Trends in Shaken Baby Syndrome Diagnoses Among Young Children Hospitalized for Abuse

Objective To investigate national trends of SBS diagnosis codes and how trends varied among patient and hospital characteristics. Methods We examined possible SBS, conrmed SBS, total SBS, and non-SBS abuse diagnoses among children age three and younger who were hospitalized for abuse between 1998 and 2014 using a secondary analysis of the National Inpatient Sample, the largest US all-payer inpatient care database (N = 66,854). A baseline category logit model was used based on a quasi-likelihood approach (QIC) with an independent working correlation structure. Results The rate (per 100,000 census population) of total SBS diagnoses was 5.4 (± 0.3) between 1998 and 2014, whereas the rate of non-SBS abuse was 19.6 (± 1.0).The rate of conrmed SBS diagnoses increased from 3.8 (± 0.3) in 1998 to 5.1 (± 0.9) in 2005, and decreased to 1.3 (± 0.2) in 2014. Possible SBS diagnoses were 0.6 (± 0.2) in 1998, increasing to 2.4 (± 0.4) in 2014. Conrmed SBS diagnoses have declined since 2002, while possible SBS diagnoses have increased. All abuse types were more frequent among infants, males, children from low-income homes, and urban teaching hospitals. Conclusions We investigated seventeen-year trends of SBS diagnoses among young children hospitalized for abuse. The discrepancy between trends in possible and conrmed SBS suggests differences in norms for diagnosing SBS, which has implications for which cases are considered AHT. Future research should investigate diagnostic processes for SBS and whether all codes associated with abusive head injuries in young children are classied as AHT. Our ndings also highlight the relativity dening and diagnosing SBS. Medical professionals nd utility in diagnosing SBS, though may be more apt to apply possible SBS diagnoses to abusive head injuries in children. Clarifying norms for SBS diagnosis and rening denitions for AHT will ensure that young children presenting with abusive head injuries are included in overall counts of AHT. This baseline data, an essential component of child abuse surveillance, will enable ongoing efforts to track, prevent, and reduce child abuse.


