Identification of Emergency Blood Donation Request on Twitter

Puneet Mathur, Meghna Ayyar, Sahil Chopra, Simra Shahid, Laiba Mehnaz, Rajiv ShahSMM4H @ EMNLP 2018 · DOI

Introduction

In many regions, blood donation systems rely heavily on personal networks and word-of-mouth. During medical emergencies, families frequently turn to social media platforms like Twitter to solicit donors. While Twitter provides a vast stream of real-time data, it is also characterized by significant noise. The primary challenge is distinguishing a genuine Emergency Blood Donation Request (EBDR)—which contains actionable details such as blood type, hospital location, and contact information—from general medical advice, promotional content, or expressions of gratitude. Automating this identification process allows critical requests to be routed to medical authorities and donor communities within the vital "golden hour."

Methodology

The researchers approached the identification of EBDRs as a supervised classification task using a Support Vector Machine (SVM). The core of the study involved evaluating four distinct categories of features to determine which signals most effectively identify urgent requests.

Table 1: Definition of feature categories and their specific attributes used in the classification model.

As shown in the feature table above, the researchers utilized:

  1. Linguistic Features (L): NLP techniques like unigrams, bigrams, and TF-IDF vectors to capture emergency-related vocabulary.
  2. User Metadata (U): Social signals such as follower counts and account status to assess the impact of user authority.
  3. Textual Metadata (T): The frequency of hashtags, URLs, and mentions.
  4. Handcrafted Features (H): Domain-specific entities including contact names, hospital names, blood groups, and patient disease descriptions.

Datasets

The study utilized a newly annotated corpus of tweets collected over two months in 2018, divided into three subsets:

  • Personal Donation Requests (PDR): 2,822 tweets curated using medical phrases from portals like the Red Cross.
  • Blood Donation Community (BDC): 5,157 tweets extracted from known donation group timelines.
  • Held-Out (HO) Dataset: 1,813 tweets combining both collection strategies to simulate real-world conditions.

Manual annotation by independent experts yielded a Cohen’s Kappa score of 0.86, indicating high reliability in the ground truth labels.

Results and Discussion

The experimental results demonstrated that content-based features significantly outperform context-based metadata. A surprising finding was that User Metadata (U) and Textual Metadata (T) actually degraded performance, suggesting that a user's follower count or hashtag usage is a poor indicator of a tweet's urgency.

Table 2: Performance metrics across different feature set combinations for the PDR, BDC, and HO datasets.

As illustrated in the results table, the combination of Linguistic and Handcrafted features (L+H) consistently yielded the highest performance across all datasets. On the Held-Out (HO) validation set, the model achieved a peak accuracy of 97.89% and an F1-score of 0.980. This synergy suggests that the model benefits from understanding both the urgent tone of the language and the presence of specific, actionable data points like blood types and phone numbers.

Conclusion and Future Work

The proposed SVM model effectively filters the noise of social media to identify life-saving requests with near-perfect accuracy. While highly effective, the system occasionally struggles with "reactionary rants" (frustration about donor shortages) and "thank you" notes that share keywords with active requests. Future research will focus on implementing automated Named Entity Recognition (NER) to extract patient details more precisely and utilizing genetic algorithms for advanced feature optimization.

Summary updated

BibTeX
@inproceedings{mathur2018identification,
  author = {Puneet Mathur and Meghna Ayyar and Sahil Chopra and Simra Shahid and Laiba Mehnaz and Rajiv Shah},
  title = {{Identification of Emergency Blood Donation Request on Twitter}},
  booktitle = {SMM4H @ EMNLP},
  year = {2018},
  doi = {10.18653/v1/w18-5907},
}
Figure