The study, published in the Journal of Religious Education, reports survey data from 54 in-service religious education teachers in Greece who completed a national professional development program known as SUB8. Run by the Institute of Educational Policy under the country’s recovery plan and co-funded by the European Union, the program trained a slice of the country’s roughly 2,700 theology-graduate RE teachers in the use of AI and digital tools, with a dedicated session on large language models such as ChatGPT. The 54 respondents represent about 19 percent of the 284 RE teachers who took the module nationally, and they answered a questionnaire covering AI literacy, digital self-efficacy, technological pedagogical content knowledge, classroom use of AI, pedagogical change, theological concerns, and professional collaboration.

The headline numbers describe a cautiously positive cohort. Self-reported AI literacy averaged 3.91 on a five-point scale, digital self-efficacy 3.82, and TPACK — the integrated measure of technological, pedagogical, and content knowledge — 4.09, the latter anchored by strong scores for selecting tools on the basis of lesson content and for designing AI-assisted micro-teaching units. Reported classroom use of AI averaged 3.84 but showed the widest spread of any section, with differentiation and material production at the top and AI-assisted assessment at the bottom, a pattern the author reads as a cohort spanning eager early adopters and more hesitant practitioners. Notably, the literacy items were bimodal: a confident majority clustered at the high end while a distinct minority sat two points down the scale, even inside a self-selected, freshly trained group.

The most distinctive section, however, was the one no standard AI-in-education instrument contains: theological concerns. The strongest consensus in the entire questionnaire outside the collaboration items went to the statement that AI requires critical pedagogical management, endorsed at 4 or 5 by 46 of the 54 teachers — 85 percent — for a mean of 4.35. Agreement was also high that AI fails to capture the depth of religious experience, while the claim that AI necessarily produces shallow learning fell just below the scale midpoint, the only item in the instrument to do so. In other words, these teachers worry about AI’s epistemological reach without conceding that it automatically cheapens what students learn.

The open-ended answers sharpen that picture into three recurring themes. The first is practical: teachers repeatedly credited AI with cutting preparation time, improving lesson quality, and enabling genuine differentiation for students of varying needs and backgrounds. The second is a mix of the epistemic and the mundane — concerns about accuracy, reliability, and the “lack of understanding of depth” in AI outputs, alongside absent interactive whiteboards, unreliable school internet, and tools that exist almost exclusively in English. The third theme, and the most consistent across responses, is a firm professional boundary: AI must be a supplementary tool, subordinate to the teacher’s judgment and to the relational, experiential, and theological dimensions of the subject. One respondent put it plainly: AI “does not replace the role of the teacher, nor the word of God.”

Why religious education poses a qualitatively different AI problem than algebra or geography is the study’s conceptual core. Greek RE is a constitutionally anchored subject taught by university-trained theologians, and its curricula since 2016 have shifted from catechetical instruction toward learner-centered religious literacy. Within the Orthodox framing the study draws on, knowledge of God is understood as personal, relational, and experiential — classically expressed in the patristic distinction between nous, the noetic faculty of direct spiritual apprehension, and dianoia, discursive reason. AI systems operate entirely within the register of dianoia. A language model can simulate a dialogue with a saint, the training material argues, but it cannot embody the uniqueness of a person on a spiritual path, which is precisely why its output in this domain risks simulating encounter without delivering it.

The theoretical scaffolding comes from Shulman’s pedagogical content knowledge, extended through Mishra and Koehler’s TPACK framework and its recent AI-informed variants, which add ethical awareness and critical judgment about algorithmic systems to the classic triad. The survey’s own numbers illustrate where the seams are: teachers scored highest on integrating theological content into digital environments and lowest on designing student-centered activities with AI, and their confidence in using technology far outstripped their confidence in resolving technical problems, which at 3.24 was the weakest single item in the instrument. That gap between fluent use and critical, student-centered deployment mirrors a broader finding in the literature that teachers’ ethical and evaluative AI knowledge lags behind their technical familiarity.

Two findings stand out for anyone designing teacher training. The first is the near-unanimous endorsement of professional development itself — a mean of 4.89 for the statement that training is essential, with 50 of 54 respondents giving it the top score, and 4.81 for the value of exchanging good practices. Even teachers with more than two decades of classroom experience endorsed structured support as strongly as their younger colleagues. The second is subtler: the teachers’ theological framing of AI’s limits converges almost verbatim with the framing the SUB8 material itself taught, including the St. Basil chatbot exercise. The author is careful about what that convergence means. Because the survey was administered in the weeks immediately after training, it cannot show whether these convictions arise independently; the defensible reading is that an explicitly theological account of AI’s limits can be deliberately cultivated through professional development, and that participants internalized and reproduced it.

That caution is one of several the study applies to itself, and they matter for how the results should be quoted. The sample is small and self-selected — teachers who voluntarily enrolled in an AI-focused course are plausibly more digitally engaged and more favorably disposed than the roughly 2,700 Greek RE teachers as a whole. All data are self-reported and unverified by classroom observation, and the timing of measurement, at peak post-training enthusiasm, may inflate the competence scores. The study therefore claims to describe this specific professional development context, not Greek religious education teachers in general, and it makes no causal claims; it is a descriptive survey, not an experiment.

What the study can legitimately claim is a place in a very small comparative literature. The only prior empirical work on AI and religious education surveyed teachers in Germany and Poland, finding a similar ambivalence: appreciation of AI’s affordances for preparation and differentiation paired with insistence on the teacher’s irreplaceability. The recurrence of that conditional-acceptance profile across Catholic, Protestant, and now Orthodox settings hints — on still thin and methodologically heterogeneous evidence — that it may be a transnational feature of RE professionals rather than a local artifact, a hypothesis the author says would need harmonized comparative instruments to test properly. The practical upshot for program designers is concrete: combine technical skills with explicit attention to AI’s epistemological limits in spiritually sensitive subjects, build sustained collegial exchange into the program, and address the infrastructure and language barriers that even an enthusiastic cohort still reports. The deeper question the study leaves open is whether the teachers’ principled stance — AI as tool, never as encounter — would hold among colleagues who never sat in the training room, and whether it survives contact with the classroom once the post-training glow fades.

Subject of Research: Teacher perceptions of artificial intelligence in Greek religious education

Article Title: Artificial intelligence in religious education: teacher perspectives from a national professional development context

Article References: Komninou, I. (2026). Artificial intelligence in religious education: teacher perspectives from a national professional development context. Journal of Religious Education. https://doi.org/10.1007/s40839-026-00308-4

Image Credits: AI Generated

DOI: 10.1007/s40839-026-00308-4

Keywords: artificial intelligence, religious education, teacher survey, TPACK, professional development, Greece, Orthodox education, AI literacy, theology of education, Artificial, intelligence, religious

Nurse innovative behavior was measured as a process in which nurses generate, promote, and implement new methods or solutions in response to clinical nursing problems and patient needs. The 10-item instrument used in the study assessed three components: generating ideas, obtaining support, and realizing ideas. Participants scored an average of 37.93 on a scale ranging from 10 to 50, with a standard deviation of 6.50.

Researchers used a stratified online survey conducted from June to July 2026. Eligible participants were registered nurses working in clinical roles, with at least one year of clinical experience, at two tertiary hospitals and four Grade-A tertiary hospitals in Yueyang. Four institutions were general hospitals and two were specialized hospitals.

Nursing units were stratified into internal medicine, surgery, and other categories, with recruitment targets of at least 40 nurses from each category per hospital. The team collected 892 questionnaires and excluded 42 that failed predetermined quality checks, including completion in less than five minutes, implausible demographic inconsistencies, and selection of the same answer for all 65 scale items. The resulting sample of 850 represented a valid-response rate of 95.3%.

A 12-item, seven-point Artificial Intelligence Literacy Scale assessed awareness, use, evaluation, and ethics, with higher scores indicating greater literacy. A 22-item scale measured AI self-efficacy through assistance, anthropomorphic interaction, comfort with AI, and technical skills. Organizational innovation climate was measured with 21 items covering organizational innovation incentives, resource provision, and management practices.

