A systematic review on eHealth technology personalization approaches

Summary Despite the widespread use of personalization of eHealth technologies, there is a lack of comprehensive understanding regarding its application. This systematic review aims to bridge this gap by identifying and clustering different personalization approaches based on the type of variables used for user segmentation and the adaptations to the eHealth technology and examining the role of computational methods in the literature. From the 412 included reports, we identified 13 clusters of personalization approaches, such as behavior + channeling and environment + recommendations. Within these clusters, 10 computational methods were utilized to match segments with technology adaptations, such as classification-based methods and reinforcement learning. Several gaps were identified in the literature, such as the limited exploration of technology-related variables, the limited focus on user interaction reminders, and a frequent reliance on a single type of variable for personalization. Future research should explore leveraging technology-specific features to attain individualistic segmentation approaches.


INTRODUCTION
eHealth technologies such as internet-based interventions and mobile apps offer opportunities to make healthcare more effective and efficient and increase health and well-being, but there is room for improvement in terms of their effectiveness. 1,24][5] These approaches offer an alternative to a "one-size-fits-all" approach by adapting eHealth technologies to individual users or user groups in the following way: the eHealth technology gathers data about certain characteristics, such as user behavior, demographics, preferences, or interaction with the eHealth technology.These data include data on physical activity from wearables, diseases from electronic health records, or self-reported data through, for example, questionnaires.Computational algorithms are often used to process data on these variables in order to segment users into groups ranging from very small (one person) to large groups (e.g., all females), depending on the number of segmentations.These segmentations are matched with adaptations of features of a technology such as the content, graphical appearance, functionalities, behavior change strategy, or channel, 6,7 resulting in a personalized eHealth technology.
Personalization can be applied to eHealth technologies in various ways.To illustrate, an example is the integration of tailored nutrition information messages in the ''Happy Me'' smartphone application designed for obesity prevention. 8To deliver tailored messages, users are segmented into four groups (precontemplation, contemplation, preparation, action) based on the transtheoretical model of change. 8ccording to this model, users in these four different groups need different kinds of support to change their eating habits.To illustrate, if someone is thinking about change but has not committed yet (contemplation stage), messages should help them understand the importance of healthy eating and encourage them to start, whereas a user who is in the action stage benefits more from messages that support their new habits, offering suggestions on how to eat, social support, and reinforcement management.Thus, the computational method for personalization uses theory-driven if-then rules (i.e., conditional logic) based on shared characteristics of a group of users.Another example is CURATE.AI, a system designed to optimize medication dosing for individuals with hypertension and type II diabetes. 9With CURATE.AI, the treatment for each patient is personalized in a data-driven manner based on their unique response to medications.The system utilizes specific data from an individual patient, such as their response to drugs and dosages, to determine a personalized treatment, which means that the system adapts to the specific characteristics of each patient, leading to a more personalized approach to treatment.As demonstrated by these examples, the ways in which eHealth technologies can be personalized and tailored are diverse.
There is considerable variation-not only in the ways in which these technologies can be adapted to user variables but also in the types and number of variables used to segment eHealth users.For example, adapting the delivery channel can be matched to segments based on users' intensive exercise instructions 297 ).An example is MyPlan, 463 in which users fill out a questionnaire on fruit or vegetables intake or physical activity, which in turn is compared with health norms (comparative feedback).At a later stage this health behavior again is compared to their previous answers to the questionnaire (comparative feedback).
Cluster 4: Behavior + comparison with similar others (n = 33) eHealth personalization in cluster 4 can be characterized by the use of behavioral variables (such as alcohol and drug use 26 ) and demographic variables (such as age, gender, and country of origin 35 ) for user segmentation.These data are used to provide feedback in which the behavior of the user is compared to behavior of others who have similar demographic characteristics.A personalization from PNC-txt 118 that falls within this cluster compares the cannabis use of the user with cannabis use of people the same age.
