How can we better use Twitter to find a person who got lost due to dementia?

Kelvin Kam-fai Tsoi, Nicholas B. Chan, Felix C. H. Chan, Lingling Zhang, Annisa C. H. Lee, Helen Mei-Ling MengNpj Digit. Medicine 2018 · DOI

Wandering and getting lost is a common and hazardous symptom of dementia, affecting between 46% and 71% of patients. In these scenarios, time is the most critical factor for a safe recovery. While traditional news outlets have historically helped spread the word, social media offers a faster, wider-reaching alternative.

This study, published in npj Digital Medicine, applies epidemiological methods to social media data to determine how Twitter can be optimized to find missing persons with dementia. By analyzing tweet characteristics and user types, the researchers identified that official police involvement and the inclusion of photos are the strongest predictors of a successful search.

Methods and Data Acquisition

The researchers employed a retrospective case-control study design, a staple of epidemiology, but applied it to a digital dataset. The study relied on a custom-built dataset of 45,719 tweets collected between April and May 2017.

Study methodology flowchart showing data acquisition, NLP filtering, and variable classification.
  1. Data Acquisition: They used the Twitter REST API to collect tweets related to dementia and missing persons.
  2. NLP Pipeline: Natural Language Processing (NLP) was used to extract names, ages, and locations from unstructured tweet text to identify unique "cases."
  3. Classification: "Cases" were individuals found within a month of the first tweet; "Controls" were those not found. After filtering for English language and relevance, the final analysis focused on 54 missing persons (40 found, 14 not found).
  4. Statistical Modeling: Univariate and multivariate logistic regression models were used to calculate Odds Ratios (OR) for various factors, including user type (police vs. media) and tweet content (photos, links, hashtags).

Key Insights and Results

The study found that while social media is a powerful tool, its success depends heavily on specific "high-value" features. Two factors stood out as the strongest predictors of success:

  • The "Authority Effect": Original tweets posted by police departments were significantly more effective than those from media organizations or the general public (OR 25.1). This suggests that official status provides a level of legitimacy that encourages higher-quality engagement.
  • Visual Processing: Including a photo was one of the strongest predictors of success (OR 34.3). Images provide richer information that is easier for the public to recall and match against people they see on the street compared to textual descriptions.
Probability of finding a missing person over time, comparing the impact of photos (a) and police involvement (b).

Overall, 76.9% of the identified missing persons were found within a month. For those found, the average time-to-recovery was 21.4 hours. Interestingly, hashtags did not significantly improve the chances of finding someone, likely due to a lack of standardization (e.g., using #MissingPerson vs. #Missing[Name]), which fragments search results.

Limitations and Future Directions

The authors acknowledge several hurdles, including a small sample size of 54 unique cases and the difficulty of automated image analysis on low-resolution social media photos. Furthermore, the results are specific to Twitter's architecture; other platforms like Facebook might yield different results due to privacy settings.

To improve outcomes, the authors suggest that police departments should not just tweet to their own followers but should "push" these high-value tweets (with photos) to users specifically located in the area where the person was last seen. They also advocate for standardized hashtags to improve searchability across the platform.

Summary updated

BibTeX
@article{tsoi2018how,
  author = {Kelvin Kam-fai Tsoi and Nicholas B. Chan and Felix C. H. Chan and Lingling Zhang and Annisa C. H. Lee and Helen Mei-Ling Meng},
  title = {{How can we better use Twitter to find a person who got lost due to dementia?}},
  journal = {Npj Digit. Medicine},
  year = {2018},
  doi = {10.1038/s41746-018-0017-5},
}
Figure