospitals with ≥10% inborn population were excluded also demonstrated similar procedures and diagnoses associated with highest rate of head CT, though the largest degree of interhospital imaging rate variation was seen for VP shunt (IQR = 12.5%, 50th percentile = 15.4%; range 0% to 54%), followed by seizure and encephalopathy (IQR = 11.9%, 50th percentile 11.9%; range = 0% to 28.9%) (Supplemental Table 4). The ICD-9 and 10-CM diagnosis codes for each of the nine diagnostic and procedural categories are listed in Supplemental Table 5.

To identify priority populations for standardization of head CT utilization, we devised a prioritization framework, based on diagnoses and/or procedures for head imaging that takes into account the potential magnitude for reduction in utilization (total head CTs) and variation in practice (IQR of utilization across hospitals). We defined high priority indications as those with the greatest contributions to head CT utilization (ie, top quartile of diagnoses and procedures) accompanied by wide variation in head CT utilization (ie, IQR > 10%). Diagnoses and procedures associated with highest number of head CT studies in 2021 included abnormal neurologic finding (n = 17 196), other disease of the nervous system (n = 5827), and stroke (n = 4180). However, the diagnoses and procedures associated with the widest variability in imaging rate (seizure and encephalopathy [IQR = 11.6%], VP shunt [IQR = 10.8%], and infection [IQR = 10.1%]) did not significantly overlap with the aforementioned diagnoses and procedures. Consequently, no singular diagnoses or procedure served as a clear target for future standardization efforts.

In this study of NICU admissions in US children’s hospitals, we found that overall utilization of head CT decreased from 2010 to 2021, but significant hospital-level variation remained. Further, although we observed a 12-year nationwide decline in head CT utilization, this rate reduction was primarily limited to the first half of our study period, with a subsequent stabilization in head CT rate. Despite the recent stabilization in utilization of head CT, the finding of persistent hospital-level variation indicates that further reductions in, and stewardship of, head CT utilization is possible.

Our finding of an overall decline in head CT utilization is consistent with previous studies in the literature that examined head CT rates at pediatric hospitals inclusive of all pediatric age groups,^6,18,19  although studies evaluating head CT trends within nonpediatric institutions have shown either less remarkable downward trends or stable rates.^26,27  This trend toward lower head CT utilization within pediatric hospitals may reflect an increased awareness of risks associated with CT,^1–3  the rise of alternative imaging modalities,^8,9,18,19,26  and the publication and increased utilization of clinical decision-making guidelines.¹⁵ 

We found persistent variation in hospital-level head CT utilization rates even after adjustment for patient-level characteristics (head CT rates 0% to 6.8%; Fig 2). Previous pediatric studies of head CT utilization variability have primarily focused on its use within emergency departments, often in the setting of trauma, and included all pediatric age groups.^17,18  Consistent with our findings, these studies demonstrated significant hospital-level variation in both overall CT imaging¹⁸  and, specifically, head CT imaging,^28,29  with significant variation noted even among providers within the same emergency department.³⁰  Although studies focused on inpatient imaging variation are rarer, a study by Lodwick et al examined cumulative inpatient, observation, and emergency department encounters among all pediatric age groups.⁶  Consistent with our study, researchers found that hospital-specific rates of head CT varied over 2-fold after adjustment for case-mix between 2009 and 2013 in included pediatric hospitals.⁶  That study did not focus specifically on NICU head CT utilization, nor did it focus on exclusively admitted patients. Nonetheless, the literature is consistent with our finding of wide variability in head CT utilization.

We also found that hospital variation in head CT utilization persisted across multiple studied diagnoses and procedures associated with head CT. This finding suggests that hospital-level variation in head CT utilization is because of institutional differences in clinical practice. The greatest hospital-level variation in head CT utilization was noted among patients with codes for seizure and encephalopathy, VP shunt, and infection. However, the diagnoses and procedures associated with the highest number of head CT studies (abnormal neurologic finding, other disease of the nervous system, stroke) did not significantly overlap with the aforementioned diagnoses and procedures. Consequently, although our prioritization framework was unable to identify a singular diagnoses or procedure that could serve as a clear target for standardization efforts, the pervasive hospital-level imaging variation indicates that head CT standardization efforts for a variety of diagnoses and procedures may be beneficial.

