Do not let the title of the article in this issue of Pediatrics, “Traumatic Head Injury and the Diagnosis of Abuse: A Cluster Analysis,” by Boos et al1  intimidate those of you who are not biostatisticians. This article is important for providers who evaluate children with head injuries in the hospital setting. Young children remain disproportionately affected by child maltreatment, with the majority of fatalities occurring as a result of abusive head trauma (AHT).2  There has been extensive, high-quality research into clinical and radiologic findings associated with AHT,35  and multicenter research networks have been formed to accelerate this important work. One of the earliest networks, the Pediatric Brain Injury Research Network (PediBIRN), has prospectively collected data on children <3 years of age with symptomatic head injury in the PICU setting to identify a valid and reliable clinical prediction rule (CPR) for pediatric AHT.6 

The authors present a unique and important analysis of the clinical and radiologic findings of 500 young patients with acute neurotrauma from data collected by PediBIRN.7  Their hypothesis is that well-established computer algorithms used to sort “like” data would identify subgroups of children with discreet clinical findings. Furthermore, these subgroups would have specific relationships to indicators of abuse and physician diagnoses. Why use computer algorithms when clinical variables specific to AHT are available? The application of computer algorithms to clinical data broadly characterizes these mathematical clustering methods with the goal of identifying more subtle associations less apparent to researchers and clinicians. There are examples of this approach in the medical sciences, such as identifying phenotypically different asthma groups8  or recognizing distinct clinical subtypes of Alzheimer disease.9  The use of computer algorithms makes this research unique, allowing elimination of potential researcher bias as the computer is blinded to physician diagnoses and previously published associations.

Several key points are integral in understanding the importance of the authors’ cluster analysis. First, they used an “unsupervised” cluster analysis, meaning that computer algorithms analyzed unlabeled data and sorted data into groups without a target outcome. Next, the authors used well-established cluster modeling tools.9,10  K-means is a partitioning cluster algorithm tool where each cluster has a center (eg, the color orange might be in a cluster with yellow and red but would not be in a cluster centered around blue). Divisive and agglomerative algorithms are hierarchical tools that either start with one larger cluster and divide like items into smaller groups or start with smaller groups and connect them back to a larger core group, respectively. For example, a phylogenetic tree of birds would demonstrate a closer ancestor between a falcon and an eagle rather than an ostrich.11  Each tool has strengths and weaknesses; therefore, several tools are used to sort data and analyze cluster similarities and differences. Finally, the authors assessed data partitioned by each algorithm with the “triad,” defined as the presence of 3 clinical variables: any subdural hemorrhage (SDH); any encephalopathy before presentation for care; and retinal hemorrhages (RHs) described by an ophthalmologist as dense, extensive, covering a large surface area, and/or extending to the ora serrata. The odds ratio (OR) and confidence interval are presented with respect to substantial association with high likelihood of AHT (HiAHT) (OR >10) and low likelihood of AHT (LoAHT) (OR <0.01). Many clinical findings did not strongly partition with either group but still demonstrated statistical significance.

The cluster-modeling tools optimally partitioned the data into 2 groups, one group comprised the majority of children ultimately diagnosed by a physician with AHT (therefore HiAHT). Variables that had the most significant association with the HiAHT group included signs/symptoms of significant brain dysfunction: brain hypoxemia, ischemia, and/or swelling; acute encephalopathy (persistent); respiratory compromise prior to admission (PTA); presence of SDH or fluid collection; and retinoschisis. Not surprisingly, acute encephalopathy, SDH, and extensive retinal hemorrhages (RHs) also had significant associations with triad partitioning. However, it is interesting to see that the triad, as defined by the authors, was not as sensitive a partitioning tool as the other cluster tools used in this study. Also of interest, additional injuries seen in child physical abuse, such as bruising (ear, neck, torso) and fractures (considered moderately/highly specific for abuse), were statistically significant in the cluster analysis, but did not have substantial associations with HiAHT or LoAHT.

What does this complex analysis mean? Children admitted with symptomatic traumatic brain injury (TBI) are not on a continuum in which the presentation with AHT looks similar to accidental head trauma or LoAHT. Presentations strongly associated with AHT are characterized by CNS dysfunction, including respiratory compromise PTA and persistent encephalopathy, as well as radiologic evidence of substantial brain injury and SDH that would not be described as “contact” (unilateral). Additionally, children with AHT are more likely to have extensive RHs and/or retinoschisis, fractures, bruising, and/or abdominal injury, consistent with the published literature.12  Presentations that are lower risk for AHT, seen more often with accidental head trauma, often have contact injuries seen with impact to the head, such as a skull fracture with an associated epidural or subdural hemorrhage. These associations are in line with published literature on objective clinical findings that assist in discriminating between these etiologies.35 

A single clinical finding in isolation may not discriminate between children at high or low risk for AHT. However, this analysis and the current literature illustrate the predictive probability of clinical findings seen in combination.13,14  The results of this unique cluster analysis, demonstrating the association of clinical findings of significant brain dysfunction with AHT, will assist medical providers in the hospital setting, highlighting the importance of a comprehensive medical evaluation in children with traumatic head injury.

