An artificial intelligence approach for selecting effective teacher communication strategies in autism education

The prevalence of autism spectrum conditions (ASC) has significantly increased, leading to a need for effective inclusive education practices. Current approaches lack a concrete consensus on guiding teacher-student interactions. This research proposes using artificial intelligence (AI) to model these interactions and identify effective communication strategies for children with ASC. AI's potential in education is well-recognized, but most research focuses on data from learning management systems. This study focuses on real-time classroom data, analyzing teacher communication strategies and student responses to improve educational and social outcomes. The researchers hypothesize that machine learning can assist in developing ASC-specific communication strategies. Previous machine learning research in autism has primarily focused on diagnostic tools; this study uniquely explores real-time classroom interaction data to inform pedagogical strategies.

The literature highlights the importance of tailored communication strategies in autism education for improved learning, social functioning, and long-term outcomes. The Social Communication, Emotional Regulation, and Transactional Support (SCERTS) framework is a widely used approach. Existing research emphasizes the use of visual strategies for effective communication with children with ASC. However, most studies focus on specific interventions in controlled settings, not general classroom situations. This study addresses this gap by focusing on real-time classroom interactions and a broader range of communication strategies.

The researchers collected data through structured classroom observations over five months in a special school. Three teachers, their assistants, and seven students (aged 6-12) participated. Observations included various classroom and non-classroom locations. A total of 5460 interactions were coded, capturing details such as student attributes (age, sex, P-level, SCERTS), teaching objective (academic, social, pedagogic), teaching type (instructions, modeling, etc.), context (whole class, small group, etc.), student emotional state, teacher communication strategy (verbal, gestures, physical prompts, visuals, objects), and student response (full, partial, no response). Machine learning classifiers (logistic regression, random forest, Gaussian process) were used to predict student responses based on various combinations of features. The researchers performed an ablation analysis to determine the impact of each feature category on prediction accuracy. The analysis included both single-feature and multi-feature models, incorporating past observations and student responsiveness to refine predictions. A Gaussian process classifier was employed to provide probabilities for various communication strategies, allowing for recommendations based on specific scenarios. The researchers also analyzed the long-term effects of different communication strategies by examining consecutive interactions. Statistical correlation analysis was conducted on an expanded dataset of over 2.6 million simulated observations to identify significant patterns.

The machine learning classifier successfully predicted student responses with an accuracy exceeding random chance and baseline predictions. Accuracy improved with the inclusion of student attributes and past interaction data, reaching 0.711 using a Gaussian process model with autoregression. The ablation analysis revealed that teacher communication strategies had the greatest impact on classification accuracy. Analysis of consecutive observations showed that visual prompts led to increased full student response rates in the short-term, while physical prompts had an initial positive impact followed by a decrease in engagement. Verbal communication alone showed lower effectiveness. The study found that using two communication strategies concurrently tended to be more effective than using a single strategy, particularly when excluding physical prompts. The analysis identified negative student emotional state, physical prompts, verbal communication, and encouragement/praise as features most correlated with student response. Visual prompts (pictures and objects) showed similar effectiveness according to the classifier.

The results demonstrate the potential of AI to assist teachers in selecting effective communication strategies for students with ASC. The classifier's ability to predict student responses and suggest suitable communication strategies based on various factors offers a valuable tool for educators. The findings support the existing literature on the effectiveness of visual aids and the potential negative impact of excessive physical prompts. The observed higher effectiveness of using two communication strategies compared to one aligns with the “cone experience” perspective, suggesting interactive approaches can enhance responsiveness. The study’s limitations notwithstanding, the findings highlight the potential of AI-driven tools to personalize education for students with ASC, moving beyond generalized interventions toward more targeted and effective classroom practices.

This study presents an innovative approach to autism education by applying machine learning to real-time classroom data to predict student responses to teacher communication strategies. The findings suggest that AI can provide teachers with valuable insights for tailoring their communication to improve student outcomes. Future research should focus on expanding the sample size to enhance generalizability, include additional relevant features, and explore the long-term impacts of AI-guided communication strategies in diverse classroom settings.

The study's limitations include the relatively small sample size (seven students), which may affect the generalizability of the findings. The student attributes used were educational indicators rather than standardized clinical diagnoses. The observation schedule could be expanded to include additional features (proximity, sensory sensitivities). While increasing features might improve accuracy, it requires significantly more data collection. A larger, more diverse cohort with a wider range of ASC phenotypic variability is needed for a more robust analysis of student attributes and the long-term effects of teacher communication strategies.