Our findings are important because previous studies have demonstrated that guidelines and use of alternative imaging modalities can successfully reduce CT utilization.^19,31,32  For example, Trost et al found that use of MRI for VP shunt evaluation increased from 1.7% to 19.8% within pediatric hospitals between 2007 and 2015, with a corresponding drop in the number of CTs performed.³³  This rise in MRI utilization was mainly attributable to practice changes at a minority of institutions,³³  a finding which may underlie some of the hospital variation noted within our study. Another study demonstrated increased use of ultrasound in the initial evaluation of children with superficial soft tissue face and neck infections resulted in decreased CT utilization without negatively impacting outcomes.³⁴  Overall, these studies suggest that encouragement and standardization of alternative imaging modalities in lieu of CT may reduce hospital-level imaging variability and minimize overall radiation exposure in children.

Our study also demonstrated a significant increase in MRI utilization over the study period. Given the overall decrease in hospital head CT utilization and concomitant increase in head MRI utilization, it is possible that a substitution trend is occurring as shown in some hospitals (eg, hospitals 11,12, 16, 23 in Supplemental Fig 6). This possible substitution pattern may be secondary to evolving standards of care for certain clinical conditions, such as hypoxic ischemic encephalopathy (HIE), for which MRI is now the preferred imaging modality over CT.^35,36  MRI provides the benefit of decreased radiation exposure compared with head CT, but imaging times are longer. Although most NICUs attempt nonsedated MRI,^37–40  failure of this technique or extenuating circumstances may necessitate sedation that carries additional short-term and potential long-term risks, including possible impacts on neurologic development in young children.⁴¹  Therefore, MRI is not always an innocuous imaging modality. However, it is important to note that this substitution pattern was not consistent across all hospitals. Multiple hospitals (eg, hospitals 27, 29, 31) appear to be increasing any form of advanced imaging (Supplemental Table 4). Taken together, these findings may indicate that future stewardship and quality improvement efforts must take into account both overall indications for head CT and address possible substitutions and their consequences. Further, given the increase in MRI utilization noted in this study, future research efforts examining the benefits of MRI utilization versus risk of sedation and increased costs may be beneficial and worth further investigation.

Our study has limitations. First, data were limited to freestanding pediatric hospitals participating in the PHIS database. Therefore, conclusions may not be generalizable to community, nonpediatric hospitals, where the majority of pediatric patients are treated.⁴²  This is particularly evidenced by the fact that 26 of the 31 participating hospitals in our study had ≥90% outborn populations. Pediatric hospitals may have greater access to resources such as pediatric ultrasound and MRI, which could contribute to different imaging trends between pediatric and general hospitals. Second, although we accounted for multiple confounders, there may be others that that could be influencing head CT utilization variability. Third, our head CT imaging diagnoses and procedure categorizations are based on individual ICD-9 and 10-CM codes, which may or may not have been the true underlying reason for the patient’s head CT. Further, these categorizations were determined after examining ICD-9 and 10-CM code frequency and were not identified a priori. A priori diagnoses and procedure categorizations may have led to different diagnoses categorizations. Fourth, our study did not assess clinical outcome of patients. Therefore, we are unable to draw conclusions on the impact of head CT and overall clinical response, both in terms of acute management of their condition and in terms of the potential long-term radiation-induced impact of CT. Fifth, there is the possibility for measurement error given that most of the participating hospitals within our study had >90% outborn population, and patients may have had imaging completed at referring hospital before transfer to PHIS-participating institution. To address this limitation, a sensitivity analysis was conducted in which hospitals with ≥10% inborn populations in 2021 were excluded. This sensitivity analysis demonstrated persistent imaging rate variation across hospitals and within diagnostic and procedural categories. An additional sensitivity analysis in which infants admitted at ≤7 days and >7days of life was also completed and demonstrated that 85% of patients were admitted at ≤7 days of life, and imaging trends remained consistent, thus mitigating some concern that the high outborn population may lead to a deceptively low imaging rate. Finally, it should be acknowledged that while further reductions in rate of head CT may be possible, there may also be some baseline rate of head CT that is necessary based on clinical diagnoses and patient condition.

In conclusion, our study demonstrates an overall decline in head CT utilization for admitted NICU patients in United States children’s hospitals. However, widespread variation in hospital head CT utilization remains. Targeted strategies to establish and implement standardized protocols for ordering head CT, with the goal of head CT stewardship, may decrease hospital-level head CT variation and reduce infant radiation exposure.