Closed loop learning through AI and smart devices to objectively measure the drivers of adult sleep and to make and measure interventions: Abstracts from the 16th World Sleep Congress, March 11-16, 2022 in Rome, Italy

Introduction: More than half of UK adults suffer from poor sleep but are not chronic insomniacs. Poor sleep has a negative impact on health and wellbeing – e.g. mental health, disease risk for stroke and dementia. Their ability to focus goes down, as does their sense of wellbeing. Treatment often targets ‘sleep hygiene’ – composed of 15 discrete factors. Individual sensitivity varies greatly on each of these dimensions – e.g. 100x differential in melatonin response to evening light (Philips et al, 2019). Currently there is no adaptive, targeted mechanism for providing sleep hygiene advice. I will use AI and smart devices to capture objective measures of biological variables such as movement, light exposure and behaviour and replace guesswork with personalised plans. This will help individuals to both understand and adapt their behaviour in a personalised, targeted manner. This will be the first study of its kind in sleep. Using AI to do this will create new challenges, like dealing with bias and how to explain it to all who use it, so I am also developing new tools to safeguard against these issues.

Materials and Methods: Our pilot uses AI to objectively assess sleep quantity and quality drivers in a closed loop, and to make personalised recommendations. Supervised machine learning and neural networks are used to analyse data collected from 4 smart devices in 28-day cycles from healthy adults (no insomnia or diagnosed mental health disorders). Our target variables are TST, SWS and SE. Our devices provide feature data on light, temperature, sound, stress, exercise, consumption as well as sleep behaviours like wake time consistency and napping. AI is used on baseline sleep to prioritise the influence of these drivers and to identify thresholds and sleep phenotypes. AI is used to identify target interventions based on an individual’s baseline and likely responsiveness. The intervention is delivered and monitored through an app which incorporates 12 behaviour change techniques: 5 of 6 highest impact digital behaviour change techniques (BCTs) (Webb et al, 2010) & all the best sleep hygiene for healthy adults (Murawski et al, 2018). AI will learn which interventions and BCTs get the best response, generating phenotypes based on reaction to intervention. Using AI to do this creates new challenges. For example, such novel research risks bias because differences are unknown. We recruit a diverse sample and use AI to augment data to deal with differences (stratified oversampling).

Results: We will show case our data collection and intervention delivery app as well as report initial results on driver priority, thresholds, sleep phenotypes, and plan for intervention targeting.

Conclusions: Individuals differ greatly in what drives their sleep quantity and quality, and their response thresholds. Objective results are very different to a priori subjective prioritisation. Clear sleeping behaviour phenotypes exist and can simplify intervention identification.