Large Language Model–Assisted Radiology Reporting: A Retrospective Cohort Study Using the UTAUT Framework to Analyze Workflow Integration and Efficiency Gains

Large Language Model–Assisted Radiology Reporting: A Retrospective Cohort Study Using the UTAUT Framework to Analyze Workflow Integration and Efficiency Gains | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Large Language Model–Assisted Radiology Reporting: A Retrospective Cohort Study Using the UTAUT Framework to Analyze Workflow Integration and Efficiency Gains Nelly Tan This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8904615/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 20 Apr, 2026 Read the published version in Abdominal Radiology → Version 1 posted 13 You are reading this latest preprint version Abstract Background Radiologist burnout affects approximately 40% of US radiologists. Large language models (LLMs) may improve workflow efficiency, but real-world implementation data are limited. Objective To evaluate LLM-assisted workflow impact on radiologist efficiency using the Unified Theory of Acceptance and Use of Technology (UTAUT) framework. Design, Setting, and Participants: HIPAA-compliant, IRB-approved retrospective cohort study of a single fellowship-trained abdominal radiologist at Mayo Clinic Arizona. We compared baseline (January–April 2024) and post-implementation (December 2025–February 2026) periods. A custom generative pre-trained transformer was developed using ChatGPT Enterprise with disease-specific templates. Main Outcome Measures: Inter-study interval time (proxy for interpretation time) compared using Wilcoxon rank-sum tests with Bonferroni correction (α = 0.0125). UTAUT constructs assessed: performance expectancy (efficiency), effort expectancy (training burden), facilitating conditions (infrastructure), and behavioral intention (satisfaction). Results We analyzed 609 studies (495 CT, 114 MRI). LLM assistance significantly reduced inter-study intervals for outpatient CT with contrast (23.0 vs 13.0 minutes; difference 10 minutes; p = 0.0021) and without contrast (18.5 vs 7.0 minutes; difference 11.5 minutes; p = 0.0017). No improvement occurred for MRI with contrast (14.0 vs 16.0 minutes; p = 0.2808) or without contrast (14.0 vs 7.0 minutes; p = 0.0889). The radiologist reported improved work-life balance for CT but neutral satisfaction for complex MRI templates. Training required 10 hours over 5 days. Conclusions LLM-assisted workflow reduced interpretation times for standardized CT studies but not heterogeneous MRI examinations, supporting UTAUT's emphasis on performance expectancy and task–technology fit as adoption drivers. Efficiency gains may reduce documentation burden when tools align with task complexity. Artificial Intelligence large language models radiology reporting burnout Figures Figure 1 Figure 2 INTRODUCTION Radiologist burnout has reached crisis levels worldwide with prevalence exceeding 75% in Germany and is ~ 40% in the United States; however, up to 83% of radiologists report at least one burnout symptom [ 1 , 2 ]. Rising imaging volumes, sustained workloads, and escalating time pressure continue to drive burnout [ 1 ]. As clinical demand grows without commensurate workforce expansion, radiology practices need tools that boost efficiency without eroding report quality. Large language models (LLMs) may help by shifting repetitive documentation away from radiologists [ 3 , 4 ]. Unlike diagnostic artificial intelligence (AI) systems that interpret images, LLMs convert dictated or written language into clinical histories, structured findings, and impressions [ 3 , 5 ]. Recent studies show that LLMs simplify reports, generate structured documentation, and reduce reporting time [ 6 – 8 ]. In a prospective study, AI-generated draft reports improved documentation efficiency and reduced time per study [ 8 ]. In another, coupling multimodal LLMs with speech recognition substantially reduced reporting time [ 7 ]. However, the relationship between AI adoption and burnout remains complex [ 1 , 9 ]. Although AI should reduce workload, emerging data link AI use to greater emotional exhaustion—especially among radiologists with high volumes or low AI acceptance [ 1 ]. By contrast, a nationwide study in China found that longer AI use correlated with lower burnout, suggesting that workflow-integrated tools may affect well-being differently than diagnostic AI [ 10 ]. Together, these mixed results underscore the need for real-world implementation studies that measure efficiency and identify contextual factors that shape technology acceptance. The Unified Theory of Acceptance and Use of Technology (UTAUT) offers a robust framework for evaluating technology adoption in healthcare [ 11 – 13 ]. UTAUT posits that four constructs—performance expectancy, effort expectancy, social influence, and facilitating conditions—shape intention to use and actual use. Meta-analyses identify performance expectancy as the strongest predictor of intention to use AI-enabled clinical decision support, followed by facilitating conditions and effort expectancy; trust also strongly predicts adoption [ 12 ]. Within radiology, investigators have applied UTAUT to evaluate picture archiving and communication system (PACS) adoption and found that radiologists prioritize usefulness (performance expectancy) over ease of use (effort expectancy) when adopting new technologies [ 14 , 15 ]. However, no studies have applied UTAUT to LLM-assisted workflows, leaving a gap as these tools rapidly evolve. We examined how an LLM-assisted workflow affects radiologist efficiency through a UTAUT lens by comparing computed tomography (CT) and magnetic resonance imaging (MRI) interpretation times before and after deploying iteratively refined AI models. We hypothesized that LLM assistance would increase performance expectancy by shortening interpretation times for both modalities, thereby supporting sustained use. MATERIALS AND METHODS Study Design and Ethical Approval This retrospective study Health Insurance Portability and Accountability Act (HIPAA) compliant, institutional review board (IRB) approved retrospective study analyzed workflow data from a single fellowship-trained abdominal radiologist at Mayo Clinic Arizona, an academic tertiary care medical center. The radiologist typically interprets 25–30 CT and 18–26 MRI studies per shift. The radiologist interpreted CT and MRI studies between January 1, 2024 – April 30, 2024 (baseline) and December 1, 2025 – February 5, 2026 (after LLM-assisted workflow implementation) for weekday inpatient/emergency department (ED) and outpatient CT or MRI shifts. Eligibility Criteria Inclusion criteria (1) Abdominal CT or MRI studies interpreted by the study radiologist; (2) weekday inpatient/ED or outpatient shifts; (3) studies from the baseline period (January 1 – April 30, 2024) or post-implementation period (December 1, 2025 – February 5, 2026). Exclusion criteria (1) Studies from the transitional period (January – November 2025), during which the radiologist was partially using AI for limited workflow tasks but not the complete LLM-assisted workflow; (2) reports initially drafted by radiology trainees; (3) whole-body MRI studies, which did not follow the standard abdominal imaging workflow. Fluoroscopy shifts and weekend on-call shifts were not included in the initial data extraction. UTAUT Framework Application We used the UTAUT framework to evaluate adoption of an LLM-assisted radiology workflow [ 11 – 13 ]. We defined performance expectancy as (1) change in interpretation time (primary outcome) and (2) perceived usefulness; effort expectancy as the training and cognitive effort required to review and edit AI-generated text; facilitating conditions as platform capabilities (ChatGPT Enterprise), disease-specific templates, and iterative prompt refinement; and behavioral intention as satisfaction and willingness to continue using LLM tools. LLM-Assisted Workflow Development and Training We used ChatGPT Enterprise (OpenAI, San Francisco, CA) to build a custom generative pretrained transformer (GPT). Throughout this manuscript, “LLM-assisted workflow” refers to clinical use of this custom GPT. We improved output quality and consistency by iteratively refining prompts and supplying exemplar reports [ 4 ]. Custom GPT configuration: We configured the custom GPT to support an abdominal radiologist. We instructed the model to: (1) choose a disease-specific template based on clinical context; (2) complete the relevant template sections; (3) verify the draft for accuracy and coherence and flag proposed edits with a brief rationale; (4) format output for PowerScribe using the Health Level Seven (HL7) V2 standard; (5) apply writing principles (use strong verbs, remove needless words, and limit nominalizations); (6) avoid symbols and combine template sections when needed; (7) standardize dates to MM/DD/YYYY and measurements to centimeters, converting millimeters when needed; and (8) report longitudinal data with the most recent examination first. We also supplied formatting examples for serial measurements and temporal trends. Clinical history generation: The radiologist copied clinical notes from the electronic health record into the custom GPT, which condensed the notes by extracting information relevant to the imaging indication and differential diagnosis. Findings and impression generation: The radiologist dictated findings into speech recognition software (PowerScribe, Nuance Communications, Burlington, MA) and pasted the dictated text into the custom GPT. The custom GPT generated structured findings and impressions using predefined templates [ 5 , 16 ] and auto-populated template fields from the dictated content. The radiologist then pasted the output back into PowerScribe to finalize and sign the report. Disease-specific templates: We developed structured templates for standard abdominopelvic CT and MR exams (e.g., rectal cancer staging, prostate MRI, pancreatic cancer evaluation, and endometriosis assessment). For general abdominal imaging, we used a standardized abdominal radiology template. ChatGPT custom GPTs support up to 20 knowledge sources; in our configuration, each source corresponded to one template (20 templates total). Although we could have created multiple custom GPTs (e.g., by modality), we used a single custom GPT for this study. Workflow integration Pre–LLM-assisted workflow The radiologist reviewed images, dictated the clinical history, findings, and impression separately using speech recognition, and manually formatted the report with institutional templates. Post–LLM-assisted workflow The radiologist reviewed images, entered clinical notes into the custom GPT to generate the history, dictated findings into speech recognition