An algorithm that monitors Twitter for breathing problem keywords could detect a thunderstorm asthma event nine hours ahead of official reports, its CSIRO developers say.
Scientists tested several different keyword algorithms on retrospective social media posts covering the period around the catastrophic November 2016 event in Melbourne that saw 8000 people present to emergency departments with respiratory difficulties, and ten asthma-related fatalities.
They found that three of their 18 models were able to filter through social media posts and avoid false alarms to detect a thunderstorm asthma trending even before reports made it into the media.
Using artificial intelligence and natural language processing analytics systems, they developed a four step algorithm that was able to monitor Tweets in real time and provide alerts when it detected a spike in respiratory incidents being reported by social media users.
The algorithm was designed to analyse tweets that mention variations of words such as breath, cough and hayfever as well as those containing terms such as difficult, unable or cannot.
The four steps algorithm involved initial selection of Tweets, personal health mention classification, duplicate removal and time between events.
Writing in Epidemiology journal, the developers led by Aditya Joshi from CSIRO Sydney said the main challenge of any event monitoring system was to filter out “alert swamping” so as to avoid false positive alerts.
In their modelling study they identified three datasets that were able to quickly and accurately identify mentions of respiratory problem in tweets.
When applied to the tweets of 21 November 2016 they found three sets in the algorithm would have been able to create an alert nine hours before the first official report of thunderstorm asthma was made at 6pm. Five of them were able to raise an alert ahead of media mentions of the emerging thunderstorm asthma events.
“While individual doctors or hospitals may have seen individual patients, an overall population picture of a severe event only became apparent to authorities as the evening progressed and ambulance services and emergency departments became stressed,” they wrote.
“We report a 9-hour advantage. In such a rapid onset event, the advantage is potentially valuable to health authorities, as it could assist with surge capacity preparation and also with putting out public alerts to warn asthmatics to remain indoors
They said that if the algorithm was validated in further testing, such social media monitoring algorithms could be used to give health authorities early warning of other emerging disease outbreaks such as those caused by influenza or Zika virus.
“Timeliness can be critical in the case of acute disease events. We show that harnessing open source data such as social media can be automated to alert health authorities to a potential event prior to official notification,” they concluded.