Table 4.

Details of studies that used real-time analytics with personalised micro-interventions (n 10)

Author, year	Intervention	Intervention duration; run-in period*	Subjective data	Objective data	Feasibility and acceptability	Accuracy
Everett et al., 2018	Sweetech app – uses machine learning to automatically translate raw data streams originating from the patient’s mobile phone and into insights about the individual’s life habit and provides personalised, contextual, just-in-time, just-in-place, recommendations	12 weeks, NR	Demographic information, past medical history and medications. 4-item Physical Activity States of Change Questionnaire	Weight: digital body weighing scale (Bluetooth) Waist circumference: flexible measuring tape Phone accelerometer	86 % retention; Validated System Usability Scale: median 78 %; 74 % would like to use the Sweetch app; 83 % found the app easy to use; 72 % found the functions of the app well integrated, 89 % felt that most people could learn to use the app very quickly and 77 % felt confident using the app	NR
Forman et al., 2018	OnTrack + dietary weight loss programme called Weight Watcher (WW) – uses machine learning algorithm to automatically build models of lapse behaviour, predict lapses before they occur and delivers micro-interventions (messages) when lapse risk is high	8 weeks; 2 weeks	Behavioural risk factors and lapse behaviour (21: affect, boredom, hunger, cravings, tiredness, unhealthy food availability, temptations, missed meals/snacks, self-efficacy (confidence), motivation, socialising (with or without food present), TV, negative interpersonal interactions, healthy food presence, cognitive load, food cues (advertisements), hours of sleep, exercise, alcohol consumption, planning food intake, time of the day	Height and weights: calibrated scale and stadiometer	85·1 % completed; 70·15 % opened risk alerts; Technology Acceptance Model Scales (TAMS): M = 6·14, sd = 1·58 (app was easy to use); minimal technical issues (M = 2·91 out of 7, sd = 1·24); Participants rated the app as moderately useful (M = 4·64, sd = 1·58) and enjoyable (M = 4·37, sd = 1·62), with a somewhat positive behavioural intention to use (M = 4·48, sd = 1·86)	72 % accuracy, 70 % sensitivity and 72 % specificity, 80 % negative predictive value
Forman et al., 2019a	OnTrack + WW	10 weeks; 2 weeks	Behavioural risk factors and lapse behaviour (seventeen potential lapse triggers, i.e., affect, boredom, hunger, cravings, tiredness, unhealthy food availability, temptations, missed meals/snacks, self-efficacy, socialising, watching TV, negative interpersonal interactions, cognitive load, food cues/advertisements, hours of sleep, alcohol consumption, and planning food intake. Time of day, automatically measured, served as an 18th trigger.)	Weight: Yumani Smart Scale (Bluetooth)	64·4 % completed; 46·9 % opened risk alerts; TAMS: M = 4·70, sd = 1·52	69·2 % sensitivity; 83·8 % specificity
Forman et al., 2019b	AI-optimised interventions include individually optimised (i.e., at each of the 24 intervention points, participants receive the intervention with the highest reward score for them so far, except when the system is ‘exploring’) or group-optimised (i.e., interventions are assigned based on the highest possible total reward scores, across all interventions assigned, given a predetermined amount of total intervention time across all participants for the day)	16 weeks; NR	Energy intake: participants logged all food and beverages using the Fitbit mobile phone application	Weigh: Yumani Smart Scale (Bluetooth) Physical activity: measured in minutes of moderate-to-vigorous physical activity (MVPA) using a wrist-worn activity tracker	A short survey of coaches: the portal was easy to use (M = 3·33 out of 4) and able to effectively carry out the remote coaching (M = 3·33 out of 4); 76·5 % reported that the contact frequency was satisfactory	NR
Liu et al., 2015	SmartCare – an energy-efficient long-term physical activity tracking system that follows users’ physical activity habits and gives personalised quantitative health assessment and health regime suggestion	4 weeks; NR	Users’ daily physical activities and body type: nine basic daily physical activities: walking, jogging, ascending and descending stairs, bicycling, travelling up in an elevator, travelling down in an elevator, using an escalator, and remaining stationary	Smartphone built-in accelerometer and magnetometer	NR	98 % accuracy in physical activity recognition
