Objectives: Data guiding abusive head trauma (AHT) diagnosis rest on case-control studies that have been criticized for circularity. We wished to sort children with neurologic injury using mathematical algorithms, without reference to physicians' diagnoses or predetermined diagnostic criteria, and to compare the results to existing AHT data, physicians' diagnoses, and a proposed triad of findings.

Methods: Unsupervised cluster analysis of an existing data set regarding 500 young patients with acute head injury hospitalized for intensive care. Three cluster algorithms were used to sort (partition) patients into subpopulations (clusters) on the basis of 32 reliable (κ > 0.6) clinical and radiologic variables. P values and odds ratios (ORs) identified variables most predictive of partitioning.

Results: The full cohort partitioned into 2 clusters. Variables substantially (P < .001 and OR > 10 in all 3 cluster algorithms) more prevalent in cluster 1 were imaging indications of brain hypoxemia, ischemia, and/or swelling; acute encephalopathy, particularly when lasting >24 hours; respiratory compromise; subdural hemorrhage or fluid collection; and ophthalmologist-confirmed retinoschisis. Variables substantially (P < .001 and OR < 0.10 in any cluster algorithm) more prevalent in cluster 2 were linear parietal skull fracture and epidural hematoma. Postpartitioning analysis revealed that cluster 1 had a high prevalence of physician-diagnosed abuse.

Conclusions: Three cluster algorithms partitioned the population into 2 clusters without reference to predetermined diagnostic criteria or clinical opinion about the nature of AHT. Clinical difference between clusters replicated differences previously described in comparisons of AHT with non-AHT. Algorithmic partition was predictive of physician diagnosis and of the triad of findings heavily discussed in AHT literature.