The Nurse Innovative Behavior Scale used 10 items to assess idea generation, support seeking, and implementation. Cronbach’s alpha values in the study were 0.90 for AI literacy, 0.96 for AI self-efficacy, 0.92 for organizational innovation climate, and 0.87 for innovative behavior. Confirmatory factor analysis indicated good model fit, with a comparative fit index of 0.978, Tucker–Lewis index of 0.971, root mean square error of approximation of 0.063, and standardized root mean square residual of 0.028.

All four principal variables were positively correlated, with coefficients ranging from 0.609 to 0.797 and P values below 0.01. AI literacy and AI self-efficacy had the strongest correlation, at 0.797, while organizational innovation climate and innovative behavior had a coefficient of 0.726. A multiple linear regression identified educational level, professional title, participation in research or innovation activities, AI literacy, AI self-efficacy, and organizational innovation climate as factors associated with innovative behavior.

Together, these variables explained 60.9% of the observed variance in innovative-behavior scores; the adjusted R-squared was 0.602. Organizational innovation climate had the largest standardized regression coefficient, 0.502, followed by AI self-efficacy at 0.211 and AI literacy at 0.124. These coefficients describe the study’s multivariable model and do not establish the independent causal influence of any one factor.

A statistical mediation analysis examined whether AI self-efficacy helped explain the association between AI literacy and innovative behavior. AI literacy was positively associated with self-efficacy, and self-efficacy was positively associated with innovation. After self-efficacy was included, the direct association between literacy and innovative behavior remained statistically significant, indicating partial rather than complete mediation.

The estimated indirect effect was 0.322, with a bootstrap 95% confidence interval of 0.243 to 0.403, while the direct effect was 0.287, with a 95% confidence interval of 0.203 to 0.372. The indirect path accounted for 52.90% of the model’s estimated total effect. Because the measures were collected at one point, this path describes the fitted statistical model rather than a demonstrated sequence in which additional AI knowledge changes confidence and subsequently produces innovation.

The researchers also tested whether organizational innovation climate altered the strength of the self-efficacy–innovation relationship. The interaction between AI self-efficacy and organizational innovation climate was positive, with a coefficient of 0.051 and P < 0.01. Simple-slope analysis showed that self-efficacy was positively associated with innovative behavior at both low and high levels of innovation climate, but the coefficient increased from 0.151 at low levels to 0.253 at high levels.

Conditional indirect effects through self-efficacy were also larger at higher climate levels: 0.120 at one standard deviation below the mean, 0.161 at the mean, and 0.202 at one standard deviation above the mean. None of the bootstrap 95% confidence intervals included zero. The index of moderated mediation was 0.041, with a bootstrap 95% confidence interval of 0.009 to 0.078, indicating that the estimated indirect relationship varied across levels of organizational innovation climate.

The analysis was organized around the Ability–Motivation–Opportunity framework, under which behavior is associated with individual capability, motivation-related beliefs, and external opportunity. Within the study model, AI literacy represented an ability foundation, AI self-efficacy reflected confidence in learning, using, and adapting to AI, and organizational innovation climate represented perceived support, resources, incentives, and practical opportunity. The pattern was consistent with nurses’ reported innovative behavior being related both to their own technology-related capabilities and confidence and to the environment in which they work.

The result does not show that individual confidence is sufficient on its own: the stronger association observed under more supportive innovation climates suggests that statistical links between confidence and behavior depended partly on organizational conditions. Several participant characteristics also differed in univariate comparisons of innovative behavior. Scores varied by age, educational level, years of work experience, professional title, marital status, work content, AI-related training, frequency of AI-tool use, and participation in nursing research or innovation activities.

In the multivariable model, nurses with bachelor’s or master’s degrees, those with higher professional titles, and those who had participated in or led innovation activities had higher innovative-behavior scores than the relevant comparison groups. The sample included 280 nurses who had attended AI-related training, 420 who reported using AI tools one to three times per week, and 202 who reported daily use. Whether AI systems had been deployed in the hospital was not significantly associated with innovative behavior in the univariate analysis, although that comparison did not establish the presence or absence of an effect in more complex models.

Several limitations constrain interpretation. The cross-sectional design, with all variables measured simultaneously, prevents determination of temporal order or causation. The central measures were self-reported, so reporting bias remains possible even though the preliminary common-method-bias test did not identify an obvious single-factor problem.

The mediation and moderated-mediation analyses were not adjusted for demographic covariates, and potential clustering of nurses within hospitals was not modeled. Although the sample included both general and specialized tertiary hospitals, it was restricted to one Chinese city and to tertiary institutions, limiting generalizability to other regions and hospital levels. Future research could combine nurse questionnaires with evaluations from colleagues and head nurses, as well as objective indicators such as completed innovation projects, patents, quality-improvement outcomes, and records of AI-tool use.

Longitudinal, multicenter, and intervention studies would be needed to test whether changes in AI competence, confidence, or organizational support lead to sustained improvements in nurse innovation.

Subject of Research: Associations among AI literacy, AI self-efficacy, organizational innovation climate, and innovative behavior among clinical nurses

Article Title: Artificial intelligence literacy and innovative behavior among clinical nurses: the roles of AI self-efficacy and organizational innovation climate

Article References: Li, J., Liu, X., Cao, Y., Tang, C., & Zhou, J. (2026). Artificial intelligence literacy and innovative behavior among clinical nurses: the roles of AI self-efficacy and organizational innovation climate. Frontiers in Public Health, 14, Article 1958854. https://doi.org/10.3389/fpubh.2026.1958854

Image Credits: AI Generated

DOI: 10.3389/fpubh.2026.1958854

Keywords: clinical nursing, artificial intelligence, AI literacy, AI self-efficacy, nurse innovation, organizational climate, cross-sectional survey, tertiary hospitals, China, Artificial, intelligence, literacy

Because all variables were measured at one point, the results identify statistical associations and a plausible pathway rather than proving that AI anxiety causes fatigue or turnover intention. The researchers recruited frontline registered nurses through convenience sampling from a regional medical consortium that included one tertiary, one secondary and two primary hospitals. Eligible participants had at least one year of clinical experience; interns, trainees, nurse managers, educators, research nurses and administrators were excluded.

Of 446 questionnaires distributed, 404 were retained, giving a valid response rate of 90.58%. Most respondents worked in clinical wards, were aged 25 to 35, held at least a bachelor’s degree and had record-based employment contracts. Turnover-intention scores differed significantly by age, department and professional title, so the researchers adjusted the main structural model for those characteristics.

Four established Chinese-language questionnaires measured the central variables. The 20-item Artificial Intelligence Anxiety Scale assessed concerns spanning learning, substitution of work by AI, socio-technical blind spots and AI configuration. The 35-item Health Workers’ Occupational Fatigue Scale covered physical, cognitive-emotional, social, sleep-related and auditory fatigue.

Turnover intention was measured with six items addressing the desire to leave, motivation to seek another position and perceived opportunities to obtain alternative employment. Emotional intelligence was assessed through the 16-item Wong and Law Emotional Intelligence Scale, which includes self-appraisal, appraisal of others’ emotions, emotion regulation and use of emotion. Internal consistency was strong across the instruments, with Cronbach’s alpha values ranging from 0.818 to 0.966.

The survey data showed substantial correlations among the study’s main measures. AI anxiety and turnover intention had a Pearson correlation of 0.756, while occupational fatigue and turnover intention were correlated at 0.693. AI anxiety was also positively correlated with occupational fatigue at 0.543.

Emotional intelligence had negative correlations with turnover intention, at −0.668, and occupational fatigue, at −0.490, as well as with AI anxiety, at −0.518. The mean AI-anxiety score was 86.98 on a scale ranging from 20 to 140, above its theoretical midpoint, whereas mean occupational fatigue was 97.90 on a scale from 35 to 175, below its theoretical midpoint. Mean turnover intention was 15.55 on a six-item scale with a possible range of 6 to 24.

Structural equation modelling was used to estimate relationships among the latent constructs while controlling for age, department and professional title. The model had acceptable, though not excellent, statistical fit: the chi-square-to-degrees-of-freedom ratio was 4.323, root mean square error of approximation was 0.082, and the comparative fit index was 0.875. AI anxiety was positively associated with occupational fatigue, with a standardized coefficient of 0.611, and with turnover intention, with a coefficient of 0.530.

