The TAM model, developed by Davis et al. (31, 39), is used to measure the acceptance, adoption, and use of IT and has become very popular with many validation studies (32, 33). It relies on five different constructs, namely, perceived usefulness (PU), perceived ease of use (PEoU), usage attitude (UA), usage intention (UI), and actual usage (AU). TAM does not focus on the success of the IT application, but on studying and predicting the user's intention to use it. TAM can be used for actual system use in both subjective (self-reporting questionnaire) and objective measurement (system log) (45). Different extensions have been proposed ever since its arrival, in an effort to strengthen TAM's explanatory power. Indicative examples include the work from: a) Liang and Yeh (46), who replaced the PU with the Perceived Entertainment, in an effort to understand if the users have different feelings when interacting with smartphone-based games across different settings and locations; b) Kim et al. (47), who added the construct of Perceived Value, to explore the system/service quality on users' beliefs of hospitality industry information management systems; c) Morosan (48), who added the construct of Perceived Innovation, when using TAM for exploring biometric systems utilization by hotels; d) Lee and Wan (49), who augmented TAM with the constructs of Functional Confidence and Familiarity for acceptance of airline e-ticketing services acceptance by travelers. These works show potential for TAM evolution; one of its referenced versions is TAM 3, which contemplates a comprehensive nomological network of the determinants of IT adoption and use by individuals (50). TAM has been tested and validated numerous times, and as Benbasat and Barki pointed out (51), TAM is considered as one of the most influential theories in IS.

A total of eight models already considered in, or resulting from, research on IT acceptance (40), are the combinatory basis for the formulation of the UTAUT, i.e., TRA (43); TAM (31, 39); motivational model (MM) (52, 53); TPB (30); a model combining TAM and the TPB (54); model of PC utilization (MPCU) (55); innovation diffusion theory (IDT) (56); and social cognitive theory (SCT) (57, 58). The behavioral intention is suggested by the UTAUT as the main factor that determines the use of IT application. A total of four key constructs have direct effect on the perceived likelihood of adopting the technology, i.e., Performance Expectancy (the degree to which a user believes that by using the IT application will gain attainment in his/her job performance), Effort Expectancy (the degree of ease associated with the use of the system), Social Influence (the degree to which a user is influenced by the opinions of important others, who believe that s/he should use the IT application), and Facilitating Conditions (the degree to which a user believes that an appropriate infrastructure, in terms of organization and technical perspectives, exists to support the use of the system) (40). Age, gender, experience, and voluntariness of use are considered as moderators of the effect of predictors (40, 59). The current version is UTAUT 2 (38, 41, 59), where the use of technology by individuals is underpinned by the effect of the three additional constructs, namely, Hedonic Motive (the fun or pleasure derived from using technology), Cost/Perceived Value (trade-off between perceived benefits and monetary cost of using the IT application), and Habit (the extent to which people tend to automatically perform behaviors) (41). The inclusion of hedonic motivation construct in the UTAUT 2 was found to be more important than performance expectancy (60), whereas the integration of price value provided a measure of the IS use cost within the consumer context (59).

D&M-IS-SM was first proposed in (44) and it is a process/causal model. It is based on Shannon and Weaver's model of communication (61, 62) and the extension by Mason (63). Actually, the idea behind was based on the assumption that the process in IS resembles the one in the communication system, considering IS as a process of information production, conveying it and transmitting it to the recipients/users. Motivated by the metrics of Shannon and Weaver's communication model, D&M-IS-SM was built upon three parts of instruments for measuring IS success, i.e., technical (system quality), semantic (information quality), and effectiveness success (use, user satisfaction, individual impact, and organizational impact) (44). D&M-IS-SM was criticized for the inclusion of the Use construct and thus was later revised (64, 65), where the Service Quality and Use constructs were replaced by the Intention to Use/Use and the Individual Impact and Organizational Impact were substituted by Net Benefits. The D&M-IS-SM serves a dual purpose: it provides a classification of success measures, and it shows how the measures influence each other. The D&M-IS-SM has numerously been tested and validated since its inception, with some studies showing the strong validity of the construct, whereas some others reporting partial validity only (66, 67).

