Can Artificial Intelligence Heal Human Hearts? A randomized controlled trial on the effects of internet cognitive behavioral therapy with artificial intelligence on depression

Author Name SO Mirai (Keio University) / SEKIZAWA Yoichi (Senior Fellow, RIETI) / TAKEBAYASHI Yoshitake (Fukushima Medical University)
Creation Date/NO. November 2016 16-J-059
Research Project Research Project on Mental Health from the Perspective of Human Capital 2
Download / Links


Background: In spite of recent high expectations of internet-based cognitive behavioral therapy (iCBT), iCBT still holds limitations including effect sustainability, function improvement, and dropout. Therefore, we focused on iCBT-AI in which automatic feedback by text or animation expressing empathy or indicating instruction is given to users with respect to their input after analysis conducted by a natural language processing (NLP) system, which is one area of artificial intelligence (AI). Since there is no evidence of iCBT using AI technology so far, we evaluated its effectiveness.
Methods: 1,187 participants recruited from the website were randomly assigned into three groups; iCBT-AI, conventional iCBT without AI, and waitlist as control. Those allocated to interventional arms were encouraged to perform each exercise at least once a week for seven weeks. The primary outcome was moderate-to-severe depression defined as a PHQ-9 score of 10 or higher. Intention-to-treat analyses were performed.
Results: The dropout rate was significantly lower in iCBT-AI than iCBT (p<0.005). At post-intervention, there were no significant differences between the interventional arms and the control group although only the iCBT-AI arm indicated a marginally significant lower odds ratio compared to the control group at the three-month follow-up (0.67 (95% confidence interval 0.43 to 1.04), p=0.07). In the secondary analysis stratifying participants by the severity of depression, only iCBT-AI users with PHQ-9 between 5 and 10 (corresponding to mild depression) showed a significantly lower odds ratio of having moderate-to-severe depression (0.35 (0.14 to 0.86), p=0.02) at the three-month follow-up.
Conclusion: Although iCBT-AI has no significant short-term antidepressant effect, iCBT-AI seems to have an exclusive potential to reduce moderate-to-severe depression in the future. Further research is required.