Skin cancer is a highly prevalent global health concern, with melanoma, the deadliest subtype, posing a significant challenge. Early detection is crucial for improving survival rates, and clinical guidelines recommend full-body skin examinations by trained dermatologists and the use of dermoscopy for assessing skin lesions [1,2]. However, a growing shortage of dermatologists is exacerbating diagnostic delays [3,4]. Even today, more than half of all skin cancers are first noticed by patients themselves rather than detected during professional screenings [5], which has driven the exploration of artificial intelligence (AI) to support skin cancer detection. Empirical studies show that deep learning models can achieve diagnostic performance comparable to human experts in analyzing dermoscopic images for melanoma [6,7]. When integrated into mobile apps, these systems provide immediate risk assessments through automated, deep learning–based image analysis. Despite their potential to support early detection, adoption rates remain low, as many users continue to prefer conventional dermatological consultations [8-10].

This reluctance reflects a broader adoption gap identified in research on AI in health care—a discrepancy between the high diagnostic performance of AI systems and their limited acceptance in real-world practice [11,12]. While this gap is well-documented, less is known about the specific digital cues that drive such skepticism. Given that users increasingly rely on app store information to evaluate AI-based health apps, understanding how user-generated ratings shape acceptance becomes critical—particularly as these ratings represent a powerful yet understudied influence on decision-making in high-stakes health care scenarios.

Despite extensive research on rating valence in e-commerce, 3 key gaps remain. First, the role of rating valence in high-stakes health contexts, such as AI–based skin cancer screening apps (AISCSAs), remains largely unexplored. Second, prior research has not systematically integrated rating valence into the technology acceptance model (TAM) [13], particularly with regard to its causal effects on perceived usefulness (PU), perceived ease of use (PEOU), trust, and outcomes such as behavioral intention to use (BI) and willingness to pay (WTP). Existing studies predominantly rely on observational, nonexperimental designs, limiting causal inference. In contrast, this study uses an experimental manipulation of rating valence (positive vs negative) to directly examine its causal impact on these constructs. Third, prior work relies almost exclusively on self-reported measures and thus provides limited insight into the underlying cognitive processes, neglecting objective behavioral data such as eye-tracking.

To address these gaps, this study develops an integrative framework that combines the TAM, trust, and rating valence within a dual-process perspective of technology acceptance. By linking experimentally manipulated rating valence with both self-reported measures and eye-tracking data, we provide novel insights into how rating valence causally shapes acceptance beliefs and intentions, how users visually process app store information, and how closely self-reported evaluations align with actual attentional behavior.

Participants were recruited via university mailing lists, flyers, and student WhatsApp groups. Eligibility criteria included a minimum age of 18 years and fluency in German. Data collection took place between April 28 and May 31, 2025. A total of 95 participants were initially recruited; 19 were excluded due to insufficient eye-tracking accuracy or failed attention checks, resulting in a final sample of 76 participants (positive condition: n=42; negative condition: n=34). An a priori power analysis (G*Power; α=.05; power=0.95; effect size=0.80) indicated a required sample size of at least 35 participants per group, confirming sufficient statistical power. In response to the reviewers’ and editor’s comments, we have streamlined and refined our hypotheses. Importantly, this revision involved a structural clarification rather than changes to the underlying analyses or hypothesis testing. The updated hypotheses and their correspondence to the original versions are presented in Table S1 in Multimedia Appendix 1.

Gaze data (fixation duration) were recorded using a Tobii Pro Spark eye tracker (60 Hz), mounted below a 22-inch laptop screen (1920×1080 pixels, 60 Hz). Sessions were conducted individually in a controlled laboratory environment, with participants seated approximately 45‐55 cm from the screen [70]. A standard 9-point calibration was performed prior to the experiment, and only calibrations with a validation accuracy below 0.80° were accepted [71].