Occupational fatigue was itself associated with turnover intention at 0.489. All three path coefficients were statistically significant, and the total standardized effect of AI anxiety on turnover intention was 0.829, with a 95% confidence interval from 0.733 to 0.934. Bootstrap testing with 5,000 resamples supported partial statistical mediation by occupational fatigue.

The indirect effect of AI anxiety on turnover intention through fatigue was 0.299, with a bias-corrected 95% confidence interval of 0.129 to 0.379, excluding zero. The direct path from AI anxiety to turnover intention remained significant after occupational fatigue was included, indicating that fatigue did not explain the entire association. In practical terms, the model is consistent with a situation in which AI-related strain is related to thoughts of leaving both through higher fatigue and through some other associated route.

It does not establish that this sequence occurred over time, nor did the researchers directly measure the proposed cognitive, emotional or workload-related resource losses that might explain the association. Emotional intelligence produced a different result. In a moderated regression controlling for age, department and professional title, each additional unit of emotional intelligence was associated with 0.257 lower units of occupational fatigue, a statistically significant negative relationship.

However, the interaction between AI anxiety and emotional intelligence was not significant: its coefficient was 0.043, with a 95% confidence interval of −0.021 to 0.107 and p = 0.187. Adding this interaction increased explained variance by only 0.0017, also with p = 0.187. The overall model explained 60.2% of the variance in occupational fatigue.

Thus, higher emotional intelligence was associated with lower fatigue across the sample, but the study found no evidence that it specifically buffered nurses against fatigue associated with AI anxiety. The distinction between these findings matters for how technology-related psychological strain may be understood. AI anxiety can involve more than discomfort with unfamiliar software.

The scale used in the study captured fears related to learning, job substitution and the changing relationship between people and technical systems. Nurses may also experience uncertainty about professional identity, clinical judgment, autonomy and job security as AI becomes more deeply embedded in records, decision support, infection control and patient services. Under Conservation of Resources theory, concerns about losing valued psychological resources can itself become stressful.

Repeatedly interpreting machine-generated information, adapting routines and learning unfamiliar interfaces may add cognitive and emotional demands, while workflow changes may reduce time available for recovery. The study tested the resulting fatigue pathway statistically; it did not directly measure those proposed mechanisms. The authors interpret the results as suggesting that individual emotional skills alone may be insufficient when technological change reaches professional roles, organizational support and job security.

Their findings support considering AI literacy programs that address both technical competence and role changes, appropriate reliance on AI outputs, ethics, autonomy and uncertainty about future responsibilities. They also identify workflow redesign and fatigue-relief strategies as areas for future intervention research. These suggestions are not direct evidence that such programs will reduce turnover, because the survey did not test an intervention, measure actual resignation or track longitudinal changes.

Nor did it directly measure whether participants regarded AI as a replacement threat or whether they experienced job insecurity. Several limitations constrain how far the findings can be generalized. The cross-sectional design prevents temporal or causal conclusions: anxiety, fatigue and turnover intention were assessed together, and the data cannot show that occupational fatigue preceded or produced thoughts of leaving.

Convenience sampling through hospital communication channels may have introduced selection bias, even though participation was anonymous and supervisors were not directly involved in recruitment. The sample came from one geographic area and a closely integrated hospital consortium, and nurses working in settings outside the sampled institutions or in health systems with different digital-transformation conditions may experience different pressures. The study also lacked objective measures of AI exposure, frequency of use, technology type and implementation stage, limiting interpretation of how directly nurses’ work exposed them to AI.

Longitudinal studies that measure these factors before and after AI implementation are needed to test whether the observed pathway develops over time. The findings nevertheless identify a measurable workforce concern at a time when healthcare organizations are expanding the use of AI. Among the surveyed nurses, AI anxiety was not merely a technology-adoption attitude: it was closely associated with fatigue and stronger turnover intention, with fatigue statistically accounting for part of the connection.

Emotional intelligence was associated with lower fatigue overall but did not specifically weaken the AI-anxiety relationship, pointing toward the possibility that responses to digital transformation must extend beyond individual coping skills. Whether addressing AI-related strain can improve nurse retention, continuity of care or patient safety remains untested in this study. For now, the evidence points to fatigue as a potential intervention point, while also showing why organizational support and role clarity matter in the psychological consequences of healthcare digital transformation.

Subject of Research: Associations among AI anxiety, occupational fatigue, emotional intelligence and nurse turnover intention during healthcare digital transformation

Article Title: Artificial intelligence (AI) anxiety and nurses’ turnover intention during healthcare digital transformation: the mediating role of occupational fatigue

Article References: Zhu, L., Zang, F., Mao, G., Li, X., & Shen, X. (2026). Artificial intelligence (AI) anxiety and nurses’ turnover intention during healthcare digital transformation: the mediating role of occupational fatigue. Frontiers in Public Health, 14, Article 1982743. https://doi.org/10.3389/fpubh.2026.1982743

Image Credits: AI Generated

DOI: 10.3389/fpubh.2026.1982743

Keywords: AI anxiety, clinical nurses, occupational fatigue, turnover intention, emotional intelligence, healthcare digital transformation, structural equation modelling, nurse retention, cross-sectional survey, Jiangsu hospitals, Artificial, intelligence

The anonymous online survey was conducted from 1 April to 1 May 2026 at Bolu Abant Izzet Baysal University Faculty of Medicine. All 1,275 active students across the six-year medical program were invited, and 249 consenting participants completed the questionnaire, representing 19.5% of the eligible population. Their mean age was 21.49 years, with a range of 18 to 34; 69.1% were women, 30.5% were men, and one participant preferred not to report sex. Previous psychiatric treatment was reported by 24.9% of respondents and current treatment by 11.6%. The authors noted that the voluntary sample may have been enriched for students with greater mental-health concerns, so the results cannot be interpreted as population prevalence.

Students used AI for a mixture of academic and personal purposes. Among the 238 participants who had used AI, 84.5% selected course study or review, 71.8% selected examination preparation, 65.1% selected medical terminology or concept learning, and 62.6% selected research or literature searching. The same proportion, 71.8%, reported personal or non-educational use, while 38.2% used AI for clinical cases or reasoning practice. ChatGPT was the most commonly selected tool, followed by Google Gemini. Among users, 58.4% had been using AI for more than one year, indicating that the survey captured sustained engagement as well as possible recent adoption.

Psychological symptoms were measured with the validated Turkish 21-item Depression Anxiety Stress Scales, or DASS-21, covering the previous week. The primary outcome was a general-distress composite formed by summing all 21 items and doubling the result to match the DASS-42 metric, producing a possible range of 0 to 126. The mean was 56.97 with a standard deviation of 31.55. Internal consistency for the composite was high, with Cronbach’s alpha and McDonald’s omega both 0.951. A one-factor model fit moderately rather than strongly, while a proposed three-factor model for depression, anxiety and stress fitted better but also fell short of conventional absolute-fit benchmarks. The total was therefore treated as a shared symptom burden, not as proof that the three dimensions were identical. DASS severity categories were treated as screening descriptors, not psychiatric diagnoses or estimates of prevalence in the eligible student population.

The central regression used three blocks and ordinary least squares with HC3 robust standard errors. The first block included training year, sex and previous psychiatric treatment and explained 10.6% of the variance in general distress. Adding three categories of AI-use frequency increased explained variance by only 0.016, a change that was not statistically significant: the conventional test gave p = 0.111 and the robust global test p = 0.088. Adding the study’s three exploratory Artificial Intelligence Attitude Scale factors increased explained variance by another 0.052, producing an overall R-squared of 0.175 and an adjusted R-squared of 0.133. Thus, frequency did not improve the adjusted model as a block, whereas attitudes added exploratory explanatory information. Because the analytic models were specified after data collection and were not preregistered, all inferential results were hypothesis-generating rather than confirmatory.

