Technology acceptance of artificial intelligence for cognitive assessment in autism spectrum disorder: a structural equation modeling approach with robustness and predictive analytics

1 Introduction

Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder and one of the most challenging-to-diagnose disorders (Lord et al., 2018; Lord et al., 2020). Typical cognitive assessment suggests is through cognitive testing from the clinician using standardized cognitive tests (Baron-Cohen and Lombardo, 2017; Hyman et al., 2020). But these approaches are time consuming to process, open to subjective interpretation and not scalable. Artificial Intelligence (AI) can be used in healthcare to correct the accuracy and efficiency of diagnosis (Esteva et al., 2019; Rajpurkar et al., 2022). AI-powered approaches, from machine learning for brain segmentation (Xu et al., 2024), to NLP and multimodal analyses (Mohammadi et al., 2025; Rubio-Martín et al., 2024), show great promise for the identification of ASD biomarkers and workflow optimization (Moon and Woo, 2025; Sohl et al., 2025).

AI has been rapidly increasing in autism diagnosis (Atturu et al., 2025; Chaddad et al., 2021; Thabtah and Peebles, 2020), and there is a dire need. Much of the remaining literature has focused on technical success rather than the human aspects that may count for future adoption (He et al., 2019; Kelly et al., 2019). Moreover, existing studies have used a linear approach, which cannot fully capture the complex, non-linear and equifinal aspects of technology acceptance (Meskó and Topol, 2023; Tunalıgil, 2025). It is therefore uncertain in the field how identified accuracy, time efficiency and ease of use influence one another when it comes to the acceptance of AI and which of these factors is most likely to inspiration this acceptance dependent on the professional role and previous interaction with AI.

To fill the gap, this study attempts to comparison these two types of cognitive assessment (traditional cognitive assessment and AI-based cognitive assessment) by applying a multi-method analysis. The study sees the structural, configurational and predictive dimensions of AI acceptance amid healthcare and education professionals through the lens of the extended Technology Acceptance Model (TAM) (Davis, 1989; Marangunić and Granić, 2015; Salloum et al., 2019; Venkatesh and Davis, 2000) and the Joined Theory of Acceptance and Use of Technology (UTAUT2) (Dwivedi et al., 2021; Dwivedi et al., 2023; Venkatesh et al., 2003; Venkatesh et al., 2012).

On methodological aspects, beyond analyses of divergence, we undertake into analytical triangulation. We combine this with Partial Least Squares Structural Equation Modeling (PLS-SEM) for the estimation of complex causal relationships (Hair et al., 2019; Sabol et al., 2023; Sarstedt et al., 2019; Sarstedt et al., 2021), fuzzy-set Qualitative Comparative Analysis (fsQCA) for the recognition of configurational pathways (Fiss, 2011; Greckhamer et al., 2018; Ragin, 2009; Schneider and Wagemann, 2012), and advanced predictive analytics for out-of-sample generalizability. This approach is widely used, to guarantee reliability and validity, the common method bias assessment (Fuller et al., 2016; Kock, 2015; Podsakoff et al., 2003) is widely used. In theory, it's an extension of technology acceptance theories around neurodevelopmental diagnostics. Disturbing practical applications, it provides evidence-based recommendations on how to enhance workflow design, training, and develop guidelines of ethical use of AI (Labrique et al., 2018; Rivera et al., 2020; WHO, 2021).

The rest of this paper is orderly in the following way. The literature review and setting of the theoretical foundation is done in Section 2. The methodological framework is advanced and is appointed in Section 3. The results of Section 4 are presented in detail. Theory and practice implications are discussed in Section 5 and final recommendations are given in Section 6. An integrated research road map is mounted in Figure 1.

2 Literature review and theoretical foundation

The traditional way to identify autism spectrum disorder (ASD) is time-consuming, resource intensive and subject to different interpretations and interrater variability (Baron-Cohen and Lombardo, 2017; Hyman et al., 2020; Lord et al., 2018; Lord et al., 2020). Traditionally, these channels necessitate hours of direct contact for administration, diagnosis and treatment of ASD, which can be an absolute hurdle in the face of the wide variations in how ASD is manifested. On the other hand, Artificial Intelligence (AI) promises unparalleled possibilities for improving diagnostic accuracy, efficiency, and scalability (Esteva et al., 2019; Rajpurkar et al., 2022). New AI systems are developed based on state-of-the-art machine learning techniques for neuroanatomical brain segmentation (Xu et al., 2024) and multimodal analyses of behavioral and linguistic markers (Chaddad et al., 2021; Mohammadi et al., 2025; Rubio-Martín et al., 2024). Recent years have seen realistic evidence suggesting that such AI-assisted tools are likely to soon become a thorough part of the primary care and therapeutic experience (Atturu et al., 2025; Sohl et al., 2025).

Although AI tools and applications have developed fast across countless industrial lines and have attained technical feasibility in laboratory environments (Moon and Woo, 2025; Thabtah and Peebles, 2020), a key epistemological and empirical gap exists. The literature available has focused little on contemplating the human factors that dictate the actual uptake of technology and how technology can be applied into the learning context (He et al., 2019; Kelly et al., 2019). It's possible for the AI system to be highly technically accurate but, if identified as too clunky or too opaque, the odds are that clinicians won't be using that system. Moreover, the present research on comparison studies mostly focuses on simple linear statistical methods, which can't reflect the complicated, non-linear, and equifinal nature of technology acceptance (Meskó and Topol, 2023; Tunalıgil, 2025). Various factors can result in identical levels of AI acceptance phenomenon, which the conventional methods are powerless of detecting. A major danger in the field is the incident of complex algorithms with little or no reality in the clinical environment, which lack the consideration of end users' views (Rivera et al., 2020; WHO, 2021).