The stimulus for the eye-tracking experiment was a mock-up of an Apple app store listing for a fictitious skin cancer screening app (“SkinScan”), modeled after existing apps. To reduce familiarity bias, app name and content were adapted. Two versions of SkinScan were created, identical in all aspects except for the middle of the 3 user star ratings: the positive rating condition showed two 5-star ratings and one 4-star rating, while the negative rating condition replaced the middle 5-star with a 1-star rating to introduce the experimental manipulation (Multimedia Appendix 2)Multimedia Appendix 2. All other elements, including the overall star rating and review text, were identical across conditions. Both app versions were integrated into and presented through Tobii Pro Lab.

After providing informed consent, participants were seated in front of the eye-tracking system. They were instructed to view the app page as they normally would and to imagine evaluating its usefulness. Subsequently, a 9-point calibration procedure was conducted. Following calibration, the experiment began. No time limit was imposed, and gaze behavior was recorded continuously throughout the task. After stimulus exposure, participants completed a questionnaire on a tablet device and received compensation (see Figure 2 for the procedure of the experiment). A small pretest (n=3) using a think-aloud approach was conducted to ensure clarity and usability of the study design.

The questionnaire consisted of 29 items. The first section measured PEOU, PU, trust, and BI using validated multi-item scales, with each construct measured by 4 items (see Table S2 in Multimedia Appendix 1 for a complete list of items and scale specifications). The second section assessed the self-rated importance of app attributes (eg, price, reviews, and data protection) on 5-point Likert scales (1=not at all important and 5=very important). The third section included self-rated AI knowledge, WTP, and demographic variables (age, sex, previous experience with health apps, and education). All scales were adapted from prior research. Following established scale development procedures [72,73], content modifications were made to reflect the study’s objectives. To ensure linguistic accuracy, all items were translated using a back-translation procedure [74]. An attention-check item was included to ensure data quality. A full list of questionnaire items is provided in Table S2 in Multimedia Appendix 1.

To assess potential common method bias, the Harman single-factor test was conducted [75]. An exploratory factor analysis including all questionnaire items was performed. Parallel analysis indicated a multifactor structure, and the single-factor solution accounted for only 30% of the total variance. Moreover, the 1-factor model showed poor fit (χ²₁₃₅=4364.2; P<.001). These results suggest that common method bias is unlikely to be a significant concern in this study.

Fixation durations (in seconds) were recorded within predefined AOIs (Multimedia Appendix 3)Multimedia Appendix 3 and processed using Tobii Pro Lab. Fixation duration data were aggregated into composite categories (eg, reviews, ratings, and description) and normalized by AOI size (seconds per square pixel) to account for differences in area. Mean fixation durations per AOI were then calculated at the participant level. Participants with insufficient tracking accuracy or missing data were excluded (n=19), yielding a final sample of 76.

Prior to analysis, scatterplots were examined to assess linearity and identify influential outliers. Pearson correlations were used to test relationships between variables, and independent-samples 2-tailed t tests (including Cohen d) were conducted to compare experimental conditions. Multiple regression analyses were performed to assess predictors of BI and WTP. Mediation and moderation effects were tested using regression-based approaches with bootstrapped CIs. All analyses were conducted at a significance level of P<.05, with 95% CIs reported.

The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008. An ethics approval was granted by the University of Applied Sciences and Arts Bielefeld (2025_002), and the study was preregistered. Informed consent was required from all participants in this study. Participants received approximately US $5.83 compensation.

The final sample consisted of 76 participants. Of these, 42 of 76 (55%) participants were randomly assigned to view the positive rating condition, while 34 of 76 (45%) participants saw the negative rating condition. As shown in Table 1, participants had a mean age of 22.88 (SD 3.13; range 18-33) years and a median age of 22 (IQR 21-24) years. The majority of participants were female (53/76, 70%). Most reported having a general school leaving certificate (49/76, 64%), while 13 of 76 (17%) held a university degree. The mean level of self-rated AI knowledge was 2.65 (SD 0.91). None of the participants had prior experience with skin cancer screening apps.

