1 Introduction

Healthcare sector comprises of several areas that work together to provide quality medical services to the public. Public health by definition is “The science and art of preventing disease, prolonging life, and promoting health through the organized efforts and informed choices of society, organizations, public and private communities, and individuals” [1]. Healthcare today is facing several challenges various circumstances contribute to these challenges, such as a public health crisis and emergencies. Another challenge is the ageing population, which is a renowned factor for the development of multiple chronic diseases like cardiovascular disease, diabetes, stroke, cancer, osteoarthritis, and dementia [2]. This ageing population corresponds to a group of people with regular healthcare need. Finally, some organizational challenges like the need for more healthcare professionals and the hospitals capacity to deal with the increase in hospital visits. For instance, in Sweden, there are only 2.22 hospital beds per 1000 people [3] and the availability of 4.19 doctors per 1000 people [4].

Traditionally healthcare system is based on a reactive approach. Which is to react when symptoms appear, crisis occur then take actions. This reactive approach is damage control [5]. Supporting patients after they became symptomatic with the disease has an adverse effect on the healthcare system. Healthcare involves different processes e.g., screening, prevention, diagnosis, and treatment [6]. In this reactive design, the patient is assigned a passive role which is to wait for a decision throughout these healthcare processes this reduces patient empowerment and minimizes the possibilities of self-management.

A definition of health was introduced by Huber as “The ability to adapt and to self-manage, in the face of social, physical and emotional challenges” [7]. With the rapid increase in mobile technologies [8] and the transformation of healthcare practices, individuals are becoming more aware and concerned about their health [9]. A survey conducted in U.S showed that 62% of smartphone owners search the internet for health-related information [10]. The recent development and availability of smart devices pave the way for the era of digital health.

Digital health incorporates eHealth and mobile health (mHealth) [11] to provide digital transformation in healthcare. eHealth does not only mean electronic health but instead, it indicates many other terms [12]. eHealth provides health services and information delivered or enhanced through the internet and communication technologies, to improve healthcare locally, regionally, and worldwide by using information and communication technology [13]. mHealth provides healthcare services using mobile and wireless technology. mHealth combines wearable sensing technology and medical technology for providing healthcare services [14]. Within digital health, mHealth incorporates all available resources of telecommunication and wireless technology for the delivery of healthcare and health information. WHO global observatory for eHealth (GOe) defined mHealth as “medical and public health practice supported by mobile devices, such as mobile phones, patient monitoring devices, personal digital assistants, and other wireless devices” [15].

Healthcare is shifting from reactive to proactive [16] with the support of Artificial Intelligence (AI), mHealth and Internet of things (IoT). Proactive healthcare is to predict risks and react with supportive actions with the focus on the user-centric approach rather than the traditional hospital-centric approach [17]. Every health problem can be easier to manage in earlier stages. Proactive healthcare can allow supportive actions before a crisis [17]. Predicting and preventing a situation on time enables care which empowers and gives the user an active role.

mHealth provides pervasiveness which minimizes the dilemmas of healthcare delivery. It provides healthcare to anyone at any time anywhere. With the emergence of cutting-edge technologies [18] in healthcare and AI, self-management or self-care is becoming more relevant. It will contribute to shifting healthcare to a new norm with a focus on the user.

Artificial Intelligence with “problem-solving, reasoning and decision making [19] in mobile health can support healthcare by augmenting human capabilities to develop awareness, promote wellbeing and self-care of healthy behaviors [20]. As healthcare is evolving and with the introduction of healthcare 4.0 to be more proactive and user-centric [17] AI can help by applying reasoning and negotiation to the available health data and recognize patterns to introduce automation and finally to make decisions.

With healthcare becoming more proactive and user-centric true potential lies with AI to contribute to that by introducing AI-enabled proactive mHealth. AI with automated decision-making and predictive analytics can provide timely preventive measures and transform future health applications [21]. Proactive mHealth applications can predict (early detect) and prevent a situation with adaptive supportive actions to improve personalization and contribute to a healthy lifestyle.

This review paper investigates existing studies with a focus on AI-enabled proactive mHealth. We also examine studies which establish proactive mHealth with properties of prediction, prevention, and personalization. To our knowledge, there is no prior study which presented a review with the focus on proactive mHealth. The methodology of a systematic literature review (SLR) [22] is used for this review. We opted for an SLR because it allows rigorous analysis of a topic with lower inclination.