To fill this missing link, several connected theoretical approaches have been combined in this study, thus providing a solid base for our multi-method analysis. The Technology Acceptance Model (TAM) (Davis, 1989), a widely recognized and examined model in healthcare contexts (Marangunić and Granić, 2015; Salloum et al., 2019), is adapted to include single aspects of AI diagnostics, including algorithm biases and professional responsibility (Venkatesh and Davis, 2000). To measure performance expectancy and effort expectancy, the Unified Theory of Acceptance and Use of Technology (UTAUT2) is included; additionally, critical contextual moderators, known to affect technology acceptance differently for HCPs and the education community, are also measured – such as prior exposure to AI and professional role (Dwivedi et al., 2021; Dwivedi et al., 2023). In addition, the application of Artificial Intelligence to decline extraneous cognitive load but preserve germane cognitive load during complex diagnostic assessments, is significant to the theory of cognitive load theory. Finally, the comparison of traditional and AI based assessment is showed based on the principles of the psychometric theory of assessment that is, validity and reliability which ensures the consistency of perceived accuracy to the underlying theories of assessment. In addition, we use the method of fuzzy-set Qualitative Comparative Analysis (fsQCA), which takes the complex and non-linear nature of technology adoption into account. It is important to emphasize that fsQCA is not a qualitative method, it is a quantitative method of set theory that is aimed at linking configurational pathways and equifinality. To improve the theoretical explanation of our analysis, we have also remarkably strengthened the theoretical underpinnings of these configurational pathways by connecting them to our core theoretical frameworks such as the TAM (accuracy and ease of use are connected to the performance and effort expectancies), UTAUT2 (where prior AI exposure represents a facilitating condition) and the Cognitive Load Theory (where the integration of AI lowers the amount of extraneous load, thereby maintaining the germane load). This comprehensive theoretical integration guarantees both variance-based analysis as well as configurational analysis a stark theoretical foundation. This speculatory synthesis makes sure that the conceptual model encompasses the technological possibilities, as well as the psychometric rigor needed for high-stakes diagnostic situations.

This work extends these theories and creates a united conceptual model (Figure 2) for analyzing perceptions and factors of AI acceptance in the context of cognitive assessment of ASD. The model includes both direct and mediating processes, as well as directing processes which can be used to capture the more nuanced and non-linear processes of technology adoption.

Such comparative perceptions are searched in the technological prowess of AI in handling involved data and streamlining clinical processes. AI systems could identify markers that are not linear and are difficult for human observers to pick up (Mohammadi et al., 2025; Xu et al., 2024), solving the classic problem of inter-rater variability (Hyman et al., 2020; Lord et al., 2020). AI brain segmentation and multimodal analyses are adept level accurate (Moon and Woo, 2025; Rubio-Martín et al., 2024; Thabtah and Peebles, 2020) and reduce cognitive load (Atturu et al., 2025; Sohl et al., 2025). Hence, we hypothesize that AI-based methods are much more recognized accurate (H1). Moreover, traditional assessment requires to take a lot of time (Hyman et al., 2020; Lord et al., 2020), while AI can automate tasks in an unattainable speed by humans (Atturu et al., 2025; Chaddad et al., 2021; Esteva et al., 2019; Sohl et al., 2025), which indicates that AI-based approaches are much more time-efficient (H2). Furthermore, AI can also facilitate the saving of the burden of procedures, according to the principles of TAM (Davis, 1989; Marangunić and Granić, 2015; Salloum et al., 2019; Venkatesh and Davis, 2000) and requirements for care integration in primary care (Kelly et al., 2019; Sohl et al., 2025). Hence, we hypothesize that AI based methods have a much lower complexity of use (H3).

That is, it is imagined that these comparative perceptions will have a direct influence on the dependent variable, AI Acceptance, which is measured by behavioral intention. The basis for selection is mostly around accuracy (Kelly et al., 2019) for high-stakes diagnostic applications, supported by theories of TAM (Davis, 1989; Dwivedi et al., 2023; Meskó and Topol, 2023; Rivera et al., 2020; Venkatesh et al., 2003; World Health Organization, 2021). Consequently, identified accuracy has a positive impact on the acceptance of AI (H4). Meanwhile, ease of use will lower barriers to AI uptake (Dwivedi et al., 2021; Dwivedi et al., 2023; Marangunić and Granić, 2015; Venkatesh et al., 2003; Venkatesh et al., 2012) and we hypothesize that ease of use will have a positive impact on AI acceptance (H5). Additionally, time efficiency has a positive impact on clinical productivity (Atturu et al., 2025; Labrique et al., 2018; Rivera et al., 2020; Sohl et al., 2025; Venkatesh et al., 2012), suggesting that time efficiency can have a positive effect on the taking of AI (H6).

The core structural relationships are conditional on important variables to account for difficulty and non-linear aspects of technology adoption. As mentioned before, the more the contact with AI, the lower the uncertainty level, which combinations the relationship between accuracy and acceptance (Dwivedi et al., 2023; Meskó and Topol, 2023; Tunalıgil, 2025; Venkatesh et al., 2003; Venkatesh et al., 2012). Thus, the relationship between accuracy and acceptance (H7) is moderated by the previous exposure to AI. The indirect effect of accuracy is on ease of use, accurate systems will help users spend time and effort to principal the systems (Davis, 1989; Fiss, 2011; Greckhamer et al., 2018; Salloum et al., 2019; Venkatesh and Davis, 2000). So, use is a mediating variable between the accuracy-acceptance relationship (H8). Finally, healthcare bringers and educators accentuate accuracy and usability, respectively (Dwivedi et al., 2023; Labrique et al., 2018; Marangunić and Granić, 2015; Salloum et al., 2019; Venkatesh et al., 2012). Hence, to model these structural relations, the professional role is measured in the present study.

Predictive power of the framework is assessed, which is in line with the current perspective of PLS-SEM which emphasizes prediction-oriented modelling and out-of-sample generalizability (Hair et al., 2019; Sabol et al., 2023; Sarstedt et al., 2019; Sarstedt et al., 2021). Hence, the model also is projecting of AI acceptance (H10).