Cluster 5: Behavior + recommendations (n = 29) Personalization in cluster 5 can be characterized by the use of behavioral data (such as frequency of binge eating 416 ) for segmentation, combined with one or more other types of variables (such as stress and fatigue 221 and occupational exposures 436 ).Adaptations in this cluster are mainly related to recommendations, such as advice and adapted difficulty levels.An example is the provision of an exercise routine on one of seven exercise levels based on exercise experience (behavioral) and pain (health). 442   Cluster 10: Psychological variables + feedback (n = 23) Cluster 10 can be characterized by the use of psychological variables for segmentation, such as mood and daily stress, 304 monitoring or blunting coping style, 289 and symptoms of social anxiety. 318Adaptations are mainly related to feedback and channeling of the eHealth technology.An example of a personalization within this cluster is Kelaa, 303 in which users' stress, well-being, and resilience (psychological variables) are used for segmentation.Users receive feedback on these psychological variables and advice on what they can do about this.Another example is CBTpsych.com, 318in which the user ranks thoughts and behaviors on how much they are relevant for their social anxiety that is used to decide on the course of treatment (sequence).
Cluster 11: Environment + recommendations (n = 28) Cluster 11 includes personalization that use environmental variables (such as the season, 466 the weather, 171 and cannabis use in the users' peer network 118 ), regularly combined with preferences (such as the users' tastes in food, 466 exercise preferences, 451 and preference for advice with or without medication 210 ).These variables are mainly translated in advice, such as providing recommended HIV testing facilities near the users' location 377 and providing exercise recommendations in line with the preferences of the user. 451Another example is the ANODE program, 466 which generates menus (advice) on the basis of the users' preferences and the season (environmental).
Cluster 12: User interaction + reminders (n = 12) Personalization in cluster 12 can be characterized by using only user interaction information for user segmentation.Examples are whether the user had logged any challenges 300 and whether the electronic diary within the eHealth technology was used. 264Adaptations mainly include reminders and very infrequently tunneling.Examples are reminders to use the eHealth technology and to enter measurement values 242 and highlighting website components the user did not navigate to (tunneling). 85uster 13: Health + recommendations (n = 48) Personalization in cluster 13 can be characterized by the use of health variables for user segmentation, such as the user's fitness level, 156 ALDH2 genotype, 62 and symptoms and medical history. 210These health data are mainly used to provide recommendations, such as providing dietary advice 275 and providing adapted entry levels regarding the intensity of the exercise regimen. 275Other examples are sending medication reminders at the user's specified dosing time 340 and safety alerts when glucose or blood pressure levels or weight exceed a certain threshold. 231ich computational methods are utilized to match user segmentation variables with technology adaptations?
To gain more insight into the computational methods used per distinct cluster, an overview of the computational methods per cluster can be found in Figure 5 below.We identified six clusters in which a highly prevalent computational method was employed.Cluster 2 (demographics + identification) predominantly utilized variable substitution techniques (89.47%).In cluster 4 (behavior + comparison with similar others), comparison algorithms were predominantly employed (93.94%).For cluster 7 (determinants + behavior change strategy), classification-based methods were the primary computational approach (95.83%).In cluster 8 (preferences + channeling), user-directed systems emerged as the dominant method (70.77%).Similarly, in cluster 9 (determinants + channeling), classificationbased methods were predominantly utilized (78.57%).Lastly, cluster 12 (user interaction + reminders) relied on dynamic algorithms (100%).Clusters 1, 3, 5, 6, 10, 11, and 13 employ a variety of computational methods for personalization.Classification-based methods appear to be widely adopted in these clusters, but also other methods are utilized, such as the comparison algorithms in cluster 1 (n = 7), dynamic algorithms in cluster 5 (n = 2), conversion in cluster 6 (n = 2), variable substitution in cluster 10 (n = 5), and context-aware computing in cluster 13 (n = 8).Clusters 3 and 11 also utilized more advanced computational methods, such as association-rule learning, reinforcement learning, and predictive modeling.