2 Methodology

A systematic literature review (SLR) is conducted to investigate the current state-of-the-art in AI-enabled proactive mHealth. An SLR is “a systematic, explicit, [comprehensive,] and reproducible method for identifying, evaluating, and synthesizing the existing body of completed and recorded work produced by researchers, scholars, and practitioners” [22]. SLR process has different stages identification, screening, eligibility, and selection. The inclusion and exclusion criteria provide the basis of selection.

2.1 Search Strategy

The search strategy is to create a search string and look for all the available studies within the relevant databases. In this study, we used the sources found in Table 1. The outcome of the search provides studies for each stage of the SLR. A significant aspect of an SLR process is string formation. A more precise string can provide better comprehensive result.

Table 1. Databases for the search.
Full size table

2.2 Search String and Stages of SLR

Table 2 shows the search string used for this SLR. Each scope provided synonyms which are grouped to form the search string. The focus is to be more precise in finding studies and grasping more by using refined keywords.

Table 2. Search string table.
Full size table

After search string formation, a comprehensive search is conducted within the databases. The first stage of SLR is the identification stage to recognize the studies for the screening stage. Using the search string available studies are collected from different sources (Table 1).

In the second stage of screening, gathered studies from the identification stage are examined. Screening is based on the criteria for inclusion and exclusion (Table 3). Initial screening is performed by analysis of the title and abstract of each study. The outcome of the first stage of screening formulates a list of studies. This list is passed to another stage of screening by analysis of introduction and conclusion. This more specific list is selected for the third stage of eligibility where full-text analysis is performed. The studies are then finally chosen for investigation in the final stage of selection. From the final list, analysis is done and discussed by providing key finding and limitations.

3 Criteria Inclusion and Exclusion of Studies

Table 3 presents the criteria for inclusion and exclusion of studies. A study is selected if it includes any of the given inclusion criteria and excludes all the exclusion criteria (Fig. 1).

Table 3. Criteria table.
Full size table

4 Outcome and Results

Fig. 1.
figure 1

Results PRISMA flowchart

Full size image

PRISMA flowchart shows each process of the review. After defining the criteria, we started with extensive search to look for available studies using our search string.

  • In the identification stage, we found 1256 studies and 15 more studies were included from references. These 15 studies were identified as relevant by keyword search performed on each of the references in the papers selected.

  • By initial identification, we excluded 997 studies in stage 1 which were not based on our criteria of inclusion. 274 studies were selected to be screened. After reading the title and abstract we excluded 197 studies and only selected 77 studies for stage 2. In stage 2 of screening, studies were examined by reading the introduction and conclusion. At this stage, another 32 studies were excluded from the list.

  • Finally, for the last stage before selection, only 45 studies were selected for full-text analysis.

  • After final stage, only 17 studies were selected, and 28 studies are excluded.

Table 4. Selected studies with analysis.
Full size table

5 Analysis of Selected Studies

Table 4 presents the list of studies selected for this review paper. Each study is evaluated based on several criteria. The first column “studies” lists the identifier (author name) of the selected study. The second column “selection criteria” is based on Table 3 criteria of inclusion and exclusion. Column “outcome category” provides detail about the targeted outcome of the study. Next three columns assess the studies based on inclusion criteria IC2 and IC4 from Table 3. Last two columns provide details about the type of AI technique used and the data sources considered.

Most of the studies are focused on a particular group of people or chronic disease management [25, 27, 28, 34]. The studies targeted diabetes [27, 34], cardiovascular disease [25], chronic obstructive pulmonary disease (COPD) [28], nutrition [26], stress management [38], weight management [37], sleep [33], depression [36] and to promote physical activity [23, 29].

Only 4 studies are targeted towards individuals that do not possess any diagnosis of disease [23, 26, 29, 33]. Most of the studies illustrate the importance of personalization however studies [23, 29, 33] considered adaptive user preference into consideration.

Only two studies [29, 31] gave insight on details regarding the AI techniques used for prediction. Studies [29, 33, 34] considered wearables as a data source. A brief introduction of each of the selected papers is presented below in terms of purpose, outcome, methodologies, and results.

Nitish Nag et al. [23] described the importance of proactive health and significance of timely interventions by introducing a health navigation map which have a goal as desired state and the current state of the user as a starting point. They also described a P5 cybernetic multimedia health recommender system which has a personalized model connected with sensors to predict a new situation, and then provide a precise solution. Many different data sources are considered important. Some tools used for recognition are Clarifai and EventShop. The need for AI techniques and models for predicting events is depicted but no analysis on the available techniques to use. The primary focus of the study is on nutrition recommendation using multimedia sources.