3 Advanced methodology

3.1 Research design & participants

In this study, however, the quantitative-dominant mixed-methods approach with analytical triangulation – namely consuming the Partial Least Squares Structural Equation Modeling (PLS-SEM), fuzzy-set Qualitative Comparative Analysis (fsQCA) and predictive analytics – will be applied in order to address gaps in the epistemological domain of the acceptance of artificial intelligence for mental assessment of ASD. This study is cross-sectional with counterbalancing, which decreases the possible response bias (Podsakoff et al., 2003) and allows for a fuller examination by several epistemological angles (Fiss, 2011; Sarstedt et al., 2021).

All health care and education professionals in Jordan working with the assessment of ASD are the target population. Jordan is a strategic middle salary setting for using advanced technologies in resource-limited systems (Labrique et al., 2018; WHO, 2021). Inclusion criteria included working 2 years or more with ASD, having prior experience using traditional assessment methods, documenting uses of AI tools in experimental or educational workflows (not just passive exposure), appropriate license, and voluntary consent. Administrative-only jobs and expired licenses, pilot involvement and language issues were all reasons for exclusion.

Stratified convenience sampling was worked to ensure representation by profession and across settings. Sample size was validated through several criteria (Table 1). The PLS-SEM “10-times rule” mandatory 65 participants (Hair et al., 2019; Kock, 2015). Cohen recommended the numbers of subjects to be 34 for t-tests and 76 for multiple regression analyses conducted using GPower. It was based on Monte Carlo analysis and was the same as 71. For fsQCA, the number of necessary cases was 50 (Schneider and Wagemann, 2012) and for multi-group analysis (PLS-MGA), the number of cases was 20 per group. After deleting 3 outliers during data screening, the final number of respondents for analysis (n* = 77) included 44 healthcare professionals (57.1%) and 33 education professionals (42.9%). This sample done all the mentioned requirements, and the smallest of the sub-samples (N = 7) was acceded for the analysis using the method of robust bootstrapping (Sarstedt et al., 2019).

Table 1. Sample size justification across multiple statistical approaches.

Sample adequacy was evaluated using multiple complementary criteria, including heuristic rules, statistical power analysis, Monte Carlo simulation, and methodological recommendations for configurational analysis. The final sample size (n = 80) exceeded the minimum requirements for all planned analyses, supporting reliable estimation of the PLS-SEM model, fsQCA configurations, predictive analytics (PLSpredict), mediation and moderation analyses, and robustness testing. *The smallest subgroup in the multi-group analysis contained only 7 observations, which is below the conventional guideline of 20 observations per group. Therefore, the subgroup analysis was treated as exploratory and interpreted cautiously, with robust bootstrapping applied.

3.2 Instrument development & psychometrics

Development of the instrument was done completely rigorous 6 phases. Phase 1 generated 60 items from established scales (Davis, 1989; Kelly et al., 2019; Meskó and Topol, 2023; Venkatesh et al., 2003, 2012; WHO, 2021), adapted to ASD (Hyman et al., 2020; Lord et al., 2020). Content validity was assessed in Phase 2 by 15 experts. Next, Phase 3 included cognitive interviewing, Phase 4 included pilot testing (n = 20), Phase 5 included the application of Item Response Theory (IRT), and Phase 6 included the final validation of the 33 items, 9-construct questionnaire (Table 2).

Table 2. Comprehensive instrument structure with advanced psychometrics.

TA, traditional assessment accuracy; TE, traditional assessment time efficiency; TU, traditional assessment ease of use; AA, AI-based assessment accuracy; AE, AI-based assessment time efficiency; AU, AI-based assessment ease of use; AC, AI acceptance; PE, prior AI exposure; EC, ethical concerns; α, Cronbach's Alpha; CR, composite reliability; AVE, average variance extracted; VIF, full collinearity variance factor; IRT, item response theory; b = Item Difficulty Parameter; a = Item Discrimination Parameter. Bold values indicate the principal observed estimates/statistics highlighted for emphasis in the corresponding table. Statistical significance is determined by the reported p-values and decision/interpretation columns and is not implied by bold formatting alone.

All the constructs had very good psychometric things. Cronbach's α (0.81–0.93) and Composite Reliability (0.81–0.93) exceeded 0.70 (Hair et al., 2019). This showed that Average Variance Extracted (AVE) values for the instruments ranged between 0.59 and 0.79 which exceeds the cut-off score of 0.50 for acceptable convergent validity. VIF values of the results were all 1.00–1.22 suggesting that there was no common method bias or multicollinearity (Kock, 2015). IRT parameters confirmed items regarding difficulty and discrimination (Embretson and Reise, 2025).

Construct authority was done through EFA, CFA, and IRT. The loading values of the indicators were greater than 0.708 (Hair et al., 2019). The Fornell-Larcker criterion, the HTMT ratios <0.85 and the cross-loadings (Sarstedt et al., 2019) were requested for the appraisal of discriminant validity. This was accompanied by Dijkstra-Henseler's rho A (Hayes 2013), test-retest ICCs, and McDonald's omega. The complete indicator loadings and cross-loadings matrix is presented in Table 5.

3.3 CMB, data collection & ethics

Common Method Bias (CMB) was manipulated for both procedurally (Podsakoff et al., 2003, Lindell & Whitney, 2001) and statistically (Table 3). Based on Harman's single factor test (PCA) the first factor constituted about 28.4% (<50%) (Podsakoff et al., 2003). Neither the marker variable technique nor the relationships were very strong. Full collinearity VIF (2.87 < 3.30) (Kock, 2015) and the latent method factor test (ΔCFI = 0.008 < 0.010) confirmed no CMB (Fuller et al., 2016).

Table 3. Common method bias (CMB) assessment.

CMB, common method bias; PCA, principal component analysis; VIF, variance inflation factor; ΔCFI, change in comparative fit index. Four complementary diagnostic procedures were employed to assess potential common method bias. All statistical indicators satisfied the recommended thresholds, indicating that common method bias is unlikely to threaten the validity of the measurement model or the subsequent structural analyses.