As Madriana et al. (68) suggest, the selection of the suitable technology acceptance framework can be based on the availability of the variable(s) in the framework with evidence from previous studies that can be best predictors for the behavior intention, the validation of the framework (based on the popularity/citations), and the fecundity (“fertility” of the theory to generate new model and hypotheses). From a comparative analysis across the TAM, UTAUT, and D&M-IS-SM reported in (68), TAM exhibits the highest popularity and the most fecundity. With regard to the evidence from previous studies, TAM was preferred in evaluating users' adoption of mobile health apps for nutrition and active living, such as in (69–74). Furthermore, in the D&M-IS-SM, there is a very broad concept of Net Benefits; this poses some challenges in clearly and carefully defining the stakeholders and context in which such net benefits are to be measured (64, 65). Moreover, as D&M-IS-SM is focused on the quality of service, a consideration should be taken about the changes in the practice of service and in the conceptual view of service quality that the use of information technology introduced nowadays. The initial view of the service quality was on the services provided by the information system to its users. However, nowadays, the service is materialized via interactions between the user and the IT provider in a serving function; hence, the information and system quality are viewed as a resource owned by the IT provider, who can clearly influence the user's intentions to continue interacting with the IT provider (75).

Based on the aforementioned evidence, the TAM model has been selected here as a starting point to build a modified version (mTAM) and to explore the role of the PROTEIN app as the technology artifact that could help its users to change their behavior and adopt a healthier living, in terms of improved nutrition and activity.

Several digital mobile health technologies for self-management of nutrition (76) and physical activity (77) have been proposed in the digital market place. Nevertheless, a lack of expert nutrition and physical activity advice provided by current mobile apps is identified, igniting further concerns regarding personalized advice tailored to specific user preferences (including dietary and allergies, PA for heart health, etc.). For instance, a content analysis of more than 50 weight loss mobile apps showed that more than 60% of the apps lack expert recommendations (78). This fact leads to the demand for high-quality and evidence-based mobile app design/development process, considered as a co-creative process between designers, developers, researchers, experts/clinicians, and end-users (79).

To address the aforementioned, a holistic approach is suggested with the introduction of the PROTEIN smartphone app. The latter is developed within the framework of an ongoing European H2020 research and innovation project (2019–2022), namely, “PROTEIN: PeRsOnalized nutriTion for hEalthy livINg” (https://protein-h2020.eu/). The latter involves a multidisciplinary team of 20 partners from a total of 11 European countries, including partners from industry, research, and technology organizations, aiming to promote a healthy lifestyle by combining the latest technologies, such as smartphones and sensors, to offer personalized nutrition and physical activity plans (80). Already available via Google Play Store¹ and currently supporting eight different languages (i.e., English, Italian, Dutch, German, Portuguese, Greek, French, and Spanish), the PROTEIN app incorporates information provided in user profiles. This includes anthropometric, biochemical, physiological, and physical activity data and personal preferences and objectives, including eating and/or physical activity habits, and proposes personalized healthy-eating and physical activity advice (through personalized nutrition and physical activity plans) (Figure 1). An additional feature enables specialized advice to be provided to individuals living with or at risk of chronic diseases (e.g., obesity, cardiovascular disease, and diabetes). The PROTEIN app leverages automated decision-making tools, including profiling and AI, to propose advice on personalized meal plans and physical activities. In particular, by including a deep learning component to generate/update recommendations for the user's dietary intake (taking into consideration their dietary preferences, daily intake, lifestyle and physiological variables), the PROTEIN app provides recommendations for intakes within the expert-established targets (81). Furthermore, PROTEIN app supports additional functionalities, such as nearby restaurant recommendations, offering a range of price points and food types; customization of user's shopping cart using ingredients of eating plans proposed by the system, and incorporation of new meals into the system by recognizing images. This capability provides automated food volume estimation from just a single food image (82), facilitating the data input to the PROTEIN app and contributing toward the development of an accurate dietary assistant application. PROTEIN app acknowledges the effectiveness of health-related games to engage users in dealing with their health (83, 84) and informing them about deterioration of symptoms in diseases (85), and thus, it also incorporates different dietary games. In this way, PROTEIN app fosters adherence to the suggested plans, making the app fun to use, and providing the means for encouragement and/or (re)education on nutritional value of the various food groups toward the adoption of a healthy and balanced diet.