Shubhi A et al. [24] explained the importance of proactive health monitoring using wearable sensors. The study focuses on health monitoring by first using a machine learning model with different datasets and person’s electronic health record to identify the disease risk, and monitoring need. The paper only focuses on recommending the type of wearable device to the user based on the prediction of disease risk.

Michael V et al. [25] discusses the opportunities of providing proactive support for cardiovascular disease (CV) by early detection. The paper reviews the current state-of-the-art in mHealth to support CV. The main emphasis of the study is towards disease prevention with early detection. The study reviews available systems with mHealth to prevent CV disease no particular AI-techniques and are elaborated.

Mike B et al. [26] introduces eNutrition in the context of eHealth to provide designed nutrition to the users. It was argued that available systems are not personalized and cannot advise the user on what to eat, to keep track of the health information while giving recommendations. An example of an AI-engine is presented which takes the input of activity, preference, body data and consumption to give timely alerts to the user on what to eat and what to avoid including the portion size. The paper discusses available models and the gap but no information on modelling the AI-engine.

Adrian A et al. [27] is ongoing research targeting diabetic and mental health patients to improve physical activities using adaptive interventions. The applications hold a user-centered design (UCD) and provide interventions as an outcome of physical activity after 6 months. Step counts are gathered using Google fit and Apple HealthKit. Reinforcement learning is used as a technique for adaptive interventions.

Yvonne J et al. [28] discusses the importance of self-management in chronic patients for improving life and health outcomes. The main focus is on chronic obstructive pulmonary disease (COPD). The paper aims to describe in detail the whole process of developing UCD digital interventions.

Talko B et al. [29] uses a machine learning model to provide personalized coaching advice for an individual. They used step counts as a data source to train eight different machine learning models. Random forest algorithm performed best with the mean accuracy of 0.93, range 0.88–0.99 and the mean F1-score 0.90 with range 0.88–0.94. The proof-of-concept showed, machine learning can automate the process of physical coaching by predicting individual ability to reach their desired goal with timely interventions. Although the results of the study are promising but using step counts as a data source limits the system for interventions.

Robert S et al. [30] defines mHealth future trends with machine learning and deep learning in the current patient-centric model. A model of analytics is provided from descriptive analytics to predictive analytics. The paper provides insights to mHealth 2.0 with machine learning tools and patient-centric mHealth application design. The study lacks any implementation with these models.

Kashif N et al. [31] presented a review of available mHealth applications with prediction models using machine learning. A cloud-based secure architecture of a mHealth system is proposed which provides a platform for data acquisition and storage. They used CV as a use-case for training a model to provide outcome-based on the seriousness of the disease. The study gives great insight on available machine learning predictive models used in mHealth applications.

Bertalan M et al. [32] discusses wellbeing as a fundamental part of digital health. They introduced the definition of proactiveness as people are becoming more aware and have an active role in their health, they will be able to enter healthcare at an early stage. The paper emphasizes that digital health keeps people healthy instead of treating them when sick. The future depends on using AI systems, but the paper does not give any details on the type technologies to use for providing proactive health.

J. Rojas et al. [33] presented a personalized intervention system using mHealth and AI. They highlighted the importance of just-in-time interventions for continuous treatment and personalization based on lifestyle, environment, and genes. Two fundamental issues related to interventions are further discussed. They proposed a model aiming interventions for sleep. A Multilayer perceptron model was used instead of decision trees with the accuracy of the classifier as 88% and F1_score 0.54.

Mirza M et al. [34] used wearable devices for self-management of diabetes. The paper discusses the importance of early detection of diabetes. Data sources includes heart rate, breathing rate and volume, activity and manual data like blood glucose, BMI, and sex. AI model using adaptive-neuro fuzzy inference was proposed. The results showed good accuracy for the early detection of diabetes.

Usman I et al. [35] introduces earlier medicine and its importance. They discuss the contribution of AI in the initiative of “prevention is better than cure”. In this study level of interventions is divided as actionable, accurate, timely and individualized.