software, pasted the dictated text into the custom GPT to generate structured findings and impressions, pasted the output back into the speech recognition software, reviewed and edited the report, and finalized it. Variables Primary exposure Use of LLM-assisted workflow (yes vs no), defined as the complete workflow described above including clinical history generation, structured findings and impression generation via custom GPT, and copy-paste integration with speech recognition software. Primary outcome Inter-study interval time (minutes), defined as the interval from report finalization of one study to the next study report finalization based on timestamps in PowerScribe dictation software. This measure serves as a proxy for interpretation time, as exam open time to finalized time could not be reliably extracted from available data. Stratification variables Examination modality (CT vs MRI), patient class (outpatient vs inpatient/ED), and use of IV contrast (with vs without)." Data Sources and Measurement We extracted data from PowerScribe dictation software (Nuance Communications, Burlington, MA). Inter-study interval time was calculated automatically from report finalization timestamps. Examination modality (CT vs MRI), patient class (outpatient vs inpatient/ED), and IV contrast use were extracted from the radiology information system. LLM-assisted workflow status was determined by study period: all studies in the baseline period (January–April 2024) were classified as no LLM assistance, and all studies in the post-implementation period (December 2025–February 2026) were classified as LLM-assisted. Data extraction was performed by the study radiologist. Bias Several potential sources of bias were considered. Using a single radiologist across both periods controlled for inter-reader variability but introduced potential performance bias from learning effects or increased familiarity with the workflow over time. Data extraction was performed by the study radiologist, which may introduce information bias; however, timestamps and categorical variables were extracted directly from electronic systems, minimizing subjective interpretation. Temporal confounding from unmeasured differences between study periods (e.g., case complexity, institutional workflow changes) could not be excluded. We did not adjust for potential confounders such as time of day, day of week, or case complexity, as this was beyond the scope of this study. Study Size No formal sample size calculation was performed. This was a convenience sample comprising all eligible studies interpreted by the study radiologist during the defined baseline and post-implementation periods. The study periods were selected to capture a stable baseline before any AI integration (January–April 2024) and a period of consistent LLM-assisted workflow use (December 2025–February 2026), yielding 295 baseline and 314 post-implementation studies. Qualitative assessment The radiologist provided qualitative feedback based on personal experience and informal reflection on perceived usefulness (performance expectancy), ease of use (effort expectancy), and overall satisfaction with the LLM-assisted workflow. We also assessed perceived work–life balance, including ability to complete work within scheduled shifts and energy level after shift completion. Statistical analysis We summarized inter-study interval time as medians and interquartile ranges (IQRs) because the distribution was right-skewed. We compared groups using Wilcoxon rank-sum tests. To test whether the association between the LLM-assisted workflow and inter-study interval time differed across subgroups, we performed pre-specified stratified analyses by patient class (outpatient vs inpatient/ED), modality (CT vs MRI), and IV contrast use (with vs without). Subgroup analyses by patient class, modality, and IV contrast use were exploratory. We applied a Bonferroni correction for four primary tests (outpatient CT with contrast, outpatient CT without contrast, inpatient/ED CT with contrast, and MRI overall), setting α = 0.05/4 = 0.0125. We calculated 95% confidence intervals for median differences using bootstrap resampling with 500 replications. We excluded exams with missing inter-study interval time or missing LLM-assisted workflow status. All tests were two-sided. No multivariable adjustment for potential confounders (e.g., case complexity, time of day, fatigue) was performed; this was beyond the scope of this exploratory study. No studies were excluded for missing inter-study interval time or missing LLM-assisted workflow status during the included study periods. No formal sensitivity analyses were conducted. All tests were two-sided. We performed analyses in Stata/IC 15.1 (StataCorp, College Station, TX). RESULTS Study cohort characteristics We initially assessed 1,236 radiology reports from CT and MRI weekday shifts spanning January 2024 – April 2024, January 2025 – December 2025, and January 2026 – February 2026. See Fig. 1 Patient flow diagram for inclusion and exclusion. We analyzed 609 studies: 495 CT (81.3%) and 114 MRI (18.7%). Outpatient exams accounted for 70.6% (430/609), and inpatient/ED exams accounted for 29.4% (179/609). The radiologist used the LLM-assisted workflow for 55.3% of studies (337/609) and the non–LLM-assisted workflow for 44.7% (272/609). IV contrast was used in 79.6% of studies (485/609); 20.4% (124/609) were noncontrast. See Table 1 for study characteristics by implementation period. No studies in the final cohort had missing data for inter-study interval time, modality, patient class, IV contrast status, or LLM-assisted workflow status. Performance expectancy: inter-study interval time Across key subgroups, LLM assistance generally shortened examination reading times. Median time decreased for outpatient CT with contrast (23.0 vs 13.0 minutes; median difference 10 minutes; p = 0.0021), outpatient CT without contrast (18.5 vs 7.0 minutes; difference 11.5 minutes; p = 0.0017), and inpatient/ED CT with contrast (15.0 vs 11.0 minutes; difference 4 minutes; p = 0.0469). In contrast, outpatient MR with contrast showed no significant difference and trended longer with LLM assistance (14.0 vs 16.0 minutes; difference − 2 minutes; p = 0.2808). Outpatient MR without contrast also trended shorter with LLM assistance but did not reach statistical significance (14.0 vs 7.0 minutes; difference 7 minutes; p = 0.0889).Overall, outpatient contrast-enhanced CT inter-study intervals shifted left with LLM assistance, with shorter times for most cases and modest attenuation of the right tail (x-axis truncated at the 99th percentile), suggesting fewer prolonged reads (Table 2 ; Fig. 1). Effort expectancy: iterative training and cognitive load Iterative training required approximately 10 hours over 5 working days to optimize the custom GPT for disease-specific templates [ 4 ]. After training, the radiologist reported less effort to review and edit AI-generated text than to dictate and format CT reports manually. For MRI, however, the radiologist reported greater effort because complex templates required extensive cross-checking across multiple fields. Facilitating conditions: technical infrastructure ChatGPT Enterprise supported the LLM-assisted workflow in a HIPAA-compliant environment. Within custom GPT constraints, we integrated up to 20 disease-specific templates as knowledge sources. For nonspecific indications, the radiologist used a general abdominal template [ 5 , 16 ]. Behavioral intention: qualitative satisfaction Based on informal feedback, the radiologist reported high satisfaction with the LLM-assisted workflow for CT, citing improved ability to finish on time and sustained energy after shifts. This experience supported continued use of LLM assistance for CT interpretation. For MRI, the radiologist reported satisfaction with straightforward templates (e.g., prostate, liver, and IPMN surveillance MRI) but neutral satisfaction with the most complex templates (e.g., defecography, rectal cancer staging) because verification increased cognitive load and some findings fit poorly within the template structure. DISCUSSION This single-radiologist study at Mayo Clinic Arizona demonstrates that LLM-assisted workflow integrates differently across modalities, patient class, and use of IV contrast. Our hypothesis that LLM assistance would reduce interpretation time across both modalities was partially supported. The custom GPT was associated with improved efficiency for standardized CT studies but not for heterogeneous MRI examinations. Specifically, outpatient CT studies showed significant reductions in inter-study interval time (8-minute and 11-minute reductions for contrast-enhanced and noncontrast studies, respectively), while aggregate MRI data showed no improvement. These results reinforce UTAUT's emphasis on usefulness of the technology as ascribed by performance expectancy as a core driver of technology acceptance [ 11 , 12 ]. The 10-minute reduction in outpatient contrast-enhanced CT inter-study interval time (p = 0.0021) and 11.5-minute reduction in outpatient noncontrast CT (p = 0.0017) suggests strong performance expectancy and potential value for repetitive, standardized workflows, though these findings require confirmation in larger, multi-radiologist studies. This finding aligns with prior studies showing reduced documentation time for structured exams [ 7 , 8 ]. By contrast, the lack of significant improvement for MRI with (median difference, -2 minutes; p = 0.2808) or without IV contrast (median difference + 7 minutes, p = 0.0889) highlights a key limitation: efficiency gains depend on task–technology alignment [ 13 ]. MRI interpretation demands greater complexity and heterogeneity than CT, and the most complex disease-specific templates require cross-checking, measurement verification, and nuanced judgment that current LLMs cannot reliably automate [ 3 , 4 ]. As MRI case complexity varies widely, workflow variability may blunt consistent efficiency gains. This mismatch underscores the importance of task–technology fit for implementation success [ 11 , 13 ]. Implementation also required strong facilitating conditions, including a HIPAA-compliant enterprise platform, iterative prompt refinement, disease-specific templates, and institutional support for workflow adjustments [ 4 , 13 ]. These elements align with UTAUT’s focus on infrastructure and organizational resources as predictors of technology use [ 11 , 12 ]. Performance expectancy and facilitating conditions consistently predict clinician acceptance of AI across settings [ 17 ]. However, iterative training improved output while exposing a barrier: customization consumes time and expertise. Broader adoption may therefore require validated prompts and shareable templates