25th January 2013: Derivation of a Clinical Prediction Rule for Non-Accidental Head Injury


Where can I find this paper?


What is this paper about (what is the research question)?

Can we identify children whose head injury is non-accidentally occurring using the presence or absence of certain clinical criteria?

Summary of the Paper

Design: multicentre prospective observational cross-sectional study

Objective: to identify and measure relationships between clinical variables and non-accidental head injury at time of PICU admission

Tests of Interest: a list of clinical and radiological findings

Reference Standard: a priori definition criteria for non-accidental head injury

Primary outcome: test characteristics with sensitivity, specificity and reliability, combined into a decision tree

Population: PICU patients across 14 participating sites, recruited between Feb 2010 and August 2011

  • Inclusion: children <3yrs of age admitted to PICU for treatment of symptomatic acute closed traumatic head injuries
  • Exclusion: pre-existing brain malformation identified on CT, absence of acute head trauma, head trauma from RTC.

Results: 209 patients recruited, of whom 45% (95) met one or more of the criteria for non-accidental head injury defined a priori. Of these patients, abuse was admitted by the perpetrator in 14 cases and moderately-strongly suspected (2+ extracranial injuries) in 53 cases.

20 variables were both reliable and descriminating of which 13 were based on information available at or near to the time of PICU admission. Binary recursive  partitioning identified five variables present at or near the time of PICU admission which, when used alone or in combination, identified 92 (97%) of those meeting the a priori defined criteria;

  1. acute respiratory compromise prior to admission
  2. seizures or acute encephalopathy
  3. bruising of the ear, neck or torso
  4. interhemispheric or bilateral subdural haemorrhage or fluid collection
  5. skull fracture other than isolated linear non-diastatic parietal fracture

For the five-part rule:

Sensitivity 0.97 (95% CI 0.90-0.99)

Specificity 0.27 (95% CI 0.20-0.37)

Positive predictive value 0.53 (95% CI 0.45-0.60)

Negative predictive value 0.91 (95% CI 0.75-0.98)

LR+ 1.33 (95% CI 1.18-1.50)

LR- 0.12 (95% CI 0.04-0.37)

Authors’ Conclusions:

Once validated, the rule could be used by paediatric intensivists to calculate an evidence-based, patient-specific estimate of abuse probability that can inform – not dictate – early decisions to launch or forego an evaluation for abuse.

On the study design

The methodology here is quite complicated. Essentially, the authors have decided in advance some criteria which cause or allow a high probability of non-accidental trauma as the cause of head injury in their PICU population (let’s call these reference critera). They have then measured the presence or absence of historical, examination or radiological findings and, while measuring the reliability of their assessment of the presence or absence of the findings, have correlated the presence or absence of the finding with non-accidental trauma as defined by the reference criteria. The individual findings were divided into “early” or “late” to help determine those likely to be present at the time of PICU admission, and the most reliable and discriminating criteria identified.

These were then combined into a decision rule which was applied to the population again, and test characteristics calculated. Phew!

There are a few problems here, most of which the authors identify. The major issue is true of all non-accidental injuries; there is no certainty in the diagnosis, no gold standard, and as anyone working in child protection knows identifying children who have been intentionally or neglectfully injured is a game of probability in the absence of confession (and even then not always a certainty). Vigilance is key; overdiagnosis causes a massive workload for paediatricians and enormous stress for families and patients (who are often separated during investigations), underdiagnosis fails the child and the family, often with tragic, deadly consequences.

The criteria used as a reference standard (table 2) are a reasonable surrogate for a gold standard.

230113 Table2

The other problem we have is that this is a very select population; by definition, these are the more severely injured children and it is likely that a significant proportion of non-accidentally head injured children will not require PICU. This immediately affects the extent to which we can generalise the findings to our significantly different PED population.

What were the results and what does this mean?

Figure 2 shows the decision tree the authors developed from the most reliable and discriminating variables.

250113 Fig2

The authors have then calculated the test characteristics for the decision tree as shown in table 6.

250113 Table6

As we can see by looking at the decision tree, the tool is far more sensitive than specific, thereby acting as a better tool to rule out NAI than ruling it in. In fact, the sensitivity is perhaps not as good as we would like (look at the 95% confidence interval – it could be as low as 90%). We can see that 3 patients classified a priori as high risk were misclassified by this tool as low risk.

The negative likelihood ratio is small, suggesting that a negative result in this population (with a low pre-test probability) produces a very low post-test probability, but again the confidence interval is quite wide.

The authors concede that adding any further variables would make the rule too complicated to be practical, which seems reasonable, but it does leave us wondering whether this rule will be fit for purpose with a potentially low NPV and sensitivity.

What can we take from this paper into clinical practice?

While this decision tree might, once validated, help to rule-out non-accidental head injury in the PICU population, the patients here are just too different from the PED population for this to be useful.

In addition, our job in PED is to resuscitate these children and while it is essential that NAI is always on the mind of the PED doctor, for these patients stabilisation and management of the acute injuries must take priority. Does this rule add anything to our PED assessment? I don’t think so – these are not the patients in whom I want to think carefully about NAI as my priorities are different; they have immediate clinical needs and NAI can be considered in more depth later. The patients in whom I want to rule NAI out are altogether less unwell. Any one of these five findings necessitates further PED assessment of the child from a perspective other than NAI.

It’s not a bad paper – it just doesn’t help us in PED.

More questions to ask

  • How does this rule perform in other PICUs (validation)?
  • How sure do we want to be of ruling out NAI (what sensitivity level should we accept in a tool like this)?
  • Does it have any predictive value for patients in the PED? Are there other cues which do have useful rule-in or rule-out potential in the less seriously injured head injury patients in PED?

Follow us on twitter: @PEMLit

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s