DISCUSSION
This study aimed to identify and categorize diverse personalization approaches offering comprehensive insights into the strategies employed for eHealth personalization.With the 412 studies that were included in the current review, we were able to identify 13 clusters of personalization approaches that show similarities in either or both the type of segmentation variables and how the eHealth technology was adapted to these user segments.Overall, we found that most personalized eHealth technologies used behavioral segmentation variables such as alcohol consumption and physical activity, which is in line with a previous meta-analysis on eHealth tailoring. 486Similar to previous descriptions of how personalization can be applied in the eHealth design, 6,487 we found that eHealth technologies are mainly adapted by providing feedback to the user.In contrast, other clusters that we identified in the current systematic review are, to our knowledge, not so evident in eHealth literature, such as user interaction + reminders, determinants + channeling, and psychological variables + feedback.

Content adaptations (n) Description
Comparative feedback (116)  Comparison of current or past data with previous data (ipsative feedback), guidelines, norms (normative feedback), recommendations or individuals who have successfully adopted the target behavior.
Advice (91)  Based on the users' data, an advice is given through the eHealth technology (e.g., how to change the target behavior, advice on which meals to prepare).
Reflective feedback (41)  Feeding back data from the user in which a value is given to the data (e.g., right or wrong answers or perceptions about a certain topic, reinforcement of positive coping strategies).
Interpretative feedback (29)  Feeding back data to the user with an interpretation of the users' data (e.g., money spent on cigarettes based on smoking behavior or risk level for disease(s)).
Feedback (28)  Feeding back data from the user without an interpretation or comparison, such as totals, means, or directly feeding back data from the user (e.g., answers to questions: ''You indicated that .'' or mean number of steps per day).
Adapted difficulty level (18)  Content of the eHealth technology provides a difficulty level that is adapted to the data from the user, e.g., step goal or intensity of exercises.Delivery timing (18)  Deliver content or parts of the intervention at a specific time point, not related to exceeding a certain threshold (e.g., deliver parts of the intervention at their preferred time).
Frequency (10)  Adapt the frequency that the eHealth technology is delivered to the user.
Alarms (10)  Inform users about exceeding a certain threshold (e.g., increased risk for relapse).
Sequence (5)  Adapt the order in which different parts of the intervention are delivered to the user.