that reduce early setup effort [ 4 ]. Finishing work on time with sustained energy represents a clinically meaningful outcome beyond time savings and directly targets major drivers of burnout, including workload and time pressure [ 1 ]. The favorable CT experience supported continued willingness to use LLM tools [ 12 ]. This observation contrasts with reports that diagnostic AI can worsen burnout by increasing cognitive demands [ 1 ]. In a cross-sectional study of 6,726 Chinese radiologists, AI use was associated with higher odds of burnout (OR, 1.20; 95% CI, 1.10–1.30), driven primarily by emotional exhaustion and most pronounced among radiologists with high workload or low AI acceptance [ 1 ]. Conversely, a nationwide Chinese study of 522 radiology staff found that longer AI use correlated with lower burnout (Pearson r = − 0.112; p < 0.05), supporting the value of tools that remove rather than add tasks [ 10 ]. A recent review characterized AI’s impact on radiologist burnout as a “black box,” citing conflicting and inconclusive evidence [ 9 ]. Our MRI findings similarly suggest that adoption without clear performance benefit may not improve satisfaction; tools that add complexity without offsetting value may frustrate rather than relieve. This study has limitations. A single radiologist at one institution limits generalizability across settings, experience levels, and workflows. The iterative training process was not standardized or reproducibly documented. We assessed social influence informally and did not formally measure it. We did not adjust for confounders such as case complexity, time of day, fatigue, or learning effects. We used inter study interval as a proxy for interpretation time because exam open to final time could not be reliably extracted; this metric includes between case gaps, system lag, and other noninterpretive activities. Subgroup analyses were exploratory, and we did not conduct sensitivity analyses or assess long term sustainability or potential deskilling with prolonged LLM use. Several biases are possible. The historical comparison risks temporal confounding because case mix, workflow, and radiologist experience may have changed between baseline (January to April 2024) and postimplementation (December 2025 to February 2026). Without case-complexity data, we could not assess this bias. Using interstudy interval rather than interpretation time adds noise from noninterpretive tasks and likely biases results toward the null. Given the single-radiologist design and unadjusted historical comparison, these findings are hypothesis-generating and warrant prospective, multi-radiologist studies with quality assessment. Generalizability These findings apply mainly to fellowship-trained abdominal radiologists at academic centers with HIPAA-compliant enterprise LLM platforms and support for workflow customization. Generalizability is limited by the radiologist’s prior AI and prompt-refinement experience, the abdominal imaging–specific GPT templates, manual transfer between dictation and the LLM platform, and local factors such as case mix, staffing, and reporting expectations. As a single-center U.S. academic study, results may not extend to other subspecialties, community or international practices, or health systems with different documentation requirements. Future directions Standardized prompts and shareable templates may reduce setup effort and improve reproducibility across institutions [ 4 ]. Integrating multimodal LLMs that analyze imaging data alongside text represents an important next step [ 18 , 19 ]. Comprehensive workflow solutions, rather than isolated tools, may maximize clinical value [ 20 ]. Prospective studies should evaluate patient outcomes, diagnostic accuracy, and report quality. Further work should also examine how LLM tools affect radiologist well-being and burnout, given conflicting evidence [ 1 , 9 , 10 ]. Finally, robust governance frameworks will be essential for safe deployment and ongoing monitoring [ 16 , 20 ]. CONCLUSION LLM-assisted workflow using a custom GPT was associated with efficiency gains for standardized abdominal CT studies in this single-radiologist experience but not for heterogeneous MRI exams. These findings reinforce UTAUT’s emphasis on performance expectancy as a central driver of technology acceptance. The reduction in CT reporting time, together with improved work–life balance, supports continued use of LLM tools when they add clear value. The absence of benefit for aggregate MRI data highlights the importance of task–technology fit. Current LLMs add the most value for structured, repetitive tasks and less for complex, variable workflows. With UTAUT-guided implementation, LLM-assisted workflows can reduce documentation burden and improve efficiency while preserving essential human oversight. Declarations Funding: This study received no external funding. The author had full access to all data and final responsibility for the decision to submit for publication. AI Use: OpenEvidence, ChatGPT, Copilot, and Consensus AI were used for literature review, drafting, and refining of the manuscript. These tools were not used for statistical analysis or data interpretation. Author Contribution All authors, N.T., whose names appear on the submission made substantial contributions to the conception or design of the work; or the acquisition, analysis, or interpretation of data; or the creation of new software used in the work; drafted the work or revised it critically for important intellectual content; approved the version to be published; and agree to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. Acknowledgement The author thanks Mary Khouri for her help with formatting the paper and logical aspect of the paper. References Liu H, Ding N, Li X, Chen Y, Sun H, Huang Y, et al. Artificial Intelligence and Radiologist Burnout. JAMA Netw Open 2024;7:e2448714. https://doi.org/10.1001/jamanetworkopen.2024.48714 . Hassankhani A, Amoukhteh M, Valizadeh P, Jannatdoust P, Ghadimi DJ, Sabeghi P, et al. A Meta-analysis of Burnout in Radiology Trainees and Radiologists: Insights From the Maslach Burnout Inventory. Acad Radiol 2024;31:1198–216. https://doi.org/10.1016/j.acra.2023.10.046 . Bhayana R. Chatbots and Large Language Models in Radiology: A Practical Primer for Clinical and Research Applications. Radiology 2024;310:e232756. https://doi.org/10.1148/radiol.232756 . Busch F, Hoffmann L, Dos Santos DP, Makowski MR, Saba L, Prucker P, et al. Large language models for structured reporting in radiology: past, present, and future. Eur Radiol 2024;35:2589–602. https://doi.org/10.1007/s00330-024-11107-6 . Elkassem AA, Smith AD. Potential Use Cases for ChatGPT in Radiology Reporting. Am J Roentgenol 2023;221:373–6. https://doi.org/10.2214/AJR.23.29198 . Doshi R, Amin KS, Khosla P, Bajaj S, Chheang S, Forman HP. Quantitative Evaluation of Large Language Models to Streamline Radiology Report Impressions: A Multimodal Retrospective Analysis. Radiology 2024;310:e231593. https://doi.org/10.1148/radiol.231593 . Gertz RJ, Beste NC, Dratsch T, Lennartz S, Bremm J, Iuga A-I, et al. From dictation to diagnosis: enhancing radiology reporting with integrated speech recognition in multimodal large language models. Eur Radiol 2025. https://doi.org/10.1007/s00330-025-11929-y . Huang J, Wittbrodt MT, Teague CN, Karl E, Galal G, Thompson M, et al. Efficiency and Quality of Generative AI–Assisted Radiograph Reporting. JAMA Netw Open 2025;8:e2513921. https://doi.org/10.1001/jamanetworkopen.2025.13921 . Parikh JR, Lexa FJ. Radiologist burnout: AI’s true black box. Eur Radiol 2026. https://doi.org/10.1007/s00330-025-12278-6 . Fang X, Ma C, Liu X, Deng X, Liao J, Zhang T. Burnout crisis in Chinese radiology: will artificial intelligence help? Eur Radiol 2024;35:1215–24. https://doi.org/10.1007/s00330-024-11206-4 . Ammenwerth Elske. Technology Acceptance Models in Health Informatics: TAM and UTAUT. Stud. Health Technol. Inform., IOS Press; 2019. https://doi.org/10.3233/SHTI190111 . Dingel J, Kleine A-K, Cecil J, Sigl AL, Lermer E, Gaube S. Predictors of Health Care Practitioners’ Intention to Use AI-Enabled Clinical Decision Support Systems: Meta-Analysis Based on the Unified Theory of Acceptance and Use of Technology. J Med Internet Res 2024;26:e57224. https://doi.org/10.2196/57224 . Thanthrige A, Lu B, Sako Z, Wickramasinghe N. Determinants of Health Care Technology Adoption Using an Integrated Unified Theory of Acceptance and Use of Technology and Task Technology Fit Model: Systematic Review and Meta-Analysis. J Med Internet Res 2025;27:e64524–e64524. https://doi.org/10.2196/64524 . Duyck P, Pynoo B, Devolder P, Voet T, Adang L, Vercruysse J. User acceptance of a picture archiving and communication system. Applying the unified theory of acceptance and use of technology in a radiological setting. Methods Inf Med 2008;47:149–56. Duyck P, Pynoo B, Devolder P, Voet T, Adang L, Ovaere D, et al. Monitoring the PACS Implementation Process in a Large University Hospital—Discrepancies Between Radiologists and Physicians. J Digit Imaging 2010;23:73–80. https://doi.org/10.1007/s10278-008-9163-7 . Bhayana R, Biswas S, Cook TS, Kim W, Kitamura FC, Gichoya J, et al. From Bench to Bedside With Large Language Models: AJR Expert Panel Narrative Review. Am J Roentgenol 2024;223:e2430928. https://doi.org/10.2214/AJR.24.30928 . Scipion CEA, Manchester MA, Federman A, Wang Y, Arias JJ. Barriers to and facilitators of clinician acceptance and use of artificial intelligence in healthcare settings: a scoping review. BMJ Open 2025;15:e092624. https://doi.org/10.1136/bmjopen-2024-092624 . Nam Y, Kim DY, Kyung S, Seo J, Song JM, Kwon J, et al. Multimodal Large Language Models in Medical Imaging: Current State and Future Directions. Korean J Radiol 2025;26:900. https://doi.org/10.3348/kjr.2025.0599 . Seah JCY, Tang JSN, Tran A. Drafting the Future: The Dawn of AI Report Generation in Radiology. Radiology 2025;316:e243378. https://doi.org/10.1148/radiol.243378 . Wichtmann BD, Paech D, Pianykh OS, Huang SY, Seltzer SE, Brink J, et al. Leadership in radiology in the era of technological advancements and artificial intelligence. Eur Radiol 2025;36:548–52. https://doi.org/10.1007/s00330-025-11745-4 . Tables Table 1. Study cohort characteristics by implementation period Values are n (% within period). Characteristic Baseline (n=295) Post-implementation (n=314) Studies 295 (100.0%) 314 (100.0%) CT 236 (80.0%) 259 (82.5%) MR 59 (20.0%) 55 (17.5%) Inpatient/ED 81 (27.5%) 98 (31.2%) Outpatient 214 (72.5%) 216 (68.8%) No contrast 58 (19.7%) 66 (21.0%) With contrast 237 (80.3%) 248 (79.0%) Table 2. Unadjusted time to read examinations (minutes) stratified by key subgroups and LLM assistance Subgroup No LLM-assisted, n LLM-assisted, n Total, n No LLM-assisted, median (IQR), min LLM-assisted, median (IQR), min Median difference (No − LLM-assisted), min 95% CI (bootstrap) Wilcoxon rank-sum p Outpatient–CT–With contrast 135 130 265 23.0 (11.0–35.0) 13.0 (10.0–26.0) 10 [-0.28, 10.28] 0.0021 Outpatient–CT–No contrast 28 38 66 18.5 (8.0–44.0) 7.0 (6.0–11.0) 11.5 [-7.72, 17.72] 0.0017 Outpatient–MR–With contrast 47 40 87 14.0 (10.0–24.0) 16.0 (13.0–23.5) -2 [0.33, 9.67] 0.2808 Outpatient–MR–No contrast 4 8 12 14.0 (12.0–42.5) 7.0 (4.5–31.0) 7 [-35.15, 45.15] 0.0889 Inpatient/ED–CT–With contrast 47 72 119 15.0 (9.0–29.0) 11.0 (8.0–20.5) 4 [-3.64, 13.64] 0.0469 Footnote: Values are median (IQR). Median difference is calculated as No LLM-assisted − LLM-assisted (positive values indicate shorter reading times with LLM assistance). 