To address the research question of how a single negative rating affects behavioral intentions, independent-samples comparisons were conducted to examine the effect of rating valence on trust, PU, PEOU, BI, and WTP (Table 2). The results indicated no significant differences between positive and negative rating conditions for any of the examined variables (all Ps>.05). Thus, H1a-e are not supported.

To investigate the research question concerning the applicability of the TAM framework in connecting rating valence to behavioral intentions, Pearson correlation coefficients were calculated to examine the strength of the relationships between the hypothesized constructs (Figure 3). PEOU was positively correlated with PU (H2b; r=0.31; P=.007) and trust (H2a; r=0.27; P=.02). Trust also showed a strong positive correlation with PU (H2c; r=0.63; P<.001). Regarding BI, all proposed constructs were significantly positively correlated: PEOU (H3a; r=0.37; P=.001), trust (H3b; r=0.62; P<.001), and PU (H3c; r=0.68; P<.001). In addition, Pearson correlation analyses revealed a small but significant positive association between trust and WTP (H4b; r=0.17; P<.001). In contrast, WTP was not significantly correlated to PU (H4c; r=0.11; P=.35) or PEOU (H4a; r=0.04; P=.76).

To further examine the predictors of BI, a multiple linear regression analysis was conducted including rating valence (condition), PU, PEOU, and trust. The overall model was significant (F_4,71=22.36; P<.001), explaining a substantial proportion of variance in behavioral intention (R²=0.56; adjusted R²=0.53). PU emerged as the strongest predictor of BI (β=.65; P<.001). Trust also showed a significant positive effect (β=.37; P=.01), while PEOU had a marginal positive influence (β=.27; P=.06). In contrast, rating valence (condition) did not significantly predict BI (β=−.20; P=.18). Overall, the findings closely resemble the individual correlations, with the exception that the effect of trust weakens once PU is included, pointing toward a potential mediation effect.

To examine whether rating valence and the TAM-related constructs predict WTP, a multiple linear regression analysis was conducted including condition, PU, PEOU, and trust. The overall model was not significant (F_4,71=0.77; P=.55) and explained only a negligible proportion of variance in WTP (R²=0.04; adjusted R²=−0.01). None of the predictors showed a significant effect on WTP, including rating valence (β=−.20; P=.32), PU (β=.02; P=.85), PEOU (β=.00; P=.99), and trust (β=.10; P=.39). Overall, the regression results largely mirror the correlation findings. However, the initially observed association between trust and WTP did not remain significant in the regression model, suggesting that this relationship is weak and not robust when controlling for other variables.

To examine whether the effect of rating valence on BI is mediated by PEOU, PU, and trust (H5a-c), mediation analyses were conducted (see Table 3 and dashed lines in Figure 3). Results revealed no significant indirect effects for any of the mediators, as all CIs included 0. Additionally, no significant total or direct effects of rating valence were observed. Thus, H5 is not supported.

To examine the research question of where users look for information when browsing for a skin cancer screening app, we analyzed how participants allocate visual attention across the app store interface. During data preprocessing, individual AOIs were aggregated into composite AOIs. Specifically, all review-related AOIs were combined into a reviews-AOI, all rating-related AOIs into a ratings-AOI, icon and screenshots into a design-AOI, and information and description into a description-AOI. To control for differences in AOI size, fixation durations were normalized and expressed as seconds per pixel² (s/px²). As shown in Table 4, the description-AOI attracted the longest fixation durations (mean 0.16, SD 0.10 s/px²), followed by the reviews-AOI (mean 0.11, SD 0.07 s/px²) and the ratings-AOI (mean 0.04, SD 0.04 s/px²). In contrast, the price-AOI (mean 0.02, SD 0.03 s/px²) and the data protection-AOI (mean 0.02, SD 0.02 s/px²) received the least attention.

To examine the research question of whether the self-rated importance of app attributes is reflected in users’ gaze patterns, Pearson correlations were computed between participants’ ratings and normalized fixation durations on the corresponding AOIs (Table 5). Results revealed a significant positive correlation for reviews (r=0.28; P=.01), while all other relationships were nonsignificant (Ps>.05). Thus, only H6b is supported.

