DEVELOPMENT OF CHATBOT FOR PRE-DIAGNOSIS AND RECOMMENDATION OF ANXIETY DISORDER USING DIET AND SENTENCE TRANSFORMER MODELS

Abstract

 Previous research on chatbots for pre-diagnosis and recommendation of anxiety disorders has been limited to therapy aids.  Comparing NLU DIET and LogisticRegressionClassifier models, this chatbot system calculates anxiety levels using GAD-7, DASS, and STAIT/STAIS-5 methods along with Sentence Transformer (SBERT) for semantic similarity.

Intent classification testing yielded 95% accuracy for NLU DIETClassifier and 99% for LogisticRegressionClassifier. The Dialog Model achieved 68% accuracy with TEDPolicy. Testing involved 35 randomly selected respondents, including students and workers. From their interactions, the SBERT recommendation model scored 30% MAP, 26% for the Indobert base and paraphrase-multilingual-MiniLM-L12-v2 models.

The average satisfaction and performance rating for the chatbot system was 3.7 out of 5. This research addresses the need for a prototype chatbot for pre-diagnosis and recommendation of anxiety disorders, with the best NLU model being LogisticRegressionClassifier at 99% accuracy and the dialog model at 68%. However, the recommendation system still has a low MAP due to the use of non-valid clinical data as references, suggesting room for improvement in future research.