Marianne M et al. [20] describes the importance of decision-making in mHealth using AI. They established that smartphones and wearables hold the key in providing health for anyone at any time with timely interventions. The paper focuses on decision-making with AI in terms of using Just-in-time adaptive interventions (JITAIs). Importance of decision rules which takes user current state as input and then decide when and what to deliver as an intervention. The study also provides the whole architecture of using AI with JITAIs. They proposed that AI-technique like reinforcement learning (RL) improves precision by learning optimal decision rules and RL also adapt and adjust to the user preferences.

Michelle N et al. [36] provided a machine learning intervention model to predict mood, emotions, motivational states, and activities for patients with depression. Phone data was used e.g., GPS and phone calls. The model with regression and decision trees proved to be accurate for predicting location but performed below par when it comes to predicting emotions. More parameters are needed for better accuracy.

Ramesh M et al. [37] developed an algorithm with adaptive boosting to focus on changes in weight loss through personal health behavior. The model used finds relevant stories based on demographics and emotional tone. The model provided an accuracy of 84%–98%. The outcome showed that there is an increase in efficacy for weight loss but the medium for the stories did not affect behavioral change.

Morrison LG et al. [38] provided a naive Bayes classifier for predicting when to send push notifications to the user and the best time for intervention. The study focused on push notifications for stress management. The results showed no major difference on outcomes of daily notifications and intelligent timely notifications.

6 Discussion

6.1 Principle Findings

This review systematically investigates existing studies with the focus on providing proactive health using mHealth and AI.

Fig. 2.
figure 2

Levels of proactive health

Full size image

The Fig. 2 provides the parameter for identifying the level of proactiveness from each study. A “low” implies that the study only introduces the importance of proactive health. A “medium” level indicates that the study provides details about managing a situation or disease beforehand to avoid health risks. Finally, a “high” level presents the prediction and prevention capability to promote wellbeing.

Table 5. Key findings table.
Full size table

Table 5 presents key findings with different columns. Level of proactiveness is divided into three parameters as shown in the Fig. 2. Remaining columns provide details about different key findings that contribute to the discussion in this paper.

Our main finding is that digital interventions and recommender systems determine the basis of providing proactive mHealth. Most of the studies found were focused on disease management [25, 27, 28, 34]. With this review, we also found different implementations for proactive health. UCD is also given importance due to the reason that proactive health is user-centric, and it is away from the hospital settings.

Only one study [20] highlighted the importance of decision-making with AI and mHealth. No particular study focuses on automated decision-making with AI and implemented predictive analytics. Decision-making in mHealth has a significance. AI with automated decision-making can be an important factor when it comes to prediction and prevention of a situation. An AI-engine can provide automated decision-making to help users make healthier choices promptly.

Fig. 3.
figure 3

JITAI points [39]

Full size image

With this review, we also found the importance of timely interventions for proactive health. Only two studies [20, 33] highlighted the importance of JITAI. Figure 3 provide details about JITAI challenges. Decision points show the time of intervention with “when”. Intervention options give structure of intervention with “what” and tailoring variables shows the target assigned as “whom”. So, moving forward challenge lies with when to intervene, with what information and to whom.

Only one study [23] highlighted the importance of P4 principles [40] for designing health systems which is personalization, prediction, prevention and participatory.

Our review found 9 different use-cases of interventions: diabetes, CV, COPD, nutrition, stress and weight management, sleep, depression, and physical activity.

We also found that existing studies used insufficient data sources. A few studies [29, 33, 34] used wearables but no study considered real-time data to represent the current state of the user. Wearables can play a vital role to provide more information on the current state of the user that can be very handy when it comes to interventions.

No existing system targeted the general population to improve wellbeing and promote health with proactiveness. A few examples targeted individuals [23, 26, 29, 33] but did not consider multiple factors e.g., user preferences and surroundings.

Nearby surroundings or environment factor was not considered by any study. Threats in the nearby environment can affect any individual. Timely intervention with a prediction and prevention mechanism can promote health and save lives. Accuracy of such a system depends on multiple data sources and sensors.

The selected studies also ascertained that mHealth implementation with sensors and mobile apps are used for providing digital interventions.

This review also provided insight into the type of AI-techniques used for mHealth applications to establish proactive health. A study [31] reviewed prediction models for mHealth applications. Depending on the use-case different studies included techniques for prediction. Only a few studies [23, 27, 29, 33, 36, 38] implemented a model and provided detailed results. A survey study was found [41] focusing on mHealth interventions using machine learning but without focusing on proactive aspect.