Per one-standard-deviation increase in an attitude factor, Benefit was associated with a 4.90-point lower general-distress score, with a 95% confidence interval from −9.14 to −0.65 and a false-discovery-rate-adjusted q value of 0.036. Risk was associated with a 4.98-point higher score, with a confidence interval of 0.87 to 9.09 and q = 0.036. Use was associated with a 4.69-point higher score, with a confidence interval of 0.26 to 9.12 and q = 0.038. The apparently conflicting Benefit and Use relationships complicate any simple “helpful versus harmful” interpretation. A student might value AI’s educational benefits while also using it more frequently and perceiving greater risks; alternatively, distress and academic pressure could influence both attitudes and willingness to engage.

The findings were directionally similar across the depression, anxiety and stress subscales: Benefit had a negative coefficient and Risk and Use had positive coefficients in each model. Statistical support was much weaker, however, after correction for the 15 additional focal tests. Only the inverse relationship between Benefit and Anxiety survived that correction, with a coefficient of −2.39 points, a 95% confidence interval of −3.91 to −0.87, a nominal p value of 0.002 and an adjusted q value of 0.032. AI-use frequency remained nonsignificant at the block level for all three outcomes. These subscales came from the same participants and do not constitute independent replications, while the strong correlations among depression, anxiety and stress also limit the interpretation of separate domain-specific effects.

Measurement uncertainty was substantial. The 13-item attitude scale had three proposed domains—Benefit, Risk and Use—and was originally developed for Turkish pre-service teachers rather than medical students. Its three-factor structure performed much better than a one-factor alternative, with standardized loadings from 0.520 to 0.938. Internal consistency ranged from 0.775 to 0.931 for Cronbach’s alpha and from 0.774 to 0.933 for McDonald’s omega. Yet its absolute fit was mixed, with a comparative fit index of 0.898, Tucker–Lewis index of 0.872, root mean square error of approximation of 0.107 and standardized root mean square residual of 0.070. The wording was translated by one investigator from the English version rather than formally adapted for cross-cultural use. Nine of the 13 administered items differed lexically or in phrasing from the published Turkish form, and one difference may have broadened the item’s content. The authors therefore treated the factor scores as exploratory and conditional on the structural evaluation.

Several design features also limit what can be inferred. The study was cross-sectional, single-center, voluntary and based entirely on self-report. AI frequency did not capture session duration, task complexity, learning quality, verification of outputs or degree of reliance, and nearly four-fifths of respondents used AI at least a few times a week, leaving limited contrast at the lower end of exposure. Students with distress or strong views about AI may also have been more likely to participate. Training-year differences cannot demonstrate improvement over time because each cohort was assessed once; the survey overlapped a first-year midterm period, and individual workload and assessment timing were not recorded. The results also do not demonstrate that any single AI tool is responsible for the observed patterns. Prior research cited by the study found that mental-health problems predicted later AI dependence among adolescents, making reverse causation a plausible explanation for the positive Use association, but the current data cannot test that pathway.

The results point to a need for longitudinal, multicenter research rather than a simple verdict on AI in medical education. Future studies would need formally adapted attitude instruments, validated well-being measures, objective or diary-based records of AI engagement, and information about assessment timing and specific educational tasks. Qualitative work could help explain how students combine perceived benefits, risks and active engagement. Until stronger evidence is available, the survey’s most defensible conclusion is that AI use was widespread among these students, but its frequency alone did not provide stable evidence of an overall relationship with distress. The exploratory attitude associations require replication and cannot be interpreted as causal effects, psychological diagnoses or evidence that greater AI exposure is inherently advantageous or harmful.

Subject of Research: A cross-sectional survey of AI use patterns, attitudes and psychological distress among 249 medical students at one Turkish university.

Article Title: Artificial intelligence use patterns, attitudes, and psychological distress among medical students: a cross-sectional survey

Article References: Sari, M., İmrek, Y., & Tufan, A. E. (2026). Artificial intelligence use patterns, attitudes, and psychological distress among medical students: a cross-sectional survey. Frontiers in Medicine, 13, Article 1965749. https://doi.org/10.3389/fmed.2026.1965749

Image Credits: AI Generated

DOI: 10.3389/fmed.2026.1965749

Keywords: artificial intelligence, medical students, psychological distress, AI attitudes, medical education, DASS-21, cross-sectional survey, mental health, digital learning, survey research, Artificial, intelligence

The study drew from GEO series GSE59856, GSE106817, GSE112264, GSE113486 and GSE124158, which together contained pancreatic cancer and healthy-control serum profiles measured with 3D Gene human miRNA microarray platforms. The sample distribution was not uniform across studies: GSE59856 contained 150 controls and 100 pancreatic cancer samples, GSE106817 115 controls and 200 cancer samples, GSE112264 50 controls and 41 cancer samples, GSE113486 40 controls and 100 cancer samples, and GSE124158 30 controls and 275 cancer samples. Such variation motivated the researchers to harmonize identifier conventions, impute missing values, align features across platforms and apply ComBat batch correction, an empirical-Bayes procedure designed to reduce nonbiological differences among datasets. The analysis started with normalized expression matrices already processed by the original studies rather than raw fluorescence measurements.

After log transformation, the researchers used the limma statistical package to compare cancer and control samples. The researchers used a limma differential expression workflow, with feature selection based on a false discovery rate threshold below 0.05 and an absolute log2 fold-change threshold of at least 2. The researchers treated the 148 resulting differentially expressed microRNAs as the input features for predictive modeling. Principal component analysis showed separation between the two groups, with the first two components accounting for 17.5% and 14.6% of variance, respectively. The researchers presented this as evidence of discriminatory information in the integrated expression profiles, but the retrospective design, harmonization of archival data and lack of uniformly available clinical variables limit how directly that separation can be interpreted in an unselected patient population.

For classification, the researchers partitioned the combined data into an 80% training set and a 20% test set, retaining proportional representation of cancer and control samples. The training set was further divided into stack-training and validation subsets. Five base algorithms—Extreme Gradient Boosting, Random Forest, Support Vector Classifier, Naïve Bayes and k-nearest neighbors—generated probabilities for the cancer class. The researchers combined those outputs through a Bayesian stacking ensemble, estimating non-negative weights by minimizing binary cross-entropy with the Sequential Least Squares Programming algorithm and constraining the weights to form a convex combination. Synthetic Minority Over-Sampling Technique was applied only within the training partitions to address the moderate class imbalance, while the test set was not oversampled. The researchers also introduced controlled label noise into 3% of training labels as an auxiliary robustness exercise; they emphasized that this was not an estimate of the real-world diagnostic misclassification rate.

On the held-out set, the ensemble’s receiver operating characteristic curve rose steeply at low false-positive rates, consistent with the reported AUC of 0.97. Its sensitivity of 94% indicates that the model identified 94% of the cancer samples in that test set, while its specificity of 99% reflects the proportion of controls correctly classified. The reported F1 score was 96%. These figures were obtained from samples drawn from the same five archived studies used to build the broader dataset, however, rather than from a strictly independent external cohort. The researchers explicitly noted a potential source of information leakage because differential expression analysis was performed before the train-test split. Although the authors characterized this as a biological feature-discovery step rather than predictive selection, they said future work should use nested cross-validation or fold-wise feature selection to test robustness more rigorously.

To make the model more transparent, the researchers applied SHAP, or Shapley Additive Explanations, to the XGBoost component of the ensemble. SHAP decomposes individual predictions into feature contributions and can rank variables by their average absolute influence across samples. The analysis used 200 samples from the SMOTE-balanced training distribution as a reference background and calculated values for all features in the test set. The resulting ranking highlighted 20 microRNAs, including hsa-miR-125a-3p, hsa-miR-6839-5p, hsa-miR-1290, hsa-miR-614, hsa-miR-4436a and hsa-miR-497-5p. hsa-miR-125a-3p ranked first and was downregulated in cancer samples, with a log2 fold change of −2.944 and an adjusted p value below 0.001. Low expression of this microRNA was associated with a stronger contribution toward a cancer classification.

