Abstract
OBJECTIVE: Although autism spectrum disorder (ASD) is a relatively common, well-known but heterogeneous neuropsychiatric disorder, specific knowledge about characteristics of this heterogeneity is scarce. There is consensus that IQ contributes to this heterogeneity as well as complicates diagnostics and treatment planning. In this study, we assessed the accuracy of the Autism Diagnostic Observation Schedule (ADOS/2) in the whole and IQ-defined subsamples, and analyzed if the ADOS/2 accuracy may be increased by the application of machine learning (ML) algorithms that processed additional information including the IQ level. METHODS: The study included 1,084 individuals: 440 individuals with ASD (with a mean IQ level of 3.3 ± 1.5) and 644 individuals without ASD (with a mean IQ level of 3.2 ± 1.2). We applied and analyzed Random Forest (RF) and Decision Tree (DT) to the ADOS/2 data, compared their accuracy to ADOS/2 cutoff algorithms, and examined most relevant items to distinguish between ASD and Non-ASD. In sum, we included 49 individual features, independently of the applied ADOS module. RESULTS: In DT analyses, we observed that for the decision ASD/Non-ASD, solely one to four items are sufficient to differentiate between groups with high accuracy. In addition, in sub-cohorts of individuals with (a) below (IQ level ≥4)/ID and (b) above average intelligence (IQ level ≤ 2), the ADOS/2 cutoff showed reduced accuracy. This reduced accuracy results in (a) a three times higher risk of false-positive diagnoses or (b) a 1.7 higher risk for false-negative diagnoses; both errors could be significantly decreased by the application of the alternative ML algorithms. CONCLUSIONS: Using ML algorithms showed that a small set of ADOS/2 items could help clinicians to more accurately detect ASD in clinical practice across all IQ levels and to increase diagnostic accuracy especially in individuals with below and above average IQ level.