HOUDINI-BD

  • Research type

    Research Study

  • Full title

    High-resolution Ongoing Unsupervised Data-efficient Identification of Novelties within Individuals – Bipolar Disorder

  • IRAS ID

    353941

  • Contact name

    Filippo Corponi

  • Contact email

    f.corponi@imperial.ac.uk

  • Sponsor organisation

    Imperial College London

  • Duration of Study in the UK

    1 years, 8 months, 28 days

  • Research summary

    Bipolar disorder (BD) is a common and severe mental health condition associated with early mortality and significant societal costs. BD manifests with recurring illness episodes, featuring profound disturbances in sleep-wake rhythms and energy levels alongside extreme mood swings. These episodes can severely disrupt lives, yet they are notoriously difficult to predict, delaying treatment and worsening outcomes. Current prevention strategies depend on self-reported symptoms and infrequent clinical visits, which often fail to detect early warning signs. Wrist-worn wearable devices have become increasingly popular and are equipped with sensors capturing data—e.g., physical activity, light exposure, and skin temperature—that is informative about the different phases of BD. These devices can monitor individuals continuously as they go about their daily lives. By leveraging artificial intelligence (AI), it is possible to identify patterns in wearable data, creating new opportunities for remote monitoring of BD. While past attempts to use AI in BD research have shown promise, they have not yet generated clinical applications, partly due to the heterogeneity of BD and the limited availability of large, diverse datasets for training AI models. This study will collect wearable data from patients with BD (N=40) at high risk of a new episode over 12 months. AI models will be developed to establish personalised baselines for each subject, detecting BD episodes as deviations from their typical patterns, e.g. disruptions in sleep-wake cycles. Just as a student should learn English before specializing in Shakespeare, I will first pre-train my models on existing wearable data datasets, regardless of the population they were collected from, and then refine them for BD. This principle underlies the success of large language models like ChatGPT.

  • REC name

    South West - Central Bristol Research Ethics Committee

  • REC reference

    26/SW/0051

  • Date of REC Opinion

    16 Apr 2026

  • REC opinion

    Further Information Favourable Opinion