Data collection (Jan-Apr 2025) was done on a secure Qualtrics platform with IP verification and attention checks. Formal Ethics Statement: Primary data was poised from human respondents (Healthcare professionals and education professionals) to survey their perception of AI tools. The study was allowed by the Institutional Review Board (IRB) of Al-Balqa Applied University under the guidelines of the Declaration of Helsinki and WHO AI ethics guidelines (Rivera et al., 2020; WHO, 2021). All individual subjects included in the study had given their learnt consent. All participants had their anonymity, encrypted storage, and confidentiality respected. (Note: Any previous metadata indicating ‘No human studies are presented’ has been amended correctly to show this primary human subject's study.

3.4 Advanced statistical analysis framework

The multi-layered analytical framework (Figure 3) is a combination of variance based, configurational, predictive and robustness analysis.

Primary Analysis: PLS-SEM. Analysis was oversaw using SmartPLS 4.0 (Hair et al., 2019; Sabol et al., 2023; Sarstedt et al., 2019; Sarstedt et al., 2021) as it is recommended for prediction-oriented studies and complex models. For bootstrapping (5,000 subsamples) the BCa method was used. The PLS in this study has been consistently altered by attenuation bias and named as PLS (PLSc). The evaluation of the measurement model and the structural model was done by following guidelines (Hair et al., 2019; Sarstedt et al., 2021). Bootstrapped secondary effects were conducted for mediation and permutation-based approaches were employed for moderation and multi-group analysis (PLS-MGA).

Secondary Analysis: fsQCA. fsQCA can discover configurational pathways, and recognize causal complexity and equifinality (Fiss, 2011; Ragin, 2009; Schneider and Wagemann, 2012). The direct method was used in calibration. To realize theoretical coherence, necessity and sufficiency analyses were performed and the intermediate solution, as described by Fiss (2011), Ragin (2009) and Schneider and Wagemann (2012) were reported.

Predictive Analytics. The out-of-sample prediction performance of PLSpredict is associated with a linear regression prediction benchmark using k-fold cross validation (k = 10) (Hair et al., 2019). To validate the machine learning models (Random Forest, GBM, SVR) they are used SHAP values for interpretability in Python. Importance Performance Map Analysis (IPMA) was used to acquire priorities for intervention.

Robustness Analyses. Weighted stability was gotten using Ten Berge's factor score regression (Ten Berge et al., 1999), Bayesian SEM with MCMC estimation, permutation tests and leave-one-out cross validation. Results were regular across subsample analysis, across alternative specifications, and CMB control models (Podsakoff et al., 2003).

Multi-Group Analysis and Exploratory Machine Learning Benchmarks: The total sample size (n = 77) and the conforming smallest sub-group (n = 7) are relatively small, but PLS-MGA was used because it does not necessitate that the underlying variables are normally distributed and is suitable for comparing small sub-groups. With small number of samples though, the statistical power of any complex Machine Learning (ML) models is also constrained. For these reasons, the applications of the ML models (Random Forest, GBM, SVR) and comparison of ML benchmarks (Figure 6b) are not considered as definitive predictive generalizations, but only exploration analyses to discover the stability of algorithms. As a result, the benchmarks for the ML are shown in the “Exploratory Limitations” context to moderate the conclusions about out-of-sample ML performance.

4 Results

4.1 Preliminary analyses

4.1.1 Data screening and assumption testing

Data screening verified that data met the statistical assumptions of multivariate analysis and PLS-SEM prior to primary analysis. The results are shown in the form of a table (Table 4). Little's MCAR test was not significant (χ² = 42.31, p = 0.287) which indicates that disappeared data was completely at random as per Little. Three observations were excluded using Mahalanobis distance and the final sample for analysis was n = 77. To correct any previous differences, it is directed out that this final sample size (n = 77) was always used as an exact dataset for all further statistical tests, regardless of any earlier mentioned sample size (e.g., n = 80) in the methodological justifications. Multivariate normality was found to be okay (Mardia's skewness = 1.87, kurtosis = 3.42, p > .05). Scatterplots were used to show the linearity. Multicollinearity was absent (VIF = 1.06–2.87 < 5.00; Hair et al., 2019). Data was found to be right as the tests of homoscedasticity (Breusch–Pagan χ² = 8.45, p = 0.392) and independence of errors (Durbin–Watson = 1.98) were met.

Table 4. Comprehensive data screening and assumption testing.

MCAR, missing completely at random; VIF, variance inflation factor. Data screening was conducted prior to model estimation to verify compliance with the assumptions required for multivariate analysis and PLS-SEM. Missing data were confirmed to be completely at random, a small number of outliers were removed, and all remaining assumptions—including normality, linearity, homoscedasticity, multicollinearity, and independence of errors—were satisfactorily met. Consequently, the final analytical sample (n = 77) was considered appropriate for the subsequent PLS-SEM, fsQCA, predictive analytics, and robustness analyses.

Table 5. Indicator loadings and cross-loadings matrix.

Bold values represent the indicator loading on its assigned construct. All outer loadings exceeded the recommended threshold (0.708), confirming satisfactory indicator reliability. Furthermore, each indicator loaded more strongly on its intended construct than on any other construct, providing strong evidence of discriminant validity based on the cross-loadings criterion.

4.2 Measurement model assessment (PLS-SEM)

4.2.1 Indicator reliability

All indicator loadings were higher than 0.708 (range: 0.803 to 0.915), meaning that they are sufficiently high and reliable as recommended by PLS-SEM (Hair et al., 2019; Sarstedt et al., 2021). The cross loadings results revealed that all indicators were more correlated with the construct they were supposed to measure than others, which was a good proof of discriminability of the constructs.

All 33 indicators exceeded the 0.708 threshold. Maximum cross-loadings were substantially lower than primary loadings across all nine constructs, confirming discriminant validity.