FUNDING: No external funding.

COMPANION PAPER: A companion to this article can be found online at www.pediatrics.org/cgi/doi/10.1542/peds.2021-051742.

AHT

abusive head trauma

CNS

central nervous system

OR

odds ratio

PediBIRN

Pediatric Brain Injury Research Network

PTA

prior to admission

RH

retinal hemorrhage

SDH

subdural hemorrhage

TBI

traumatic brain injury

1
Boos
SC
,
Wang
M
,
Karst
WA
,
Hymel
KP
;
Pediatric Brain Injury Research Network (PediBIRN) Investigators
.
Traumatic head injury and the diagnosis of abuse: a cluster analysis
.
Pediatrics.
2022
:
149
(
1
):
e2021051742
2
Palusci
VJ
,
Covington
TM
.
Child maltreatment deaths in the US National Child Death Review Case Reporting System
.
Child Abuse Negl.
2014
;
38
(
1
):
25
36
3
Maguire
S
,
Pickerd
N
,
Farewell
D
,
Mann
M
,
Tempest
V
,
Kemp
AM
.
Which clinical features distinguish inflicted from non-inflicted brain injury? A systematic review
.
Arch Dis Child.
2009
;
94
(
11
):
860
867
4
Kemp
AM
,
Jaspan
T
,
Griffiths
J
, et al
.
Neuroimaging: what neuroradiological features distinguish abusive from non- abusive head trauma? A systematic review
.
Arch Dis Child.
2011
;
96
(
12
):
1103
1112
5
Kelly
P
,
John
S
,
Vincent
AL
,
Reed
P
.
Abusive head trauma and accidental head injury: a 20-year comparative study of referrals to a hospital child protection team
.
Arch Dis Child.
2015
;
100
(
12
):
1123
1130
6
Hymel
KP
,
Willson
DF
,
Boos
SC
, et al;
Pediatric Brain Injury Research Network (PediBIRN) Investigators
.
Derivation of a clinical prediction rule for pediatric abusive head trauma
.
Pediatr Crit Care Med.
2013
;
14
(
2
):
210
220
7
Pediatric Brain Injury Research Network
.
Welcome to the CDR Implementation Trial
.
Available at: https://www.pedibirn.com. Accessed October 3, 2021
8
Haldar
P
,
Pavord
ID
,
Shaw
DE
, et al
.
Cluster analysis and clinical asthma phenotypes
.
Am J Respir Crit Care Med.
2008
;
178
(
3
):
218
224
9
Alashwal
H
,
El Halaby
M
,
Crouse
JJ
,
Abdalla
A
,
Moustafa
AA
.
The application of unsupervised clustering methods to Alzheimer’s disease
.
Front Comput Neurosci.
2019
;
13
:
31
10
Everitt
BS
,
Landau
S
,
Leese
M
,
Stahl
D
.
Cluster Analysis
, 5th ed.
Chichester, England
:
John Wiley & Sons
;
2011
11
Mayr G
G
.
The deep divergences of neornithine birds: a phylogenetic analysis of morphological characters
.
Cladistics.
2003
;
19
(
6
):
527
533
12
Choudhary
AK
,
Dies Suarez
P
,
Binenbaum
G
, et al
.
Consensus statement on abusive head trauma: additional endorsements
.
Pediatr Radiol.
2019
;
49
(
3
):
421
13
Maguire
SA
,
Kemp
AM
,
Lumb
RC
,
Farewell
DM
.
Estimating the probability of abusive head trauma: a pooled analysis
.
Pediatrics.
2011
;
128
(
3
):
e550
e564
14
Vinchon
M
,
de Foort-Dhellemmes
S
,
Desurmont
M
,
Delestret
I
.
Confessed abuse versus witnessed accidents in infants: comparison of clinical, radiological, and ophthalmological data in corroborated cases
.
Childs Nerv Syst.
2010
;
26
(
5
):
637
645

Competing Interests

FINANCIAL DISCLOSURE: The authors have indicated they have no financial relationships to this article to disclose.

POTENTIAL CONFLICT OF INTEREST: Drs George and Harper’s institution has received payment for expert witness testimony that has been provided in cases of suspected child abuse in which the physicians have been subpoenaed to testify.