Several other leading candidates were upregulated in the integrated data. hsa-miR-1290 had a log2 fold change of 2.774, hsa-miR-614 had 2.250 and hsa-miR-128-2-5p had 2.522, each with an adjusted p value below 0.001. hsa-miR-497-5p was among the downregulated features, with a log2 fold change of −2.665 and an adjusted p value below 0.001. The overlap with previously reported pancreatic-cancer markers was interpreted as supporting biological plausibility: earlier work had associated serum miR-1290 with early pancreatic cancer and miR-125a-3p with pancreatic and biliary cancer. The researchers also emphasized the limits of that comparison. A high SHAP value describes how strongly a feature influences this particular model; it does not demonstrate that the microRNA causes cancer, acts upstream in tumor development or would independently improve patient outcomes.

The team next examined predicted or reported targets for the 20 SHAP-ranked microRNAs using DIANA-miRPath v4.0, miRDB, miRTarBase 2022, miRBase v22.1 and KEGG resources. Twelve microRNAs had DIANA-associated targets connected to pathways relevant to pancreatic cancer, including genes involved in cell-cycle control, MAPK signaling, apoptosis, DNA-damage responses, angiogenesis and inflammation. The highlighted targets included CDK4, CDK6, CCND1, CDKN1A, MAPK1, RAF1, MAP2K1, TP53, BCL2L1, E2F3, VEGFA and RELA. The researchers reported that hsa-miR-497-5p targeted several components of the MAPK pathway, while hsa-miR-125a-3p was linked to CDK4 and VEGFA. Six candidates not represented in DIANA were examined through miRDB, where their predicted targets were associated with cytoskeletal remodeling, ubiquitination, transcriptional regulation and metabolic adaptation. These database relationships generate mechanistic hypotheses but do not experimentally verify the proposed interactions.

Pancreatic cancer remains difficult to detect early because it can progress without obvious symptoms, while benign pancreatic disorders can resemble it clinically and radiologically, and definitive diagnosis may involve invasive tissue sampling. That context makes minimally invasive blood markers attractive, but the new results do not establish a ready screening test. The study used a case-control design comparing pancreatic cancer samples with healthy controls rather than a prospective screening population, and it did not evaluate performance against benign pancreatitis, other cancers or patients with other pancreatic diseases. Disease stage, tumor subtype, demographic characteristics, treatment history and other clinical variables were not consistently available across the five public datasets, so the resulting signature cannot be interpreted as specific to a stage, subtype or treatment group. Its reported diagnostic performance may therefore be optimistic in routine care, where disease prevalence, spectrum and sample handling differ from the research setting.

Translation would also require a reproducible laboratory assay, clinically meaningful decision thresholds and prospective testing in cohorts that represent the intended screening population. The 20-microRNA panel may need to be reduced to a smaller set to limit cost and complexity, and its performance should be compared with, or combined with, established markers such as CA19-9 rather than assumed to replace existing testing. Orthogonal methods, including RT-qPCR or digital PCR, would be needed to confirm that the prioritized microRNAs can be measured reliably in serum. Independent cohorts, standardized collection and processing, assays against benign pancreatic disease and functional experiments would then test whether the candidates retain their predictive value and whether their proposed target relationships have biological meaning. The study therefore offers a detailed and reproducible computational shortlist rather than a validated pancreatic-cancer diagnostic: its strongest contribution is the combination of multi-cohort data, transparent model attribution and pathway-based context, while its clinical promise remains to be tested.

Subject of Research: Computational identification of circulating microRNA candidates for pancreatic cancer detection

Article Title: Explainable artificial intelligence identifies circulating microRNA candidate biomarkers for liquid biopsy-based pancreatic cancer detection

Article References: Mandal, S., Ansari, M. A., & Chanu, W. P. (2026). Explainable artificial intelligence identifies circulating microRNA candidate biomarkers for liquid biopsy-based pancreatic cancer detection. Frontiers in Artificial Intelligence, 9, Article 1810551. https://doi.org/10.3389/frai.2026.1810551

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1810551

Keywords: pancreatic cancer, liquid biopsy, circulating microRNAs, explainable artificial intelligence, machine learning, SHAP, biomarker discovery, serum diagnostics, Explainable, artificial, intelligence, identifies

The central argument is that artificial intelligence and Universal Design for Learning should be treated as mutually constraining rather than as interchangeable technologies that can simply be combined. UDL is an evidence-informed approach to reducing barriers by designing flexible curricula around learner variability, with three principles: multiple means of engagement, representation, and action and expression. AI can operationalize those principles by supplying tools that adapt the pace or difficulty of learning, provide immediate explanations, caption media, translate content, generate alternative examples and support multiple forms of assessment. UDL, in turn, determines which applications are pedagogically appropriate by asking whether they expand students’ opportunities to engage, receive information and demonstrate understanding. An efficient system that narrows those options may perform well technically while contradicting the principles it is intended to support.

Five theoretical components explain how the proposed model works. Systems theory emphasizes that leadership, policy, infrastructure, technology and professional practice must align for sustainable adoption. Institutional theory explains why universities may imitate peer organizations, comply with external expectations or seek legitimacy by adopting similarly worded policies. Responsible AI supplies an ethical foundation based on fairness, transparency, accountability, privacy and human oversight. The model also adds a critical counter-current to these accounts. It warns that AI can encode historical bias, transfer educational judgment to proprietary systems, narrow what counts as measurable learning and extend commercial platform logics into university teaching. Critical oversight is therefore not a separate final stage. It operates across the full cycle, from leadership decisions and tool selection to faculty practice, student assessment and the metrics used to judge success.

The first pillar is institutional leadership and strategic governance. University leaders would establish a clear AI-UDL strategy, embed UDL and accessibility requirements in curriculum approval, allocate resources and define indicators for equitable implementation. The pillar also requires leaders to distinguish between tools supported by pedagogical evidence and systems adopted mainly because competing institutions possess them. That distinction is particularly important in a centralized system where national priorities, accreditation expectations and regional examples can create pressure for universities to resemble one another. The second pillar, AI-enabled UDL, maps applications onto all three UDL principles. For engagement, adaptive pathways, intelligent tutoring systems and conversational assistants can offer individualized support. For representation, text-to-speech, automatic captioning, translation and simplified explanations can provide alternative ways to access course material. For action and expression, AI-supported feedback, digital portfolios, adaptive quizzes and multimedia submissions can enable students to demonstrate learning in different formats while maintaining consistent academic standards.

None of these capabilities works automatically. The study specifically notes that machine-translation quality requires faculty review for disciplinary accuracy and cultural appropriateness before translated materials reach students. Predictive learning analytics may identify first-year students who need additional advising or instruction, but systems trained on historical institutional data can reproduce inequities associated with disability, language, gender or socioeconomic background. A dashboard does not neutrally portray a student: it constructs a profile that can shape the support, scrutiny or opportunities that person subsequently receives. The third pillar, the responsible AI ecosystem, consequently connects tool use to fairness audits, procurement rules, data governance, cybersecurity, algorithm evaluation and human review. Institutions would also need to examine whether vendors’ accessibility claims have been independently verified, whether algorithmic recommendations receive human scrutiny before affecting advising, and whether core teaching, assessment or advising functions have become dependent on one commercial platform.

Faculty capacity building forms the fourth pillar because institutional technology programs cannot deliver their intended benefits when instructors lack the knowledge or confidence to use them well. The study defines professional learning around five connected domains: AI literacy, inclusive pedagogy, responsible AI, critical evaluation of claims about artificial intelligence, and digital assessment design. Training on these subjects should be continuous rather than delivered as isolated workshops, with universities providing protected time, peer mentoring and accessible technical support. Lecturers are encouraged to redesign authentic assessment and multi-format activities rather than depend on software that attempts to detect generative AI use. Students also need instruction in responsible use, critical evaluation and understanding of their data rights. This need is particularly timely because the evidence reviewed by the study indicates that Gulf students already use generative tools for drafting, revision and exam preparation, sometimes ahead of consistent university guidance.

The fifth pillar, continuous evaluation, closes the feedback loop by examining learning, equity, accessibility, faculty outcomes and the effects of institutional AI policies. Recommended indicators are disaggregated by factors including disability, gender, language and socioeconomic background, allowing institutions to examine whether an apparently helpful tool produces different effects across groups. The framework also proposes three forms of critical oversight: an annual deprofessionalisation audit to identify grading judgments, risk flags or curriculum decisions that have moved from faculty to algorithms; a vendor-dependency register to make platform lock-in visible; and a periodic purpose check to ask whether institutional metrics have quietly narrowed the definition of successful learning. These mechanisms connect the framework to responsible governance because an institution cannot evaluate equity adequately if it monitors only overall averages, engagement metrics or aggregate performance.