4.2.2 Internal consistency and convergent validity

The internal consistency of Table 6 is very good. Cronbach's alpha (0.831–0.936) and Dijkstra–Henseler's rho A (0.839–0.942) beat 0.70. Composite reliability (0.887–0.949) surpassed 0.70. The convergent validity was confirmed on the AVE values of more than 0.50 which was 0.725–0.792.

Table 6. Measurement model assessment: reliability, convergent validity, and discriminant validity.

α, Cronbach's Alpha; ρ_A, Dijkstra–Henseler's rho; CR, composite reliability; AVE, average variance extracted; √AVE, square root of the average variance extracted (Fornell–Larcker Criterion); HTMT, heterotrait–monotrait ratio; VIF, variance inflation factor. All constructs satisfied the recommended thresholds for internal consistency reliability, convergent validity, discriminant validity, and collinearity diagnostics. The square root of the AVE for each construct exceeded its correlations with all other constructs (see Table 7), confirming discriminant validity. Bold values indicate the principal observed estimates/statistics highlighted for emphasis in the corresponding table. Statistical significance is determined by the reported p-values and decision/interpretation columns and is not implied by bold formatting alone.

4.2.3 Discriminant validity

Discriminant validity was encouraged by the Fornell-Larcker criterion (Table 7) which showed that √AVE values were superior than all the inter-construct correlation values. The HTMT ratio remained below 0.85 (max = 0.741). The conclusions of full collinearity VIF values (1.13–1.42 < 3.30) verified that there was no common method bias (Kock, 2015).

Table 7. Discriminant validity assessment using the Fornell–Larcker criterion and HTMT ratios.

Bold diagonal values represent the square root of the Average Variance Extracted (√AVE) according to the Fornell–Larcker criterion, whereas the off-diagonal values represent inter-construct correlations. The maximum HTMT ratio for each construct remained below the recommended threshold of 0.85, confirming satisfactory discriminant validity across all latent constructs.

4.3 Structural model assessment

4.3.1 Collinearity and path coefficients

The results of the structural model are shown in Table 8 and Figure 4. VIF values (1.06–1.42 < 3.30) indicated no multicollinearity (Hair et al., 2019). Perceived AI-Based Assessment Accuracy (AA) directly influenced AI Acceptance (AC) the most (β = 0.412, p < .001, f² = 0.284). Perceived Ease of Use (AU; β = 0.324, p = .001) and Time Efficiency (AE; β = 0.281, p = .001) also positively influenced AC, aligning with UTAUT2 (Venkatesh et al., 2012). Prior AI Exposure (PE; β = 0.243, p = .001) positively affected AC, while Ethical Concerns (EC; β = −0.184, p = .003) negatively impacted it.

Table 8. Structural model assessment: path coefficients, significance, effect sizes, and model Fit indices.

β = Standardized path coefficient; SE = Standard Error; CI = Bias-corrected 95% bootstrap confidence interval; f² = Cohen's effect size (0.02 = small, 0.15 = medium, 0.35 = large); Q²_predict = Predictive relevance obtained using the blindfolding procedure; SRMR = Standardized Root Mean Square Residual; NFI = Normed Fit Index; R² = Coefficient of Determination. Bold values indicate the principal observed estimates/statistics highlighted for emphasis in the corresponding table. Statistical significance is determined by the reported p-values and decision/interpretation columns and is not implied by bold formatting alone.

4.3.2 Model fit and predictive power

Table 9 shows exceptional model fit (SRMR = 0.068 < 0.080; NFI = 0.82 ≥ 0.80). The model explained 67% of the variance in AC (R² = 0.670). Stone–Geisser's Q² (0.480 > 0.250) and PLSpredict's Q²predict (0.410) proved large predictive relevance. RMSE (0.384) and MAE (0.296) were less than the standards of linear regression and it indicates that the model has a better predictive ability.

Table 9. Model Fit indices, predictive power, and effect size assessment.

SRMR, standardized root mean square residual; NFI, normed fit index; R², coefficient of determination; Q², Stone–Geisser predictive relevance; Q²_predict, out-of-sample predictive relevance (PLSpredict); RMSE, root mean squared error; MAE, mean absolute error; f², Cohen's effect size; VIF, variance inflation factor. Bold values indicate the principal observed estimates/statistics highlighted for emphasis in the corresponding table. Statistical significance is determined by the reported p-values and decision/interpretation columns and is not implied by bold formatting alone.

4.4 Hypotheses testing summary

A summary of the results of the hypothesis testing can be found in Table 10. There was evidence for all 10 hypotheses. AI-based methods were also highly significantly more accurate (H1: d = 0.71), more timesaving (H2: d = 1.61), and more user-friendly (H3: d = 1.07) than traditional methods. Direct effects (H4–H6), moderation by prior exposure (H7: β = 0.214, p = .022), partial mediation by ease of use (H8: VAF = 55.3%), multi-group differences (H9: Δβ = 0.128, p = .018), and predictive relevance (H10: Q²predict = 0.410) were all statistically supported.

Table 10. Comprehensive hypotheses testing results.

SE, standard error; CI, bias-corrected bootstrap confidence interval; PLS-MGA, partial least squares multi-group analysis; VAF, variance accounted for; Q²_predict, out-of-sample predictive relevance. Effect sizes were interpreted according to Cohen's guidelines (0.02 = small, 0.15 = medium, 0.35 = large). All direct, indirect, moderation, and predictive hypotheses were statistically supported, indicating strong empirical support for the proposed conceptual framework.

4.5 Mediation analysis

There was evidence of partial complementary mediation as the results of bias-corrected bootstrapping (5,000 resamples) showed that AI-Based Ease of Use (AU) partially mediated the AA → AC relationship (Indirect β = 0.241, p = .003, 95% CI [0.091, 0.414], VAF = 55.3%). The indirect pathway via Time Efficiency (AE) also played a role (β = 0.183, p = .010, VAF = 42.1%). The total effect of AA on AC was highly significant (β = 0.760, p < .001) (Table 11).

