목적: This study aims to prove the usability of large language models (LLMs) for tinnitus Cognitive Behavioral Therapy (CBT) analysis and THI score prediction. 방법:Cognitive Behavioral Therapy (CBT) for tinnitus alleviates psychological discomfort caused by severe tinnitus symptoms. During CBT, the patients will have various homework assignments, including writing daily diaries and self-monitoring. Most of these homework assignments are hand-written, textual data. This paper proposes that tinnitus therapeutics can utilize Large Language Models (LLMs) to analyze CBT and predict the outcomes of CBT treatments to manage high caseloads. We anonymized patient data and examined it with GPT-2- based-embedding, GPU-accelerated dimensionality reduction, and clustering process to observe how patients themselves changed their misconceptions and developed less unnecessary excessive emotional discomfort and how their Tinnitus Handicap Inventory (THI) scores were improved after the CBT treatment. We also discussed clustering results as a part of the demonstrations that LLMs can give us insights into the CBT. Then, we augmented textual patient data in three ways to minimize augmentation bias with a corresponding penalty to overcome the constraints of limitation of the number of datasets. The augmentation algorithm we employed is three combinations, each with a different penalty level. We created three unique datasets with those three different augmentation algorithms. We trained the Google T5 Transformer with the augmented data to predict the THI score outcomes at the end of the CBT sessions. We measured the performance using the ROUGE-L metric during the training and validation. The generated THI scores by Google T5 were converted from strings to floats to measure RMSE performance, which proved that the LLM could predict the outcome of CBT treatment with CBT data. 결과:As the complexity of the level of augmentation increases, the error rates also increase in general. We individually trained the same Google T5 LLM per augmented dataset and compared the prediction outputs. As text augmentation and typo complexity increase, the RMSE drops slightly but maintains around 0.0138~0.005, and severe numerical augmentation also increases ROUGE-L losses a little from 0.7514 to 0.8600, which means that Google T5 LLM is very suitable for generalization and can predict the outcome of the tinnitus CBT treatment based on CBT text entries and partial patient information. 결론:Google T5 LLM was able to generalize variations of the augmented dataset and still predict the correct outcome of the treatment of the CBT. Even though data augmentation with a small number of data would bring a risk of overfitting issues, our approach proved that LLM-based tinnitus therapeutics are crucial to managing a high caseload of tinnitus patients. However, future work should include a more extensive and more diverse dataset for the real-world setting.