A scoping review of methodologies for applying artificial intelligence to physical activity interventions

Highlights

• •Artificial intelligence (AI) models were generally effective for physical activity (PA) promotion (16 studies), outcome prediction (7 studies), and pattern recognition (1 study).
• •Twelve studies found AI-driven interventions, such as mobile apps, recommendation systems, and chatbots improved PA outcomes compared to traditional approaches.
• •An increasing trend was observed of adopting state-of-the-art deep learning and reinforcement learning models over standard machine learning.
• •Six key areas were identified for future AI adoption: personalized interventions, real-time monitoring and adaptation, multimodal data integration, evaluating effectiveness, expanding access, and preventing injuries.
• •Exploring emerging AI-driven strategies is essential for optimizing PA interventions and promoting public health.

Abstract

Purpose

This scoping review aimed to offer researchers and practitioners an understanding of artificial intelligence (AI) applications in physical activity (PA) interventions; introduce them to prevalent machine learning (ML), deep learning (DL), and reinforcement learning (RL) algorithms; and encourage the adoption of AI methodologies.

Methods

A scoping review was performed in PubMed, Web of Science, Cochrane Library, and EBSCO focusing on AI applications for promoting PA or predicting related behavioral or health outcomes. AI methodologies were summarized and categorized to identify synergies, patterns, and trends informing future research. Additionally, a concise primer on predominant AI methodologies within the realm of PA was provided to bolster understanding and broader application.

Results

The review included 24 studies that met the predetermined eligibility criteria. AI models were found effective in detecting significant patterns of PA behavior and associations between specific factors and intervention outcomes. Most studies comparing AI models to traditional statistical approaches reported higher prediction accuracy for AI models on test data. Comparisons of different AI models yielded mixed results, likely due to model performance being highly dependent on the dataset and task. An increasing trend of adopting state-of-the-art DL and RL models over standard ML was observed, addressing complex human–machine communication, behavior modification, and decision-making tasks. Six key areas for future AI adoption in PA interventions emerged: personalized PA interventions, real-time monitoring and adaptation, integration of multimodal data sources, evaluation of intervention effectiveness, expanding access to PA interventions, and predicting and preventing injuries.

Conclusion

The scoping review highlights the potential of AI methodologies for advancing PA interventions. As the field progresses, staying informed and exploring emerging AI-driven strategies is essential for achieving significant improvements in PA interventions and fostering overall well-being.

Graphical Abstract

1. Introduction

Regular physical activity (PA) plays a crucial role in promoting overall health, with benefits including reduced risk of chronic diseases,¹ improved mental health,² and enhanced cognitive function in older adults.³ These benefits of PA are supported by several systematic reviews and meta-analyses, which have consistently found regular PA to be associated with improved health outcomes across a range of measures.^4,5 Yet, physical inactivity is pervasive—a sizable portion of the population falls short of the recommended Physical Activity Guidelines.⁶ Furthermore, simply meeting these guidelines might not be sufficient to mitigate the health risks associated with sedentary behavior.⁷ Therefore, it is essential to encourage people to increase PA levels and reduce sedentary behavior for optimal health outcomes.

Artificial intelligence (AI) is a broad term referring to machines or computers that mimic human intelligence.⁸ Machine learning (ML), a subdomain of modern AI, creates rules based on training data.⁹ Deep learning (DL), a subset of ML, employs artificial neural networks to model complex patterns in large-scale data.^10,11 Reinforcement learning (RL), another ML subdomain, draws on psychology and engineering concepts to enable autonomous learning in dynamic environments.^12,13 These have become essential tools in health sciences, with diverse applications ranging from disease outbreak prediction to medical imaging, patient communication, and behavioral modification.14, 15, 16, 17, 18, 19 Over the past decade, there has been a surge of scientific literature adopting AI in health research.^20,21 These investigations have applied an extensive array of AI models—from “shallow” ML algorithms like decision trees and k-means clustering to “deep” neural networks and RL²²—utilizing various data sources (e.g., clinical, observational) and types (e.g., tabular, text, image).²³

In the context of PA interventions, AI holds the potential to revolutionize the way health professionals design and implement these programs. Integrating AI in PA interventions has facilitated personalized and adaptive recommendations, real-time feedback, and data-driven insights that promote adherence and optimize outcomes. The potential of AI to promote PA manifests in several ways. It can customize interventions by interpreting individual data and recommending specific fitness routines, bolster positive behaviors by providing real-time feedback, and expedite broader interventions by pinpointing demographic groups that would gain most from increased PA.^24,25 These capabilities are further amplified by the advent of AI-powered wearable devices and mobile applications that track, monitor, and analyze PA data, extending personalized advice for enhancing fitness levels.²⁶ Another emerging trend is the advent of AI-driven virtual coaches and chatbots that provide motivation and guidance through personalized exercise regimens.²⁷ The interdisciplinary nature of AI and its potential for transforming PA interventions underscores the importance of scoping reviews that synthesize the available literature and methodologies to facilitate further advancements in the field.

Two recent studies reviewed the applications of AI in health behavior promotion. A systematic review by Oh et al.²⁸ found that AI chatbots showed promise in increasing PA but called for standardization in designing and reporting interventions. Aggarwal et al.²⁹ reported the high efficacy of AI chatbots in promoting healthy lifestyles but emphasized the need for robust randomized control trials to establish definitive conclusions. This article represents the first methodological review of the applications of AI in PA interventions, and it has 2 primary objectives. First, we aimed to summarize and categorize AI methodologies employed in the existing PA literature with the intent of identifying trends, synergies, and patterns that could guide future investigations. Second, we provided a concise overview of fundamental AI techniques in the realm of PA, enhancing clarity and promoting their incorporation into PA research. The uniqueness of this review lies not in the intervention health outcome, but rather in its focus on applying diverse AI methodologies within the realm of PA.