Delivery channel (4)
Adapting the type of channel through which the intervention is delivered based on user's data (e.g., via text or avatar).
We identified several gaps in the literature.First, previous studies have found that eHealth usage is often limited, which is regularly attributed to digital skills. 488Yet, none of the included studies used technology-related variables for user segmentation.Technology-related variables encompass a user's skills and experience with different forms of technology, such as digital literacy and experience with VR technologies. 7This absence of technology-related variables for user segmentation is notable, given the potential of these variables to increase adherence and engagement to eHealth technologies. 489,490Moreover, previous studies on adherence to eHealth technologies have found that often only young, educated women are adherent to eHealth technologies, 491 which might be explained by higher eHealth literacy levels. 492Given the promising prospects that personalization holds for increasing both engagement and effectiveness, 493 it is important to explore how personalizing to these technology-related variables can improve eHealth technologies.Technology-related variables offer opportunities to focus not only on adapting the content of an eHealth technology but also on other aspects such as providing a simple design or incorporating text-to-speech engines for low eHealth literacy users. 494In essence, addressing these technology-related variables in personalization strategies could bridge the gap between current usage patterns and the potential for increased engagement and effectiveness across a broader population of users.
Second, user interaction reminders were the least used in the included literature.A possible explanation for this is that the first versions of tailored (non-digital) health communication used printed materials and did not collect data on whether, for example, a user had read certain information or not (e.g., 495 ).In addition, when user interaction segmentation was used in the included studies, only dynamic algorithms were used, such as whether or not a user visited a particular page or not.A more personalized approach toward user interaction can be to segment users according to how their usage changes over time or how they respond to certain behavior change strategies. 496More advanced computational methods can be used that better exploit the opportunities for capturing the users' eHealth interaction in real-time through log data. 497For example, reinforcement learning can be utilized to analyze user interaction patterns to determine the most effective types and frequencies of reminders.By continuously observing how users respond to different types of reminders, reinforcement learning algorithms can learn which strategies lead to increased user engagement and adherence to the eHealth technology and adapt the eHealth technology to the individual user accordingly.
Third, the largest part of the included studies only used one type of variable in their eHealth technology for user segmentation.The implementation of personalization based on a single variable fails to capture the complexity inherent in individual behaviors, characteristics, and the process of behavioral change.It is important to recognize that the type and number of variables depend on individual differences, 7 so there is no ''one-size-fits-all'' approach or standard set of variables that can be used within each technology.Yet, adopting a multi-faceted approach to user segmentation holds promise in enhancing the efficacy of personalized eHealth interventions.To illustrate, health-related segmentation variables allow for adaptations like personalized recommendations.Integrating these variables with preferences, such as whether the user prefers recommendations with or without medications, might enhance these adaptations.Another approach involves the use of existing behavioral patterns combined with user preferences to adapt suggestions given to the user for increasing physical activity and improving dietary behaviors. 479 last gap in literature is related to the computational methods used in the different personalization clusters.We found that the current computational approaches mainly use classification-based methods, which are primarily suited for rather stable characteristics of users or eHealth technologies designed for short-term use.However, there are clusters with more variability in the computational methods employed, Table 5. Types and number of adaptations to the behavior change strategy and their descriptions

Adaptations to the behavior change strategy (n) Description
Stage matching (32)  Adapting the behavior change strategy based on the process of change of the user, such as messaging tips and tricks around the users' quit date and motivational messages for users who did not set a quit date, or focus on benefits of behavior change in the precontemplation stage.
Target (24)  Adapting the way in which the eHealth technology changes (determinants) of behavior, such as a focus on either increasing knowledge or self-efficacy or providing suggestions for overcoming barriers that the user identified.
Framing (8)  Adapting the eHealth technology by using words, other type of content, or in-depth or concise information in such a way that several aspects of what is described are implicitly highlighted with the assumption that this improves the behavior change strategy.

Graphical adaptations (n) Description
Similarity (7)  Adapting graphical aspects (e.g., avatars, background pictures) of the eHealth technology so that the user identifies with graphical aspects of the eHealth technology.

Structure of data representation (2)
Represent data from the user in such a way that the graphical representations (e.g., tables and figures) are adapted to the data of the eHealth user.
such as cluster 3 (behavior + feedback) and cluster 11 (environment + recommendations).This variability likely arises from the dynamic nature of the variables involved, like behavior and environment, which change over time, and the adaptive elements, like recommendations and feedback.For clusters displaying similar patterns, it could be beneficial to evaluate other computational methods.For instance, cluster 7 (determinants + behavior strategy) includes determinants that may change over time, suggesting that alternative computational approaches may offer added value.Thus, although classification-based computational methods might be more suitable in some contexts, it may not be sufficient for more dynamic characteristics or for eHealth technologies designed for long-term use.The use of dynamic algorithms enhances these approaches by collecting segmentation variables over time and adapting to changes in these variables.Yet, these dynamic computational methods do rely on general rules that have to be decided in advance by the designer of the technology.Therefore, these computational methods seem mainly appropriate in contexts in which the matching between segmentations and adaptations remains stable and when the designer of the technology can anticipate what adaptations should be matched with which user segmentations in a meaningful way.More advanced machine learning methods offer opportunities to further refine how segmentations are matched with adaptations in real time.
For instance, reinforcement learning techniques allow for adaptations based on segmentation variables collected through the eHealth technology.To illustrate, users can provide feedback on the messages they receive, which in turn is used to ''learn'' what adaptations are suitable for the individual user, and thus improve the matching of segmentations with adaptation strategies.Moreover, predictive models can be trained on data specific to each individual user (such as finding predictors for certain events that may only account for an individual user).These models learn patterns and behaviors unique to each user, allowing for highly personalized recommendations, predictions, or interventions that take into account individual variation.