Table 11. Advanced mediation analysis using bias-corrected bootstrapping.

AA, AI-based assessment accuracy; AU, AI-based assessment ease of use; AE, AI-based assessment time efficiency; AC, AI acceptance; VAF, variance accounted for. Indirect effects were estimated using 5,000 bias-corrected bootstrap resamples. Mediation was considered significant when the 95% bootstrap confidence interval did not include zero. The observed VAF of 55.3% indicates partial complementary mediation, suggesting that AI-based assessment accuracy influences AI acceptance both directly and indirectly through perceived ease of use. The indirect pathway through time efficiency was also statistically significant but exhibited a comparatively smaller mediating contribution.

4.6 Moderation analysis

The AA → AC relationship is greatly persuaded by the prior AI exposure (Interaction β = 0.214, p = .022), as demonstrated in Figure 5a. This relationship was largest for those with a high level of prior exposure (β = 0.626, p < .001) and lowest for those with low level of prior exposure (β = 0.198, p = .075) (simple slopes analysis). This directs that experience contributes to the means of appreciating the accuracy of AI (Figure 5b).

4.7 Multi-group analysis (MGA)

Healthcare professionals and education professionals can obtain information on MGA in Table 12. Leveling was done up to configural and metric level and partially up to scalar level by overseeing an invariance test. The path AA → AC was significantly different between groups (Δβ = 0.132, p = .042) and the path in the healthcare group was more significant (β = 0.481) than the path in the education group (β = 0.349). No significant difference in AU → AC & AE → AC paths.

Table 12. Measurement invariance assessment and multi-group analysis (healthcare vs. Education Professionals).

χ²_diff, Chi-square difference test; ΔCFI, change in comparative fit index; ΔRMSEA, change in root mean square error of approximation; PLS-MGA, partial least squares multi-group analysis; β, standardized path coefficient. Bold values indicate the principal observed estimates/statistics highlighted for emphasis in the corresponding table. Statistical significance is determined by the reported p-values and decision/interpretation columns and is not implied by bold formatting alone.

It is vital to note, however, an important statistical limitation of the present analysis: The extremely small sizes of the subgroups (e.g., smallest n = 8 in some stratifications) are correlated with a low statistical power and high risk of overfitting. Thus, the AA → AC difference is statistically significant but should be interpreted in the light of this multi-group comparison. These results should be taken as exploratory and indicative and further studies with larger and balanced samples in each of the groups should be achieved to confirm the between-group difference.

4.8 Robustness checks

Robustness checks, using other estimation methods, are presented in Table 13. The AA → AC path was quite stable with the Covariance-Based SEM (β = 0.389), Bayesian SEM (posterior mean = 0.421), non-parametric bootstrapping, permutation analysis, subsample validation and common method bias controls. The mean β of all methods was 0.407 ± 0.016 (all p < .001), indicating very good stability.

Table 13. Comprehensive robustness checks across alternative estimation methods.

PLS-SEM, partial least squares structural equation modeling; CB-SEM, covariance-based structural equation modeling; ML, maximum likelihood estimation; MCMC, Markov Chain Monte Carlo; CrI, Bayesian credible interval; CMB, common method bias. Bold values indicate the principal observed estimates/statistics highlighted for emphasis in the corresponding table. Statistical significance is determined by the reported p-values and decision/interpretation columns and is not implied by bold formatting alone.

4.9 fsQCA analysis

The results from the fuzzy-set Qualitative Comparative Analysis (fsQCA) to identify pathways to high AI acceptance are presented in Table 14. Three conformations were a satisfactory consistency (>0.80). Solution 1 with Consistency of 0.892 had high approval with high accuracy, ease of use and low ethical concerns and high prior exposure. Solution 2 (0.846) revealed that, in cases where ethical issues are not prominently present, high accuracy and efficiency are more significant than ease of use. Solution 3 (0.817) indicated that, even though the accuracy is lower, factors such as high usability, efficiency, high exposure and a high level of ethical awareness of the result foster acceptance. The complete solution accounted for 61.1% of cases (Overall Consistency = 0.863), and showed causal complexity and equifinality (Fiss, 2011; Ragin, 2009).

Table 14. fsQCA configurations leading to high AI acceptance.

●, presence of the condition; ⊗, absence of the condition; AA, AI-based assessment accuracy; AU, AI-based assessment ease of use; AE, AI-based assessment time efficiency; PE, prior AI exposure; EC, ethical concerns. Raw Coverage indicates the proportion of cases explained by each configuration, whereas Unique Coverage reflects the proportion explained exclusively by that configuration. Following established fsQCA guidelines, solution consistency values above 0.80 indicate sufficient causal configurations, while solution coverage reflects the empirical relevance of the overall configurational model.

4.10 Predictive analytics

AI-Based Assessment Accuracy (AA) was mapped in the Importance–Performance Map Analysis (IPMA), resulting in Importance = 0.410 and Performance = 78.5 (Figure 6a), and hence placed in the “High Importance–High Performance” quadrant. This quadrant (Importance = 0.233, Performance = 72.1) saw a drop in the score of AU as well. Time Efficiency (AE) was in the “High Importance–Low Performance” quadrant (Importance = 0.273, Performance = 58.6), suggesting that this needs to be improved. Ethical Concerns (EC) changed Low Importance and Low Performance (LI/LP) while Prior Exposure (PE) was low in importance and high in performance. In Figure 6b, the out-of-sample performance of PLSpredict is associated with the machine learning benchmarks. PLS-SEM obtained RMSE = 0.384 and MAE = 0.296 which is better than linear regression and competitive with Random Forest and SVM. The Q²predict of 0.410 was definite to have a high out-of-sample predictive accuracy.

Table 15 presents the training, tuning and cross validation specifications for the algorithms used to ensure transparency of the machine learning benchmarks.

Table 15. Machine learning algorithm specifications and cross-validation details.

