Improving healthy adult sleep with AI

Abstract

Short sleep — defined as less than seven hours of Total Sleep Time (TST) — and the age-related decline of Slow-Wave Sleep (SWS) are prevalent among mid-life adults, with significant health implications including elevated risk of chronic disease, cognitive impairment, and mortality. The main intervention for those who are not chronic insomniacs is sleep hygiene (SH), a long list of behavioural and environmental practices. Although SH interventions can be effective, they lack objective measurement, personalised prioritisation, and integration of behaviour change techniques (BCTs) that are effective in similar contexts. Artificial Intelligence (AI) could bridge this gap, but AI systems lack trust infrastructure to support user adoption.

We design, develop, and test a novel sociotechnical system to address these limitations, integrating smart-device data, a new Neuro||Symbolic AI engine, a BCT-enabled user interface delivering personalised SH recommendations, and a monitoring module. Trust is discerned through the new Leap of Faith (LoF) matrix and demonstrated and deserved trust metrics: Demonstrated Action Follow-through Index (DAFTI) and Deserving Of Trust Index (DOTI), as well as Demonstrated Intention Relative Trust Index (DIRTI).

The system was tested in a pilot study with ten healthy mid-life adults over three months. Results demonstrate that it is feasible to deliver objective, personalised, trusted, AI-informed BCT-supported SH interventions that improve sleep duration. Clinically meaningful improvements were observed: average TST increased by 24 minutes [95% confidence interval (CI): 12, 36], SWS by 8 minutes [-3, 18], and REM by 12 minutes [4, 19].

This thesis contributes to the fields of computer, sleep, and behaviour science by establishing the feasibility of the novel system, and makes other original theoretical, methodological, engineering, and interdisciplinary contributions. The LoF matrix provides real-time transparency to model users when making a decision: they should be inclined to trust where the models agree, and understand that a leap of faith is required where they disagree. DIRTI, DAFTI and DOTI link user intention, action, and outcome, real time. The implementation offers a replicable template for interventions that combine AI-driven personalisation with real-time delivery. Future work should further establish method reliability, assess causality and long-term adherence, and consider real-world deployment challenges.

Date of Award	25 Jun 2025
Original language	English
Awarding Institution	University of Bath
Supervisor	Julian Padget (Supervisor), George Stothart (Supervisor) & Elizabeth Coulthard (Supervisor)