To address the research question of how rating valence affects gaze patterns, fixation duration on the manipulated review-AOI was compared across rating conditions. Absolute fixation durations were used, as the AOI size was identical across conditions. Results showed no significant difference between the positive (mean 6.17, SD 4.15 seconds) and negative conditions (mean 6.21, SD 3.98 seconds; P=.97; d=−0.01). Although participants in the negative condition showed a slight tendency toward longer viewing times, this effect was not statistically significant. Thus, H7 is not supported.

To test whether visual attention moderates the relationship between rating valence and key constructs (H8a-d), moderation analyses were conducted using fixation durations on the ratings-AOI (Table S3 in Multimedia Appendix 1). Results indicated no significant interaction effects between rating valence and fixation durations on PU, PEOU, trust, or BI (all Ps>.05). Thus, H8 is not supported.

Analyses of individual differences revealed no significant effects of demographic or experiential variables on the main outcome measures (trust, PU, PEOU, BI, and WTP; all Ps>.05). Notably, both age and prior experience showed limited variance within the sample, which restricts the interpretability of their relationships with the outcome variables. In particular, none of the participants reported prior experience with skin cancer screening apps. Furthermore, self-reported AI knowledge was not significantly associated with any of the outcome variables (all Ps>.05). Similarly, no significant differences were observed across sex or education levels (all Ps>.05)

Furthermore, we explored potential sex differences in self-rated app attribute importance and gaze patterns. Overall, female and male participants evaluated most app attributes similarly. No significant sex differences were observed for the importance of price, data protection, or evidence (all Ps>.05). However, female participants rated the importance of reviews (d=0.63; P=.01) and design (d=0.69; P=.02) significantly higher than male participants (Table 6).

With regard to gaze patterns, female participants demonstrated longer fixation durations than male participants across all AOIs; however, this difference reached statistical significance only for ratings (d=0.47; P=.03). No significant sex differences were observed for fixation durations on data protection, description, reviews, design, or price (all Ps>.05; Table 6).

This study investigated the acceptance of AISCSAs by integrating the TAM with eye-tracking and the experimental manipulation of rating valence. Importantly, by systematically varying rating valence as an external heuristic cue, the study enabled a direct test of how such cues causally influence the formation of PU, PEOU, and trust, as well as BI and WTP. Our findings reveal a hierarchy of visual attention dominated by textual descriptions and social proof while also identifying a significant gap between self-reported attribute importance and actual gaze behavior. Although the core TAM constructs (PU, PEOU, and trust) emerged as robust predictors of BI, the experimental manipulation of rating valence did not significantly alter acceptance beliefs or attention patterns in this specific laboratory context.

Contrary to our expectations, rating valence did not significantly influence trust, PU, PEOU, BI, or WTP. Likewise, the negative rating condition did not increase visual attention to the review section. Thus, neither H1 nor H7 was supported. At first glance, these findings appear inconsistent with research on negativity bias and online review effects, which often show that negative information attracts more attention and exerts a disproportionately strong influence on judgments [24,60]. However, our findings are more plausibly interpreted as evidence of a boundary condition than as evidence against the relevance of rating valence per se.

A first explanation concerns the subtlety of the manipulation. In this study, rating valence was altered only minimally, replacing 1 of 3 visible ratings with a single 1-star evaluation, while all other cues remained favorable. Such a minor discrepancy may have been insufficient to shift appraisals of a high-stakes health app. This is consistent with meta-analytic evidence by Purnawirawan et al [76], which demonstrates that valence effects are not uniform; rather, they are highly dependent on the degree of extremity and the evaluative context, such as product type and brand familiarity. In high-stakes environments, a single negative cue may lack the weight necessary to override a predominantly positive information environment. In addition, in high-stakes contexts, users may engage in more motivated and systematic processing, thereby reducing the relative impact of simple heuristic cues such as rating valence.