2. Methods

This scoping review was carried out following the guidelines of the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) extension for scoping reviews (PRISMA-ScR).³⁰ The review was registered on PROSPERO (registration No. CRD42023389962).

2.1. Study selection criteria

Studies that met all of the following criteria were included in the review: (a) study design: experimental studies (e.g., randomized controlled trials, pre–post interventions, and cross-over trials); (b) analytic approach: use of AI, including ML, DL, and RL in measuring, predicting, or intervening in PA-related outcomes; (c) study subjects: humans of all ages; (d) outcomes: PA measures (e.g., daily steps, duration, frequency, or intensity of activity), which could include structured exercise as well as other forms of PA; (e) article type: original, empirical, and peer-reviewed journal publications; (f) time window of search: from the inception of an electronic bibliographic database to February 20, 2023; and (g) language: articles written in English.

Studies were excluded from the review if they met any of the following criteria: (a) studies with a focus on outcomes unrelated to PA (e.g., diet, sleep); (b) studies employing a rule-based (“hard-coded”) approach instead of example-based ML, DL, or RL; (c) non-English language articles; and (d) letters, editorials, study or review protocols, case reports, or review articles.

2.2. Search strategy

A keyword search was performed in 4 electronic bibliographic databases: PubMed, Web of Science, Cochrane Library, and EBSCO. The search algorithm included all possible combinations of keywords from the following 3 groups: (a) “artificial intelligence”, “computational intelligence”, “machine intelligence”, “computer reasoning”, “machine learning”, “deep learning”, “neural network”, or “reinforcement learning”; (b) “exercise”, “motor activity”, “sport”, “physical fitness”, “physical exertion”, “physical activity”, “physical inactivity”, “sedentary behavior”, “sedentary lifestyle”, “inactive lifestyle”, “active living”, “active lifestyle”, “outdoor activity”, “walk”, “walking”, “running”, “bike”, “biking”, “bicycle”, “bicycling”, “cycling”, “stroll”, “strolling”, “active transport”, “active transportation”, “active transit”, “active commuting”, “travel mode”, or “physically active”; and (c) “intervention”, “program”, “trial”, “treatment”, “effect”, or “impact”. The medical subject heading terms “artificial intelligence” and “exercise” were included in the PubMed search. Supplementary Material 1 presents the search algorithm used in PubMed, Web of Science, Cochrane Library, and EBSCO. Two authors (JS and JW) independently screened the title and abstract for the articles found through the keyword search, obtained potentially relevant articles, and reviewed their full texts. The inter-rater agreement between these 2 authors (JS and JW) was evaluated using Cohen's kappa (κ = 0.83). Disagreements were settled through conversation.

2.3. Data extraction and synthesis

Using a standardized data extraction form, the following methodological and outcome variables were collected from each study: author(s), year of publication, country/region, study design, sample size, sample characteristics, proportion of females, age range, AI model(s) used, intervention design, intervention duration, type of input data, AI tasks and applications, outcome measures, intervention effectiveness, and perceived usefulness of AI technologies.

2.4. Methodological review

We categorized the AI methodologies employed by the included studies based on their position in the AI hierarchy. This structure includes ML as a subfield of AI, and DL and RL as subfields of ML. Within the ML models, methods were organized into 2 subcategories: unsupervised and supervised learning. Within the DL models, methods were classified into tabular data modeling, computer vision, and natural language processing. RL models were examined for their applications in creating virtual agents to interact with users, such as chatbots and virtual health coaches. Instead of listing every model used by the included studies, we concentrated on the popular models employed by 2 or more studies.

3. Results

3.1. Identification of studies

Fig. 1 presents the PRISMA flow diagram. An initial keyword search identified 7350 articles, and after the removal of duplicates, 6651 unique articles remained for screening based on their titles and abstracts. Of these, 6601 articles were considered irrelevant and subsequently excluded from the review. The study selection criteria were then applied to the remaining 50 articles, leading to the exclusion of 26 studies for various reasons, such as lack of AI technology adoption (n = 8), absence of a PA intervention (n = 8), being a commentary instead of original empirical research (n = 3), focusing on biomechanics (n = 5), and focusing on injury prevention (n = 2). Ultimately, 24 studies were found to be relevant and were included in the review.24, 25, 26, 27^,31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50

Fig. 1.

PRISMA flow diagram. AI = artificial intelligence; PRISMA = the Preferred Reporting Items for Systematic Reviews and Meta-Analyses.

3.2. Study characteristics

Table 1 reports the characteristics of the 24 studies included in the review. The earliest study³¹ was published in 2012, with single publications in 2013,³² 2015,⁵⁰ and 2020.²⁷ Two studies^33,34 were published in 2017, 3 studies in 2018,^24,35,36 5 studies in 2021,^26,37, 38, 39, 40 and 10 studies in 2022.^25,41, 42, 43, 44, 45, 46, 47, 48, 49 These studies were conducted in various countries, including 8 studies^24,31, 32, 33^,35,36,38,50 in the USA, 6 studies^39,43,44,47, 48, 49 in China, 3 studies^37,41,45 in Japan, 2 studies each in Korea^25,46 and Australia,^26,27 and 1 study each in Israel,³⁴ UK,⁴² and Denmark and Norway.⁴¹ The sample sizes in these studies were relatively small, ranging from 10 to 461 participants. Specifically, 16 studies^24,27,33, 34, 35, 36, 37, 38, 39^,41, 42, 43^,46,47,49,50 had a sample size between 10 and 99, 7 studies^{25,26,31,32,44,45,48} between 100 and 199, and 1 study⁴⁰ had 461 participants. The studies targeted different populations: 7 studies^{25,26,32,35,42,45,50} focused on healthy adults, 2 studies^31,33 on healthy children and adolescents, 1 study³⁶ on healthy young adults, and 1 study²⁷ on healthy middle-aged or older adults. Additionally, 2 studies^46,47 involved children and adolescent patients, 5 studies^24,39, 40, 41^,49 examined adult patients, 3 studies^34,37,38 targeted middle-aged and older adult patients, and 3 studies^43,44,48 targeted older adult patients. Among the studies, 7 studies25, 26, 27^,31,39,47,49 employed a pre–post design, while 17 studies^24,32, 33, 34, 35, 36, 37, 38^,40, 41, 42, 43, 44, 45, 46^,48,50 utilized an randomized controlled trial design.