95% CIs for median differences were estimated using nonparametric bootstrap (500 replications). P-values are from two-sided Wilcoxon rank-sum tests. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 20 Apr, 2026 Read the published version in Abdominal Radiology → Version 1 posted Editorial decision: Revision requested 21 Feb, 2026 Reviews received at journal 20 Feb, 2026 Reviews received at journal 19 Feb, 2026 Reviewers agreed at journal 18 Feb, 2026 Reviews received at journal 18 Feb, 2026 Reviewers agreed at journal 18 Feb, 2026 Reviewers agreed at journal 18 Feb, 2026 Reviewers agreed at journal 18 Feb, 2026 Reviewers agreed at journal 18 Feb, 2026 Reviewers invited by journal 18 Feb, 2026 Editor assigned by journal 18 Feb, 2026 Submission checks completed at journal 18 Feb, 2026 First submitted to journal 17 Feb, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8904615","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":593900179,"identity":"8f23e65d-a3de-445f-8b16-114ece0ad3b3","order_by":0,"name":"Nelly Tan","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA1UlEQVRIiWNgGAWjYPACCxkG9gYo+wABtTwQSoKHgQemlHgtEglEarFnP3uAuaBCgsfg5uvEzzwVDHJ8NxLwa+HhyUtgnnEGqOV27mZpnjMMxpIEtTDkGDDztoG1bJDmbWNI3EBQC/8boJZ/IIed3fwbqKWesBYJkC0NQC03eLeBbEkwIKjlxhuDwzzHJHgkz+Rus5xzRsJw5pkH+LWw9+cYPuapsZHjO3528403FTbyfMcJ2AICB0CEApBkAsYOYeVwIN/AwMD4gwQNo2AUjIJRMHIAAJl4P2nOeqc9AAAAAElFTkSuQmCC","orcid":"","institution":"Mayo Clinic Arizona","correspondingAuthor":true,"prefix":"","firstName":"Nelly","middleName":"","lastName":"Tan","suffix":""}],"badges":[],"createdAt":"2026-02-17 22:53:17","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8904615/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8904615/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1007/s00261-026-05524-y","type":"published","date":"2026-04-20T15:58:47+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":103178855,"identity":"5255ed71-ba28-4a5e-a7be-71bdb7fb8455","added_by":"auto","created_at":"2026-02-22 17:07:01","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":38496,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eStudy flow diagram\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-8904615/v1/2a067adbf1ce16f2340976e2.png"},{"id":103178856,"identity":"01860438-df31-4700-83e1-9858ad974a3c","added_by":"auto","created_at":"2026-02-22 17:07:01","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":134119,"visible":true,"origin":"","legend":"\u003cp\u003eKernel density estimates of exam reading time (minutes) for outpatient contrast-enhanced CT studies, comparing With Model and No Model. The curve with the model shifts left relative to no model, indicating shorter reading times across most cases and a modest reduction in the right tail. The x‑axis is truncated at the 99th percentile to limit the influence of extreme values.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8904615/v1/a2f277d79afad8fa95a4b866.png"},{"id":107927873,"identity":"66d546d9-f8d0-444c-b566-769b31f41e53","added_by":"auto","created_at":"2026-04-27 16:05:43","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":377652,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8904615/v1/1629dd11-360d-4499-807c-6177cd66c383.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Large Language Model–Assisted Radiology Reporting: A Retrospective Cohort Study Using the UTAUT Framework to Analyze Workflow Integration and Efficiency Gains","fulltext":[{"header":"INTRODUCTION","content":"\u003cp\u003eRadiologist burnout has reached crisis levels worldwide with prevalence exceeding 75% in Germany and is ~\u0026thinsp;40% in the United States; however, up to 83% of radiologists report at least one burnout symptom [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Rising imaging volumes, sustained workloads, and escalating time pressure continue to drive burnout [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. As clinical demand grows without commensurate workforce expansion, radiology practices need tools that boost efficiency without eroding report quality. Large language models (LLMs) may help by shifting repetitive documentation away from radiologists [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Unlike diagnostic artificial intelligence (AI) systems that interpret images, LLMs convert dictated or written language into clinical histories, structured findings, and impressions [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Recent studies show that LLMs simplify reports, generate structured documentation, and reduce reporting time [\u003cspan additionalcitationids=\"CR7\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. In a prospective study, AI-generated draft reports improved documentation efficiency and reduced time per study [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. In another, coupling multimodal LLMs with speech recognition substantially reduced reporting time [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eHowever, the relationship between AI adoption and burnout remains complex [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Although AI should reduce workload, emerging data link AI use to greater emotional exhaustion\u0026mdash;especially among radiologists with high volumes or low AI acceptance [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. By contrast, a nationwide study in China found that longer AI use correlated with lower burnout, suggesting that workflow-integrated tools may affect well-being differently than diagnostic AI [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Together, these mixed results underscore the need for real-world implementation studies that measure efficiency and identify contextual factors that shape technology acceptance.\u003c/p\u003e \u003cp\u003eThe Unified Theory of Acceptance and Use of Technology (UTAUT) offers a robust framework for evaluating technology adoption in healthcare [\u003cspan additionalcitationids=\"CR12\" citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. UTAUT posits that four constructs\u0026mdash;performance expectancy, effort expectancy, social influence, and facilitating conditions\u0026mdash;shape intention to use and actual use. Meta-analyses identify performance expectancy as the strongest predictor of intention to use AI-enabled clinical decision support, followed by facilitating conditions and effort expectancy; trust also strongly predicts adoption [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eWithin radiology, investigators have applied UTAUT to evaluate picture archiving and communication system (PACS) adoption and found that radiologists prioritize usefulness (performance expectancy) over ease of use (effort expectancy) when adopting new technologies [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. However, no studies have applied UTAUT to LLM-assisted workflows, leaving a gap as these tools rapidly evolve. We examined how an LLM-assisted workflow affects radiologist efficiency through a UTAUT lens by comparing computed tomography (CT) and magnetic resonance imaging (MRI) interpretation times before and after deploying iteratively refined AI models. We hypothesized that LLM assistance would increase performance expectancy by shortening interpretation times for both modalities, thereby supporting sustained use.\u003c/p\u003e"},{"header":"MATERIALS AND METHODS","content":"\u003cp\u003eStudy Design and Ethical Approval\u003c/p\u003e \u003cp\u003eThis retrospective study Health Insurance Portability and Accountability Act (HIPAA) compliant, institutional review board (IRB) approved retrospective study analyzed workflow data from a single fellowship-trained abdominal radiologist at Mayo Clinic Arizona, an academic tertiary care medical center. The radiologist typically interprets 25\u0026ndash;30 CT and 18\u0026ndash;26 MRI studies per shift. The radiologist interpreted CT and MRI studies between January 1, 2024 \u0026ndash; April 30, 2024 (baseline) and December 1, 2025 \u0026ndash; February 5, 2026 (after LLM-assisted workflow implementation) for weekday inpatient/emergency department (ED) and outpatient CT or MRI shifts.\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eEligibility Criteria\u003c/h2\u003e \u003cp\u003e \u003cstrong\u003eInclusion criteria\u003c/strong\u003e \u003cp\u003e(1) Abdominal CT or MRI studies interpreted by the study radiologist; (2) weekday inpatient/ED or outpatient shifts; (3) studies from the baseline period (January 1 \u0026ndash; April 30, 2024) or post-implementation period (December 1, 2025 \u0026ndash; February 5, 2026).\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eExclusion criteria\u003c/strong\u003e \u003cp\u003e(1) Studies from the transitional period (January \u0026ndash; November 2025), during which the radiologist was partially using AI for limited workflow tasks but not the complete LLM-assisted workflow; (2) reports initially drafted by radiology trainees; (3) whole-body MRI studies, which did not follow the standard abdominal imaging workflow. Fluoroscopy shifts and weekend on-call shifts were not included in the initial data extraction.\u003c/p\u003e \u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eUTAUT Framework Application\u003c/h3\u003e\n\u003cp\u003eWe used the UTAUT framework to evaluate adoption of an LLM-assisted radiology workflow [\u003cspan additionalcitationids=\"CR12\" citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. We defined performance expectancy as (1) change in interpretation time (primary outcome) and (2) perceived usefulness; effort expectancy as the training and cognitive effort required to review and edit AI-generated text; facilitating conditions as platform capabilities (ChatGPT Enterprise), disease-specific templates, and iterative prompt refinement; and behavioral intention as satisfaction and willingness to continue using LLM tools.