A second explanation is that app store ratings may have limited informational value in digital health contexts. Several studies show that consumer ratings are poor proxies for clinically relevant attributes such as quality, safety, and privacy. For example, Hyzy et al [77] found that ratings and downloads do not reliably reflect professional assessments, while de Chantal et al [78] reported a clear gap between user ratings and expert evaluations. Similarly, Levine et al [79] demonstrated that star ratings correlate only weakly with critical dimensions such as privacy, security, and interoperability, with even highly rated apps often performing poorly in these areas. This suggests that rating information may not provide users with meaningful signals for evaluating clinically relevant aspects of health apps. As a result, ratings may play a limited role in shaping acceptance beliefs.

From a dual-process perspective, the results indicate that the manipulated rating did not cross the threshold required to redirect processing from a default heuristic impression (type 1) to more effortful elaboration (type 2). While negative information can trigger deeper scrutiny [62], this is most likely when it is salient, coherent, and clearly diagnostic. Here, the cue may have been noticeable but not sufficiently consequential to reorganize attention or alter technology beliefs. This interpretation also fits the absence of any moderation by fixation durations on the ratings-AOI (H8): without a sufficiently strong evaluative signal, there was little for visual attention to amplify. Importantly, the null findings for H1 and H7 should not be interpreted as evidence that rating valence is irrelevant in digital health. Rather, they indicate that isolated, weakly negative rating cues are insufficient to meaningfully influence acceptance beliefs or increase visual attention in a high-risk medical AI context.

Despite the absence of direct effects of rating valence, the internal structure of the extended TAM was strongly supported. The positive relationships between PEOU, PU, and trust (H2a-c), as well as their associations with BI (H3a-c), are consistent with prior research, demonstrating the robustness of the TAM in digital health contexts [37,40,80]. The regression results further refine this picture by highlighting a differentiated role of the constructs: PU emerged as the strongest predictor of BI, followed by trust, while PEOU showed only a marginal effect. This pattern suggests that usefulness constitutes the primary decision criterion, whereas ease of use operates mainly indirectly via PU.

The clear hierarchy of predictors reveals a specific decision-making logic in high-stakes AI health contexts. The dominance of PU suggests that for potentially life-impacting technologies such as AISCSAs, users follow a strongly instrumental evaluation logic, prioritizing diagnostic utility over interface convenience. In such contexts, perceived functional value appears capable of compensating for usability barriers, indicating that effectiveness outweighs ease of use. This finding is consistent with prior research in health care contexts, showing that PU is typically the strongest predictor of behavioral intention [34,40].

At the same time, our findings highlight the important role of trust in AI-based health care. While PU emerged as the strongest predictor of BI, trust remains as an independent predictor and is particularly relevant in contexts characterized by uncertainty and limited user expertise. In such settings, users rely on systems whose diagnostic processes are not easily interpretable, making trust a necessary condition for acting on PU. In this sense, trust may function as an enabling mechanism or boundary condition that determines whether PU translates into behavioral intention. This interpretation aligns with trust theory, which emphasizes its importance under conditions of uncertainty and risk [81,82], and with prior extensions of the TAM that position trust as a core belief construct [34,83-86].

Interestingly, the marginal effect of PEOU challenges the traditional emphasis on usability as a key driver of technology acceptance [13] and aligns with prior research, showing that PU is typically the dominant predictor, particularly in high-stakes contexts [34,37]. In this study, PEOU appears to play a secondary role once a system is perceived as useful and trustworthy. Users may be willing to tolerate complexity if it is associated with higher perceived competence or clinical rigor. This suggests that simplicity is not the primary objective in AI-based health care apps, with reliability and effectiveness outweighing usability considerations.