The PROTEIN app also provides a “Help center” portal for questions and/or suggestions to improve the functionalities of the PROTEIN app. However, it is important to underline that the PROTEIN app is not a medical device, meaning that information and advice that PROTEIN app provides does not constitute medical advice. Finally, complying with the General Data Protection Regulation (EU GDPR) (https://gdpr.eu), the PROTEIN app allows the users to download and have access to their data, withdraw from the application at any time and delete all previously collected data.

Since the PROTEIN app is co-designed with the users and implemented for their benefit, the analysis of their perceptions when using the PROTEIN app is crucial, as it could shed light upon the role of specific structural characteristics and functionalities that further foster their intention to use the PROTEIN app for scaffolding their behavior change toward healthier living. Consequently, the PROTEIN app can have a positive footprint in people's nutritional and physical activity habits. In this vein, an experimental study was set up as a framework to acquire user's perspectives about the use of the PROTEIN app and model them, accordingly, as explained next.

A total of ninety-three adult participants in total were asked to download and use the PROTEIN app within a 2-month period (15 April -−15 June 2020) and then to participate in an online PROTEIN survey (refer to Supplementary Material – Online Survey), to give feedback about their use of the PROTEIN app. The type of sample design followed was the non-probability sampling design, using as sampling technique the convenience sampling, i.e., each respondent is selected for inclusion in the sample based on the ease of the access. Since the subjects must be smartphone users, the sample frame adopted was social media platforms/channels (e.g., Twitter, Instagram, Facebook), emails, and web advertisements, supported with flyers. During the registration process of the PROTEIN app, the participants were asked to fill in the required field and access the “Terms of Use” and “Privacy Policy” pages to complete the registration process, being, thus, informed about the purposes of the PROTEIN project and the processing of their personal data. Once registered, users were sent an email to activate their account and allow them access to the PROTEIN app.

The online PROTEIN survey was designed and developed by a multidisciplinary team within the PROTEIN Consortium, based on a “co-creation” process (86), in which nutrition and physical activity experts, researchers, end-users, designers, and developers explore a problem and produce solutions together, considering their different approaches, needs, and perspectives (87, 88). In this way, an open, active, and creative process, where all relevant stakeholders are engaged, was followed.

For the design of the online PROTEIN survey, the following requirements were considered: i) least time consuming, i.e., no more than 20 min for completing it; ii) simple, objective, and quantitative structure, i.e., as little free text as possible; iii) inclusion of demographic information; and iv) feedback acquisition on what is more/less appealing to users. After review rounds within the PROTEIN Consortium, the final version of the online PROTEIN survey resulted in an efficient version. The online survey consisted of 28 questions, divided into sections, including rating scale, multiple-choice, dichotomous, and open-ended questions (refer to Supplementary Material – Online Survey). The online PROTEIN survey was provided to users utilizing the SurveyMonkey platform (SVMK Inc., CA, USA).

From the 93 participants in the study, 85 (refer to full demographic characteristics in the Results section) provided online consent and fully completed the online PROTEIN survey; hence, only their responses were kept for the modeling analysis, described next.

Taking into consideration the TAM model (refer to theoretical background section), this study adopts the PU, PEoU, UA, and UI from TAM model and introduces two additional constructs, i.e., the “Perceived Novelty” (PN) and the “Perceived Personalization” (PP), toward the model output of “Behavior Change” (BC). This results in a modified TAM (mTAM) framework, as depicted in Figure 2.

To explore the contribution of each mTAM construct to the BC, a series of hypotheses is introduced. Davis et al. (39) mention that both PU and PEoU directly impact usage intention (UI) via the influence of UA. Moreover, the study of Holden and Karsh (89) states that the PU and the PEoU positively impact UA, showing that when users are confident that the adoption of novel technologies helps to improve their work performance, they exhibit more positive UA for adopting these novel technologies. In addition, if users perceive that a novel technology is easier to learn and that they do not need to apply much effort on the technology, then they have a more positive UA toward adopting this technology. This is also supported by the findings of Chopra et al. (90) that stress the influence of PU and PEoU to the UA, when evaluating the acceptance of dietary service applications by smartphone users using TAM. In the same vein, Mohammadi and Isanejad (91) found an influence of PEoU to UA, when evaluating the acceptance of IT in activity organization. From these perspectives, Hypotheses 1 and 2 are proposed below:

Hypothesis 1 (H1): Users' perceived usefulness toward the AI-based smartphone PROTEIN app positively influences usage attitude.