Table 1.

Characteristics of the studies included in the review.

Number	Study	Country/ region	Intervention duration	Study design	Sample size	Sample characteristics	Female (%)	Age (year, mean ± SD, range)	AI model category	AI model
1	Trost et al. (2012)31	USA	1 occasion (5-min  or 10-min)	Pre–post	100	Healthy children and adolescents	50.0	11.0	DL	ANN
2	Bickmore et al. (2013)32	USA	8 weeks	RCT	113	Healthy adults	61.0	33.0	DL	Automated health counselor agent
3	Mendoza et al. (2017)33	USA	3–5 weeks	RCT	54	Healthy children and adolescents	64.8	9.9	ML	RF (with hidden Markov smoothing)
4	Yom-Tov et al. (2017)34	Israel	26 weeks	RCT	27	Middle-aged and older adults with diabetes	33.3	Treatment: 58.7 ± 2.1; Control: 55.1 ± 3.6	RL	Multi-armed bandit
5	Rabbi et al. (2018)24	USA	5 weeks	RCT	10	Adults with chronic back pain	70.0	31.0–60.0	RL	Multi-armed bandit
6	Zhou et al. (2018)35	USA	10 weeks	RCT	64	Healthy adults	83.0	41.1	RL	BAA (It uses RL to construct a predictive model based on a participant's historical step and goal data and adaptively uses the estimated model to generate challenging yet realistic step goals.)
7	Zhou et al. (2018)36	USA	10 weeks	RCT	13	Healthy young adults	77.0	22.2	RL	BAA
8	Maher et al. (2020)27	Australia	12 weeks	Pre–post	31	Healthy middle-aged or older adults	67.7	56.2	DL	NLP (chatbot and health coach)
9	Anan et al. (2021)37	Japan	12 weeks	RCT	94	Middle-aged and older adults with neck/shoulder pain/stiffness or low back pain	23.4	Intervention group: 41.8 ± 8.7; Control group: 42.4 ± 8.0	DL	NLP (chatbot and health coach)
10	Hassoon et al. (2021)38	USA	4 weeks	RCT	42	Middle-aged and older adults with cancer	90.0	62.1	DL	NLP (chatbot, health coach, and text-to-speech)
11	Huang et al. (2021)39	China	12 weeks	Pre–post	24	Young adults with hypertension	NA	18.0–45.0	ML	Dictionary learning (on sparse data)
12	Sandal et al. (2021)40	Denmark and  Norway	12 weeks	RCT	461	Adults with lower back pain-related disability	55.0	47.5	DL	NLP (chatbot and health coach)
13	To et al. (2021)26	Australia	6 weeks	Pre–post	116	Healthy adults	81.9	49.1	DL	NLP (chatbot)
14	Itoh et al. (2022)41	Japan	12 weeks	RCT	99	Adults with lower back pain	44.4	Exercise group: 47.9 ± 10.2; Conventional group: 46.9 ± 12.3	DL	NLP (chatbot)
15	Kurt et al. (2022)42	UK	8 weeks	RCT	85	Healthy adults	75.3	≥25.0	ML	k-means clustering
16	Lin et al. (2022)43	Taiwan, China	12 weeks	RCT	16	Older fragile adults	75.0	84.4	ML	AIoT, DT
17	Meng et al. (2022)44	China	24 weeks	RCT	150	Older fragile adults	56.0	≥65.0	ML	ETC, GB, k-NN, LDA, LR, LGBM, DT, SVM, RF, XGBoost, and stacking
18	Nakata et al. (2022)45	Japan	12 weeks	RCT	141	Healthy adults	26.2	Intervention group: 42.3 ± 9.4 Control group: 44.0 ± 9.1	DL	NLP (chatbot and health coach)
19	Oh et al. (2022)46	Korea	3 weeks	RCT	24	Children and adolescents with obesity	16.7	13.2	DL	CNN
20	Park et al. (2022)25	Korea	2 weeks	Pre–post	176	Healthy adults	47.7	Intervention group: 36.67 ± 8.03; Control group: 38.28 ± 7.04	DL	CNN
21	Sun et al. (2022)47	China	12 weeks	Pre–post	41	Children and adolescents with autism	NA	3.0–6.0	ML	SVM
22	Wang et al. (2022)48	China	24 weeks	RCT	171	Older fragile adults	53.8	1. Baduanjin group: 71.8 ± 3.8; 2. Baduanjin, strength, and endurance group: 70.7 ± 3.7 3. Strength and endurance group: 70.7 ± 3.5	ML	LGBM, GB, XGBoost, ETC, DT, RF, LDA, LR, and stacking
23	Wu et al. (2022)49	China	4 weeks	Pre–post	40	Adults with CFS and healthy adults	67.5	CFS group: 38.15 ± 12.05 Control group: 32.85 ± 12.31	ML	SVC and RF
24	Rabbi et al. (2015)50	USA	3 weeks	RCT	17	Healthy adults	47.0	28.3 ± 6.96	ML	MAB

Abbreviations: AI = artificial intelligence; AIoT = artificial intelligence of things; ANN = artificial neural network; BAA = behavioral analytics algorithm; CFS = chronic fatigue syndrome; CNN = convolutional neural network; DL = deep learning; DT = decision tree; ETC = extra tree classifier; GB = gradient boosting; k-NN = k-nearest neighbor; LDA = linear discriminant analysis; LGBM = light gradient boosting machine classifier; LR = logistic regression; MAB = multi-armed bandit; ML = machine learning; NA = not applicable; NLP = natural language processing; RCT = randomized controlled trial; RF = random forest; RL = reinforcement learning; SVM = support vector machine; XGBoost = eXtreme gradient boosting.

