A deep learning approach for identifying cancer survivors living with post-traumatic stress disorder on Twitter

Nur Hafieza Ismail, Ninghao Liu, Mengnan Du, Zhe He, Xia HuBMC Medical Informatics Decis. Mak 2020 · OpenReview

TL;DR

This paper introduces a deep learning approach using Convolutional Neural Networks (CNNs) to identify cancer survivors experiencing Post-Traumatic Stress Disorder (PTSD) by analyzing their public Twitter posts. The method leverages a novel filtering strategy and a pre-trained depression lexicon to efficiently identify relevant tweets, aiming to provide an early warning system for this vulnerable population.

What Problem Does This Paper Address?

Surviving cancer is a significant life event that can lead to complex emotional and psychological challenges, including PTSD. Traditional methods for diagnosing mental health conditions are often time-consuming, expensive, and rely on self-reporting or clinical assessments. The authors highlight that cancer survivors are at a higher risk for developing PTSD due to the traumatic nature of diagnosis, treatment, and the fear of recurrence. However, there's a lack of readily available, quantitative data to help identify these individuals, hindering timely intervention. Social media platforms like Twitter offer a vast, real-time source of user-generated content that can reflect psychological states. The challenge lies in sifting through the noise and identifying genuine expressions of PTSD among cancer survivors, a task that is manual and resource-intensive.

Methods

The researchers propose a multi-stage approach to tackle this problem. First, they collect tweets using "cancer" as a keyword. To efficiently narrow down the dataset, they employ a filtering strategy using keywords related to "cancer-free" and "PTSD." This initial filtering aims to reduce the manual annotation burden.

The core of their method involves a Convolutional Neural Network (CNN) for text classification. The CNN is trained to learn representations from the tweet data. Key aspects of their CNN architecture include:

  • Word Embeddings: Tweets are represented as sequences of word embeddings, initialized with random weights and learned during training.
  • Convolutional Layers: These layers extract local features and patterns from the word embeddings.
  • ReLU Activation: A standard non-linear activation function used after convolution.
  • Max Pooling: This layer reduces the dimensionality of the feature maps and helps control overfitting.
  • Sigmoid Output Layer: Produces a probability score between 0 and 1, indicating the likelihood of a tweet belonging to the PTSD-positive class.

A crucial element is the "knowledge transfer" step. They leverage a pre-existing "depression lexicon" (a list of words associated with depression symptoms) as a proxy to filter out irrelevant tweets. This is based on the observation that depression and PTSD often co-occur, and the lexicon helps identify negative sentiments related to health. The extracted tweets are then manually annotated to create a ground truth dataset.

Datasets

The primary dataset used in this study consists of tweets collected from Twitter. The authors describe a process of crawling tweets using the keyword "cancer" over a three-month period (August-October 2019). They then filtered these tweets using keywords related to "cancer survivor" (e.g., 'cancer survivor', 'cancer-free', 'I had cancer') and a depression lexicon.

For the PTSD-negative group, they used tweets from the "Twitter User Gender Classification" dataset from Kaggle, mixed with cancer survivor tweets expressing positive sentiment. Both the PTSD-positive and PTSD-negative datasets were balanced to 5,000 tweets each for training and evaluation.

Key Insights and Ideas

  • Social Media as a Proxy for Mental Health: The study reinforces the idea that social media can be a valuable, albeit noisy, source for understanding mental health states, offering a low-cost alternative to traditional methods.
  • Leveraging Existing Lexicons for New Problems: The innovative use of a "depression lexicon" as a proxy for identifying PTSD-related tweets demonstrates a practical application of knowledge transfer in a domain where specific lexicons might be scarce.
  • CNNs Excel at Textual Nuances: The CNN architecture proved effective in capturing the subtle linguistic patterns and representations within tweets that indicate PTSD symptoms in cancer survivors.
  • Early Detection Potential: The proposed system acts as an "alarming system," suggesting its potential for early detection and intervention for cancer survivors at risk of PTSD.

Key Results

The proposed CNN model achieved a high accuracy of 91.29% in identifying cancer survivors with PTSD. This performance significantly outperformed baseline methods, including Naive Bayes (NBC), Support Vector Machine (SVM), and Multi-Layer Perceptron (MLP).

  • CNN Accuracy: 91.29%
  • Baseline Accuracies: NBC (86.5%), SVM (49.0%), MLP (49.99%), CNN n-gram (63.28%)

The study also found that using their custom-curated vocabulary and filtering strategy performed better than using only the depression lexicon (67.03% accuracy). The CNN model was also able to provide probability scores for new tweets, aiding in the interpretation of its predictions.

Limitations and Future Work

The authors acknowledge several limitations and suggest avenues for future research:

  • Misclassification with Similar Terms: The model struggled when words like "tumor" or "cyst" were used instead of "cancer," indicating a need to expand the keyword set to include more diverse cancer-related terms.
  • Limited Data for Specific Cases: The presence of only a few tweets from some participants made it challenging to capture their full context. A larger dataset for training, potentially incorporating cancer-free with PTSD lexicons, is needed.
  • Contextual Information: The current model relies solely on textual content and does not consider other contextual information like user demographics (gender, age) or historical posts. Future work could incorporate these elements.
  • Broader Scope: The authors suggest expanding the model to identify other conditions like suicidal ideation or PTSD treatment side effects. They also propose exploring multimodal approaches (audio, image) for better diagnosis.
  • Caregiver Identification: The study notes that PTSD can also affect cancer survivors' caregivers and suggests that the framework could be adapted to identify PTSD in this population, though defining caregiver characteristics would be a prerequisite.

Summary updated

Abstract

Background Emotions after surviving cancer can be complicated. The survivors may have gained new strength to continue life, but some of them may begin to deal with complicated feelings and emotional stress due to trauma and fear of cancer recurrence. The widespread use of Twitter for socializing has been the alternative medium for data collection compared to traditional studies of mental health, which primarily depend on information taken from medical staff with their consent. These social media data, to a certain extent, reflect the users’ psychological state. However, Twitter also contains a mix of noisy and genuine tweets. The process of manually identifying genuine tweets is expensive and time-consuming. Methods We stream the data using cancer as a keyword to filter the tweets with cancer-free and use post-traumatic stress disorder (PTSD) related keywords to reduce the time spent on the annotation task. Convolutional Neural Network (CNN) learns the representations of the input to identify cancer survivors with PTSD. Results The results present that the proposed CNN can effectively identify cancer survivors with PTSD. The experiments on real-world datasets show that our model outperforms the baselines and correctly classifies the new tweets. Conclusions PTSD is one of the severe anxiety disorders that could affect individuals who are exposed to traumatic events, including cancer. Cancer survivors are at risk of short-term or long-term effects on physical and psycho-social well-being. Therefore, the evaluation and treatment of PTSD are essential parts of cancer survivorship care. It will act as an alarming system by detecting the PTSD presence based on users’ postings on Twitter.
BibTeX
@inproceedings{ismail2020deep,
  author = {Nur Hafieza Ismail and Ninghao Liu and Mengnan Du and Zhe He and Xia Hu},
  title = {{A deep learning approach for identifying cancer survivors living with post-traumatic stress disorder on Twitter}},
  booktitle = {BMC Medical Informatics Decis. Mak},
  year = {2020},
  url = {https://openreview.net/forum?id=dB-AWiIcWzD},
}
Figure