Deep Learning to Predict Caregiver’s Distress Level in Palliative Care from Natural Language Conversations

Objectives: Addressing caregivers’ burden requires the proper identification of their distress level. The distress levels can be severe, high, moderate and low. Through appropriate identification of the intensity of the distress level, caregivers might be provided with support to cope with the issues. Conversations between caregivers and palliative care professionals can be one easy but effective way to identify the level of distress of caregivers. Artificial intelligence (AI), especially the deep learning models, can be helpful to efficiently identify the level of distress of the caregivers. In this paper, a deep learning model has been developed to predict caregivers’ distress levels from the natural language conversations between caregivers and the palliative care experts. Materials and Methods: The paper proposes a deep learning model called the convolutional neural network (CNN) which has been trained on 200 natural language conversations between family caregivers and palliative care experts. The natural language conversation data was generated hypothetically using ChatGPT. For training and testing the CNN model with the generated data, data pre-processing has been done. The CNN model has five layers: Input layer, embedding layer, 1D convolution layer, global max pooling layer and a fully connected dense layer. The model accepts pre-processed conversation data as input and classifies it in one of the four distress levels of a caregiver, namely severe, high, moderate or low distress level. Moreover, for a more fine-grained classification of caregivers’ distress, the proposed CNN model has been integrated with the Bidirectional Encoder Representations from Transformers (BERT) model. Results: The CNN model is effective in broader classification of caregivers’ distress levels and shows 97.56% accuracy. The refined CNN model, integrated with BERT, the model achieved 98.42% accuracy on the hypothetical data generated for this study. External validation using real-world caregiver conversations is required before clinical implementation. The proposed model’s performance was illustrated using visual aids like word clouds and samples of its predicted output. Conclusion: In palliative care, the proposed AI-driven paradigm provides a scalable way to determine caregiver’s suffering, encouraging prompt assistance and better treatment results. Its usefulness has to be improved and validated by further research using actual data.