3.3. Outcome measures

Table 2 summarizes the intervention design and outcome measures for studies related to AI applications. These studies spanned 4 areas: PA promotion alone, PA promotion plus healthy diet promotion, PA pattern recognition, and PA outcome prediction. In the pool of included studies, 15 studies24, 25, 26, 27^,34, 35, 36, 37, 38^{,40,41,43,45,46,50} applied AI technology for PA promotion and 7 studies^31,33,39,44,47, 48, 49 for outcome prediction. Additionally, 1 study⁴² concentrated on pattern recognition, and 1 study³² combined the promotion of PA and a healthy diet. The data formats analyzed in these studies spanned tabular, text, image, video, and audio. Specifically, 13 studies^24,31,33, 34, 35, 36^,39,43,44,47, 48, 49, 50 worked with tabular data, while 8 studies^26,27,32,37,40, 41, 42^,45 employed text data. Furthermore, 2 studies^25,46 used image/video data, while 1 study³⁸ used text and audio data. Regarding the AI tasks, 6 studies^{31,33,39,44,48,49} involved classification, 5 studies^{25,31,35,36,47} used regression, and 12 studies^{24,26,27,32,34,37,38,40,41,43,45,50} employed text generation. Moreover, 1 study⁴² implemented text clustering, and 1 study³⁸ implemented text and speech generation. Various outcome measures of PA were used in these studies, such as PA type (e.g., cycling, walking), intensity, duration, energy expenditure, and quantity (e.g., daily step counts). Additional outcomes encompassed diet consumption, musculoskeletal symptoms, blood pressure, pain intensity, quality of life, physical function, fatigue syndrome, sleep quality, blood glucose levels, social communication impairments, restricted and repetitive behaviors, and frailty status.

Table 2.

Intervention design and outcome measures in the studies included in the review.

Number	Study	Intervention design	AI applications	Type of input data	AI task	Outcome measures
1	Trost et al. (2012)31	1 arm: 12 standardized PAs	Outcome prediction	Tabular	Classification, regression	PA type and EE
2	Bickmore et al. (2013)32	4 arms: Arm 1: Daily steps Arm 2: Fruit and vegetable consumption Arm 3: Both interventions Arm 4: Control group	PA promotion, healthy diet promotion	Text	Text generation	PA quantity (step counts) and fruit/vegetable consumption
3	Mendoza et al. (2017)33	2 arms: Arm 1: Bicycle training group Arm 2: Control group	Outcome prediction	Tabular	Classification	PA type (cycling vs. other PA types)
4	Yom-Tov et al. (2017)34	2 arms: Arm 1: Daily personalized feedback messages and weekly summaries. Arm 2: Identical once-per-week reminders to exercise	PA promotion	Tabular	Text generation	PA type and PA quantity (step counts)
5	Rabbi et al. (2018)24	2 arms: Arm 1: Generic PA recommendations from an expert Arm 2: Mobile phone-based auto-personalized PA recommendations (MyBehaviorCBP)	PA promotion	Tabular	Text generation	PA duration (walking and non-walking PAs), pain level
6	Zhou et al. (2018)35	2 arms: Arm 1: Automated adaptively-personalized daily step goals Arm 2: Constant goal of 10,000 steps per day	PA promotion	Tabular	Regression	PA quantity (step counts)
7	Zhou et al. (2018)36	2 arms: Arm 1: Automated adaptively-personalized daily step goals Arm 2: Constant goal of 10,000 steps per day	PA promotion	Tabular	Regression	PA quantity (step counts)
8	Maher et al. (2020)27	Pre–post 1 arm: AI chatbot use	PA promotion	Text	Text generation	Step counts
9	Anan et al. (2021)37	2 arms: Arm 1: AI-assisted PA promotion program Arm 2: Usual care	PA promotion	Text	Text generation	Musculoskeletal symptoms (neck/shoulder stiffness/pain and low back pain)
10	Hassoon et al. (2021)38	3 arms: Arm 1: Voice-assisted AI coaching delivered via a smart speaker (MyCoach) Arm 2: autonomous AI coaching delivered via text messages (SmartText) Arm 3: PA educational materials as the control	PA promotion	Text, audio	Text and speech generation	Step counts
11	Huang et al. (2021)39	1 arm: hypertension patients	Outcome prediction	Tabular	Classification	Blood pressure
12	Sandal et al. (2021)40	2 arms: Arm 1: Use of the SELFBACK app Arm 2: Usual care as the control	PA promotion	Text	Text generation	Lower back pain, coping ability, fear-avoidance belief, cognitive and emotional representations of illness, health-related quality of life, PA level
13	To et al. (2021)26	1 arm: AI chatbot use	PA promotion	Text	Text generation	Step counts
14	Itoh et al. (2022)41	2 arms: Arm 1: Receiving education and exercise therapy via a mobile messaging app Arm 2: No intervention	PA promotion	Text	Text generation	Work productivity, pain intensity, quality of life, fear of movement, and depression
15	Kurt et al. (2022)42	2 arms: Arm 1: Qigong training Arm 2: No intervention	Pattern recognition (clustering)	Text	Text clustering (based on TF-IDF)	Physical and mental well-being, sleep quality, physical energy, fatigue, positivity, stress, connections to self and nature
16	Lin et al. (2022)43	2 arms: Arm 1: 12-week AIFASE system exercise intervention Arm 2: No intervention	PA promotion	Tabular	Text generation	Body composition, muscle strength, physical function tests, and health-related quality of life
17	Meng et al. (2022)44	3 arms: Arm 1: Eight-form Tai Chi group (TC) Arm 2: The strength and endurance training group (SE) Arm 3: A comprehensive intervention combining both TC and SE	Outcome prediction	Tabular	Classification	Post-experimental frailty status of frail older adults
18	Nakata et al. (2022)45	2 arms: Arm 1: Use of CALO Mama Plus Arm 2: No intervention	PA promotion	Text	Text generation	Body weight
19	Oh et al. (2022)46	2 arms: Arm 1: AI-based gesture recognition game (SUKIA) Arm 2: Nintendo Switch	PA promotion	Image/video	Image generation	Calorie consumption, VO2max, 6-min walking test, body mass index, and perceived exertion
20	Park et al. (2022)25	2 arms: Arm 1: A ML-based motion-detecting mobile exercise coaching application Arm 2: Video streaming	PA promotion	Image/video	Regression (keypoint-based movement detection)	Quality of life and lower back pain
21	Sun et al. (2022)47	2 arms: Arm 1: Experimental group participating in the mini-basketball training program Arm 2: Control group	Outcome prediction	Tabular	Regression	Social communication impairments and restricted and repetitive behaviors
22	Wang et al. (2022)48	3 arms: Arm 1: Baduanjin training Arm 2: Strength and endurance training Arm 3: Both interventions	Outcome prediction	Tabular	Classification	Levels of perceived frailty and physical fitness
23	Wu et al. (2022)49	2 arms: Arm 1: Chronic fatigue syndrome patients group Arm 2: Healthy control group	Outcome prediction	Tabular	Classification	Fatigue syndrome and sleep quality
24	Rabbi et al. (2015) 50	2 arms: Arm 1: receive MyBehavior's personalized suggestions Arm 2: receive nonpersonalized suggestions	PA promotion	Tabular	Text generation	Walking duration

