Haley, Fintan, Andrews, Jacob, Moghaddam, Nima and Turner, Alex
(2025)
Forecasting depressive symptom deterioration using wearable sensor data and LSTM models: a longitudinal analysis from the RADAR-MDD Study.
DClinPsy thesis, University of Nottingham.
Abstract
Major Depressive Disorder (MDD) is a highly prevalent and disabling mental
health condition, with early detection of symptom deterioration offering
opportunities for timely intervention and improved outcomes. Advances in
wearable technology and machine learning have prompted growing interest in
passive monitoring as a scalable means of identifying individuals at risk of
worsening depression. This thesis investigates the feasibility of forecasting
depressive symptom deterioration using wearable sensor data and Long Short-
Term Memory (LSTM) models, leveraging data from the RADAR-MDD study—a
large, multi-site, longitudinal cohort of individuals with a history of MDD (N =
623).
The primary aim was to develop a predictive model capable of identifying
clinically significant deterioration, defined as a ≥5-point increase on the Patient
Health Questionnaire-8 (PHQ-8), using passive data streams (e.g., sleep, step
count, heart rate) collected via Fitbit devices. Despite rigorous data
preprocessing, normalization, and class-balancing procedures, the LSTM model
failed to correctly identify any cases of symptom deterioration, achieving an
AUC-ROC of 0.50, F1 score of 0.00, and high specificity driven by substantial
class imbalance. A secondary logistic regression analysis using a reduced
feature set similarly failed to demonstrate predictive utility, suggesting that the
limitations lay not in the modelling technique but in the quality and
completeness of the data.
Contributing factors to poor model performance included significant data
sparsity—particularly in sleep data—declining participant adherence over time,
and the bidirectional nature of certain behavioural indicators (e.g., sleep
duration). Furthermore, exploratory analyses revealed significant cultural
variation in PHQ-8 response patterns across the UK, Spain, and the Netherlands, raising concerns about the universal applicability of standardised
depression measures and global prediction models.
The findings highlight the current limitations of population-level machine
learning approaches in forecasting depression and support a shift toward
idiographic, context-sensitive modelling strategies. The thesis also examines
the psychological, ethical, and practical implications of continuous monitoring
technologies in mental health, emphasising the need for user-centred, culturally
informed, and ethically robust design. Recommendations for future research
include improved data collection strategies, adaptive feedback systems, and the development of person-specific early warning systems that align more closely with clinical realities.
| Item Type: |
Thesis (University of Nottingham only)
(DClinPsy)
|
| Supervisors: |
Moghaddam, Nima
Turner, Alex |
| Keywords: |
Major Depressive Disorder, Wearable Devices, Machine Learning, LSTM, Digital Mental Health, RADAR-MDD, Prediction Models, Cultural Variation |
| Subjects: |
W Medicine and related subjects (NLM Classification) > WM Psychiatry |
| Faculties/Schools: |
UK Campuses > Faculty of Medicine and Health Sciences > School of Medicine |
| Item ID: |
82500 |
| Depositing User: |
Haley, Fintan
|
| Date Deposited: |
10 Dec 2025 04:40 |
| Last Modified: |
10 Dec 2025 04:40 |
| URI: |
https://eprints.nottingham.ac.uk/id/eprint/82500 |
Actions (Archive Staff Only)
 |
Edit View |
Tools
Tools
