Investigating the Generalisability of Machine Learning Algorithms for Classifying Mental Illnesses in Low- and Middle-Income Countries Using Multimodal Data from Smartphones and Social Media

Mental health disorders, particularly neurodevelopmental and behavioral conditions, are a major health concern globally, with high prevalence rates in low- to middle-income countries like India. Access to mental health care is often limited in these countries due to factors such as poor health infrastructure, lack of skilled mental health workers, and social stigma. As a result, many patients fail to receive timely intervention, which can worsen their conditions and increase challenges in managing mental health over time. The widespread use of smartphones and social media presents an opportunity to address these challenges. These digital platforms generate vast amounts of data on daily activity, social interaction, and online behavior, which can provide valuable information about individuals' mental health. The purpose of this study is to investigate how such digital information can be harnessed to classify mental disorders, with a particular focus on ADHD and eating disorders, especially in regions where traditional healthcare services are scarce or difficult to access.

This PhD study investigates methodologies to analyze and detect patterns associated with a range of mental and neurodevelopmental disorders using both social media and smartphone sensor data. The first study uses machine learning algorithms like CNN and Word2Vec to classify mental health conditions including anxiety, autism, schizophrenia, depression, bipolar disorder, and borderline personality disorder based on information gathered from social media platforms such as Twitter and Reddit. Large datasets were collected from Reddit and Twitter and analyzed to develop models capable of identifying patterns associated with these disorders. The results showed consistent model performance across platforms, with positive evaluation scores such as precision, recall, and F1-score validating the effectiveness of the classification methods.

The second study addresses the shortcomings of lab-based research in assessing neurodevelopmental disorders like ADHD by using smartphone sensor data for a more objective assessment. Accelerometer, location tracking, application usage, and smartphone interaction data (e.g., keyboard and touch gestures) were collected from 43 participants, 21 with ADHD and 22 without. Analysis revealed significant differences in attention-related activity patterns between the groups, with variations in attentiveness and interaction patterns serving as strong indicators of ADHD. These findings suggest that smartphone sensor data can effectively classify ADHD-related behaviors and provide insights into attentiveness patterns. The third study focuses on the assessing patterns associated with of eating disorders using smartphone sensor data. Accelerometer and location information, along with application use, keyboard interactions, and touch movements, were collected from 45 participants: those without eating disorders, those with moderate disorders, and those with severe disorders. Statistical analysis revealed unique behavioral patterns, particularly during meal sessions, such as longer periods of screen inactivity and abnormal touch movements in individuals with severe eating disorders. Machine learning models were able to detect these patterns, accurately classifying eating disorders by severity. This study also explored practical and methodological challenges in smartphone-based mental health research in low- and middle-income countries, including participant recruitment amid limited smartphone availability, variability in device performance and connectivity, and cultural and geographic differences in digital behavior and expression of mental health. Strategies such as adaptive data collection methods and culturally sensitive study designs were discussed to improve the reliability, accessibility, and impact of smartphone-based mental health research in resource-scarce environments.

This PhD study demonstrates that digital data from social media and smartphones can aid in the classification of neurodevelopmental and mental disorders, with particular emphasis on ADHD and eating disorders. These methodologies offer non-invasive, affordable, and scalable means of detecting health-related patterns. The findings indicate that digital data can play a key role in improving early identification and supporting interventions, especially in resource-limited settings where access to mental health professionals and clinical resources is constrained.