Large individual folds of a dataset are of additional importance, and we need to explicitly acknowledge the extreme risk of sample overfitting when accomplishing 10-fold cross-validation on a dataset of n = 77, with each fold containing ∼7 cases. A small fold size will drastically restrict how stable the ML training and validation splits are. The comparison of PLSpredict with these machine learning benchmarks (Figure 6b) is therefore described here as an exploratory exercise, with the point of showing trends in the algorithms and not a definitive evaluation of out-of-sample generalizability. The sturdy predictions made with this study are based primarily on the results of the PLS-SEM and PLSpredict; the results from the ML benchmarks are taken with caution, because there is a known risk of overfitting data when using ML with a small taster.

5 Discussion

The analysis framework used in this study is advanced and is based on multi methods to explore health care and education professionals who work with autistic spectrum disorder (ASD) patients' attitudes towards the acceptance of AI-based cognitive assessment tools. Answers showed high levels of empirical evidence that the AI based assessment technique is seen as notably more accurate to use, time efficient and user friendly when compared to the conventional diagnostic technique. Important to note, however, are that these findings are based on perceptions of these tools and not on their actual, objective system performance. The structural model accessible significant explanatory power revealed that professed accuracy was the highest significant determinant of AI acceptance (β = 0.412, p < .001) followed by ease of use (β = 0.324, p = .001) and time efficiency (β = 0.281, p = .001).

A multi-Method approach was used in the study of the getting of AI based cognitive assessment for people with autism spectrum disorder (ASD) by professionals. Specifically, results indicate that AI methods are seen as significantly superior when they arrive at accuracy, time efficiency and simplicity of use by professionals. Therefore, the structural model was highly explanatory (R² = 0.67) and the strongest predictors of the taking of AI were perceived accuracy (β = 0.412, p value <.001), ease of use (β = 0.324, p value = .001), and time efficiency (β = 0.281, p value = .001). Further analysis using mediation discovered that ease of use has a partial mediation effect on the accuracy-acceptance relationship, which explained 55.3% of the total effect. Besides mediation, the results of the moderation analysis signaled that the relationship between these variables is stronger when participants have prior exposure to AI. In the end, fsQCA produced three equifinal configurational pathways to high acceptance and the model pointed out to be highly predictable (Q²predict = 0.410).

The study figures on the Technology Acceptance Model (TAM) and Unified Theory of Acceptance and Use of Technology (UTAUT2) to the specialized setting of ASD diagnosis. While these frameworks have been applied in healthcare information systems thus far (Marangunić and Granić, 2015; Salloum et al., 2019), there is no prior example of adapting it for AI-based neurodevelopmental diagnostics. To exhibit this special context, we reinterpreted the concept of performance expectancy as perceived accuracy and perceived time efficiency, as both are diagnostic priorities (Davis, 1989; Venkatesh and Davis, 2000). Furthermore, prior coverage to AI and ethical issues tackle the autonomous nature of AI and unique ethical issues around bias and accountability (Venkatesh et al., 2003, 2012). Lastly, this spreads the nomological network by illustrating how central relationships in the context of high-stakes can be regulated by domain-specific factors (Dwivedi et al., 2021; Dwivedi et al., 2023).

The study is based on two theories: cognitive assessment theory and technology acceptance research. Psychometric concepts and adoption frameworks are examined and used as a foundation for accepting the interaction of the traditional concepts of measurement and perceptions of technology. Thus, apparent accuracy emerged as the most significant predictor, suggesting that professionals view AI tools as instruments for measurement, which must meet strict psychometric criteria. In order to describe this mechanism in detail, the Cognitive Load Theory (CLT) demonstrated to be pertinent as it provided insights on how AI can impact the cognitive load of clinicians to make the system easier to use, a key mechanism to translate accuracy to acceptance and optimize germane load when accomplishment complex tasks.

This methodology is a large contribution with the use of fsQCA. Varance-based methods capture average effects, on the other hand, conversely, fsQCA can find causal complexity and equifinality (Fiss, 2011; Ragin, 2009). For instance, when three different configurations are found, it demonstrates that a similar level of acceptance is achieved from diverse paths – and that challenges assumptions of uniformity for assumption. From a configurational point of view, this shows that depending on the context, the results for accuracy and ease of use appear to be complementary and that the results of exceptional time efficiency compensate for modest ease of use when ethical concerns are minimal (Greckhamer et al., 2018; Schneider and Wagemann, 2012). This therefore helps to advance theoretical understanding in that it reveals that the relationship between the determinants is more complex than a linear model.

While the findings provide possible guidelines for administrators to deliberate when implementing AI tools, the results need to be interpreted with care because this study was conducted at a single point in time across several different contexts and involved professed rather than actual system performance, and the single country (Jordan) context in which the study was conducted. As depicted in Figure 7, such an integrated hybrid diagnostic model might be a blend of AI analytics and traditional psychometrics and clinical domain knowledge in a constant learning model.

Therefore, it would be beneficial for organizations to look at tools that are seen as being high accuracy and easy to use. Conversely, for institutions with limited resources, a perception of time-savings may outweigh the level of ease of use if there are systems in place to control issues of transparency and ethics in the governance process (Rivera et al., 2020; WHO, 2021). In addition, paths suggesting strategies identified in fsQCA reveal that implementation strategies may have to be modified to fit professional groups (Labrique et al., 2018).

However, having confirmed the apparent time efficiency of AI in this setting (Cohen's d = 1.61), there is a lot in the way of workflow enhancement and a diminution in wait time. In the medical field, AI could help automate the initial screening phase for systems facing high volumes, thus freeing up time for clinical teams to delve deeper into more important formulations (Atturu et al., 2025; Sohl et al., 2025). However, for thriving implementation in practice, AI functionality should be designed around specific targets and there remains a need for clinician–patient interaction (He et al., 2019; Kelly et al., 2019). Phased integration assistances ease into this by enabling professionals to become familiar with the system step by step.