In contrast, WTP was not explained by the included variables. As WTP was measured using an open-ended response format, the variable was not normally distributed. To account for this, additional nonparametric correlation analyses (Spearman ρ) were conducted, yielding the same pattern of results. WTP was not significantly associated with PU or PEOU, and its relationship with trust remained weak (H4a and H6c not supported). This indicates that WTP represents a distinct and less stable evaluative judgment that cannot be fully explained by the traditional TAM constructs [23,28]. Instead, monetary evaluations appear to be driven more strongly by psychological and normative factors, such as perceived risk, fairness, and emotional or social considerations [27]. Accordingly, while the TAM provides a robust framework for explaining acceptance intentions, it appears less suitable for capturing economic decision-making without incorporating additional constructs.

The lack of significant mediation (H5) and moderation (H8) effects is directly tied to the nonsignificant main effect of rating valence. Without an initial “push” from the experimental condition, the “amplification effect” of visual attention, where fixations serve to increase the weighting of attended information [65,66], had no baseline signal to amplify. These findings underscore the robustness of the TAM framework while highlighting that single-cue manipulations may be insufficient to shift these deeply interrelated constructs.

The eye-tracking results reveal a clear attentional hierarchy: users primarily focused on the app description, followed by reviews and ratings, while price and data protection received little attention. This pattern suggests that, in high-stakes contexts such as AI-based health apps, users prioritize information that helps them understand the app’s functionality and potential clinical value, indicating more deliberate and analytical processing (type 2). This finding is consistent with prior research showing that textual information plays a central role in app evaluation, despite developers often emphasizing visual design [54,87].

At the same time, the secondary attention given to reviews and ratings highlights that social proof remains relevant, particularly under uncertainty. These cues are easily interpretable and visually salient, making them effective heuristics for initial evaluation and risk reduction [19,21,60]. However, the fact that they received comparatively less attention than the description suggests that social proof complements rather than dominates decision-making in this context. Overall, the findings indicate that users integrate both analytical and heuristic processing, with greater weight placed on substantive, information-rich content.

In contrast, the data protection-AOI and the price-AOI received minimal attention. The limited focus on these elements is consistent with prior eye-tracking research, showing that users often overlook pricing information and security indicators in digital environments [58,88,89]. Several factors may explain this pattern. First, many health apps follow freemium models, where pricing information is either absent or only becomes relevant after installation. Second, although users frequently report privacy as a major concern [90,91], the complexity and legalistic nature of privacy policies may discourage active engagement. In addition, users may rely on implicit trust in platform-level regulation or assume a baseline level of data protection for health-related apps [92]. Finally, the low visual salience of these sections likely contributes to their neglect, as prior research shows that subtle privacy indicators are easily overlooked, whereas more prominent cues attract attention [56]. Taken together, these findings provide behavioral evidence for the well-documented “privacy paradox”: despite expressing concerns about data protection, users allocate little attention to related information during decision-making.

A central finding of this study is the limited alignment between self-reported importance and actual visual attention, with the exception of reviews (H6b). While participants who rated reviews as important also allocated more attention to them, no such relationship was observed for design, price, evidence, or data protection (H6a and H6c-f). This discrepancy suggests that self-reports and gaze data capture different aspects of the decision-making process. In line with prior research, eye-tracking reflects implicit attentional processes that are not accessible through retrospective self-reports [41].

From a dual-process perspective, this divergence can be interpreted as a misalignment between type 1 and type 2 processing [14]. Self-reported importance ratings primarily reflect deliberative, reflective evaluations (type 2), often shaped by normative expectations or socially desirable responses, particularly in sensitive contexts such as health care. In contrast, eye-tracking captures spontaneous, heuristic-driven allocation of attention during actual interaction (type 1). This gap is particularly pronounced for data protection. Although participants rated privacy as highly important, it received little visual attention during the task. This pattern is consistent with prior findings on the “privacy paradox,” where expressed concerns do not translate into corresponding behavior [57,58]. Taken together, these findings highlight that technology acceptance cannot be understood as a purely rational process, as suggested by traditional TAM frameworks. Instead, it reflects an interplay between reflective evaluations and implicit attentional processes. By capturing these behaviorally grounded mechanisms, eye-tracking provides a valuable complement to self-report measures and offers a more nuanced understanding of how users evaluate health-related AI systems.