Introduction
Every year, 30 per 100,000 infants are diagnosed with abusive head trauma in US hospitals, [1] leading to mild to moderate behavioral and cognitive problems among abuse survivors, and lifelong disability or death in severe instances. [2] Pediatric abusive head trauma (AHT) includes injuries to children's skulls or intracranial contents that occur when perpetrators throw, drop, suffocate, choke, and violently shake young children.
[3] The actual incidence of AHT remains unknown due to the secrecy, stigma, and shame associated with child abuse, [4] , [5] eliciting researchers to surveil AHT using 15 diagnosis codes recommended by the Centers for Disease Control (CDC). [6] Though the annual incidence of AHT likely exceeds the use of AHT diagnosis codes within hospitals, inpatient data are ideal for public health surveillance; hospital datasets provide consistent and reliable sampling frames (i.e., all children hospitalized for AHT), and uniform measures of AHT (i.e., diagnosis codes from International Classi cation of Disease, Clinical Modi cation [ICD-CM]). [7] AHT is the o cial term for intentional abusive head injuries in children, yet researchers and medical professionals formerly ascribed these injuries to shaken baby syndrome (SBS). SBS is di cult to substantiate, relying on a speci c mechanism for injury, i.e., shaking. The narrow de nition of SBS along with legal and medical controversies surrounding SBS contributed to the American Academy of Pediatrics (AAP) and CDC's decision to adopt AHT terminology instead of SBS. 6,[8] SBS is often framed as a questionable diagnosis in both mainstream media and research, [9] - [10] [11] although nearly 40% of AHT hospitalizations include a diagnosis code for shaken baby syndrome (SBS). [12] The continued reliance on SBS codes suggests that medical professionals nd utility in the diagnosis, yet little is known about the incidence of SBS diagnosis codes per year or over time. These knowledge gaps are problematic for two reasons. First, researchers do not know whether trends in SBS diagnoses correspond with trends in AHT diagnoses, which have stabilized or declined over time, depending on the study. 12, [13] Evidence that SBS diagnoses are changing over time would provide AAP and CDC con dence about the effectiveness and accuracy of their messaging about AHT. Second, researchers do not know the extent to which SBS is diagnosed every year nor the hospital and patient characteristics associated with the diagnosis. Such evidence would highlight inconsistencies between current and desired diagnostic practices and help policymakers align current practices with their guidelines. In response, we examined trends in SBS diagnoses (ICD-9 code 995.55) between 1998 and 2014.
Another issue related to SBS diagnosis pertains to accuracy. Medical professionals who diagnose SBS rely largely on a unique pattern of injuries they attribute to shaking, including subdural hematoma, retinal hemorrhages and encephalopathy, otherwise known as the so-called "triad". [14] , [15] However, researchers do not know whether the SBS diagnostic code (995.55) captures all abusive head injuries related to the "triad." In absence of this information, some pediatric abusive head injuries may remain unclassi ed as either SBS or AHT. The CDC de nitions for both probable and de nite AHT, for example, exclude retinal hemorrhages and convulsions without a seizure disorder, two symptoms typically associated with SBS. [16] This is problematic considering that up to 40% of AHT hospitalizations with an SBS diagnosis would be considered non-AHT abuse without the SBS diagnostic code. 12 In response, we used diagnostic codes from ICD-9 to develop measures of possible and con rmed SBS. For possible SBS, we included ICD-9 codes for retinal hemorrhages and convulsions without a seizure disorder, whereas our measure of con rmed SBS included only the ICD-9 code for shaken baby syndrome (995.55) among young children hospitalized for child abuse.
In all, few, if any, researchers have examined the prevalence of SBS diagnosis codes or how the use of codes are trending over time. In the absence of this data, researchers will not know the extent to which SBS is diagnosed, or the patient and hospital characteristics associated with SBS diagnosis. Therefore, the purpose of our study was to investigate trends in SBS diagnosis codes among a nationally representative sample of children age three and younger who were hospitalized for abuse. Using 1998-children hospitalized for both possible and con rmed SBS, calculated the annual rate of non-SBS abuse and possible and con rmed SBS diagnosis codes, and investigated how trends varied among patient and hospital characteristics.
Through our ndings on trends in SBS diagnosis, we can determine whether the use of the SBS code (995.55) has declined since the AAP's recommendation in 2009, and whether the SBS code accounts for all possible SBS diagnoses, which are not currently captured within the 15 AHT diagnosis codes.

Sample
Data from 1998-2014 from the National Inpatient Sample (NIS) 17 were used for this retrospective study. The NIS is the largest all-payer, inpatient care database in the US and is maintained by the Healthcare Utilization project. We identi ed our sample of children three years of age or younger who were hospitalized for abuse using the following ICD-9 diagnosis and external cause of injury codes, which describe the intent, mechanism, and circumstances of an injury: Abuse by father/step-father/boyfriend (E967.0), abuse by mother/step-mother/girlfriend (E967.2), abuse by all other relatives (E967.3 -E967.7), abuse by non-relative (E967.8), neglect (995.52), physical abuse (995.54), psychological abuse (995.51), shaken baby syndrome (995.55), sexual abuse (995.53), and other child abuse not speci ed (995.50, 995.59, E967.1, & E967.9). We examined SBS diagnoses among young children hospitalized for any form of abuse for two reasons. First, though associations between physical abuse (e.g. SBS) and non-physical forms of abuse may seem surprising, physical and non-physical abuse co-occurs among 30 to 90% of abused children. [18] - [19][20] [21] Boxer and Terranova found that 50% of hospitalized children experienced more than one abuse type, with physical abuse, emotional abuse, neglect, and sexual abuse sharing modest to moderate correlations. [22] Second, healthcare providers may misdiagnose or fail to diagnose earlier instances of child physical abuse preceding severe incidences of abuse. [23] , [24] These ndings suggest that young children with SBS diagnoses may also present to the hospital with a non-physical form of abuse.
Although data from 2015 and 2016 are available, we did not use these data due to the switch from ICD-9-CM to ICD-10-CM diagnosis standards in 2015. This change led to substantive shifts in abuse coding, meaning that changes in abuse frequency could not be discerned from changes in the coding standard. All analyses incorporated sampling weights to provide nationally representative estimates.