Based on these findings, there is a need for extensive training because of the controlling impact of prior AI experiences. Therefore, individuals with less experience tend to have lower accuracy-acceptance relationships and focused training could keep levels of reference for the decision on AI (Meskó and Topol, 2023; Tunalıgil, 2025). Besides, conceptual understanding, output interpretation and bias recognition should be included in the programs. Lastly, competency development should emphasize the use of AI as a decision support tool, complementing the professional judgment, and the continuing pieces of education should involve hands-on experience under supervision (Esteva et al., 2019; Rajpurkar et al., 2022).

Providing strong governance, covering issues of bias, privacy, consent and accountability, may be necessary in the face of these negative effects of identified ethical concerns (β = −0.184, p = .003) (Rivera et al., 2020; WHO, 2021). However, regular algorithmic bias audits should be part of the governance process, especially because of the contentious issue of the heterogeneous presentations of ASD (Hyman et al., 2020; Lord et al., 2020). The organizations should think about executing the use of explainable AI (XAI), which can provide reasons for recommendations in a coherent manner. Finally, there must be clear guidelines regarding professional responsibility and liability of AI systems and clinicians with respect to the accountability of AI systems.

Policymakers should strike a balance between the speediness of implications of AI deployment and its safety, effectiveness and equity. These recommendations, which are largely based on the experience of one country (Jordan) and the acuities of professionals with no objective clinical outcomes, indicate that regulation frameworks should pay attention to guidelines for the psychometric validation, clinical utility and post marketing surveillance (Conway et al. 2021; WHO 2021). Concurrently, health policy should consider prioritizing equity in access particularly in population groups that are typically lessened, including people within populations lacking traditional health care services (Labrique et al., 2018). Moreover, reimbursement policies might eventually adjust accommodate assessments enhanced by AI. Last, policies should foster workforce development for the attainment of AI literacy to overcome some digital openings which widen the ASD diagnostic gaps.

There are some limitations with these influences though. First, one cannot make inferences about temporal dynamics from cross sectional designs, so longitudinal designs need to be employed. Second, sampling took place in only one national context (Jordan) and it is not viable to transfer the policy and workflow recommendations to other national contexts with other cultural orientations and healthcare systems. Third, the professed attributes of the AI systems were used, and the actual usage and objective diagnostic accuracy measures were not determined; more importantly, actual usage data and objective diagnostic accuracy measures should be encompassed in future research. Fourth, fsQCA pinpointed the configurational complexity without superseding variance-based methods.

Future directions of research are proposed. Firstly, longitudinal studies to survey up initial perceptions and relate to the changes over time is important. Secondly, this would help to inform culturally meaningful implementation strategies based on comparative analyses of various health care systems. Third, studies of the effect of AI on diagnostic accuracy and clinical outcomes should be conducted to term its clinical utility (Chaddad et al., 2021; Moon and Woo, 2025). Fourthly, experiments where the design of the AI interface was replaced would ultimately render it more user friendly. Fifth, the study of interprofessional dynamics of AI adoption should be discovered. Lastly, ethical, legal and social implications (ELSI) would assist in responsible governance and consider innovation and protection of vulnerable populations (Meskó and Topol, 2023; WHO, 2021).

6 Conclusions

This study adds to the understanding of perceived AI acceptance in ASD cognitive assessment using a difficult multi-method approach using PLS-SEM, fsQCA and predictive analytics. It contributes to three keyways, namely: It is the first to relate TAM/UTAUT2 in the field of neurodevelopmental diagnostics, illustrating how domain-specific factors interact with known factors. Secondly, it ties the gap between theory and practice of cognitive assessment by technology acceptance concept, uncovering how health professionals perceive and use the AI tools, not only as a technological tool, but also as a psychometrically valid tool. Third, it shows configurational involvedness, providing evidence that there are multiple pathways to high acceptance with fsQCA. Results show that there is strong support for AI methods in terms of professed superiority both in perceived accuracy, efficiency, and usability, with perceived accuracy being the most important factor. Furthermore, the model showed a high explanatory power (R2 = 0.67), high predictive relevance (Q2predict = 0.410) and superb robustness of the explanations of these perceptions.

For Healthcare Administrators: First and foremost, make sure to focus on AI tools that are perceived to have a strong diagnostic capability and user-friendly interface. Also, offer strong guidance to facilitate phased integration. Finally, establish ethical governance in terms of bias, privacy and accountability. Researchers: Understand the pros and cons of using AI in research; know how to estimate AI-generated ideas and ideas generated by others using AI. To Policymakers: Develop regulatory frameworks to saddle innovation, keep people safe and prioritize equitable access and workforce development to cultivate AI competence. For AI Developers: Ensure that AI systems are not only accurate but also easy to use and consider incorporating explainable AI components and engage with end-users throughout the development process to facilitate seamless integration.

In conclusion, the latent for using AI in ASD cognitive evaluation is transformative, presenting opportunities to potentially improve diagnostic accuracy, efficiency, and accessibility based on professional perceptions. As demonstrated in this research, there are three aspects to the positive implementation of this technology, namely: Accuracy perceptions, Usability, Efficiency and ethical safeguards. This research helps to define these dynamics and provides evidence for guiding implementation. In the future, it will be essential that AI enhances, not replaces, human abilities, improving speed, accuracy, and compassion of persistent care. The journey to this vision is proposed to be the hybrid diagnostic model, presented in this study as a conceptual model to inform the way forward in integration, not an achieved clinical outcome. Although empirical validation of the actual clinical effectiveness and objective diagnostic accuracy of these AI tools is still needed in the clinical context, this conceptual model – with both AI analytics and traditional psychometrics and clinical judgment – provides a structured approach for future practice. With thorough consideration and application, AI cognitive assessment tools can fulfill these promises and significantly enhance the lives of those with ASD.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Cristina Dumitru, Polytechnic University of Bucharest, Romania

Reviewed by: Muryanti Muryanti, Universitas Islam Negeri Sunan Kalijaga Fakultas Ilmu Sosial dan Humaniora, Indonesia

Verónica Freire, National University of Chimborazo, Ecuador