Unravelling the Collective Diagnostic Power Behind the Features in the Autism Diagnostic Observation Schedule

Background: Autism is a group of heterogeneous disorders defined by deficits in social interaction and communication. Typically, diagnosis depends on the results of a behavioural examination called the Autism Diagnostic Observation Schedule (ADOS). Unfortunately, administration of the ADOS exam is time-consuming and requires a significant amount of expert intervention, leading to delays in diagnosis and access to early intervention programs. The diagnostic power of each feature in the ADOS exam is currently unknown. Our hypothesis is that certain features could be removed from the exam without a significant reduction in diagnostic accuracy, sensitivity or specificity.

Objective: Determine the smallest subset of predictive features in ADOS module-1 (an exam variant for patients with minimal verbal skills).

Methodology: ADOS module-1 datasets were acquired from the Autism Genetic Resource Exchange and the National Database for Autism Research. The datasets contained 2572 samples with the following labels: autism (1763), autism spectrum (513), and non-autism (296). The datasets were used as input to 4 different cost-sensitive classifiers in Weka (functional trees, LADTree, logistic model trees, and PART). For each classifier, a 10-fold cross validation was preformed and the number of predictive features, accuracy, sensitivity, and specificity was recorded.

Results & Conclusion: Each classifier resulted in a reduction of the number of ADOS features required for autism diagnosis. The LADtree classifier was able to obtain the largest reduction, utilizing only 10 of 29 ADOS module-1 features (96.8% accuracy, 96.9% sensitivity, and 95.9% specificity). Overall, these results are a step towards a more efficient behavioural exam for autism diagnosis.