Identifying Mental Health Conditions in Cancer Patients Using a Hybrid TCN-BiGRU-DATT Deep Learning Model

Abstract

Cancer patients suffer a tremendous psychological burden, but a substantial portion of that is not captured in the clinic due to the episodic nature of screening and reliance on disclosure. Social media text can be a valuable tool for identifying mental health issues as many patients are able to communicate their fears, sadness, and uncertainty openly in anonymous online communities. In this work, the authors introduce a hybrid deep learning (DL) model, called TCN-BiGRU-DATT, for detecting mental health problems in cancer patients’ Reddit narratives. We created an annotated collection of over 23,000 cancer patient stories from 74 targeted subreddits, which were curated from first-person cancer stories. Six labels were used for the development of the models, namely: anxiety, bipolar disorder, depression, post-traumatic stress disorder (PTSD), schizophrenia, and no-condition. Temporal convolutional, bidirectional recurrent, and dual attention layers were fed with after class balancing Bidirectional Encoder Representations from Transformers (BERT) embeddings. The model achieved 97.32% accuracy, 97.20% macro F1-score and 98.00% macro area under the curve on the balanced test set. Best performance was found in bipolar disorder, PTSD, schizophrenia while anxiety and no condition were more challenging due to the overlap in language with normal cancer-related distress. The results show that both temporal and attention-based modeling approaches can aid in scalable detection of mental health signals in social media text from cancer patients.