PsychSynth: Advancing Mental Health AI Through Synthetic Data Generation and Curriculum Training

Abstract

The number of Mental-health help seekers are on rise over recent years, but the medical practitioners are limited due to which the healthcare system is heavily loaded and thereby the outcome is not promising. Advanced AI Technology is expected to ease this problem, however completely relying on the tools is reported to not work with the help seekers. Hence technology to assist help seekers and allow the healthcare system to screen the needy ones from the large pool of distress candidates should make the system efficient a nd effective.H owever for establishing the AI supportive system, two challenges exists: one is the large pool of dataset with diversity, and other is the AI model to respond to a context-aware situations. Expecting large dataset for training model from medical community needs further infrastructure support to make it digitally available with embedded annotations. Relying patiently for the dataset will miss the opportunity to serve the distress patients. Hence synthetic dataset generation and its validation from medical experts is an alternative for training robust and reliable model. Besides, context-aware curriculum inspired AI based summarizer model is found appropriate to adopt for this use-case where relevant features meant for diagnosing the problem is extracted from the improvised input text. The proposed curriculum trained AI model helps in transforming the improvised text inputs fed from the distress individuals to a summarized version representing domain expert form, embedded with symptoms related features for further classification. The synthetic data-set g eneration through OpenAI's GPT-40 models and Nemotron models are further evaluated with BERT based classifier m odels a nd curriculum based AI model. The training of the classifier m odels are also evaluated for synthetic and real-world dataset, which was scrapped from Reddit forum. Around 800 stream of real-world posts were evaluated from the medical experts and their findings related to sympotoms and annotations were employed to fine-tune the classifier and summarizer m odel. It was found t hat the fine-tuned models and training of BERT models from the merged dataset composed of synthetic ones with the medical practitioners annotated dataset were found to perform better than others. The summarizer model fetching shorter version of domain expert output enhanced the classification accuracy by 5 % for the real-world data. The effort is a step towards developing AI assistant to screen large posts of submissions from distress individuals and arrange for the necessary connects for the needy ones with the medical experts. The models and pruned datasets are made freely available for further usage to the researchers community.

Ähnliche Arbeiten