Abbreviations: AI = artificial intelligence; AIFASE = artificial intelligence of things-based feedback assistive strengthening ergometer; EE = energy expenditure; ML = machine learning; PA = physical activity; RCT = randomized controlled trial; TF-IDF = term frequency-inverse document frequency; VO_2max = maximal oxygen consumption.

3.4. Main findings

Table 3 summarizes the estimated effects and main findings of the studies included in the review. Three primary findings have emerged.

Table 3.

Estimated effects and main findings of the studies included in the review.

Number	Study	Intervention effectiveness	Usefulness of AI technologies
1	Trost et al. (2012)31	NA	1. Prediction of PA type: + (ANN accuracy: pre = 81.3%, post = 88.4%) 2. Prediction of PA energy expenditure: + (ANN RMSE: pre = 1.1, post = 0.9) 3. Classification of PA intensity: + (ANN exhibited better classification accuracy than the Freedson-Trost and Treuth prediction models)
2	Bickmore et al. (2013)32	Daily steps: +	Automated health intervention software designed for efficient re-use is effective at modifying health behavior
3	Mendoza et al. (2017)33	1.Daily commutes by cycling: + (mean percentage increased by 44.9%) 2. MVPA: + (daily MVPA increased by 21.6 min/day from Time 1 to Time 2 compared to controls)	Identification of cycling behavior: + (accuracy = 99.9%)
4	Yom-Tov et al. (2017)34	Between-arm comparisons (treatment arm compared to control arm): Blood glucose levels: –	Prediction of the messages promoting PA: +
5	Rabbi et al. (2018)24	Between-arm comparisons (treatment arm compared to control arm): Daily walking: +	Effectiveness of MyBehaviorCBP (MyBehaviorCBP is a mobile phone app that uses ML on sensor-based and self-reported PA data to find routine behaviors and automatically generates PA recommendations that are similar to existing behaviors): +
6	Zhou et al. (2018a)35	Within-arm comparisons: Treatment arm: Daily steps: – (by 390 steps) Control arm: Daily steps: – (by 1350 steps)	Effectiveness of automated, personalized daily step goals recommendation system: +
7	Zhou et al. (2018b)36	Within-arm comparisons: Treatment arm: Daily steps: + (by 700 steps) Control arm: Daily steps: – (by 1520 steps)	Effectiveness of the CalFit app (CalFit uses RL to generate personalized daily step goals that are challenging but attainable): +
8	Maher et al. (2020)27	Weekly MVPA time: + (baseline: 206.1 min vs. post: 315.9 min)	Effectiveness of the AI chatbot (health coach): +
9	Anan et al. (2021)37	Between-arm comparisons (treatment arm compared to control arm): 1. Severity of neck/shoulder pain/stiffness and low back pain: + (OR = 6.36, 95%CI: 2.57–15.73; p < 0.001). 2. Subjective assessment of pain/stiffness: + (OR = 43.00, 95%CI: 11.25–164.28; p < 0.001).	Effectiveness of the AI chatbot (health coach): +
10	Hassoon et al. (2021)38	Within-arm comparisons: MyCoach arm: Daily steps: + (by 3618.2) SmartText: Daily steps: + (by 1619.0) Control arm: Daily steps: + (by 886.1)	Effectiveness of the AI chatbot (health coach): +
11	Huang et al. (2021)39	VO2/heart rate is a powerful, new prognostic indicator for predicting aerobic exercise efficacy	The dictionary learning algorithm outperforms conventional classifiers in predicting individual responsiveness to aerobic exercise intervention based on the time series of metabolic indicators.
12	Sandal et al. (2021)40	Between-arm comparisons (treatment arm compared to control arm): 1. Mean difference in RMDQ score = 0.79 (95% CI: 0.06–1.51; p = 0.03). 2. Percentage of patients reporting improvement in the RMDQ score = 52% vs. 39% (OR = 1.76; 95%CI: 1.15–2.70; p = 0.01).	Effectiveness of the AI chatbot (health coach): +
13	To et al. (2021)26	Daily steps: + (by 627 steps) Total PA time: + (by 154.2 min/week) Meeting PA guidelines: + (OR = 6.37, 95%CI: 3.31–12.27)	Effectiveness of the AI chatbot: +
14	Itoh et al. (2022)41	Between-arm comparisons (treatment arm vs. control arm): Symptoms of low back pain: 3.2 vs. 3.8 (between-group difference = – 0.5, 95%CI: – 1.1 to 0.0; p = 0.04); Quality of life: 0.068 vs. 0.006 (between-group difference = 0.061, 95%CI: 0.008–0.114; p = 0.03); Fear of movement: −2.3 vs. 0.5 (between-group difference = – 2.8, 95%CI: −5.5 to −0.1; p = 0.04).	Effectiveness of the AI chatbot (for treating chronic back pain): +