\u003c/p\u003e\n\u003ch3\u003eLLM-Assisted Workflow Development and Training\u003c/h3\u003e\n\u003cp\u003eWe used ChatGPT Enterprise (OpenAI, San Francisco, CA) to build a custom generative pretrained transformer (GPT). Throughout this manuscript, \u0026ldquo;LLM-assisted workflow\u0026rdquo; refers to clinical use of this custom GPT. We improved output quality and consistency by iteratively refining prompts and supplying exemplar reports [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e].\u003c/p\u003e\n\u003ch3\u003eCustom GPT configuration:\u003c/h3\u003e\n\u003cp\u003eWe configured the custom GPT to support an abdominal radiologist. We instructed the model to: (1) choose a disease-specific template based on clinical context; (2) complete the relevant template sections; (3) verify the draft for accuracy and coherence and flag proposed edits with a brief rationale; (4) format output for PowerScribe using the Health Level Seven (HL7) V2 standard; (5) apply writing principles (use strong verbs, remove needless words, and limit nominalizations); (6) avoid symbols and combine template sections when needed; (7) standardize dates to MM/DD/YYYY and measurements to centimeters, converting millimeters when needed; and (8) report longitudinal data with the most recent examination first. We also supplied formatting examples for serial measurements and temporal trends.\u003c/p\u003e\n\u003ch3\u003eClinical history generation:\u003c/h3\u003e\n\u003cp\u003eThe radiologist copied clinical notes from the electronic health record into the custom GPT, which condensed the notes by extracting information relevant to the imaging indication and differential diagnosis.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eFindings and impression generation:\u003c/h2\u003e \u003cp\u003eThe radiologist dictated findings into speech recognition software (PowerScribe, Nuance Communications, Burlington, MA) and pasted the dictated text into the custom GPT. The custom GPT generated structured findings and impressions using predefined templates [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e] and auto-populated template fields from the dictated content. The radiologist then pasted the output back into PowerScribe to finalize and sign the report.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eDisease-specific templates:\u003c/h3\u003e\n\u003cp\u003eWe developed structured templates for standard abdominopelvic CT and MR exams (e.g., rectal cancer staging, prostate MRI, pancreatic cancer evaluation, and endometriosis assessment). For general abdominal imaging, we used a standardized abdominal radiology template. ChatGPT custom GPTs support up to 20 knowledge sources; in our configuration, each source corresponded to one template (20 templates total). Although we could have created multiple custom GPTs (e.g., by modality), we used a single custom GPT for this study.\u003c/p\u003e\n\u003ch3\u003eWorkflow integration\u003c/h3\u003e\n\u003cp\u003e \u003cstrong\u003ePre\u0026ndash;LLM-assisted workflow\u003c/strong\u003e \u003cp\u003eThe radiologist reviewed images, dictated the clinical history, findings, and impression separately using speech recognition, and manually formatted the report with institutional templates.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003ePost\u0026ndash;LLM-assisted workflow\u003c/strong\u003e \u003cp\u003eThe radiologist reviewed images, entered clinical notes into the custom GPT to generate the history, dictated findings into speech recognition software, pasted the dictated text into the custom GPT to generate structured findings and impressions, pasted the output back into the speech recognition software, reviewed and edited the report, and finalized it.\u003c/p\u003e \u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eVariables\u003c/h2\u003e \u003cp\u003e \u003cstrong\u003ePrimary exposure\u003c/strong\u003e \u003cp\u003eUse of LLM-assisted workflow (yes vs no), defined as the complete workflow described above including clinical history generation, structured findings and impression generation via custom GPT, and copy-paste integration with speech recognition software.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003ePrimary outcome\u003c/strong\u003e \u003cp\u003eInter-study interval time (minutes), defined as the interval from report finalization of one study to the next study report finalization based on timestamps in PowerScribe dictation software. This measure serves as a proxy for interpretation time, as exam open time to finalized time could not be reliably extracted from available data.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eStratification variables\u003c/strong\u003e \u003cp\u003eExamination modality (CT vs MRI), patient class (outpatient vs inpatient/ED), and use of IV contrast (with vs without).\"\u003c/p\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eData Sources and Measurement\u003c/h2\u003e \u003cp\u003eWe extracted data from PowerScribe dictation software (Nuance Communications, Burlington, MA). Inter-study interval time was calculated automatically from report finalization timestamps. Examination modality (CT vs MRI), patient class (outpatient vs inpatient/ED), and IV contrast use were extracted from the radiology information system. LLM-assisted workflow status was determined by study period: all studies in the baseline period (January\u0026ndash;April 2024) were classified as no LLM assistance, and all studies in the post-implementation period (December 2025\u0026ndash;February 2026) were classified as LLM-assisted. Data extraction was performed by the study radiologist.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eBias\u003c/h2\u003e \u003cp\u003eSeveral potential sources of bias were considered. Using a single radiologist across both periods controlled for inter-reader variability but introduced potential performance bias from learning effects or increased familiarity with the workflow over time. Data extraction was performed by the study radiologist, which may introduce information bias; however, timestamps and categorical variables were extracted directly from electronic systems, minimizing subjective interpretation. Temporal confounding from unmeasured differences between study periods (e.g., case complexity, institutional workflow changes) could not be excluded. We did not adjust for potential confounders such as time of day, day of week, or case complexity, as this was beyond the scope of this study.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eStudy Size\u003c/h2\u003e \u003cp\u003eNo formal sample size calculation was performed. This was a convenience sample comprising all eligible studies interpreted by the study radiologist during the defined baseline and post-implementation periods. The study periods were selected to capture a stable baseline before any AI integration (January\u0026ndash;April 2024) and a period of consistent LLM-assisted workflow use (December 2025\u0026ndash;February 2026), yielding 295 baseline and 314 post-implementation studies.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eQualitative assessment\u003c/h2\u003e \u003cp\u003eThe radiologist provided qualitative feedback based on personal experience and informal reflection on perceived usefulness (performance expectancy), ease of use (effort expectancy), and overall satisfaction with the LLM-assisted workflow. We also assessed perceived work\u0026ndash;life balance, including ability to complete work within scheduled shifts and energy level after shift completion.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eStatistical analysis\u003c/h2\u003e \u003cp\u003eWe summarized inter-study interval time as medians and interquartile ranges (IQRs) because the distribution was right-skewed. We compared groups using Wilcoxon rank-sum tests. To test whether the association between the LLM-assisted workflow and inter-study interval time differed across subgroups, we performed pre-specified stratified analyses by patient class (outpatient vs inpatient/ED), modality (CT vs MRI), and IV contrast use (with vs without). Subgroup analyses by patient class, modality, and IV contrast use were exploratory. We applied a Bonferroni correction for four primary tests (outpatient CT with contrast, outpatient CT without contrast, inpatient/ED CT with contrast, and MRI overall), setting α\u0026thinsp;=\u0026thinsp;0.05/4\u0026thinsp;=\u0026thinsp;0.0125. We calculated 95% confidence intervals for median differences using bootstrap resampling with 500 replications. We excluded exams with missing inter-study interval time or missing LLM-assisted workflow status. All tests were two-sided. No multivariable adjustment for potential confounders (e.g., case complexity, time of day, fatigue) was performed; this was beyond the scope of this exploratory study. No studies were excluded for missing inter-study interval time or missing LLM-assisted workflow status during the included study periods. No formal sensitivity analyses were conducted. All tests were two-sided. We performed analyses in Stata/IC 15.1 (StataCorp, College Station, TX).\u003c/p\u003e \u003c/div\u003e"},{"header":"RESULTS","content":"\u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003eStudy cohort characteristics\u003c/h2\u003e \u003cp\u003eWe initially assessed 1,236 radiology reports from CT and MRI weekday shifts spanning January 2024 \u0026ndash; April 2024, January 2025 \u0026ndash; December 2025, and January 2026 \u0026ndash; February 2026. See Fig.\u0026nbsp;1 Patient flow diagram for inclusion and exclusion. We analyzed 609 studies: 495 CT (81.3%) and 114 MRI (18.7%). Outpatient exams accounted for 70.6% (430/609), and inpatient/ED exams accounted for 29.4% (179/609). The radiologist used the LLM-assisted workflow for 55.3% of studies (337/609) and the non\u0026ndash;LLM-assisted workflow for 44.7% (272/609). IV contrast was used in 79.6% of studies (485/609); 20.4% (124/609) were noncontrast. See Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e for study characteristics by implementation period. No studies in the final cohort had missing data for inter-study interval time, modality, patient class, IV contrast status, or LLM-assisted workflow status.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003ePerformance expectancy: inter-study interval time\u003c/h2\u003e \u003cp\u003eAcross key subgroups, LLM assistance generally shortened examination reading times. Median time decreased for outpatient CT with contrast (23.0 vs 13.0 minutes; median difference 10 minutes; p\u0026thinsp;=\u0026thinsp;0.0021), outpatient CT without contrast (18.5 vs 7.0 minutes; difference 11.5 minutes; p\u0026thinsp;=\u0026thinsp;0.0017), and inpatient/ED CT with contrast (15.0 vs 11.0 minutes; difference 4 minutes; p\u0026thinsp;=\u0026thinsp;0.0469). In contrast, outpatient MR with contrast showed no significant difference and trended longer with LLM assistance (14.0 vs 16.0 minutes; difference\u0026thinsp;\u0026minus;\u0026thinsp;2 minutes; p\u0026thinsp;=\u0026thinsp;0.2808). Outpatient MR without contrast also trended shorter with LLM assistance but did not reach statistical significance (14.0 vs 7.0 minutes; difference 7 minutes; p\u0026thinsp;=\u0026thinsp;0.0889).Overall, outpatient contrast-enhanced CT inter-study intervals shifted left with LLM assistance, with shorter times for most cases and modest attenuation of the right tail (x-axis truncated at the 99th percentile), suggesting fewer prolonged reads (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e; Fig.