Harnessing AI to Optimize Thought Records and Facilitate Cognitive Restructuring in Smartphone CBT: An Exploratory Study

Abstract

Background: Effective cognitive restructuring (CR) requires identification of automatic thoughts that underlie experienced emotions. However, accurate recording of thoughts and emotions is challenging when CR is provided in internet cognitive-behavior therapy (iCBT). This study investigated the potential use of the artificial intelligence (AI) including the natural language processing (NLP) to facilitate CR offered in iCBT. Methods: We applied the Japanese Text-to-Text Transfer Transformer (T5), one of the most advanced Large Language Models for the NLP,to records of thought-feeling pairs provided by participants in two randomized controlled trials of iCBT. We conducted threefold cross-validated prediction of self-reported feelings based on recorded thoughts. We examined the validity of the predictions by checking them against the human expert judgments and by the efficacy when the thought records were subjected to CR. Results: 1626 participants provided 4369 though-feeling records. The overall prediction accuracy was 73.5%. The self-reported feelings matched the human expert judgments more frequently when they were correctly predicted by the T5 than not (90% vs 37.5%, 95%CI of difference: 34.8 to 70.2%). When subjected to CR, the correctly predicted thought-feeling pairs led to greater reductions in negative feelings than the incorrectly predicted pairs (− 1.54 vs − 1.43 on a scale of 0 to 5, 95%CI of difference: 0.03 to 0.19). Conclusions: A new CR module of an iCBT application can incorporate this model and advise the users to revisit and revise their automatic thoughts to reflect their feelings more accurately. Whether such an iCBT application can ultimately lead to greater reductions in depression is to be examined in a future randomized trial.