Limitations of the study
A limitation of this systematic review is that the included studies sometimes described the way in which users were segmented differently, which may have affected the clustering results.For example, some studies describe the segmentation variable as ''risk of cardiovascular disease'' (health variable), whereas another study describes the same in more detail (what constitutes this risk of cardiovascular disease), such as

Functionality adaptations (n) Description
Self-monitoring (3) eHealth technology either provides functionalities for self-monitoring or does not provide this functionality for the user.
Support (1)  eHealth technology either provides functionality that allows for social support or does not provide this functionality for the user.
Text message reminders (1) eHealth technology either sends text message reminders to the user or does not send these text message reminders.Dynamic algorithms (54)  Computational approach that adapts and responds to changes in segmentation variables.
A booster was employed for participants who return to negative behavior after exhibiting positive behavior 203 (cluster 1); reminders to use the website and after not logging in for 7 days (cluster 12).
User-directed systems (47)  Computational method that allows the user to directly control the system's behavior, functions, and features.
During the sign-up process, users are prompted to specify their gender, first name, and their preferred conversation style in Dutch.This style choice includes distinctions between formal and informal conversation forms, which are adapted based on specific display rules 276 (cluster 8).
Conversion (27)  Applying a conversion factor or formula for an estimation based on the given segmentation variables.
The reported alcohol consumption is converted to the caloric value, and the maximum reported alcohol intake is converted to the blood alcohol concentration along with potential consequences 54 (cluster 3).
Context-aware computing (23)  Using contextual information, such as location and time to personalize the intervention delivery to the segmentation variables.
Two additional messages were sent at the participants' heaviest typical drinking times 42 (cluster 1).
Predictive modeling (13)  Computational method that is used to predict future outcomes based on segmentation variables.
During the initial two weeks, EMA surveys were gathered to personalize the lapse prediction algorithm.Risk alerts were then activated at the onset of the third week.A decision tree algorithm was employed to predict the likelihood of a lapse report in the subsequent EMA survey.The algorithm, based on both group and individual data, classified responses into four risk categories: no risk, low risk, medium risk, and high risk, aiding in timely intervention 267 (cluster 9).

Reinforcement learning (4)
Adapting the matching of segmentations and adaptations based on the collected segmentation variables to continuously refine the personalization approach.
After clustering user behaviors, MyBehavior uses an exploit-explore strategy to automatically generate suggestions based on users' past physical activities and food intake 479 (cluster 3).
(Continued on next page) ''physical activity,'' ''smoking'' (behavioral), ''age'' (demographic), and ''history of cardiovascular disease'' (health).It is essential that in future research, this is reported more explicitly. 498The data that were collected from users should be described, followed by a description of the computational method used, in this example, to calculate their risk of cardiovascular disease.Furthermore, due to the interdisciplinary nature of eHealth research, we chose to include databases that collectively cover a wide range of disciplines, including medicine, biomedical sciences, psychology, engineering, technology, and social sciences, while refraining from multiple databases within a single discipline.Although this approach mitigates potential bias toward a particular field, it is important to acknowledge the risk of omitting valuable records.Furthermore, our inclusion criteria focused specifically on evaluation studies with randomization, aiming to encourage users to make healthier choices without significantly altering their dietary habits 483 (cluster 3).
Figure 5. Overview of computational methods by cluster potentially overlooking more exploratory studies.For instance, it is likely that more advanced computational methods, which may still be in the exploratory phase, were not fully represented in our analysis.Lastly, the coding process embraced a collaborative approach, with segmentation and adaptation data partly double-coded (5%) and discrepancies resolved through consensus discussions.This method not only provided an understanding of the strategies used for personalization but also allowed for the incorporation of both existing frameworks and emergent themes, enriching the qualitative analysis.Although this approach facilitated comprehensive insights, it is important to acknowledge that the absence of formal inter-rater reliability assessment introduces some uncertainty regarding coding consistency.