15	Kurt et al. (2022)42	Between-arm comparisons (treatment arm compared to control arm): Sleep quality: + Feeling able to cope with life: + Life energy: + Stress level: –	K-mean clustering is a useful ML method to recognize text clusters and infer common topics
16	Lin et al. (2022)43	Hip flexor strength: + Semi-tandem stand: + Tandem stand: +	Effectiveness of the AIoT ergometer to deliver customized physical training prescriptions to improve the physical performance of long-term care facility residents: +
17	Meng et al. (2022)44	Effectiveness of the hybrid exercise program: Improvement in 10-m maximum walking speed: + Improvement in 6-min walk test: + Reduction in frailty: +	Stacking model prediction: + (accuracy = 67.8%, F1-score = 71.3%)
18	Nakata et al. (2022)45	Within-arm comparisons Treatment groups: Body weight: - Control groups: Body weight: – Between-arm comparisons (treatment arm compared to control arm) Body weight: –	Effectiveness of the AI chatbot (health coach): +
19	Oh et al. (2022)46	Super Kids Adventure (SUKIA, an artificial intelligence-based gesture recognition game application) showed superior effects on calorie consumption, VO2max, and Rating of Perceived Exertion compared to Nintendo Switch group	Effectiveness of the gesture recognition Kinect-type game: +
20	Park et al. (2022)25	Between-arm comparisons (treatment arm vs. control arm) SF-36 score: 9.10 vs. 1.09 (p < 0.01) Lower back pain score: 0.96 vs. – 0.26 (p < 0.01)	Effectiveness of the AI motion detection tool to provide real-time feedback: +
21	Sun et al. (2022)47	NA	1. Prediction of PA intervention-related social communication impairments using ML models: + (MSE = 0.188, RMSE = 0.434, R2 = 0.83) 2. Prediction of PA intervention-related restricted and repetitive behaviors using ML models: + (MSE = 0.051, RMSE = 0.226, R2 = 0.86)
22	Wang et al. (2022)48	Improvement in fitness: + Reduction in frailty: +	Prediction using stacking model: + (accuracy = 75.5%, precision = 77.1%)
23	Wu et al. (2022)49	Improvement in fatigue syndrome: + Improvement in sleep quality: + Improvement in body health statement: +	Health outcome recognition using ML: + (mean accuracy = 80.5% ± 9%, highest accuracy = 90%)
24	Rabbi et al. (2015) 50	Walking duration: +	Effectiveness of MyBehavior: +

Note: “+” denotes an increase, and “–” denotes a reduction.

Abbreviations: 95%CI = 95% confidence interval; AI = artificial intelligence; AIoT = artificial intelligence of things; ANN = artificial neural network; CNN = convolutional neural network; DL = deep learning; EE = energy expenditure; GB = gradient boosting; ML = machine learning; MVPA = moderate-to-vigorous-intensity physical activity; MSE = mean square error; NA = not applicable; OR = odds ratio; PA = physical activity; RL = reinforcement learning; RMDQ = Roland-Morris disability questionnaire; RMSE = root mean square error; SF-36 score = The medical outcomes study short form 36-item health survey; SOJ = Lyden's sojourn method; SVM = support vector machine; VO₂ = oxygen consumption; VO_2max = maximal oxygen consumption.

First, AI models, including ML and DL, were generally effective in predicting and analyzing PA outcomes and patterns. For instance, AI models were found to be useful in predicting PA energy expenditure,³¹ classifying PA intensity,³¹ and identifying cycling behavior.³³

Second, most studies comparing AI applications in PA interventions with traditional methods reported favorable outcomes. For example, AI-powered health intervention software, mobile phone apps, and recommendation systems effectively modified health behaviors, such as daily steps and moderate-to-vigorous-intensity PA.^{24,32,35,36,50} AI chatbots (health coaches) were also reported to be effective in various interventions.^{26,27,37,38,40,41,45}

Third, recent studies have increasingly adopted state-of-the-art DL and RL models and have applied AI technologies for various tasks in PA interventions. For instance, Artificial Intelligence of Things technologies delivered customized physical training prescriptions to improve physical performance in long-term care facility residents.⁴³ AI-based motion detection tools provided real-time feedback for patients with lower back pain.²⁵ Gesture recognition Kinect-type games were found to have superior effects on calorie consumption and maximal oxygen consumption compared to other gaming interventions.⁴⁶