Rabbi et al., 2015	MyBehaviour – (1) uses a combination of automatic and manual logging to track physical activity (e.g., walking, running, gym), user location, and food, (2) automatically analyse activity and food logs to identify frequent and nonfrequent behaviours and (3) generate personalised suggestions that ask users to either continue, avoid or make small changes to existing behaviours to help users reach behavioural goals	3 weeks, NR	Activity tracking and manual food logging either by selecting food items from a database or directly input energy information from nutrition labels. (Users can take photos of food as reminders to input energy intake)	Accelerometer and GPS	According to the suggestion-rating survey, participants in the experimental group had a significantly higher intention to follow personalised suggestions than those in the control group in following generic suggestions	NR
Goldstein et al., 2020	Two OnTrack versions – OnTrack-short (OT-S) (8 lapse trigger questions at each EMA survey) and OnTrack-long (OT-L) (17 lapse triggers questions at each EMA survey). When an EMA survey was completed, the algorithm classified responses as no risk (when a prediction was ‘no lapse’), low risk (probability of lapse > 40 %), medium risk (probability of lapse between 40 and 70 %) or high risk (probability of lapse > 70 %)	10 weeks; 2 weeks	Seventeen lapse triggers: affect, sleep, fatigue, hunger, motivation to adhere to a diet, cravings, boredom, temptation, cognitive load, confidence, socialising, television, negative interpersonal interactions, presence of tempting foods, food advertisements, planning food, alcohol, time/	NR	84·3 % completed; 65·4 % average EMA survey adherence in OT-S and 60·5% in OT-L	79·8% accuracy of lapse prediction (79·7% in OT-S v. 79·9% in OT-L); 74·5% sensitivity in OT-S v. 77·7% in OT-L; 83·1 % specificity (84·4% in OT-S v. 81·7% in OT-L)
Spanakis et al., 2017b	Think Slim – uses machine learning to predict unhealthy eating behaviour and allow users to report potential unhealthy eating promoting factors (emotions, activities, etc.). Emphasis is given to providing feedback before possible unhealthy eating events (i.e., warn users in the appropriate time manner using a classification algorithm) and to construct groups of eating behaviour profiles (using a clustering algorithm)	8 weeks; 1 week	Fifteen lapse triggers: date, food craving, seven emotions each measured on ten-point VAS scale (worried, angry/annoyed, stressed/tense/relaxed/at ease. Cheerful/happy, sad/depressed, bored), specific craving, location, activity, specific eating, thoughts regarding eating, food intake image/	NR	70·5 % completed	NR
Stein et al., 2017	Lark’s AI health – uses machine learning to power a Chatbot that mimic health professionals’ empathetic health counselling	16 weeks; NR	Weight loss, meal quality, physical activity and sleep data were collected through user input Data points were user-entered values for age, gender, height, weight, dietary intake, with self-reported anthropometric data and Web-reported diet intake/	Sleep and physical activity, partly through automatic detection by the user’s mobile phone. User engagement was assessed by duration and amount of app use	44·0 % active users by end of the intervention; In-app user trust survey: average scores for satisfaction, disappointment if not offered and health outcome were 7·9, 8·3 and 6·73	NR
Zhou et al., 2020	CalFit app – mobile phone app which delivers daily step goals using push notifications and allows real-time physical activity monitoring	10 weeks; 1 week	Socio-demographic information, self-reported medical history, Barriers to Being Active Quiz (twenty-one questions on a ten-point Likert scale on seven sub-areas: lack of time, social influence, lack of energy, lack of willpower, fear of injury, lack of skill, and lack of resources), International Physical Activity Questionnaire – Short Form/	Phone accelerometer	77·5 % retention	NR

NR, not reported; IMU, inertial measurement unit.

^*Included within intervention;