Hypothesis 2 (H2): Users' perceived ease of use toward the AI-based smartphone PROTEIN app positively influences usage attitude.

A crucial characteristic of any innovation is its novelty. Wells et al. (92) tried to frame the PN of an IT innovation as a salient affective belief in the nomological network related to the technology adoption. They conclude that the PN is an affective belief that plays a significant role in the adoption of IT innovations. So, if users perceive that novelty, it influences the adoption process, representing new alternatives to the existing ones, then they have a more positive UA toward adopting this technology. This was supported by the work of Robinson et al. (93), who found that new trends in technology engage the users and transfer them the notion that they use new products and technology before others, i.e., they become the leaders in acceptance of new ideas and products. In fact, novelty indirectly affects user's intentions and behavior, especially regarding new technological innovation (94). In this line, Hypothesis 3 is proposed as follows:

Hypothesis 3 (H3): Users' perceived novelty toward the AI-based smartphone PROTEIN app positively influences usage attitude.

Personalized approaches seem to be needed to promote healthy behaviors. Personalization concepts associated with behavior change in mobile health can be categorized into four different dimensions, namely: users (related to user-related characteristics, e.g., personality, profile, need for cognition, and perception of social norms), system functionalities (related to the functionalities that can be found in apps, e.g., reminders, gamification features), information (related to the way information is transmitted, e.g., type of feedback provided), and mobile app properties (e.g., the aesthetics of the app) (95). In addition, a recent meta-analysis study that aimed to evaluate the impact of personalized mobile apps and fitness trackers on lifestyle behaviors (i.e., physical activity, diet, smoking and alcohol consumption) revealed that studies must explore the integration of multiple data from different sources and include personalized features other than content (96). In fact, smartphone users seem highly receptive to personalized content and services; personalized recommendations, for instance, can generate better responses according to the user personal preferences and interests (97). Some studies have also suggested that users' attitudes and behaviors can be affected by personalization aspects (98, 99). In this vein, it is very probable that users' PP affects UA toward UI and then to BC. Thus, we hypothesize the following (H4):

Hypothesis 4 (H4): Users' perceived personalization toward the AI-based smartphone PROTEIN app positively influences usage attitude.

Davis et al. (39) and Robbins (100) state that UA directly influences UI, meaning that users' real usage behaviors are determined by UI, and that their UI are determined by their individual UA. In addition, Heberlein and Black (101) suggest that if UA and UI are more specific, then the relationship between the two is more distinct. From these perspectives, it becomes clear that if the users have a more positive UA toward a system, then they will have a higher UI toward its use. This is also supported by the work of Chen et al. (102), who explored the individual's attitude toward and intention to download and use dietary and fitness apps, and the findings of Okumus et al. (103), who explored the factors that influence customers' acceptance of smartphone diet apps when ordering food at restaurants. Hence, the Hypothesis 5 is proposed:

Hypothesis 5 (H5): Usage attitude of the users toward the AI-based smartphone PROTEIN app positively influences usage intention.

Usage intention defines the degree to which an individual is willing to engage in a particular behavior (43). When individuals' UI is strong, the probability that they will engage in that behavior is higher, and consequently, UI and real behaviors present a strong correlation. Furthermore, UI has been considered the best variable to predict an individual's behavior (43). This view is also adopted in the work of West et al. (104), who analyzed how diet and nutrition-related mobile apps lead to behavior change, and in the work of Requero et al. (105), where the enhancement of the correspondence between attitudes and behavioral intentions was explored as promoter of healthy eating. In this vein, Hypothesis 6 is proposed, as follows:

Hypothesis 6 (H6): Usage intention of the users toward the AI-based smartphone PROTEIN app positively influences behavior change.