3.5. Methodological review

3.5.1. Overview of AI

AI endeavors to automate intellectual tasks traditionally done by humans.⁵¹ Spanning 2 developmental periods, symbolic AI (1950s–1980s) was rule-centric but found complex tasks challenging.^9,52 Modern AI capitalizes on ML and DL methods, extracting features through layered artificial neurons to address intricate tasks, especially with large, unstructured datasets.^10,11,53 Evolving from concepts like dynamic programming, RL emerged in the late 20th century as a pivotal AI area dedicated to learning in dynamic environments.^12,13,54,55

3.5.2. AI vs. traditional statistical approaches

While AI models prioritize predictive accuracy, even if it compromises interpretability, traditional statistical models emphasize understanding variable relationships.56, 57, 58 AI leverages training, validation, and test sets to ensure peak performance and counteract overfitting.^59,60 In comparison, traditional methods focus on performance metrics without distinct data partitioning.⁶¹

3.5.3. Machine learning

ML bifurcates into unsupervised and supervised learning.⁶² The former delves into unlabeled datasets to uncover latent patterns, encompassing clustering, association, and dimensionality reduction tasks.63, 64, 65, 66 Specific studies, like those by Kurt et al.⁴² and Huang et al.,³⁹ employ techniques like k-means clustering and dictionary learning for PA pattern recognition and outcome prediction. Kurt et al.⁴² applied k-means clustering to identify the optimal number of clusters in a term frequency-inverse document frequency (TF-IDF) array. The optimal number of clusters was 4, each representing a specific health benefit derived from lung-strengthening Qigong practices.⁴² Huang et al.³⁹ used a dictionary learning-based classifier to predict patient responsiveness to an aerobic exercise intervention by addressing issues related to redundant information and noises in the cardiopulmonary exercise testing data.

Supervised ML, conversely, relies on labeled data to delineate input-output mappings.⁶⁷ Various studies have adopted diverse algorithms: Wang et al.⁴⁸ employed linear discriminant analysis, eXtreme Gradient Boosting, and light gradient boosting machine on measures including 10-meter maximum walk speed, grip strength, timed up-and-go test, and 6-minute walk test to determine the post-experimental frailty state of older adults, while Meng et al.⁴⁴ utilized the K-nearest neighbors algorithm to predict the frailty status of older participants following an intervention by utilizing pre-intervention features. Moreover, Sun et al.⁴⁷ utilized the support vector machines algorithm to predict mini-basketball training program outcomes concerning social communication impairments and restricted and repetitive behaviors in preschool children with autism spectrum disorders. Also notable is that Lin et al.⁴³ employed a decision tree-based remote healthcare warning system to monitor physical conditions, including excessive exercise intensity, during interventions tailored for older adults. Mendoza et al.³³ employed a random forest algorithm to create a mapping between minute-level accelerometer and global positioning system features and observed PA. The algorithm was trained using data from participants who performed 5 typical types of childhood PA.³³

3.5.4. Deep learning

A sophisticated subset of ML, DL employs multi-layered artificial neural networks to decipher vast, intricate datasets, particularly images and texts.⁶⁸ This capability, which is superior to traditional ML methods, makes DL especially potent when dealing with “big data”. For PA interventions, DL models, like convolutional neural networks and natural language processing models, have showcased immense promise. They address challenges related to health behavior, real-time communication, and decision-making, empowering researchers to devise highly tailored PA interventions. Trost et al.³¹ utilized artificial neural networks to predict PA type and energy expenditure in children and adolescents by using features like percentiles and autocorrelation extracted from various time windows of accelerometer data. Oh et al.⁴⁶ implemented a convolutional neural network-based gesture recognition algorithm in the SUKIA system to detect different upper extremity and lower body movements for a game application. Similarly, Park et al.²⁵ incorporated convolutional neural network technology into the LikeFit app to guide users through exercises by accurately detecting motion on a mobile device. natural language processing models, meanwhile, have revolutionized human-machine communication,⁶⁹ as is exemplified in works by Maher et al.²⁷ and Anan et al.³⁷ Maher et al.²⁷ employed an AI chatbot called Paola to assist and motivate participants in making dietary and PA changes while monitoring body composition and blood pressure. Anan et al.³⁷ utilized an exercise-based AI-assisted interactive health promotion system, a mobile messaging app chatbot, to provide daily exercise instructions to workers with neck/shoulder stiffness/pain and low back pain in a 12-week randomized controlled trial. Subjective pain severity was evaluated using a scoring scale at baseline and post-intervention.³⁷

3.5.5. Reinforcement learning

RL, a distinctive AI domain, educates agents to make choices by interacting with environments, receiving feedback as rewards or penalties.^70,71 In PA interventions, RL algorithms personalize recommendations to foster behavior change and enhance health outcomes. Adapting from user responses, RL methodologies, which are exemplified by the multi-armed bandit and behavioral analytics algorithm, provide personalized dynamic experiences. For example, Rabbi et al.²⁴ utilized a multi-armed bandit with a knapsack approach to promote the issuance of simple exercise recommendations while also allowing for the exploration of other options within a 60-min time frame to adhere to a predefined target. Zhou et al.³⁵ implemented the behavioral analytics algorithm in the CalFit app to aid users in achieving their fitness objectives. The behavioral analytics algorithm develops predictive models based on participants’ past steps and goal data. These models are then utilized to dynamically generate challenging step goals, which optimize future PA.^35,36

4. Discussion

In this study, we conducted a scoping review of the applications of AI in PA interventions. A keyword search in digital bibliographic databases identified numerous studies that employed diverse ML, DL, and RL models to enhance various PA prevention and promotion aspects. In general, studies found AI models effective in identifying meaningful patterns of PA behavior and relationships between specific factors and intervention outcomes. The majority of studies comparing AI models with conventional statistical approaches found the former to achieve higher prediction accuracy on test data. Some studies comparing the performances of different AI models revealed mixed results, likely indicating the high contingency of model performance on the dataset and task it was applied to. An accelerating trend of adopting state-of-the-art DL and RL models over standard ML to address challenging tasks in human-machine communication, behavior modification, and decision-making was observed. We concisely introduced the popular ML, DL, and RL models and summarized their specific applications in the studies included in the review.

