Predicting Loneliness Using Machine Learning and Self-Logged Behavioural Data

Loneliness is a growing public health concern, particularly among older adults, and has been linked to adverse physical and mental health outcomes. This study presents a machine learning approach to predict levels of loneliness using behavioural and emotional data collected from 124 participants through a mobile phone application over a 71-day period. The dataset includes 27 features derived from self-logged information such as wellbeing scores, mood fluctuations, and time spent in various home locations. Feature selection was applied to identify the most discriminative indicators, with classification and regression models evaluated using both Support Vector Machine (SVM), and Random Forest (RF). We applied feature selection to identify the most discriminative indicators and evaluated both Support Vector Machine (SVM) and Random Forest (RF) models for classification and regression. The highest classification accuracy - 69.19% on a 7-point loneliness scale - was achieved using a five-fold SVM with the top 13 features. In the regression task, the best performance was observed using 26 features, resulting in a minimum Mean Squared Error (MSE) of 0.6752. These findings indicate that a selected subset of behavioural and emotional features can offer a meaningful estimation of loneliness levels. This has potential to inform the design of real-time, personalised digital tools aimed at identifying and supporting individuals at risk of loneliness.