Stata version 14 (StataCorp) was used for the statistical analysis. Descriptive statistics were calculated for demographics and the related questions from the online survey (refer to Supplementary Material – Online Survey). To explore the magnitude of the influence between the mTAM constructs, a linear regression analysis was adopted and the corresponding parameters, i.e., coefficient of determination (R²), adjusted R² (adj R²= 1 - [(1-R2)(N-1)N-k-1]≤ R2, where N is the number of points in the data sample and k is the number of independent regressors (variables), excluding the constant), regression coefficient with standard error (SE), t ratio (the estimate divided by the SE), p - value, and 95% confidence interval (CI), were estimated. Values of p < 0.05 and |t| > 1.96 were set as thresholds for adopting statistically significantly difference from 0 of the coefficient's estimate at the 95% CI, hence for adopting the corresponding hypothesis. Moreover, the F-statistics, i.e., the ratio of the mean sum of squares of the model to that of the residual, were also estimated for measuring how the ratio of the explainable mean variance to the unexplainable mean variance is statistically >1. The corresponding degrees of freedom for the model and residual are given in parentheses.

The regression analysis has been selected instead of the structural equation modeling (SEM) (106–110), as the latter represents and relies upon the causal assumptions of the researcher, and its credibility depends on the credibility of the causal assumptions in each application (111). Moreover, there is a semantic difference between the coefficients originating with a regression where no causal assumptions are made vs. from a SEM that makes strong and weak causal assumption (111). In this way, with the adopted regression analysis, any bias from the classification to strong and weak causal assumptions in the construction of the mTAM framework is avoided.

The reliability and validity of the items used in the proposed mTAM framework have been tested, accordingly. In particular, for the reliability analysis to examine the internal consistency and stability of grouped variables, denoting how closely related a set of items are as a group, the Cronbach's α was estimated as a =MC¯var¯+(M-1)C¯, where M is the number of grouped items, C¯ and var¯ are the average covariance between item - pairs and average variance, respectively. Values of a ≥ 0.70 denote acceptable internal consistency (112). With regard to the validity analysis, the confirmatory factor analysis (CFA) (113) was adopted, and the factor loading (FL) with the related p-value, average variance extracted (AVE), and composite reliability (CR) parameters were estimated. According to Nunnally (114), if FL > 0.5, AVE > 0.5, and CR > 0.7, then the research variables exhibit good convergent validity.

The PROTEIN app was evaluated by eighty-five (n = 85) volunteers coming from eight European countries (i.e., Greece, Portugal, Germany, UK, Belgium, Italy, the Netherlands, and Serbia) with average age of 49.3 years (±std = 16.3 years; age range = [25−44] years), who responded anonymously to the online survey (average missing values <4). The highest percentage of the respondents, 32%, were in the 35–44 age groups followed by 30% of 25–34 years old, then 16% of 18–24 years old, 12.5% of 44–54, 7% of 55–64, and the least, 2.5% was for 65 and over years old. Moreover, the distribution of male/female was 56% male and 41.5% female. In terms of body weight and body height, the means were 72.5 kg (±std = 16.5 kg) and 164 cm (±std = 8.4 cm). The majority of users (74%) were classified as general public and therefore did not belong to any clinically defined group (such as users under medical/nutrition supervision).

The survey responses of the PROTEIN app users provided some insights into their general opinions of the app. In particular:

The aforementioned perspectives are propagated through the usage attitude and intention as means for activity and eating behavior change, as reflected in the adopted mTAM and the related hypothesis testing results as presented below.

Figure 3 depicts the path relationships between mTAM constructs (Figure 2) and the related items via the values of adj R² (^* p < 0.05;^** p < 0.01; ^*** p < 0.001), as they were derived from the performed regression analysis. Moreover, Table 1 tabulates all the corresponding results regarding different regression paths that relate items with separate constructs (nos. 19), and the regression paths between constructs that support the hypothesis testing outcome (nos. 1015), along with the corresponding Cronbach's α (where applicable). The effect of sex, age range, and BMI was tested for all the regression paths tabulated in Table 1. From this analysis, statistically significant differences related to sex were found for the regression paths of nos. 14 (p = 0.028, 0.013, 0.019, and 0.017, respectively). Moreover, statistically significant differences related to age range and BMI were only found for the regression paths of no 12 (p = 0.041) and no 13 (p = 0.002), respectively.

