Prediction models for highly scalable technology-assisted differential diagnostics of autism spectrum disorder
Abstract Background Diagnosing autism spectrum disorder (ASD) in adulthood is time-consuming and markedly complicated by the requirement to distinguish between ASD and differential diagnoses also associated with social interaction difficulties, such as borderline personality disorder (BPD)—a distinction for which currently no valid screening or diagnostic tool exists. While technology-assisted diagnostics (TAD) has emerged, existing algorithms have focused on classifying between ASD and no diagnosis, not fully addressing clinical reality. Objective Therefore, we assessed the feasibility of TAD for differential diagnostics by classifying between ASD and BPD in this proof-of-concept study. Methods We collected a rich multimodal dataset of reciprocal interactions, specifically dyadic conversations (n=120 interaction partners). From this data, we extracted more than 800 features, allowing us to capture the core area of defining symptoms for both conditions: social interactions. These features include speech patterns, facial expressions, movement quantity and interpersonal synchrony. We used these features to train and stack linear support vector machines to classify between ASD-involved, BPD-involved and comparison interaction partners. Findings Base models capturing facial expressions during speaking and listening, speech patterns, synchronisation of facial expressions and movement quantity all performed above chance when differentiating between ASD-involved and BPD-involved interaction partners. Stacking all base models containing conceptually related features further increased accuracy, with our algorithm achieving nearly 82% of balanced accuracy, solely based on 20 min of conversation. Conclusions Our proof-of-concept study shows the immense potential of TAD for differential diagnostics: data collection only requires microphones and webcams while feature-extraction is automated, making this approach highly objective, scalable and user-friendly. Clinical implications Our TAD algorithm shows the potential of multimodal, behavioural data for differential diagnostics. On the basis of clinical validation such an algorithm has the potential to streamline differential diagnoses of ASD in the future, enabling faster and more accurate diagnostic assessment and ultimately reducing patient distress by shortening the wait for an appropriate treatment plan.