JMIR Publications

ABSTRACT

Background:

With an increasing number of Major Depressive Disorder (MDD) patients, an efficient and accurate tool for screening and auxiliary diagnosis is in need. This process is often done by mental status examination and assessment scales. In recent years, Artificial Intelligence (AI) has been used on some auxiliary diagnosis such as pulmonary nodules diagnose. We want to use AI to help us to do the screening and auxiliary diagnosis, explore effective stimulus form and comprehensive makers.

Objective:

To find an AI method to do the screening and auxiliary diagnosis of MDD on videos based on specially designed multi-stimulus diagram, which is not scale way, and find effective stimulus and comprehensive markers.

Methods:

We designed a novel way to make the prediction, which included questionnaire, question and answering (Q&A), mental imagery description and video watching. A total of 89 persons participated in our project. We extracted multi-modal features which included vision, audio and text and trained a deep-learning based classification model to give the screening and extent estimation. We trained models for stimulus combinations and got the corresponding performance to evaluate the effectiveness. We used visualization method to see what the AI models learn and compare with correlation analysis results to find valuable markers.

Results:

Top screening accuracy and extent accuracy was 94.12% and 76.47% respectively. Q&A performed best in the listed stimuli. Based on the comparison between the attention learned and traditional correlation results, the agreement exists and there were some typical symptoms of each group.

Conclusions:

In the field of computational psychiatry, deep learning based multimodal-analyzing can help to do the screening and auxiliary diagnosis of MDD, depression related Q&A is the most effective stimulus and AI learned attention which is consistent to human’s prior experience.