Our exploratory data suggest sex-specific patterns in app evaluation, with female participants assigning greater importance to reviews and design and showing longer fixation durations on ratings. This aligns with prior research, indicating that female participants tend to engage in more comprehensive information processing and pay closer attention to user-generated content when evaluating digital products [60,67,93]. Beyond sex, recent studies highlight the broader relevance of sociodemographic factors in shaping AI acceptance in health. For instance, Aras et al [68] show that younger individuals, female participants, and those with higher income levels are more likely to adopt AI in health care. Similarly, Méndez-Suárez et al [69] find that positive attitudes toward AI-based products are more prevalent among younger individuals and those with higher socioeconomic status. Complementing these findings, Kauttonen et al [94] demonstrate that individuals with moderate knowledge exhibit the highest acceptance of AI apps in health, whereas both low and high knowledge levels are associated with more cautious evaluations.

The absence of significant effects for AI knowledge and prior experience in this study likely reflects the relative homogeneity of the student sample, which limits variability in key characteristics. This highlights an important boundary condition of the findings. Future research should therefore examine more diverse populations, particularly older users and individuals with varying socioeconomic backgrounds and levels of digital literacy, to better capture how individual differences shape trust and acceptance in AI-based health care contexts.

This study, while offering valuable insights, is subject to several limitations that should be acknowledged. First, although the stimulus used in the eye-tracking experiment ensured a high degree of ecological validity by closely resembling an actual app store page, the experimental environment and the task itself remained artificial and hypothetical; for instance, participants did not autonomously decide to search for an AISCSA. This setting may not fully capture the complexity and dynamics of real-world app evaluation and decision-making. Moreover, the absence of real consequences, such as actually downloading or using the app, could have influenced participants’ engagement, attention patterns, and responses. Second, all participants were young university students with a narrow age range and no prior experience with skin cancer screening apps. This homogeneity limits the generalizability of the findings to other populations, especially older users or individuals with specific health concerns, who may differ in their levels of trust, digital literacy, and motivation. Therefore, the findings should be interpreted with caution, particularly when generalizing to more diverse or clinically relevant populations. Third, the internal consistency of some constructs, such as PU, PEOU, and trust, was relatively low, with Cronbach α values falling slightly below the conventional 0.70 threshold. While these levels are often considered acceptable in exploratory research within emerging fields, they suggest a moderate reliability, which should be taken into account when interpreting the strength of the observed relationships. Future research could benefit from further refining these scales to enhance their psychometric robustness in the context of AI-based health apps.

Addressing a key gap in prior research, this study extends the TAM by experimentally manipulating rating valence as an external variable and examining its effects on trust and core acceptance constructs. While prior research has largely relied on observational data, this approach enables a direct test of how external information influences the TAM-related belief structures.

The findings confirm that PU, PEOU, and trust remain significant predictors of behavioral intention in the context of AISCSAs. However, rating valence showed no direct effect on user acceptance or visual attention. This suggests that in high-stakes health contexts, users are less influenced by simple social-proof cues and instead prioritize substantive, information-rich content.

The primary implication for developers and marketers of health-related AI tools is that relying solely on positive ratings to drive adoption may be an insufficient strategy. Instead, efforts should be directed toward providing transparent, evidence-based information, as the description-AOI received the longest fixation duration. Building acceptance through verifiable information appears to be more critical than leveraging simple social proof in this domain. Moreover, designing a more appealing data protection section may increase user attention, rendering them better-informed consumers. Importantly, the weak or nonexistent associations between fixation durations on specific AOIs and the self-rated importance of app attributes indicate that eye-tracking captures aspects of information processing that are not directly reflected in self-reported questionnaires. These findings demonstrate that combining self-report measures with behavior-based methods, such as eye-tracking, provides a more comprehensive understanding of user acceptance processes in health-related AI systems. This dual approach captures both reflective, self-reported evaluations (type 2) and implicit, attention-driven processes (type 1), which are not accessible through questionnaires alone.