A recent study [42] presented a review of machine learning applications for big data analytics in healthcare. It explores the latest trends in using machine learning techniques for analytics in healthcare. Another contemporary survey study [43] surveys digital health role in nutrition support.

With this review, we conclude that there is no existing research conducted within proactive health with AI and mHealth to predict and prevent a situation beforehand. Which is targeted towards the general population to make them aware in a timely manner, so they do not become sick. Considering the context (surroundings/environment) of person. A system that is personalized and accounts for a person, not a patient. Automated decision-making (ADM) will be the core part of such a system.

6.2 Results and New Insights

The goal of proactive health is to make people proactive about their health, so they are aware of their surroundings and their well-being. To early detect a situation or anomaly and prevent it with actions in a timely manner for wellness promotion.

Fig. 4.
figure 4

JITAI [20]

Full size image

Just-in-time adaptive interventions as shown in Fig. 4 can consider the current state of the user (from sensors) into account and provide timely interventions. Personalization of such interventions depends on user profiling and decision-making.

The world is currently experiencing a pandemic (COVID-19) which had an impact on our everyday life [44]. Proactive mHealth with AI can automate processes and support healthcare empowering people to take care of their health before becoming sick. An example is of several people working in a closed space. A proactive system can early detect a situation (e.g., fever alert) and proactively ask for symptoms and to provide a decision (go home) also give an alert flag for a possible outbreak.

An AI-engine can provide automated decision-making on a user-level to support self-care. An example is to identify someone in the risk of getting infected based on their profile and no. of cases in the city and reach out with timely advice (stay home). ADM can provide people with initial screening and advice on what steps to take.

AI systems are becoming intelligent and more automated in healthcare, in health decision-making it is vital to have transparency. System performance and transparency doesn’t go well together. AI systems opacity and black-box decision-making approach directly influence trustworthiness. A novel concept of eXplainable AI (XAI) [45] is introduced which proposes AI-techniques that are more explainable and do not affect performance of the model. A system that is understandable by humans.

Environment factor along with health information gathered can lead to beneficial outcomes for people in danger e.g., an avalanche, to identify risks in nearby environment or an adverse effect on people’s health. A proactive system can provide interventions based on individual or community level to save lives and promote wellness.

Real-time data from wearables like heart rate, activity, SpO2 etc. and user-profile can equip the system to be adaptive with current state for timely interventions.

Adopting the precise technique depends on the type of datasets and the desired outcome. Two studies [20, 27] opted for reinforcement learning for adaptive interventions, the reason is it improves precision by learning optimal decision rules, adapts and adjusts to user preferences.

Contribution for AI-Enabled Proactive mHealth.

This review presented current state-of-the-art when it comes to AI-enabled proactive mHealth. The study identified AI algorithms and techniques used in existing studies, along with the target group. The SLR presents which studies considered proactiveness with prediction and prevention capabilities. The analysis provides insights into data sources considered for modelling. This study also explores the level of proactiveness in selected studies. The detailed review of existing research prompted new insights for future work.

7 Limitations

For this review, broader search strings AI, machine learning, deep learning and ANN are used, which might have excluded some studies with a more precise title. While assessing the selected studies reliability in terms of data inputs (no. of people that participated) is not considered. Instead, the selection is based on multiple parameters.

8 Conclusion and Future Work

This paper presents a review of existing research with AI in mHealth to establish proactive mHealth. The review shows that proactive mHealth is complex and AI-enabled digital interventions and recommendation systems form the basis of proactive mHealth. The studies examined in this review show the effectiveness of being proactive but in the context of managing a certain health condition. We conclude that there is no existing research found for AI-enabled proactive mHealth that predicts and prevents a situation beforehand and targets the general population. A system that can consider multiple parameters of wellbeing e.g., environment or surroundings, and provide timely interventions before someone becomes sick. Furthermore, it is personalized and accounts for a person, not a patient. With this review, it is also concluded that the foundation of proactive mHealth lies with AI to provide automated decision-making with predictive analytics. P5 approach to mHealth and JITAI can contribute to the system design and implementation. These all pieces fit together to form the framework of AI-enabled proactive mHealth.

Future research must consider an iterative approach to establish AI-enabled proactive mHealth first step is to select multiple parameters (attributes) and a target group. Wearables and other sensors can be vital data sources. The next step is to identify these data sources based on the parameters and collect data. Finally, choosing a precise model which can not only provide timely intervention but can learn and adapt.