\u0026nbsp;1).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003eEffort expectancy: iterative training and cognitive load\u003c/h2\u003e \u003cp\u003eIterative training required approximately 10 hours over 5 working days to optimize the custom GPT for disease-specific templates [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. After training, the radiologist reported less effort to review and edit AI-generated text than to dictate and format CT reports manually. For MRI, however, the radiologist reported greater effort because complex templates required extensive cross-checking across multiple fields.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003eFacilitating conditions: technical infrastructure\u003c/h2\u003e \u003cp\u003eChatGPT Enterprise supported the LLM-assisted workflow in a HIPAA-compliant environment. Within custom GPT constraints, we integrated up to 20 disease-specific templates as knowledge sources. For nonspecific indications, the radiologist used a general abdominal template [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section2\"\u003e \u003ch2\u003eBehavioral intention: qualitative satisfaction\u003c/h2\u003e \u003cp\u003eBased on informal feedback, the radiologist reported high satisfaction with the LLM-assisted workflow for CT, citing improved ability to finish on time and sustained energy after shifts. This experience supported continued use of LLM assistance for CT interpretation. For MRI, the radiologist reported satisfaction with straightforward templates (e.g., prostate, liver, and IPMN surveillance MRI) but neutral satisfaction with the most complex templates (e.g., defecography, rectal cancer staging) because verification increased cognitive load and some findings fit poorly within the template structure.\u003c/p\u003e \u003c/div\u003e"},{"header":"DISCUSSION","content":"\u003cp\u003eThis single-radiologist study at Mayo Clinic Arizona demonstrates that LLM-assisted workflow integrates differently across modalities, patient class, and use of IV contrast. Our hypothesis that LLM assistance would reduce interpretation time across both modalities was partially supported. The custom GPT was associated with improved efficiency for standardized CT studies but not for heterogeneous MRI examinations. Specifically, outpatient CT studies showed significant reductions in inter-study interval time (8-minute and 11-minute reductions for contrast-enhanced and noncontrast studies, respectively), while aggregate MRI data showed no improvement. These results reinforce UTAUT's emphasis on usefulness of the technology as ascribed by performance expectancy as a core driver of technology acceptance [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe 10-minute reduction in outpatient contrast-enhanced CT inter-study interval time (p\u0026thinsp;=\u0026thinsp;0.0021) and 11.5-minute reduction in outpatient noncontrast CT (p\u0026thinsp;=\u0026thinsp;0.0017) suggests strong performance expectancy and potential value for repetitive, standardized workflows, though these findings require confirmation in larger, multi-radiologist studies. This finding aligns with prior studies showing reduced documentation time for structured exams [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. By contrast, the lack of significant improvement for MRI with (median difference, -2 minutes; p\u0026thinsp;=\u0026thinsp;0.2808) or without IV contrast (median difference\u0026thinsp;+\u0026thinsp;7 minutes, p\u0026thinsp;=\u0026thinsp;0.0889) highlights a key limitation: efficiency gains depend on task\u0026ndash;technology alignment [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. MRI interpretation demands greater complexity and heterogeneity than CT, and the most complex disease-specific templates require cross-checking, measurement verification, and nuanced judgment that current LLMs cannot reliably automate [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. As MRI case complexity varies widely, workflow variability may blunt consistent efficiency gains. This mismatch underscores the importance of task\u0026ndash;technology fit for implementation success [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eImplementation also required strong facilitating conditions, including a HIPAA-compliant enterprise platform, iterative prompt refinement, disease-specific templates, and institutional support for workflow adjustments [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. These elements align with UTAUT\u0026rsquo;s focus on infrastructure and organizational resources as predictors of technology use [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Performance expectancy and facilitating conditions consistently predict clinician acceptance of AI across settings [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. However, iterative training improved output while exposing a barrier: customization consumes time and expertise. Broader adoption may therefore require validated prompts and shareable templates that reduce early setup effort [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eFinishing work on time with sustained energy represents a clinically meaningful outcome beyond time savings and directly targets major drivers of burnout, including workload and time pressure [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. The favorable CT experience supported continued willingness to use LLM tools [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. This observation contrasts with reports that diagnostic AI can worsen burnout by increasing cognitive demands [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. In a cross-sectional study of 6,726 Chinese radiologists, AI use was associated with higher odds of burnout (OR, 1.20; 95% CI, 1.10\u0026ndash;1.30), driven primarily by emotional exhaustion and most pronounced among radiologists with high workload or low AI acceptance [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Conversely, a nationwide Chinese study of 522 radiology staff found that longer AI use correlated with lower burnout (Pearson r\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;0.112; p\u0026thinsp;\u0026lt;\u0026thinsp;0.05), supporting the value of tools that remove rather than add tasks [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. A recent review characterized AI\u0026rsquo;s impact on radiologist burnout as a \u0026ldquo;black box,\u0026rdquo; citing conflicting and inconclusive evidence [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Our MRI findings similarly suggest that adoption without clear performance benefit may not improve satisfaction; tools that add complexity without offsetting value may frustrate rather than relieve.\u003c/p\u003e \u003cp\u003eThis study has limitations. A single radiologist at one institution limits generalizability across settings, experience levels, and workflows. The iterative training process was not standardized or reproducibly documented. We assessed social influence informally and did not formally measure it. We did not adjust for confounders such as case complexity, time of day, fatigue, or learning effects. We used inter study interval as a proxy for interpretation time because exam open to final time could not be reliably extracted; this metric includes between case gaps, system lag, and other noninterpretive activities. Subgroup analyses were exploratory, and we did not conduct sensitivity analyses or assess long term sustainability or potential deskilling with prolonged LLM use.\u003c/p\u003e \u003cp\u003eSeveral biases are possible. The historical comparison risks temporal confounding because case mix, workflow, and radiologist experience may have changed between baseline (January to April 2024) and postimplementation (December 2025 to February 2026). Without case-complexity data, we could not assess this bias. Using interstudy interval rather than interpretation time adds noise from noninterpretive tasks and likely biases results toward the null. Given the single-radiologist design and unadjusted historical comparison, these findings are hypothesis-generating and warrant prospective, multi-radiologist studies with quality assessment.\u003c/p\u003e \u003cdiv id=\"Sec24\" class=\"Section2\"\u003e \u003ch2\u003eGeneralizability\u003c/h2\u003e \u003cp\u003eThese findings apply mainly to fellowship-trained abdominal radiologists at academic centers with HIPAA-compliant enterprise LLM platforms and support for workflow customization. Generalizability is limited by the radiologist\u0026rsquo;s prior AI and prompt-refinement experience, the abdominal imaging\u0026ndash;specific GPT templates, manual transfer between dictation and the LLM platform, and local factors such as case mix, staffing, and reporting expectations. As a single-center U.S. academic study, results may not extend to other subspecialties, community or international practices, or health systems with different documentation requirements.\u003c/p\u003e \u003cdiv id=\"Sec25\" class=\"Section3\"\u003e \u003ch2\u003eFuture directions\u003c/h2\u003e \u003cp\u003eStandardized prompts and shareable templates may reduce setup effort and improve reproducibility across institutions [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Integrating multimodal LLMs that analyze imaging data alongside text represents an important next step [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. Comprehensive workflow solutions, rather than isolated tools, may maximize clinical value [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. Prospective studies should evaluate patient outcomes, diagnostic accuracy, and report quality. Further work should also examine how LLM tools affect radiologist well-being and burnout, given conflicting evidence [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Finally, robust governance frameworks will be essential for safe deployment and ongoing monitoring [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"CONCLUSION","content":"\u003cp\u003eLLM-assisted workflow using a custom GPT was associated with efficiency gains for standardized abdominal CT studies in this single-radiologist experience but not for heterogeneous MRI exams. These findings reinforce UTAUT’s emphasis on performance expectancy as a central driver of technology acceptance. The reduction in CT reporting time, together with improved work–life balance, supports continued use of LLM tools when they add clear value. The absence of benefit for aggregate MRI data highlights the importance of task–technology fit. Current LLMs add the most value for structured, repetitive tasks and less for complex, variable workflows. With UTAUT-guided implementation, LLM-assisted workflows can reduce documentation burden and improve efficiency while preserving essential human oversight.