Towards automated symptoms assessment in mental health

Activity and motion analysis has the potential to be used as a diagnostic tool for mental disorders. However, to-date, little work has been performed in turning stratification measures of activity into useful symptom markers. The research presented in this thesis has focused on the identification of objective activity and behaviour metrics that could be useful for the analysis of mental health symptoms in the above mentioned dimensions. Particular attention is given to the analysis of objective differences between disorders, as well as identification of clinical episodes of mania and depression in bipolar patients, and deterioration in borderline personality disorder patients. A principled framework is proposed for mHealth monitoring of psychiatric patients, based on measurable changes in behaviour, represented in physical activity time series, collected via mobile and wearable devices. The framework defines methods for direct computational analysis of symptoms in disorganisation and psychomotor dimensions, as well as measures for indirect assessment of mood, using patterns of physical activity, sleep and circadian rhythms. The approach of computational behaviour analysis, proposed in this thesis, has the potential for early identification of clinical deterioration in ambulatory patients, and allows for the specification of distinct and measurable behavioural phenotypes, thus enabling better understanding and treatment of mental disorders.

💡 Research Summary

The thesis “Towards Automated Symptoms Assessment in Mental Health” presents a comprehensive framework for objectively measuring and classifying mental‑health symptoms using data from mobile phones and wearable devices. Recognising that current psychiatric assessment relies heavily on structured interviews and self‑report questionnaires—methods that are often subjective and insufficient—the author seeks to translate quantifiable activity and physiological signals into reliable symptom markers.

Data were collected from more than 100 participants, including healthy controls, patients with bipolar disorder (BD), borderline personality disorder (BPD), and a separate cohort of schizophrenia patients. Participants wore wrist‑mounted accelerometers and used a custom smartphone app that recorded continuous activity, sleep–wake cycles, and heart‑rate (HR) data. A substantial portion of the thesis is devoted to data‑quality management: noise filtering, bias removal, handling missing entries, and segmenting the raw time series into meaningful periods (stationary activity, sleep, wake, bed‑wake) using methods such as Recursive Mean Difference Maximisation, Bayesian Blocks, Bayesian Online Change‑Point Detection, Hidden Markov Models, and Explicit‑Duration Semi‑Markov Models.

The author structures symptom analysis into three clinically relevant dimensions derived from Liddle et al.’s five‑factor model: (1) psychomotor, (2) disorganisation, and (3) mood (captured via sleep and circadian rhythms). For each dimension a rich set of features is engineered:

• Psychomotor – Euclidean Norm Minus One (ENMO), epoch‑based activity counts, metabolic‑equivalent intensity, distributional statistics of activity levels.
• Disorganisation – multiscale entropy, detrended fluctuation analysis (scaling exponent), day‑to‑day activity variability, activity persistence metrics.
• Mood – L5/M10 (least/most active 5‑hour windows), sleep‑wake segmentation, circadian amplitude, non‑parametric rest‑activity characteristics.

Feature selection combines naïve ranking, Minimum Redundancy Maximum Relevance (mRMR), and LASSO regularisation to reduce redundancy and highlight the most predictive variables. Classification is performed with logistic regression and Support Vector Machines, evaluated via leave‑one‑out cross‑validation to ensure robustness on small sample sizes.

Key performance results include:

• Differentiation between healthy controls and BD (67 % accuracy), controls and BPD (70 %), and BD vs. BPD (80 %).
• Classification of clinical states within BD: euthymia vs. mania (80 %), euthymia vs. depression (85 %), mania vs. depression (90 %).
• Personalized mood‑model regression achieving mean absolute errors of 1.36–3.32 points on standard psychiatric scales (the clinically relevant range is 4–5 points).
• When heart‑rate features are combined with locomotor features, classification accuracy improves by roughly 10 % over locomotor‑only models and by 17 % over heart‑rate‑only models, indicating complementary information from autonomic and motor systems.

The thesis also demonstrates the transferability of the approach to a schizophrenia cohort, achieving 95.3 % accuracy in distinguishing patients from matched controls, confirming that both activity and physiological signals contribute to diagnostic power.

Limitations are candidly discussed: modest sample sizes, potential bias in recruitment, the need for stronger causal links between behavioural metrics and underlying psychopathology, and the variability of sensor reliability in real‑world settings. Future work is outlined to include multi‑site, larger‑scale studies, real‑time feedback loops in mHealth platforms, and integration with clinical decision support systems to enable early detection of deterioration and personalized intervention.

In sum, this work advances the field of digital psychiatry by providing a rigorously validated, multimodal data pipeline that translates raw sensor streams into clinically meaningful symptom scores, offering a promising route toward objective, scalable, and early mental‑health monitoring.