For implementation, the study presents a four-stage roadmap: institutional readiness, capacity building, AI-enabled curriculum redesign and continuous improvement. Institutions would first assess their infrastructure, accessibility, faculty readiness and policy maturity. They would then develop governance policy, professional learning and accessibility standards, redesign programmes around UDL using responsible AI technologies, and monitor learning analytics, accessibility indicators and quality-assurance data to feed improvements back into the next cycle. A self-administered maturity model can locate an institution’s starting point by rating each pillar as beginning, in process or continuously improving. Stakeholder-specific actions include connecting national AI strategy with inclusive-education policy, setting minimum standards for transparency, fairness auditing and student-data protection, and investing in accessible Arabic-first infrastructure. Universities could form multidisciplinary governance committees involving academic leaders, disability specialists, technologists, legal and cybersecurity staff, and students.

The framework also assigns technology vendors a direct role. Vendors would be expected to meet accessibility standards as a contractual baseline subject to independent institutional audit, provide explainable and auditable systems, and invest in Arabic and multilingual model quality rather than treating those languages as secondary markets. The study draws on international comparisons to show why context matters. Canada’s binding accessibility requirements, New Zealand’s approach to cultural and data sovereignty, and the different governance arrangements in the United Kingdom, Australia, Singapore and Finland are treated as possible reference points, not directly transferable models. Kuwait’s simultaneous negotiation of AI governance and UDL, centralized national planning and need for Arabic-first infrastructure justify a context-specific framework. The comparison is conceptual and does not establish which national approach produces better student outcomes in the Gulf.

The paper is explicit about the empirical work still required. Its structured search covered Scopus, Web of Science, ERIC, Google Scholar and several policy databases. It returned approximately 1,450 records before deduplication and 1,120 after deduplication; title and abstract screening retained 210 records, while full-text screening retained 96 directly relevant sources, of which 42 appeared in the final reference list. The validation of the five pillars was conceptual, or face, validity based on comparison with reference frameworks. It did not establish content validity through a formal expert panel or construct validity through empirical testing. Applicability may also vary with institutions’ technological maturity, the synthesis drew predominantly on English-language research, and much of the supporting governance evidence originated outside the Gulf Cooperation Council. The proposed six-part research agenda calls for longitudinal or quasi-experimental studies of engagement, performance, retention and accessibility; tests of whether fairness auditing moderates disparate impacts; research tracing the pathway from leadership through faculty capability to outcomes; case studies involving disabled, multilingual and first-generation students; comparisons of institutional maturity; and expert consultation followed by factor analysis to validate the framework itself. The study’s principal contribution is therefore a testable institutional architecture, not proof that AI has already made Kuwaiti higher education more inclusive.

Subject of Research: A conceptual framework for integrating artificial intelligence, Universal Design for Learning and responsible governance in Kuwaiti higher education

Article Title: Artificial intelligence and universal design for learning in Kuwait’s higher education institutions: towards an AI-enabled framework for inclusive education

Article References: Alenezi, A. F. (2026). Artificial intelligence and universal design for learning in Kuwait’s higher education institutions: towards an AI-enabled framework for inclusive education. Frontiers in Education, 11, Article 1934276. https://doi.org/10.3389/feduc.2026.1934276

Image Credits: AI Generated

DOI: 10.3389/feduc.2026.1934276

Keywords: artificial intelligence, Universal Design for Learning, inclusive education, higher education, Kuwait, responsible AI, AI governance, learning analytics, faculty development, accessibility, Artificial, intelligence

The review, published in Journal of Neuro-Oncology, evaluated research that applied AI to Gamma Knife clinical, imaging or treatment-planning data. The authors searched PubMed/MEDLINE, Embase and Scopus through September 20, 2026. Studies outside the field of Gamma Knife radiosurgery and those not using learning-based methods were excluded, while eligible studies were synthesized according to major stages of the radiosurgery workflow. As a narrative review, the analysis maps the applications of AI rather than testing one model or intervention in a clinical trial.

Identified applications include tumor segmentation, isocenter selection, prescription and dose-distribution prediction, synthetic computed tomography generation, and assessment of treatment-plan quality. These tools address different parts of the technical planning process, from identifying anatomy to supporting decisions about where to focus treatment and how a proposed radiation dose is distributed. Several models have reported promising performance, but the review does not establish that routine use of these methods would improve patient outcomes. The evidence for each application also remains tied to the specific data, task and clinical setting in which its model was evaluated.

Post-treatment applications include longitudinal tumor monitoring and prediction of response, functional outcomes and toxicity. The reviewed toxicity models address outcomes such as peritumoral edema and radiation necrosis. Some studies have produced encouraging results within their respective datasets, but those findings do not demonstrate clinical benefit across Gamma Knife centers or among individual patients. Predictive performance alone therefore does not show that an AI tool can meaningfully improve treatment decisions, and the review emphasizes the need to demonstrate clinical benefit separately from technical performance.

Across the literature, the central limitation is the gap between promising model results and dependable clinical use. Many studies rely on small or single-center datasets, while external and prospective validation has been limited. This restricts the strength of the conclusions that can be drawn for broader clinical adoption. The review calls for multicenter validation, prospective evaluation, integration into clinical workflows, and evidence of meaningful clinical benefit before AI becomes routine in Gamma Knife radiosurgery. Until that evidence is available, the authors characterize AI as a decision-support tool that can support radiosurgical expertise rather than replace it. The current evidence therefore supports further evaluation of AI across specific Gamma Knife tasks, but not autonomous use or claims of improved safety, effectiveness or consistency in clinical practice.

Subject of Research: Artificial intelligence applications in Gamma Knife radiosurgery

Article Title: Artificial intelligence in gamma knife radiosurgery: applications in treatment planning, response prediction, and toxicity assessment

Article References: Zidan, A., El-Sherbini, A., Burns, L., & Soliman, H. (2026). Artificial intelligence in gamma knife radiosurgery: applications in treatment planning, response prediction, and toxicity assessment. Journal of Neuro-Oncology, 180(2), Article 27. https://doi.org/10.1007/s11060-026-05834-7

Image Credits: AI Generated

DOI: 10.1007/s11060-026-05834-7

Keywords: artificial intelligence, Gamma Knife radiosurgery, treatment planning, tumor segmentation, outcome prediction, toxicity assessment, radiation necrosis, medical imaging, clinical validation, decision support, Artificial, intelligence

Most neuro-oncology studies described technical performance, feasibility, retrospective analyses or planned work rather than demonstrated improvements in patient outcomes. Patient-reported outcomes, or PROs, are information reported directly by patients about symptoms, functioning and quality of life, without interpretation by a clinician. The review identifies pain, communication difficulty, perceived cognition, seizures, symptomatic adverse events, physical functioning and role functioning as priority constructs for high-grade glioma trials and clinical care.

Clinician-rated measures, including the Karnofsky Performance Status, Neurologic Function Score and Common Terminology Criteria for Adverse Events, remain central in neuro-oncology, while the Mini-Mental State Examination and Montreal Cognitive Assessment provide performance-based cognitive screening. These measures complement rather than replace PROs. Studies cited in the review found that clinicians can underestimate symptom frequency and severity, whereas cognitive impairment, aphasia and reduced insight can compromise self-reporting.

Caregiver proxies can provide information, but their interpretation introduces additional complexity. The quality of PRO data therefore constrains what AI systems can learn. Missing responses, selection bias, inconsistent assessment timing and disagreement between clinicians and patients can distort symptom trajectories, estimated treatment effects and model predictions.

Questionnaire completion commonly declines with respondent burden, clinical deterioration and logistical barriers. In a phase III trial of prophylactic cranial irradiation in extensive-stage small cell lung cancer, fewer than half the expected quality-of-life forms were returned at 12 months. Attrition is especially consequential because patients whose condition is worsening are more likely to stop responding, potentially reducing statistical power and biasing results.