\u003c/p\u003e\u003c/div\u003e"},{"header":"Declarations","content":"\u003ch2\u003eFunding:\u003c/h2\u003e \u003cp\u003eThis study received no external funding. The author had full access to all data and final responsibility for the decision to submit for publication.\u003c/p\u003e \u003cp\u003eAI Use: OpenEvidence, ChatGPT, Copilot, and Consensus AI were used for literature review, drafting, and refining of the manuscript. These tools were not used for statistical analysis or data interpretation.\u003c/p\u003e \u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eAll authors, N.T., whose names appear on the submission made substantial contributions to the conception or design of the work; or the acquisition, analysis, or interpretation of data; or the creation of new software used in the work; drafted the work or revised it critically for important intellectual content; approved the version to be published; and agree to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e\u003cp\u003eThe author thanks Mary Khouri for her help with formatting the paper and logical aspect of the paper.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eLiu H, Ding N, Li X, Chen Y, Sun H, Huang Y, et al. Artificial Intelligence and Radiologist Burnout. JAMA Netw Open 2024;7:e2448714. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1001/jamanetworkopen.2024.48714\u003c/span\u003e\u003cspan address=\"10.1001/jamanetworkopen.2024.48714\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHassankhani A, Amoukhteh M, Valizadeh P, Jannatdoust P, Ghadimi DJ, Sabeghi P, et al. A Meta-analysis of Burnout in Radiology Trainees and Radiologists: Insights From the Maslach Burnout Inventory. Acad Radiol 2024;31:1198\u0026ndash;216. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.acra.2023.10.046\u003c/span\u003e\u003cspan address=\"10.1016/j.acra.2023.10.046\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBhayana R. Chatbots and Large Language Models in Radiology: A Practical Primer for Clinical and Research Applications. Radiology 2024;310:e232756. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1148/radiol.232756\u003c/span\u003e\u003cspan address=\"10.1148/radiol.232756\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBusch F, Hoffmann L, Dos Santos DP, Makowski MR, Saba L, Prucker P, et al. Large language models for structured reporting in radiology: past, present, and future. Eur Radiol 2024;35:2589\u0026ndash;602. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s00330-024-11107-6\u003c/span\u003e\u003cspan address=\"10.1007/s00330-024-11107-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eElkassem AA, Smith AD. Potential Use Cases for ChatGPT in Radiology Reporting. Am J Roentgenol 2023;221:373\u0026ndash;6. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.2214/AJR.23.29198\u003c/span\u003e\u003cspan address=\"10.2214/AJR.23.29198\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDoshi R, Amin KS, Khosla P, Bajaj S, Chheang S, Forman HP. Quantitative Evaluation of Large Language Models to Streamline Radiology Report Impressions: A Multimodal Retrospective Analysis. Radiology 2024;310:e231593. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1148/radiol.231593\u003c/span\u003e\u003cspan address=\"10.1148/radiol.231593\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGertz RJ, Beste NC, Dratsch T, Lennartz S, Bremm J, Iuga A-I, et al. From dictation to diagnosis: enhancing radiology reporting with integrated speech recognition in multimodal large language models. Eur Radiol 2025. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s00330-025-11929-y\u003c/span\u003e\u003cspan address=\"10.1007/s00330-025-11929-y\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang J, Wittbrodt MT, Teague CN, Karl E, Galal G, Thompson M, et al. Efficiency and Quality of Generative AI\u0026ndash;Assisted Radiograph Reporting. JAMA Netw Open 2025;8:e2513921. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1001/jamanetworkopen.2025.13921\u003c/span\u003e\u003cspan address=\"10.1001/jamanetworkopen.2025.13921\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eParikh JR, Lexa FJ. Radiologist burnout: AI\u0026rsquo;s true black box. Eur Radiol 2026. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s00330-025-12278-6\u003c/span\u003e\u003cspan address=\"10.1007/s00330-025-12278-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFang X, Ma C, Liu X, Deng X, Liao J, Zhang T. Burnout crisis in Chinese radiology: will artificial intelligence help? Eur Radiol 2024;35:1215\u0026ndash;24. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s00330-024-11206-4\u003c/span\u003e\u003cspan address=\"10.1007/s00330-024-11206-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAmmenwerth Elske. Technology Acceptance Models in Health Informatics: TAM and UTAUT. Stud. Health Technol. Inform., IOS Press; 2019. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3233/SHTI190111\u003c/span\u003e\u003cspan address=\"10.3233/SHTI190111\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDingel J, Kleine A-K, Cecil J, Sigl AL, Lermer E, Gaube S. Predictors of Health Care Practitioners\u0026rsquo; Intention to Use AI-Enabled Clinical Decision Support Systems: Meta-Analysis Based on the Unified Theory of Acceptance and Use of Technology. J Med Internet Res 2024;26:e57224. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.2196/57224\u003c/span\u003e\u003cspan address=\"10.2196/57224\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThanthrige A, Lu B, Sako Z, Wickramasinghe N. Determinants of Health Care Technology Adoption Using an Integrated Unified Theory of Acceptance and Use of Technology and Task Technology Fit Model: Systematic Review and Meta-Analysis. J Med Internet Res 2025;27:e64524\u0026ndash;e64524. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.2196/64524\u003c/span\u003e\u003cspan address=\"10.2196/64524\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDuyck P, Pynoo B, Devolder P, Voet T, Adang L, Vercruysse J. User acceptance of a picture archiving and communication system. Applying the unified theory of acceptance and use of technology in a radiological setting. Methods Inf Med 2008;47:149\u0026ndash;56.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDuyck P, Pynoo B, Devolder P, Voet T, Adang L, Ovaere D, et al. Monitoring the PACS Implementation Process in a Large University Hospital\u0026mdash;Discrepancies Between Radiologists and Physicians. J Digit Imaging 2010;23:73\u0026ndash;80. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s10278-008-9163-7\u003c/span\u003e\u003cspan address=\"10.1007/s10278-008-9163-7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBhayana R, Biswas S, Cook TS, Kim W, Kitamura FC, Gichoya J, et al. From Bench to Bedside With Large Language Models: \u003cem\u003eAJR\u003c/em\u003e Expert Panel Narrative Review. Am J Roentgenol 2024;223:e2430928. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.2214/AJR.24.30928\u003c/span\u003e\u003cspan address=\"10.2214/AJR.24.30928\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eScipion CEA, Manchester MA, Federman A, Wang Y, Arias JJ. Barriers to and facilitators of clinician acceptance and use of artificial intelligence in healthcare settings: a scoping review. BMJ Open 2025;15:e092624. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1136/bmjopen-2024-092624\u003c/span\u003e\u003cspan address=\"10.1136/bmjopen-2024-092624\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNam Y, Kim DY, Kyung S, Seo J, Song JM, Kwon J, et al. Multimodal Large Language Models in Medical Imaging: Current State and Future Directions. Korean J Radiol 2025;26:900. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3348/kjr.2025.0599\u003c/span\u003e\u003cspan address=\"10.3348/kjr.2025.0599\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSeah JCY, Tang JSN, Tran A. Drafting the Future: The Dawn of AI Report Generation in Radiology. Radiology 2025;316:e243378. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1148/radiol.243378\u003c/span\u003e\u003cspan address=\"10.1148/radiol.243378\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWichtmann BD, Paech D, Pianykh OS, Huang SY, Seltzer SE, Brink J, et al. Leadership in radiology in the era of technological advancements and artificial intelligence. Eur Radiol 2025;36:548\u0026ndash;52. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s00330-025-11745-4\u003c/span\u003e\u003cspan address=\"10.1007/s00330-025-11745-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"},{"header":"Tables","content":"\u003ch3\u003eTable 1. Study cohort characteristics by implementation period\u003c/h3\u003e\n\u003cp\u003eValues are n (% within period).\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eCharacteristic\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eBaseline (n=295)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003ePost-implementation (n=314)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eStudies\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e295 (100.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e314 (100.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eCT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e236 (80.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e259 (82.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eMR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e59 (20.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e55 (17.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eInpatient/ED\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e81 (27.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e98 (31.2%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eOutpatient\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e214 (72.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e216 (68.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eNo contrast\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e58 (19.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e66 (21.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eWith contrast\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e237 (80.