Evaluating Loneliness Proxy Elicitation via Digital Mental Health Platform Data Using Transformer-Based Natural Language Processing

Abstract

Introduction & Background
It has been estimated that over 9 million people in the UK feel lonely either often or all the time. Loneliness is difficult to measure due to its latent nature and associated stigma.

This work demonstrates the application of transformer-based natural language processing (NLP) to model themes of loneliness and identify potential proxy terms for loneliness within digital footprint data on Kooth, a digital mental health platform (DMHP). Kooth provides access to an online community of peers, a team of experienced counsellors, and a group of emotional well-being practitioners. The study aims to derive insight into themes related to the elicitation of loneliness by Kooth service users (SUs).

Objectives & Approach
This research aims to demonstrate how transformer-based topic modelling can be combined with digital footprint data from DMHPs to identify direct and indirect elicitations of loneliness. While previous research has evaluated the automated inference of loneliness, this work has typically been limited to social media platform analysis, and such studies rarely translate to improvements in mental health practice. By focusing on loneliness elicitation within a digital mental health forum, this study provides a nuanced understanding of how SUs express loneliness concerns, and these findings can directly inform and enhance service delivery, advancing beyond traditional social media-based loneliness analysis.

Relevance to Digital Footprints
The data comprised forum posts (n=140,027) from Kooth SUs (n=22,631). On the Kooth platform, SUs are labelled with ‘presenting issues’ - discrete categories of mental health concerns elicited either through one-on-one text-based therapeutic sessions or via forum posts.

Results
Based on the ’loneliness’ presenting issue, the dataset was divided into ‘lonely’ (n=19,720) and ‘non-lonely’ (n=120,307) posts. Initial proxy terms for loneliness were developed through consultations with practitioners identifying key concepts such as ‘feeling unheard,’ ‘feeling withdrawn,’ and ‘friendship.’ These terms informed the subsequent transformer-based topic modelling analysis, which identified additional themes that were characteristic of posts from users with loneliness as a presenting issue.

Conclusions & Implications
This analysis of digital mental health footprint data extends previous forum data research approaches while deepening our understanding of loneliness-associated language patterns.

The transformer-based modelling approach enables the evaluation of a considerable volume of data faster than manual review processes while crucially creating a meaningful bridge between research and practice, delivering actionable insights that can directly inform therapeutic interventions.