Measures
Our analysis included four measures of child abuse within our sample: Non-SBS abuse, con rmed SBS, possible SBS, and total SBS. We de ned these three categories as follows: Con rmed SBS Abuse: The presence of diagnosis code 995.55. Non-SBS Abuse: Remaining cases that did not meet the criteria of Con rmed or Possible SBS.
Total SBS Abuse: The sum of con rmed and possible incidences of SBS.
Per the Barell injury matrix, [25] Type 1 TBI diagnoses include primary intracranial injury, moderate to prolonged loss of consciousness, shaken infant syndrome, or injuries to the optic nerve pathways. We included retinal hemorrhage and convulsions not related to a seizure disorder because they indicate acceleration/deceleration of the head, 16 yet are not included in the CDC's recommended codes for AHT. 6 While traumatic brain injury from impact is not precluded, this combination of codes indicates that children were subjected to high-force acceleration injury, characteristic of SBS.

Analytic strategy
We calculated the annual prevalence of SBS hospital diagnoses from 1998 through 2014 for con rmed SBS, possible SBS, and total SBS. Our primary objective was to estimate trends in the probability of con rmed and possible SBS diagnoses. A Chi-square test was used to test for independence of possible and con rmed SBS diagnoses from patient and hospital characteristics. Differences between non-SBS abuse and total SBS were also estimated.
Trends were estimated overall and within categories of age and hospital size. Hospital size is categorized as large, medium, and small, although the NIS does not provide detailed information on how these categories are constructed. A baseline category logit model [26] was used based on a quasilikelihood approach (QIC) with an independent working correlation structure in which non-SBS abuse is the reference category. The primary sampling unit in the NIS is the hospital; there are repeated observations within each hospital and across time. To account for the misspeci cation of the correlation structure, robust standard errors were used (i.e., "sandwich" estimators). Since there is a lack of research on SBS time trends, QIC was used to determine the form of the time trend: linear, quadratic, cubic, or treating time as categorical. A reduction in QIC of two-units was chosen a priori to indicate model preference. All statistical analyses were performed using SAS version 9.4.

Sample characteristics
From 1998-2014, there were an estimated 66,854 total hospital admissions for child abuse among children three years of age and younger, with 52,562 hospital admissions for non-SBS abuse and 14,292 for SBS abuse, of which 10,083 were con rmed and 4,209 were possible SBS. Both possible SBS and non-SBS abuse admissions had signi cantly older age distributions than con rmed SBS or total SBS admissions, respectively (p < 0.0001; Table 1). Although not statistically signi cant, 63.6% of possible SBS admissions were male compared to 59.7% of con rmed SBS admissions (p = 0.0749). In contrast, 60.8% of total SBS admissions were male as compared and 57.2% of non-SBS abuse admissions (p < 0.001). There were statistically signi cant differences in the racial distributions and hospital deaths of total SBS and non-SBS abuse admissions (p < 0.0001), yet no signi cant differences for possible and con rmed SBS admissions by racial distribution (p = 0.2402) or hospital deaths (p = 0.4673). We used the default racial categories provided by NIS (Asian or Paci c Islander, Black, Hispanic, Native American, Other, White).
The NIS categorizes patient income into quartiles based on the estimated median household income of residents in the patient's ZIP code. Lower income families (1 st and 2 nd income quartiles) had a higher representation among possible SBS and total SBS admissions than con rmed SBS admissions (62.1% & 56.0% vs. 54.2%, respectively; p = 0.0005). Moderate to higher income families (2-4 th quartile) had greater representation among total SBS admissions compared to non-SBS abuse admissions (p < 0.001).
Con rmed SBS was most common in large hospitals, while possible SBS was more frequent in medium and small-sized hospital (p = 0.0082). There were no statistically signi cant differences for total SBS and non-SBS abuse by hospital size. Possible SBS and total SBS were also more frequent than con rmed SBS and non-SBS abuse in urban teaching hospitals when compared to urban non-teaching and rural hospitals (p = 0.0054, p < 0.0001).