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e248 (79.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003ch3\u003eTable 2. Unadjusted time to read examinations (minutes) stratified by key subgroups and LLM assistance\u003c/h3\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 93px;\"\u003e\n \u003cp\u003eSubgroup\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003eNo LLM-assisted, n\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003eLLM-assisted, n\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 46px;\"\u003e\n \u003cp\u003eTotal, n\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003eNo LLM-assisted, median (IQR), min\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003eLLM-assisted, median (IQR), min\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 73px;\"\u003e\n \u003cp\u003eMedian difference (No \u0026minus; LLM-assisted), min\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 78px;\"\u003e\n \u003cp\u003e95% CI (bootstrap)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003eWilcoxon rank-sum p\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 93px;\"\u003e\n \u003cp\u003eOutpatient\u0026ndash;CT\u0026ndash;With contrast\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e135\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e130\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 46px;\"\u003e\n \u003cp\u003e265\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e23.0 (11.0\u0026ndash;35.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e13.0 (10.0\u0026ndash;26.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 73px;\"\u003e\n \u003cp\u003e10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 78px;\"\u003e\n \u003cp\u003e[-0.28, 10.28]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e0.0021\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 93px;\"\u003e\n \u003cp\u003eOutpatient\u0026ndash;CT\u0026ndash;No contrast\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e28\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e38\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 46px;\"\u003e\n \u003cp\u003e66\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e18.5 (8.0\u0026ndash;44.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e7.0 (6.0\u0026ndash;11.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 73px;\"\u003e\n \u003cp\u003e11.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 78px;\"\u003e\n \u003cp\u003e[-7.72, 17.72]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e0.0017\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 93px;\"\u003e\n \u003cp\u003eOutpatient\u0026ndash;MR\u0026ndash;With contrast\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e47\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 46px;\"\u003e\n \u003cp\u003e87\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e14.0 (10.0\u0026ndash;24.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e16.0 (13.0\u0026ndash;23.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 73px;\"\u003e\n \u003cp\u003e-2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 78px;\"\u003e\n \u003cp\u003e[0.33, 9.67]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e0.2808\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 93px;\"\u003e\n \u003cp\u003eOutpatient\u0026ndash;MR\u0026ndash;No contrast\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 46px;\"\u003e\n \u003cp\u003e12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e14.0 (12.0\u0026ndash;42.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e7.0 (4.5\u0026ndash;31.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 73px;\"\u003e\n \u003cp\u003e7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 78px;\"\u003e\n \u003cp\u003e[-35.15, 45.15]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e0.0889\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 93px;\"\u003e\n \u003cp\u003eInpatient/ED\u0026ndash;CT\u0026ndash;With contrast\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e47\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e72\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 46px;\"\u003e\n \u003cp\u003e119\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e15.0 (9.0\u0026ndash;29.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e11.0 (8.0\u0026ndash;20.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 73px;\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 78px;\"\u003e\n \u003cp\u003e[-3.64, 13.64]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e0.0469\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eFootnote: Values are median (IQR). Median difference is calculated as No LLM-assisted \u0026minus; LLM-assisted (positive values indicate shorter reading times with LLM assistance). 95% CIs for median differences were estimated using nonparametric bootstrap (500 replications). P-values are from two-sided Wilcoxon rank-sum tests.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"abdominal-radiology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"aima","sideBox":"Learn more about [Abdominal Radiology](http://link.springer.com/journal/261)","snPcode":"261","submissionUrl":"https://submission.springernature.com/new-submission/261/3","title":"Abdominal Radiology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Artificial Intelligence, large language models, radiology reporting, burnout ","lastPublishedDoi":"10.21203/rs.3.rs-8904615/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8904615/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cb\u003eBackground\u003c/b\u003e\u003c/p\u003e \u003cp\u003eRadiologist burnout affects approximately 40% of US radiologists. Large language models (LLMs) may improve workflow efficiency, but real-world implementation data are limited.\u003c/p\u003e\u003cp\u003e\u003cb\u003eObjective\u003c/b\u003e\u003c/p\u003e \u003cp\u003eTo evaluate LLM-assisted workflow impact on radiologist efficiency using the Unified Theory of Acceptance and Use of Technology (UTAUT) framework.\u003c/p\u003e\u003cp\u003e\u003cb\u003eDesign, Setting, and Participants:\u003c/b\u003e\u003c/p\u003e \u003cp\u003eHIPAA-compliant, IRB-approved retrospective cohort study of a single fellowship-trained abdominal radiologist at Mayo Clinic Arizona. We compared baseline (January\u0026ndash;April 2024) and post-implementation (December 2025\u0026ndash;February 2026) periods. A custom generative pre-trained transformer was developed using ChatGPT Enterprise with disease-specific templates.\u003c/p\u003e\u003cp\u003e\u003cb\u003eMain Outcome Measures:\u003c/b\u003e\u003c/p\u003e \u003cp\u003eInter-study interval time (proxy for interpretation time) compared using Wilcoxon rank-sum tests with Bonferroni correction (α\u0026thinsp;=\u0026thinsp;0.0125). UTAUT constructs assessed: performance expectancy (efficiency), effort expectancy (training burden), facilitating conditions (infrastructure), and behavioral intention (satisfaction).\u003c/p\u003e\u003cp\u003e\u003cb\u003eResults\u003c/b\u003e\u003c/p\u003e \u003cp\u003eWe analyzed 609 studies (495 CT, 114 MRI). LLM assistance significantly reduced inter-study intervals for outpatient CT with contrast (23.0 vs 13.0 minutes; difference 10 minutes; p\u0026thinsp;=\u0026thinsp;0.0021) and without contrast (18.5 vs 7.0 minutes; difference 11.5 minutes; p\u0026thinsp;=\u0026thinsp;0.0017). No improvement occurred for MRI with contrast (14.0 vs 16.0 minutes; p\u0026thinsp;=\u0026thinsp;0.2808) or without contrast (14.0 vs 7.0 minutes; p\u0026thinsp;=\u0026thinsp;0.0889). The radiologist reported improved work-life balance for CT but neutral satisfaction for complex MRI templates. Training required 10 hours over 5 days.\u003c/p\u003e\u003cp\u003e\u003cb\u003eConclusions\u003c/b\u003e\u003c/p\u003e \u003cp\u003eLLM-assisted workflow reduced interpretation times for standardized CT studies but not heterogeneous MRI examinations, supporting UTAUT's emphasis on performance expectancy and task\u0026ndash;technology fit as adoption drivers. Efficiency gains may reduce documentation burden when tools align with task complexity.\u003c/p\u003e","manuscriptTitle":"Large Language Model–Assisted Radiology Reporting: A Retrospective Cohort Study Using the UTAUT Framework to Analyze Workflow Integration and Efficiency Gains","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-22 17:06:56","doi":"10.21203/rs.3.rs-8904615/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-02-21T13:38:07+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-02-21T03:05:58+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-02-19T08:50:39+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"266199424278534813838535000458177795964","date":"2026-02-18T22:44:50+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-02-18T18:35:32+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"99773548497550534624720903134909398020","date":"2026-02-18T18:15:19+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"134114737998008114038236372979798754923","date":"2026-02-18T15:16:43+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"98047858058243879102596764535718579766","date":"2026-02-18T13:45:06+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"274282751361712510006442555315740841761","date":"2026-02-18T13:32:25+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-02-18T13:20:20+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-02-18T10:45:02+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-02-18T10:41:14+00:00","index":"","fulltext":""},{"type":"submitted","content":"Abdominal Radiology","date":"2026-02-17T22:45:15+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"abdominal-radiology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"aima","sideBox":"Learn more about [Abdominal Radiology](http://link.springer.com/journal/261)","snPcode":"261","submissionUrl":"https://submission.springernature.com/new-submission/261/3","title":"Abdominal Radiology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"e5a59eca-a361-4644-8a8c-73d394d8abf2","owner":[],"postedDate":"February 22nd, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2026-04-27T16:03:29+00:00","versionOfRecord":{"articleIdentity":"rs-8904615","link":"https://doi.org/10.1007/s00261-026-05524-y","journal":{"identity":"abdominal-radiology","isVorOnly":false,"title":"Abdominal Radiology"},"publishedOn":"2026-04-20 15:58:47","publishedOnDateReadable":"April 20th, 2026"},"versionCreatedAt":"2026-02-22 17:06:56","video":"","vorDoi":"10.1007/s00261-026-05524-y","vorDoiUrl":"https://doi.org/10.1007/s00261-026-05524-y","workflowStages":[]},"version":"v1","identity":"rs-8904615","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8904615","identity":"rs-8904615","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}