Social media has long been a space to share life’s highlights and everyday struggles. A post about feeling overwhelmed, a discussion thread about stress or a comment venting frustration all offer glimpses into how people are coping emotionally with difficult events.
YYet it can be challenging to assess population-wide mental health needs without timely, accessible and effective indicators, noted Yinping Yang, a Senior Principal Scientist at the A*STAR Institute of High Performance Computing (A*STAR IHPC).
The COVID-19 pandemic saw 6,600 calls to Singapore's National Care Hotline within two months, and a record high of reported suicides in 2020. These highlighted the need for proactive ways to detect emerging mental health issues, as demand for support can rapidly outpace available resources in a time of crisis.
Yang and A*STAR IHPC colleagues including Senior Scientist Chitra Panchapakesan, Senior Research Engineers Nur Atiqah Othman, Brandon Loh and Mila Zhang, and Principal Scientist Raj Kumar Gupta turned to social media to fill the information gap. Working with collaborators from Singapore’s Ministry of Health (MOH), MOH Office for Healthcare Transformation (MOHT) and Institute of Mental Health (IMH), they investigated whether emotions expressed in public posts or ‘tweets’ on Twitter (now known as X) could be early indicators of rising public mental health needs.
“Mental health conditions often go undetected unless individuals actively seek help, which introduces significant underreporting,” said Mythily Subramaniam, Assistant Chairman of IMH’s Medical Board (Research) and study collaborator. “Traditional survey methods—while valuable—face logistical limits and often only capture people’s feelings after a crisis unfolds.”
To effectively decode emotions online for assessing mental health needs, the team used CrystalFeel, an in-house emotion analysis engine developed at A*STAR IHPC, to analyse 2.5 years of local tweets, filtering out advertisements and influencer content. CrystalFeel counted, measured and classified the intensity of four primary emotions—fear, anger, sadness and joy—based on language use across 140,598 tweets.
“These emotional indicators were then compared with two key outcome indicators of mental health needs: the number of mindline.sg website users showing signs of crisis, and IMH emergency visits,” said Othman and Panchapakesan.
The team found that changes in emotional expressions in tweets significantly enhanced the predictions of mindline crisis and IMH visits. Joy intensity and anger count strongly predicted IMH visits, while sadness count, joy intensity, anger count and joy count did likewise for mindline crisis states. In contrast, situational indicators, such as COVID-19 case numbers, were less effective.
“MOHT started mindline in response to COVID-19, and the stresses on all walks of life were evident,” said Robert Morris, MOHT Chief Technology Strategist. “We subsequently saw changes in psychological wellness as the pandemic waxed and waned, but this study’s extra signals from social media made the effects and their nature much clearer.”
“Predictive models based on such new tools could enable authorities to make more proactive, informed decisions in resource allocation, such as staffing plans in future crises,” added Kelvin Bryan Tan, MOH Principal Health Economist.
The A*STAR-affiliated researchers contributing to this research are from the A*STAR Institute of High Performance Computing (A*STAR IHPC).