Time trends in rate of SBS
In Table 1, we summarize the rates (per 100,000 census population [27] -, [28], [29] ) for non-SBS abuse, possible SBS, con rmed SBS, and total SBS. The overall rate of total SBS diagnoses was 5.4 (± 0.3) for every 100,000 children 3 years of age or younger from 1998-2014, whereas the rate of non-SBS abuse was 19.6 (± 1.0). The overall rate of con rmed SBS and possible SBS admissions was 3.8 (± 0.3) and 1.6 (± 0.1), respectively. The annual rate of non-SBS abuse admissions remained fairly stable, with the lowest rate of admissions in 2000 (16.6 ± 2.0) and the highest rate of admissions in 2010 (25.8 ± 4.2). However, the rate of admissions for con rmed SBS abuse increased from 3.6 (± 0.6) in 1998 to 5.1 (± 0.9) in 2005, at which point the rate steadily decreased to 1.3 (± 0.2) in 2014. Conversely, the rate of admissions for possible SBS increased gradually per year. In 1998, the rate of possible SBS admissions was 0.6 (± 0.2) and steadily increased to 2.4 (± 0.4) in 2014. Total SBS uctuated but remained relatively stable, beginning at 4.1 (± 0.7) in 1998 and ending at 3.7 (± 0.4) in 2014.
Time trends in probability of type of abuse diagnosis. We conducted a baseline category logit model to examine whether the difference in trends for con rmed SBS, possible SBS, and the reference category, non-SBS abuse, were statistically signi cant. Figures 1 -3 indicate the results from these analyses (solid line and 95% con dence band). Non-SBS abuse includes all abuse cases not meeting criteria of con rmed or possible SBS. A statistically signi cant difference was found in the overall trends of the probability of possible SBS versus con rmed SBS (Figure 1). Possible SBS gradually increased over time, whereas con rmed SBS slightly increased until 2001 and then decreased from 2002 to 2014 with the trends crossing over in 2011. In 1998, the estimated probability of possible SBS was 2.8% (95% CI: 1.7, 4.6), whereas the chance of con rmed SBS was 18.5% (95% CI: 12.5, 26.4). In 2014, the estimates for possible SBS and con rmed SBS were 9.4% (95% CI: 7.7, 11.3) and 5.3% (95% CI: 3.4, 8.2), respectively. Figure 2 shows the statistically signi cant trends for possible SBS versus con rmed SBS by age group: < 1 year old, 1 year old, and 2 or 3 years old. The decrease in the probability of con rmed SBS was largest for infants (< 1 year old), declining from 29.9% (95% CI: 23.2, 37.4) in 1998 to 3.2% (95% CI: 1.4%, 7.3%) in 2014, a decrease of 26.7%. Con rmed SBS diagnosis trends for infants (< 1 year old) overlapped with possible SBS in 2011. The decrease in con rmed SBS diagnosis trends was more attenuated for young toddlers (1 year old) than for older toddlers (2 or 13.9%], respectively). The probability of possible SBS increased for all age groups in the same time period.
The trends for con rmed and possible SBS among females and males ( Figure 3) were similar to the overall trend ( Figure 1). For males, the probability of con rmed SBS increased from 20.5% (95% CI: 13 Figure 4 shows the statistically signi cant trends for possible SBS versus con rmed SBS by hospital size (based on bed count). The most pronounced decrease in the probability of con rmed SBS was for large hospitals between 1998 and 2014 (