The review notes that standardized schedules, electronic collection, reminders, staff education, compliance monitoring and appropriate methods for handling missing data may improve data quality, while each must accommodate differences in neurological status, communication and access. Evidence that frequent symptom monitoring can improve outcomes is strongest outside neuro-oncology. In a single-institution randomized trial involving 766 patients receiving chemotherapy for metastatic cancers, electronic PRO monitoring with automated nurse alerts reduced emergency-department visits, prolonged time on chemotherapy and improved one-year survival.

The PRO-TECT cluster randomized trial, conducted across 52 practices with 1,191 patients, reported that weekly electronic PRO surveys improved physical function, symptom control and health-related quality of life at three months in diverse community oncology settings. Brain-tumor evidence has largely remained at the feasibility stage. A glioblastoma smartphone-app pilot showed that longitudinal health-outcome tracking could include caregiver-proxy entries, while a separate study involving 43 patients with glioblastoma found that electronic PROMIS and EORTC assessments could be completed in a clinical setting.

These studies demonstrate feasibility rather than the clinical benefits reported in general oncology. Classical machine learning offers a way to analyze large and heterogeneous longitudinal PRO datasets by learning patterns from data rather than following fixed rules. Early work predicting health-related quality of life after low-grade brain-tumor resection used support vector machines and random forests; the classifiers achieved areas under the curve above 0.90 for most symptom outcomes using routine perioperative inputs.

A systematic review of machine-learning approaches to cancer symptom prediction found that temporal and longitudinal modeling were underrepresented and that neuro-oncology was absent. Natural language processing can extract structured information from free-text clinical notes, another large and largely untapped part of the medical record, but oncology-specific applications in neuro-oncology remain limited. Many such systems also require extensive manual rule development and domain-specific expertise, which can restrict adaptability across settings.

Large language models may extend the range of data that can be processed. In an analysis of 51,541 neuro-oncology clinical notes from an electronic health record, GPT-4o extracted nine symptom domains with mean sensitivity and specificity of 0.97. It identified an average of 3.92 symptoms per note, and reported symptom prevalence was substantially higher than in comparable clinical trials.

Patients receiving immune checkpoint inhibitors had the highest symptom burden, illustrating how automated extraction can provide treatment-stratified observations that complement prospective PRO instruments. The analysis did not establish that extracted symptoms can prospectively detect deterioration in primary brain-tumor patients or serve as substitutes for patient reports. That distinction is fundamental: a PRO originates directly from the patient, whereas an LLM-derived interpretation is an AI-derived measure and should be labeled and evaluated separately.

Broader assessments of patient-reported outcome models indicate weaknesses in validation and implementation. A 2024 systematic review of 52 oncology studies found that fewer than 8% of models had been externally validated and fewer than 6% shared reproducible code, despite a roughly twentyfold increase in machine-learning publications using PRO data between 2011 and 2021. A separate 2025 scoping review of 94 studies using PROM data to predict outcomes across clinical fields identified only 10 external validations and three studies that had reached implementation.

It also found inconsistent handling of missing data and limited patient or clinician involvement in model development. Because the reviews covered different scopes, they do not demonstrate a direct time trend. They do show that analytical performance has often been assessed without replication in independent settings.

Neuro-oncology faces additional constraints from disease heterogeneity, modest sample sizes and symptoms that patients may be unable to report, including seizures and neglect. Caregiver reports, wearable data and clinical narratives can supply complementary information. Integrating predictive systems into care presents separate challenges from developing them.

In a review of cancer symptom-management clinical decision-support systems, only 25.5% were integrated into electronic health records and 18.2% incorporated AI, with most relying on rules rather than predictive models. The review cites evidence favoring standardized terminology, high-yield alerts and a defined response whenever an alert is issued, because frequent low-specificity warnings can produce alert fatigue. Decisions based on unreliable data, or decisions that could cause harm if made incorrectly, may require clinician verification.

A modular autonomous workflow using five LLM agents to identify cognitive concerns in unstructured notes performed comparably to, but did not outperform, an expert-refined workflow on an independent validation set. Its inspectable structure made prompt refinement easier to review. In neuro-oncology, a glioblastoma remote-monitoring study registered in 2023 pairs Apple Watch biometric data with patient- or caregiver-entered symptoms, while a French glioma study registered in 2026 will evaluate a platform for neurologic symptoms, treatment tolerance and quality of life.

These protocols describe planned research and do not establish efficacy. Ethical requirements are particularly consequential for patients whose cognition or decision-making capacity may change over time. Informed consent and clear communication are needed about when AI is used, what information it processes, how its output could affect care and what alternatives exist.

A clinician should be able to inspect the original patient or proxy report, identify its source and review clinically consequential interpretations in context. If a large language model infers a symptom from a conversation, that output must not be presented as a patient-reported outcome. Privacy and security obligations also vary by jurisdiction.

In the United States, a provision of the 21st Century Cures Act can exclude certain clinician-facing decision-support functions from the definition of an FDA-regulated medical device when specified criteria are met, including independent clinician review of the basis for a recommendation; this is not a blanket exemption for AI or electronic-health-record tools. In the European Union, the General Data Protection Regulation governs personal and health-data processing, while the EU AI Act applies a risk-based framework to qualifying medical devices. Potential bias can arise from unrepresentative study populations, nonrepresentative clinical language, proxy variables, aggregation and selective feature use.

Clinical trial populations can underrepresent women and non-White patients, and symptom documentation may also differ across racial, ethnic, gender and socioeconomic groups. AI-generated summaries can intensify this problem when they present a narrow or skewed dataset as representative of a whole population. The review highlights equity-centered frameworks that require risk-benefit assessment, formal oversight and mechanisms for redress, alongside proposals to add explicability to established bioethical principles.

Its overall assessment is that AI and PRO systems could make neuro-oncology more patient-centered and data-driven, but current evidence is constrained by small studies, missing symptom data, inconsistent measurement and weak external validation. Progress depends on multicenter collaboration, standardized PRO frameworks, interoperable data systems, reproducible validation and broader inputs such as caregiver reports and wearables. Translation into care will require interpretable, electronic-health-record-integrated tools that provide actionable information while preserving patient consent, privacy, autonomy, equity and clinician oversight.

Subject of Research: Artificial intelligence for symptom assessment and patient-reported outcomes in neuro-oncology

Article Title: Artificial intelligence for symptom assessment and patient-reported outcomes in neuro-oncology

Article References: Yogendran, L. V., Avery, C., Crooms, R. C., Ozair, A., Siegel, C., Guterman, I., Rhee, J. Y., Binder, J., Minor, M., Alder, L., Davis, J., & Sharma, A. (2026). Artificial intelligence for symptom assessment and patient-reported outcomes in neuro-oncology. Journal of Neuro-Oncology, 180(2), Article 26. https://doi.org/10.1007/s11060-026-05832-9

Image Credits: AI Generated

DOI: 10.1007/s11060-026-05832-9

Keywords: artificial intelligence, neuro-oncology, brain tumors, patient-reported outcomes, symptom monitoring, machine learning, large language models, clinical decision support, medical ethics, data bias, Artificial, intelligence

The central innovation is not a single machine-learning model but an “intelligent service hub” that coordinates several task-specific algorithms. Load forecasting used a long short-term memory network with two recurrent layers, while a deep neural network handled binary fault classification and a random forest estimated whether equipment required maintenance. A convolutional neural network extracted visual features, natural-language processing transformed maintenance records into numerical representations, and a weighted late-fusion module combined the results. The authors assigned 50% of the fusion weight to sensor and time-series features, 25% to text and 25% to images or video, reflecting the continuous availability of electrical measurements and the contextual value of reports and visual inspections.

Information moves through the framework in a closed loop. Smart meters, Internet of Things sensors, cameras, drones, operational logs and inspection records are collected, cleaned and synchronized before being sent to specialized processing modules. The service hub then combines those inputs with a knowledge base containing grid topology, operating rules, maintenance records, constraints and safety limits. Its outputs can include load forecasts, fault classifications, equipment-failure probabilities, maintenance schedules, power-rerouting recommendations, renewable-energy dispatch, demand-response adjustments and cybersecurity alerts. Results from the resulting actions are fed back into the system so that predictions and control decisions can be revised as operating conditions change.

