Disaster medicine & AI: Improving predictive modelling of human disaster behaviour with machine learning

Lead Research Organisation: University College London
Department Name: Institute of Health Informatics

Abstract

Project lay summary:
Disaster modelling would be improved if we could better understand and predict individual human behaviour in a catastrophic event. While current research predominantly focuses on improving situational awareness and resource allocation during catastrophes, there is little attention given to the actions of regular citizens in the period before first-responders arrive. Yet, extensive re-search in disaster medicine indicates that it is the actions of these ordinary citizens which have the greatest impact on overall disaster survival.

Researchers have previously demonstrated that all individuals have a 'disaster personality' - a persona that they adopt during a crisis and the behaviour that they enact. 'Disaster personalities' can be modelled in advance and therefore used to predict the reactions and movements of indi-viduals following a human/natural disaster. Data from social media platforms provides an oppor-tunity to use online digital activity to map these disaster personalities on a population scale, such insights have the potential to vastly improve the predictive modelling of disaster events and plan interventions accordingly.

Project details:
(i) Aims and Objectives: To use sentiment analysis and text mining to model online disaster reactions and improve our understanding of crisis-related human behaviour.
(ii) EPSRC's Research Area: Artificial intelligence Technologies

4 (i) Methodology & (ii) The research methodology, including new knowledge or tech-niques in engineering and physical sciences that will be investigated
1. Data Collection - A database of 'disaster reactions' will be formed from online Twitter activity produced in the immediate aftermath of a disaster. Using the Twitter API and the Tweepy package, tweets with the relevant hashtag can be extracted with the data/timestamp of when the event took place. We will also collect other relevant tweets that use alternative hashtags or no hashtags at all. Ethics approval will be sought for obtaining this data.
2. Data Analysis - We propose using clustering algorithms to form sentiment subgroups, as demonstrated in previous research that has used machine learning for opinion mining. We will undertake a temporal sentiment analysis using the most relevant methods in recent research.
3. Data Interpretation - Creating sentimental sub-clusters will allow us to identify the categories of reactions that humans have to disasters. By exploring the themes within these sub-clusters, we can observe recurring ideas that exist within different sentimental reactions. This information can inform our understanding of how humans may differ in their disaster psycholo-gy, personality and behaviour.

5. Supervisor details: Professor Ingemar Cox, Dr Vasileios Lampos, Professor Ilan Kelman

6. Companies or collaborators involved: N/A

Publications

10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/S021612/1 01/04/2019 30/09/2027
2418758 Studentship EP/S021612/1 28/09/2020 30/09/2024 Isabel Marie Straw