From a general perspective of the results presented in Table 1, it can be shown that the Cronbach's α values range from 0.7674 (group of items nos. 5, 6) up to 0.8456 (group of constructs nos. 10–13), revealing that the grouped items and constructs all have good internal consistency and reliability. In addition, the corresponding validity results, also tabulated in Table 1, show that all items have FL > 0.5, CR > 0.7, AVE > 0.5, indicating that the items/constructs within the mTAM framework have good convergent validity. All estimated values of F-statistics were significantly >1 supporting the acceptance of the related hypothesis. Moreover, in almost all cases, the estimated p - values are less than 0.001, with the exception of the regression paths no. 1 (p <0.01) and no. 5 (p <0.05), indicating extreme statistical significance. Additionally, the t ratio ranges from 2.54 (no. 5) up to 12.53 (no. 14), justifying the statistical validity of the estimated regression coefficients. Furthermore, R² varies from 0.110 (no. 5) up to 0.740 (no. 14) and the corresponding adj R² ranges from 0.100 (no. 5) up to 0.732 (no. 14), showing that, in most cases, there is a good fit of the regression models used to express the relation between the items/constructs of the proposed mTAM framework.

Following the paths of Figure 3, combined with the results of Table 1, it can be seen that the “Helps me to improve my health”, “Helps me to exercise more,” “Recommends me suitable activities,” and “Helps me plan healthier meals” items collectively contribute to the PU construct (Cronbach's α = 0.8367) of the mTAM. The corresponding adj R² values are 0.129, 0.193, 0.309, and 0.328, respectively, showing that the plan of healthier meals explains the highest part of the PU variance, compared to the “Recommends me suitable activities,” “Helps me to exercise more,” and “Helps me to improve my health” items. Moreover, the “Easy to organize supermarket shopping list” and “Easy in interaction” items both contribute to the PEoU construct (Cronbach's α= 0.7757) of the mTAM. The corresponding adj R² values are 0.327 and 0.387, respectively, revealing almost an equal contribution in the explanation of the PEoU variance. In its turn, the “Provides novel meal/activity recording” and “Differs from other related app” items collectively contribute for the PN construct of the mTAM α (Cronbach's α = 0.7674). Clearly, the latter item exhibits much higher adj R² value (0.719) when compared to the one from the former (0.100). This is anticipated, as the concept of novelty is first conveyed to the users via the comparison of the PROTEIN app with the state - of - the - art and then via the feature of the meal/activity recording. For the PP construct of the mTAM, the “Provides personalized notifications,” “Provides personalized achievements,” and “Provides complete personalized profile” items collectively contribute (Cronbach's α = 0.8071). From these items, the “Provides personalized notifications” item exhibits the highest adj R² of 0.411, followed by the “Provides complete personalized profile” (adj R²= 0.341) and the “Provides personalized achievements” (adj R² = 0.313) items. This shows that the sense of personalization is perceived by the users textitvia personalized feedback and the ability to customize the user's profile, taking into consideration the specificity of each user's preferences and sensitivities (e.g., allergies, insufficiencies).

Moving on to the hypotheses testing, the results from Table 1 justify that all hypotheses, i.e., H1-H6 (refer to Figure 2) can be accepted. In particular, there is a collective contribution of the PU, PEOU, PN, and PP constructs to the UA construct (Cronbach's α = 0.8456). In addition, the corresponding adj R² values are 0.306, 0.436, 0.224, and 0.424, respectively, showing that the PEoU explains a high part of the variation in the UA, followed by the PP, PU, and PN. There is a strong connection between the UA and UI constructs, as the corresponding adj R² value is 0.732, revealing that the establishment of a solid usage attitude will effectively lead to an established usage intention. Consequently, this solidification could feed the BC construct (adj R² = 0.445), explaining a significant part of its variance.

The aforementioned results showcase the validity of the proposed mTAM framework and related hypotheses to explain the interaction between the items and constructs that are activated during the process of PROTEIN app evaluation and its potential users' acceptance.