Despite the variety of ML, DL, and RL models utilized in PA interventions, it may only be the beginning of the trend for AI application in the big data era. Future adoptions of AI in PA research could be influenced by a broad spectrum of factors, 6 of which are discussed in the following sections.

4.1. Implications for research and practice

4.1.1. Personalized PA interventions

A significant trend in the current AI-based PA interventions is the shift towards personalization. By leveraging AI algorithms, researchers and health practitioners can analyze individual data, such as demographics, fitness levels, and personal preferences, to create customized exercise programs that cater to each person's unique needs and goals.^35,36 This approach may result in more effective and sustainable behavior change and improved health outcomes.⁴³

4.1.2. Real-time monitoring and adaptation

Another trend is the use of AI for real-time monitoring of individuals’ PA and physiological parameters, such as heart rate, energy expenditure, and movement patterns.³⁵ Researchers can develop systems that adapt exercise recommendations and interventions based on real-time data by incorporating sensor data and AI algorithms²⁴ to ensure that individuals receive the most safe and effective guidance possible throughout their exercise routines.

4.1.3. Integration of multimodal data sources

The increasing availability of diverse data sources, such as wearable devices, mobile apps, and electronic health records, presents an opportunity to combine these data streams to better understand and predict PA behaviors.²⁴ The utilization of AI algorithms facilitates the integration and analysis of these multimodal data sources, providing researchers with richer insights and more accurate predictions regarding PA and health outcomes.^44,48 Evidence of such synergies is already emerging in the literature. For instance, a recent study employed AI to leverage cardiopulmonary exercise testing data, leading to personalized aerobic exercise interventions for young hypertensive patients, ultimately resulting in improved treatment outcomes. This showcases how AI's capability for interpreting complex health data can drive personalization and effectiveness in PA interventions.³⁹

4.1.4. Evaluation of intervention effectiveness

AI algorithms can be employed to evaluate the effectiveness of PA interventions by comparing outcomes across different groups, settings, and intervention strategies.^34,42 AI models can help researchers identify the most effective interventions for specific populations and contexts, allowing for the optimization of resources and maximizing the impact of these programs on public health.

4.1.5. Expanding access to PA interventions

AI-driven solutions, such as chatbots and mobile apps, can help expand access to PA interventions, particularly for individuals who face barriers to participating in traditional programs.^26,37,38,40 These AI-powered tools can provide personalized guidance and support to users, making PA interventions more accessible and inclusive for diverse populations.²⁷

4.1.6. Predicting and preventing injuries

AI algorithms can be employed to analyze biomechanical data and movement patterns, potentially predicting and preventing injuries related to PA. AI models can mitigate injury occurrence by identifying high-risk individuals and offering targeted interventions, thus fostering safe PA participation.^44,48

4.2. Limitations of the scoping review and included studies

To our knowledge, this study is the first to systematically review AI-related methodologies employed in PA intervention literature and to project trends for future technological development and applications. Nevertheless, several limitations should be noted concerning this review and the included studies. First, the majority of research in the area primarily highlights short-term effects of AI-based PA interventions. Future studies should be designed to assess the long-term sustainability of these interventions, as this aspect remains largely unexplored. Second, the AI-based PA interventions reviewed here often lacked standardized protocols, impeding their replicability and comparison across different studies. It is crucial for future research to address this issue to enhance reliability and trust in the findings. Third, as our review focused on ML, DL, and RL methods, specific findings from individual studies (e.g., the effectiveness of an intervention or estimated associations between variables and outcomes) were not thoroughly synthesized. Fourth, the studies included in our review were diverse regarding research questions, study designs, target populations, data collection methods, sample sizes, and data quality. The chosen analytical approaches were intrinsically linked to these study-specific factors, making cross-study comparisons of model performances potentially unreliable. Fifth, conclusions about relative model performances (e.g., the prediction accuracy of logistic regression vs. support vector machines) within the same study might lack generalizability due to the interdependency between data and AI algorithms. Sixth, our search was limited to articles published in English, and we excluded gray literature. These limitations may impact the generalizability of our synthesized evidence, although the decision was driven by practical constraints associated with translating and interpreting articles in different languages. Our focus on peer-reviewed articles was intended to maintain a high level of scientific rigor and reliability. AI technologies are evolving at an accelerating rate. As a result, a review like this one may have a limited shelf life, necessitating periodic updates to remain current.

5. Conclusion

This scoping review provides a thorough analysis of AI methodologies, such as ML, DL, and RL algorithms, in the context of PA interventions. By mapping the current landscape of AI techniques in PA literature, the review discerns trends, synergies, and patterns with the intent to guide subsequent research endeavors. A comprehensive overview of popular ML, DL, and RL models is presented along with a summary of their specific applications and an identification of 6 critical areas for future AI adoption in PA interventions. These areas include personalized PA interventions, real-time monitoring and adaptation, integrating multimodal data sources, evaluating intervention effectiveness, expanding access to PA interventions, and predicting and preventing injuries. As the field advances, staying informed and exploring emerging AI-driven strategies will be essential for significant progress in PA interventions and fostering overall well-being.