Differentiating Autism Spectrum Disorder (ASD) Subtypes Using Estimates of IQ and Language

Background: While it is possible to reliably and accurately differentiate autism from non-ASD children, the differentiation of Asperger syndrome (AS) from autism is much more difficult. Although DSM-IV states that children with AS do not have clinically significant cognitive or language delay, the operationalization of these criteria are difficult to implement and evaluate.

Objectives: To evaluate the ability of a diagnostic algorithm employing both estimates of IQ and language to accurately differentiate Asperger syndrome from autism.

Methods: The sample consisted of 347 individuals with autism, and AS referred to a genetic study. Inclusion criteria were any child with an ASD but without a primary genetic or medical syndrome. All subjects were assessed using available clinical records, ADI-R, ADOS, a non-verbal IQ assessment (the Leiter International Performance Scale) and the Vineland Adaptive Behaviour Scales (VABS).  Each individual was given a best estimate (BE) diagnosis based on clinical records and the above information.  Receiving operating characteristic curves (ROC’s) were used to estimate the score on the non-verbal IQ test and the Communication scale that best differentiated AS from autism. Using these scores, a new diagnostic algorithm was devised combining the Leiter and VABS Communication scores and was compared to the BE diagnosis. Estimates of sensitivity, specificity and positive predictive value were calculated.

Results: ROC curves identified a score of 80 on the Leiter and 75 on the VABS Communication Scale as the estimates that best differentiate autism from AS (area under the curve=0.80 and 0.87 respectively). Combining both scores as categorical thresholds gives a sensitivity of 0.64, a specificity of 0.85 and a positive predictive value of 0.51.

Conclusions: To differentiate AS from autism, researchers will have to use supplementary information in addition to the ADI-R and ADOS. Using the thresholds employed here, estimates of sensitivity and specificity are quite good but because of the low prevalence of AS in this sample, the positive predictive value is only moderately high. This will be a problem for any diagnostic algorithm that attempts to distinguish AS from autism.