Future research
A focus on several key areas is needed in future research efforts to advance the field of eHealth personalization.Firstly, we found that personalization approaches are diverse, and their added value can be explained in various ways.To illustrate, eHealth technologies can be personalized based on a single type of variable (e.g., eHealth interaction or preferences), but this can also be done based on multiple different types of variables (e.g., eHealth interaction combined with user preferences).These approaches might differ in the extent to which they increase the effectiveness of eHealth technologies (e.g., using more variables might work better).As such, we argue that future research should explore what effective components of personalization strategies are by employing advanced research methods (e.g., dismantling designs or factorial designs) and moving away from randomized controlled trial (RCT) studies that only take into account whether or not an eHealth technology was personalized to gain more insight into why and for whom certain personalization strategies work.As a preliminary step, our clusters of personalization approaches highlight the divergent approaches of personalization across various applications.This insight can inform future studies aimed at investigating how personalization enhances the effectiveness of eHealth technologies, such as its impact on adherence, engagement, perceived relevance, and other relevant possible working mechanisms. 490,499,500To illustrate, we believe that for instance, demographics + identification (cluster 2) shows different working mechanisms than cluster 1 in which the user is actually tunneled toward components of the eHealth technology that are presumed to be more relevant to the user based on behavioral segmentation variables.
Second, future research should focus on further exploring how advanced computational methods can be utilized for eHealth personalization.This can be done by examining approaches in other domains outside of eHealth as well as by considering lessons learned from exploratory studies within the field of eHealth itself, which were not included in the current review.It is also important to focus on the underlying theories of these advanced computational methods to explain their added value.By understanding the theoretical foundations, researchers can better articulate why these methods enhance personalization.

Conclusion
In conclusion, the broad range of clusters of personalization approaches we identified illustrates the multifaceted nature of eHealth personalization.The clusters of personalization approaches can be used as a resource for informing the design process.By comprehending the diverse applications of personalization, designers can integrate this knowledge into the development of eHealth technologies based on specific contextual needs.However, the finding that several possibilities of eHealth personalization have not yet been fully explored, such as the use of technology-related variables for user segmentation and the use of advanced computational methods to match user segmentations with adaptations, illustrates the importance of further research.

QUANTIFICATION AND STATISTICAL ANALYSIS
The extracted data were split into distinct personalization or tailoring strategies (e.g., an eHealth technology with two personalizations was split into two rows).The variables that were used to define user segments were deductively coded into the categories 'behavioral', 'determinants', 'health', 'demographic', 'preferences', 'psychological', 'environmental', 'eHealth interaction' and 'technology' based our framework developed in a previous study. 7Adaptations were coded using a combination of deductive coding into the categories 'content', 'channeling', 'behavior change strategy', 'graphical' and 'functionalities' (using the same framework), and inductive coding to define subcategories of these categories.These coded combinations of segmentation variables and adaptations were hierarchically clustered in RStudio using the Cluster package. 501Gower distances were used since data was both binary (whether a variable was used for segmentation) and categorical (what type of adaptation was used).The number of clusters was determined using incremental clustering, meaning that, starting from two clusters, extra clusters were added until an extra cluster did not add extra information.

Figure 2 .
Figure 2. Publication years of included studies

Figure 3 .
Figure 3. Types of variables used for segmentation

Figure 4 .
Figure 4. Target behaviors for each cluster

Table
. Number of segmentation variables per personalization approach a Per personalization approach.

Table 2 .
Number of personalization approaches per eHealth technology a Per distinct eHealth technology.

Table 4 .
Types and number of channeling adaptations and their descriptions

Table 7 .
Types and number of functionality adaptations and their descriptions