Discussion
AHT is a serious form of child physical abuse that happens when caregivers throw, drop, choke, and violently shake young children. 3 These abusive head injuries were classi ed as SBS until 2009, when the AAP and CDC recommended AHT diagnoses instead of SBS. 8 Though non-fatal AHT diagnoses have stabilized or declined over time, 12,13 little is known about SBS diagnosis trends. This knowledge gap is problematic considering that 40% of AHT hospitalizations with an SBS diagnosis would not be classi ed as AHT without an SBS diagnosis. 12 Without knowing trends in SBS diagnosis codes, we cannot determine the extent to which medical professionals follow the diagnosis recommendations of CDC and AAP. Further, the patient and hospital characteristics associated with SBS diagnoses remain unknown. Having this data could help researchers and policymakers identify the factors associated with SBS diagnosis. In response, we investigated seventeen-year trends in possible SBS, con rmed SBS, total SBS, and non-SBS abuse diagnoses among young children. We also examined the patient and hospital characteristics associated with these diagnoses.
In our study, we found support for the following: 1) Non-SBS abuse is the most common form of abuse in our study; 2) Con rmed SBS diagnosis trends have declined while possible SBS diagnosis trends have increased; 3) All abuse diagnoses were more common among infant, male, or low-income children and urban teaching hospitals. Taken together, our ndings contribute to literature on SBS and non-SBS abuse diagnoses within hospitals, including diagnosis trends and characteristics associated with each abuse category. Policymakers can use our ndings to develop plans for aligning current diagnostic practices with CDC and AAP guidelines. According to our estimates, possible and con rmed SBS cases represent a fraction of the overall abuse young children experience. Our results are similar to reports on overall child maltreatment investigations, in which only 10.7% of cases comprise physical abuse. [30] Hospitalizations often represent the most severe incidences of non-fatal child abuse, which are disproportionately experienced by young children. 30 Two research teams reported unchanging rates of child abuse hospitalizations among children ages 0 to 18, with young children facing the largest share of hospitalizations between 1998 and 2016, [31] and stable overall trends in child maltreatment hospitalization among young children ages 0 to 3 from 1997 to 2009. [32] Along with these study ndings, our results provide evidence that child abuse hospitalizations remain a signi cant problem in the US.
Of our four abuse diagnosis categories, only con rmed SBS diagnoses declined between 2002 to 2014. This nding aligns with AHT literature indicating that AHT diagnoses decreased during overlapping time periods (2003 -2008; 12 2000 -2009 13 ). There may be various explanations for reductions to con rmed SBS diagnoses over time, including declines in the actual incidence of SBS. Unfortunately, our ndings on the decreasing use of SBS code 995.55 most likely does not re ect shifts in abusive caregiver practices. Given that possible SBS diagnoses increased steadily over the study period, a likelier explanation relates to changing diagnostic practices of medical professionals. Though AAP and CDC recommendations likely impacted the decline in con rmed SBS diagnosis, con rmed SBS diagnoses began decreasing in 2002, seven years before AAP's recommendation in 2009. It is possible that medical and legal controversy surrounding SBS also contributed to the decline. Numerous papers have questioned SBS as a reputable diagnosis, citing inadequate scienti c evidence that the injuries typically associated with SBS are caused solely by shaking. [33] - [34] [35] Whereas con rmed SBS diagnoses have decreased, possible SBS diagnoses have increased. By 2011, the rate of possible SBS diagnoses exceeded the rate of con rmed SBS diagnoses. This nding suggests that medical professionals are not diagnosing retinal hemorrhage and/or convulsions not associated with a seizure disorder, as con rmed SBS, even in the presence of physical abuse and Type 1 internal traumatic brain injury. If these diagnosis codes represent SBS yet are not coded as SBS, there are implications for diagnostic norms for both SBS and AHT. At present, CDC-recommended AHT codes do not include retinal hemorrhage and/or convulsions without seizure disorder (i.e. possible SBS). The implication is that our possible SBS diagnosis may capture cases that would not meet the CDC's de nition of AHT, 6 perhaps being misclassi ed as non-AHT abuse, as previously indicated by Parks and colleagues. 12 In the future, researchers should investigate the overlap between possible SBS and AHT, and compare trends in con rmed SBS, possible SBS, and AHT diagnoses over time. This research would help determine whether current de nitions of AHT capture all diagnoses associated with abusive head injuries in young children. If current AHT de nitions exclude instances of abusive head injuries, including retinal hemorrhage and/or convulsions not associated with a seizure disorder in the presence of physical abuse and Type 1 internal traumatic brain injury, policymakers could consider adding codes to measure AHT Our nding on increasing possible SBS diagnoses may also provide context to the overall process for diagnosing SBS, which Narang and Greeley describe as a complex, context-driven process without reputable diagnosis guidelines. [36] Through our ndings, there are a few directions researchers, policymakers, and medical professionals could take. If medical professionals nd utility in diagnosing SBS, it appears that clearer diagnostic guidelines are needed, especially in light of the hospital characteristics associated with possible versus con rmed SBS diagnosis. According to our study, possible SBS diagnoses were more frequent in urban teaching hospitals and large hospitals, suggesting different diagnostic protocols in these institutions. Findley and colleagues recommend the development of a national registry on SBS and protocols for diagnosing SBS along with alternative explanations for SBS-like injuries. [37] Likewise, we propose that researchers and pediatric medical providers agree to a standardized de nition and diagnostic guidelines for possible and con rmed SBS, much like the AHT guidelines proposed by CDC, which may help reduce discrepancies in diagnosis and treatment and improve options for surveillance. [38] , [39] Finally, our ndings on the patient and hospital characteristics associated with SBS diagnoses align with what is known about AHT in the literature. Like AHT, our ndings indicate that all abuse types were diagnosed more frequently among infants (<1-year-old), boys, and children from low-income households than toddlers, girls, and children from higher-income households. 12,13,[40] The prevalence of SBS and AHT of infants under the age of 1 has been attributed to infant crying and subsequent parental or caregiver stress. Perpetrators are often caregivers who shake children to stop persistent crying during the rst few months after birth. [41] The age distribution of SBS and AHT follow what is referred to as the "normal crying curve", or typical periods of increased infant crying that peaks between 6 to 10 weeks of age. [42] [43] Explanations for the prevalence of SBS among male infants may include the acoustic characteristics of male cries, societal norms related to crying in boys, and gender stereotypes. Some research suggests that male infant cries may aggravate caregivers more than female cries and that caregivers may tolerate male cries less than female cries due to societal norms around crying; [44] in the US, crying is more acceptable for girls and is considered a weakness for boys. Gender stereotypes held by physicians and medical professionals may in uence their diagnoses of abuse related injuries. For example, Ravichandiran and colleague reported that physicians initially miss the abuse of boys more often than girls, [45] suggesting that physicians may perceive injuries differently among boys and girls, possibly because boys are socialized for rough-and-tumble play and are more prone to accidental injury. [46] In all, we contribute to the literature by examining seventeen-year trends of SBS among young children hospitalized for abuse, yet there are limitations. First, our possible SBS measure may not account for all SBS victims who are hospitalized. Physicians use a variety of tools, including constellations of injuries consistent with SBS and a child's history of injuries, to diagnose SBS. We chose a combination of codes highly indicative of acceleration/deceleration injury to avoid false-positive diagnoses, but these codes are not exhaustive. Second, our analysis includes no correction for confounders in the estimation of time trends. The lack of research on these trends, however, warranted our approach of estimating simple time trends by subgroup. Third, because our estimates rely on in-patient data, they do not include victims of SBS who were not hospitalized and therefore may misrepresent the actual incidence of SBS. This surveillance issue is a common problem for all studies on child maltreatment. Finally, SBS-related diagnosis codes only describe how medical professionals code abusive head injuries and cannot describe trends related to abusive parenting practices. Surveillance of child abuse itself remains a signi cant challenge in literature on child maltreatment, and is an issue our study cannot address. 4