The experimental dataset contained 20,000 time-ordered samples assembled from public smart-grid resources. It included electrical measurements, environmental variables, renewable-generation information, maintenance records and fault labels, with text and visual inputs included for equipment monitoring. Because the measurements were time dependent, the researchers used the earliest 80% for training, the next 10% for validation and hyperparameter tuning, and the final 10% as an independent test set. They also conducted walk-forward validation with five sequential folds to assess performance under a more operationally realistic sequence of training on earlier observations and testing on later ones. Ten independent runs were reported with means, standard deviations and 95% confidence intervals.

For fault detection, the deep neural network produced probability scores for faulty and non-faulty states, which were assessed across decision thresholds using a receiver operating characteristic curve. The resulting ROC-AUC was 0.970 ± 0.012, with a 95% confidence interval from 0.961 to 0.979. The predictive-maintenance random forest combined historical maintenance data with equipment age, temperature, vibration, loading conditions and previous faults. On the 2,000-sample test set, it correctly classified 1,348 normal cases and 562 maintenance-required cases, while misclassifying 52 normal cases and missing 38 maintenance-required cases. Its precision for the maintenance-required class was 91.53%, recall was 93.67% and the F1-score was 92.59%.

The same predictive system attained 95.5% overall accuracy and 94.98% balanced accuracy, with a Matthews correlation coefficient of 0.894. The authors also evaluated a Weibull reliability model to represent failures associated with early life, random events or aging, using available component-failure histories where possible. Its initial scale and shape parameters were based on engineering assumptions when records were unavailable, and the researchers said they examined parameter sensitivity. In the operational configuration, components with a predicted failure probability above a validation-selected threshold of 0.60 were marked as requiring maintenance.

p>The load-forecasting module used 24-step sliding windows to predict the next demand or equipment-risk state. Across the reported repeated experiments, it achieved a MAPE of 4.8% ± 0.5%, with a 95% confidence interval of 4.4% to 5.2%. In an example daily profile, demand ranged from about 350 MW at night to 450 MW at midday. In another comparison, consumption ranged from 280 to 320 MW while generation ranged from 300 to 340 MW, illustrating the mismatch the optimization system was intended to manage by adjusting conventional generation, renewable use, storage and flexible demand. The paper describes this as a benchmark evaluation and does not report deployment in an operating utility.

For energy optimization, the framework worked over a 24-step horizon, maintained a 10% reserve margin and allowed demand-response adjustments of up to 15% of non-critical controllable load. The experimental baseline included conventional generation scheduling and reactive adjustment, while the optimized condition jointly coordinated load forecasts, renewable-generation predictions, storage, power allocation and demand response. Under identical test-horizon conditions, total modeled operating cost fell from $12,000 to $8,400. The largest reported saving was in conventional generation, which declined from $9,200 to $6,450 as renewable generation was prioritized and dispatch was adjusted; smaller reductions appeared in grid losses, renewable curtailment, peak demand and storage-control costs.

An ablation analysis indicated that the unified design mattered more than any isolated component. Removing sensor fusion produced the largest degradation among the component-removal tests, lowering fault-detection AUC from 0.970 to 0.931, predictive-maintenance accuracy from 95.5% to 91.6%, load-forecasting MAPE from 4.8% to 6.7%, and cost reduction from 30.0% to 23.5%. Removing energy optimization lowered the modeled savings to 16.2%, while a sensor-only model reached an AUC of 0.926 and 22.7% cost reduction. A broader ablation table reported an even larger cost reduction of 18.6% without multimodal fusion and 21.8% without the service hub, although the paper does not reconcile every numerical difference among its tables and figures.

Human oversight is built into the proposed control hierarchy. Routine, lower-risk actions—such as load balancing, energy allocation and maintenance scheduling—can be automated, whereas severe faults, major voltage deviations or cybersecurity alerts are routed to an operator dashboard for approval. The cybersecurity design contemplates false-data injection, replay attacks, denial of service, unauthorized commands and poisoned training data, using checks such as timestamp verification, role-based access, command authentication, anomaly filtering and cross-modal consistency. These safeguards remain architectural proposals within the evaluated framework. The study itself identifies noisy or incomplete sensor data, renewable intermittency, model generalization to larger grids and extreme conditions, cybersecurity exposure, implementation expense and regulatory standardization as unresolved limitations. Large-scale field testing with live streaming data will therefore be required before the reported benchmark gains can be treated as demonstrated utility-scale benefits.

Subject of Research: Evaluation of an autonomous, operator-controllable AI framework for multimodal smart-grid management

Article Title: An artificial intelligence based framework for smart power grid management using an autonomous and controllable intelligent service hub with multimodal document processing

Article References: Zeng, F., Zheng, H., Li, C., Rong, J., & Nong, Y. (2026). An artificial intelligence based framework for smart power grid management using an autonomous and controllable intelligent service hub with multimodal document processing. Frontiers in Energy Research, 14, 1910556. https://doi.org/10.3389/fenrg.2026.1910556

Image Credits: AI Generated

DOI: 10.3389/fenrg.2026.1910556

Keywords: smart grids, artificial intelligence, multimodal data, fault detection, predictive maintenance, load forecasting, energy optimization, renewable energy, demand response, grid cybersecurity, artificial, intelligence

Researchers collected responses from 20 April to 1 May 2026. Of 375 questionnaire submissions, 362 met the study’s quality-control criteria, producing an analytic retention rate of 96.5%. Separate item-level confirmatory factor analyses evaluated a four-factor structure of AI literacy comprising knowledge, skills, values and ethics. The model showed good fit and strong item loadings, and skills and values were highly correlated. Additional measurement analyses supported perceived usefulness, perceived ease of use and attitudes toward artificial intelligence as correlated but distinguishable constructs.

The fully adjusted multivariable regression associated three characteristics with mean AI-literacy scores: more frequent AI use, previous AI education and interest in AI. These findings show positive statistical relationships after adjustment for the variables included in the model, rather than experimental effects. The population was limited to first-year nursing students, and the geographically narrow, cross-sectional design prevents conclusions about causality, changes over time or how the pattern might differ in other provinces, countries, educational programmes or later stages of nursing training.

A regularised partial correlation network was then used to map conditional relationships among seven dimension-mean nodes: knowledge, skills, values, ethics, perceived usefulness, perceived ease of use and attitudes toward AI. Of 21 potential network edges, 17 were non-zero. Skills had the highest expected influence, while perceived usefulness had the highest strength. Bootstrap diagnostics supported strong stability of those two network measures under case removal, with a case-dropping stability coefficient of 0.751 for both. Near-zero intraclass correlation estimates indicated little clustering by institution, and institution-adjusted analyses preserved the main regression and network findings.

The ranking of one-step bridge connectivity depended on the network partition used. Values ranked first under the primary two-community partition, whereas perceived usefulness ranked first under the alternative four-community partition; perceived ease of use ranked second in both. The researchers concluded that AI-use frequency, prior AI education and interest were positively associated with AI literacy, while skills, perceived usefulness, values and perceived ease of use may help identify questions for prospective evaluation of AI education in undergraduate nursing programmes. The network rankings and regression associations describe the observed sample and do not prove that any dimension should be prioritised in a curriculum or that an educational intervention will improve literacy or technology acceptance.

Subject of Research: Cross-sectional analysis of AI literacy and technology-acceptance constructs among first-year nursing students in China

Article Title: Artificial intelligence literacy and technology acceptance model-related constructs among first-year nursing students: a multi-institutional cross-sectional network analysis

Article References: Wang, Y., Zhao, Y., Li, L., Yu, Y., Zhu, L., Ma, S., & Yang, Z. (2026). Artificial intelligence literacy and technology acceptance model-related constructs among first-year nursing students: a multi-institutional cross-sectional network analysis. BMC Nursing. https://doi.org/10.1186/s12912-026-05449-3

Image Credits: AI Generated

DOI: 10.1186/s12912-026-05449-3

Keywords: artificial intelligence literacy, nursing education, first-year students, network analysis, technology acceptance, perceived usefulness, perceived ease of use, AI skills, AI ethics, China, Artificial, intelligence