Using Social Behavior Profiles to Predict Autism and Schizophrenia Diagnoses

Background: Overlapping social impairments in Autism Spectrum Disorder (ASD) and Schizophrenia (SCZ) contributed to decades of diagnostic confusion that continues to this day in some clinical settings. Our previous work used discriminant function analysis (DFA) to classify the social skills profiles of typically-developing (TD) adults and those with ASD and SCZ based upon their behavior during a 3 minute social interaction. Seven discrete social behaviors were coded: interactive behaviors (e.g., asking questions), nonverbal behaviors (e.g., gaze, affect expression), paralinguistic behaviors (e.g., verbal clarity), appropriate verbal content, repetitive movement, repetitive verbal content, and amount of time speaking. DFA profiles separated ASD (p < .001) and SCZ (p = .041) from the TD group, with ASD uniquely characterized by fewer interactive behaviors and more repetitive behaviors, and SCZ characterized by greater impaired gaze and flat/inappropriate affective responses.

Objectives: The current study tested our model’s predictive accuracy in classifying both the original sample and a new sample of ASD, SCZ, and TD individuals based on their social behavior.

Methods: The original sample consisted of three adult groups: ASD (n=54), SCZ (n=54), and TD (n=56) comparable on age (group mean range: 25.69-28.67), gender (87% male), race (80-91% Caucasian), and IQ (group mean range: 103-106). DFA was used to classify a second unmatched sample (ASD n=28, TD n=31, SCZ n=16).

Results: Seventy percent of ASD and 70% of TD but only 33% of SCZ cases were correctly classified using the original sample. When applied to the second sample, TD classification accuracy increased to 87%, but decreased to 46% for ASD and 18% for SCZ. To explore this classification inaccuracy, we examined comparability in the two samples. The ASD cases in the second sample were younger than the initial sample (p=.010)—many were recruited from a university setting—and exhibited better interactive behaviors and poorer paralinguistic behaviors than the original sample (p’s<.02). The second sample of SCZ individuals had more females and lower IQ (p’s<.045) than the first sample, and demonstrated more repetitive movement (p=.01).

Conclusions: Group separation based on social behavior between ASD, TD, and SCZ individuals in an initial sample accurately predicted 70% of ASD cases but only 33% of SCZ cases. This disparity suggests that unique features of ASD (e.g., repetitive behavior) may aid in classifying ASD correctly, whereas differentiating SCZ from ASD based solely upon observable social behavior is less successful. Further, when the classification system was applied to a new sample, it proved accurate only at classifying TD individuals, with reduced classification accuracy for ASD and SCZ individuals. This may have occurred because the second sample of clinical participants was notably distinct from the original sample—ASD participants were younger and more skilled, and the SCZ group had lower IQ and more females. The failure of the model to predict clinical status within a novel sample suggests that social behavior within these clinical groups varies as a function of sample heterogeneity, and thus the model may be of limited utility in clinical settings, particularly those that evaluate diverse clinical populations.