A Deterministic N-of-1 Pipeline for Multimodal Digital Phenotyping in ADHD and Bipolar Disorder

Marques Teixeira, Rodrigo (2025) A Deterministic N-of-1 Pipeline for Multimodal Digital Phenotyping in ADHD and Bipolar Disorder. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (1MB) | Preview
Preview	PDF (Configuration Manual) Download (357kB) | Preview

Abstract

Attention-Deficit/Hyperactivity Disorder (ADHD) and Bipolar Disorder (BD) are characterised by marked intra-individual variability in sleep, activity and cardiovascular signals. Digital phenotyping offers a promising opportunity to capture such temporal patterns, but existing work often lacks deterministic pipelines, transparent heuristics, leakage-resistant evaluation, and long-term N-of-1 designs.

This research presents a fully deterministic, ten-stage N-of-1 pipeline operating over eight years of multimodal wearable data. State of Mind (SoM) — a user-reported three-class measure of daily subjective stability exported from Apple Health — serves as the primary supervised target. The Personal Behaviour Stability Index (PBSI), a transparent physiological composite derived from sleep, heart-rate and activity metrics, is incorporated only as an auxiliary feature. The modelling framework includes: (i) ML6 baseline with Logistic Regression under calendar-based cross-validation; (ii) ML7 baseline with LSTM for sequence modelling; (iii) ML6-Extended comparing ten classical algorithms (Random Forest, XGBoost, LightGBM, Gradient Boosting, SVM, Gaussian Naive Bayes, KNN); and (iv) ML7-Extended evaluating four deep learning architectures (LSTM, GRU, Conv1D, CNN-LSTM). All stages incorporate strict anti-leak safeguards, drift detection (ADWIN, KS) and SHAP interpretability.

Although the full dataset spans 2017–2025, predictive modelling was restricted to 77 SoM-labelled days (2024–2025) due to sparse user-reported labels. Results show that Random Forest achieved the best classical performance (macro-F1 ≈ 0.52), while LSTM was the best sequence model (macro-F1 ≈ 0.50). All models showed moderate performance, reflecting the challenging nature of predicting user-reported stability from sparse labels.

These findings demonstrate the feasibility of predicting a user-defined behavioural stability construct (SoM) under strict anti-leak conditions. The deterministic pipeline may serve as a template for future N-of-1 digital phenotyping studies with reproducible and transparent analytical workflows.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Mattos, Agatha UNSPECIFIED
Subjects:	R Medicine > RC Internal medicine > RC0321 Neuroscience. Biological psychiatry. Neuropsychiatry Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence R Medicine > RA Public aspects of medicine > RA790 Mental Health
Divisions:	School of Computing > Master of Science in Artificial Intelligence for Business
Depositing User:	Ciara O'Brien
Date Deposited:	03 Sep 2026 08:39
Last Modified:	03 Sep 2026 08:39
URI:	https://norma.ncirl.ie/id/eprint/9781

Actions (login required)

View Item