Conclusion
Our study ndings demonstrated that while con rmed SBS has decreased since 2002, possible SBS has increased. The discrepancy between trends in possible and con rmed SBS suggests differences in norms for diagnosing SBS, which has implications for which cases are considered AHT and which are not. Future research should investigate diagnostic processes for SBS and whether all codes associated with abusive head injuries in young children are being classi ed as AHT. Our ndings also highlight the relativity de ning and diagnosing SBS. According to our ndings on con rmed SBS diagnoses, medical professionals nd utility in the diagnosis, though may be more apt to apply possible SBS diagnoses to abusive head injuries in children given AAP and CDC recommendations and controversy surrounding SBS diagnoses. Clarifying norms for SBS diagnosis and re ning de nitions for AHT will ensure that all young children presenting with abusive head injuries included in overall counts of AHT. This baseline data, an essential component of child abuse surveillance, will enable ongoing efforts to track, prevent, and reduce child abuse.

Availability of data and materials
The data that support the ndings of this study are available for purchase from the Healthcare Utilization Project. Data are available, however, from the authors upon reasonable request and with permission of the Healthcare Utilization Project.
Authors' contributions AC, RO, BB, and CPA contributed signi cantly to the study conceptualization and design. AC contributed to manuscript writing. CPA and BB acquired the data, performed data cleaning and analysis, and contributed to manuscript writing. RO provided editing and guidance throughout manuscript revisions. MR conducted the literature review and contributed to manuscript writing. All authors read and approved the nal manuscript.
Ethics approval and consent to participate Not applicable. NIS is a publicly available dataset.

Consent for publication
Not applicable.

Competing interests
The authors declare that they have no competing interests.

Funding
Not applicable.