The stepped-wedge cluster-randomized design was chosen for this trial because of its clear methodological benefits. Unlike parallel designs, all clusters eventually receive intervention, enhancing acceptability in clinical settings where prior evidence suggests potential benefit [28,29]. The phased implementation also mirrors real-world logistical constraints while allowing for both within- and between-cluster comparisons, thereby improving statistical efficiency and power [28,30]. Additionally, repeated measurements over time enable adjustment for secular trends, strengthening causal inference [28-31]. These characteristics make the stepped-wedge design especially suitable for pragmatic evaluations of health care interventions.

This nationwide, multicenter, stepped-wedge cluster-randomized trial (SW-CRT) follows the CONSORT-AI (Consolidated Standards of Reporting Trials-Artificial Intelligence) and SPIRIT-AI (Standard Protocol Items: Recommendations for Interventional Trials-Artificial Intelligence) guidelines [32,33]. It will evaluate whether AI can improve CXR interpretation in routine clinical care across Brazil. Cities will act as clusters, comprising public, private, and philanthropic health care units to ensure broad socioeconomic and geographic representation [28,30,31,34]. All clusters will begin in the control condition (standard physician interpretation) and will cross over to the intervention (AI-assisted physician interpretation) at randomly assigned monthly intervals. Potential contamination across clusters will be minimized by selecting independent clinical units with nonoverlapping staff and workflows.

A batched stepped-wedge approach allows staggered initiation, accommodating operational constraints while minimizing contamination or learning effects. Once a cluster transitions, it remains in the intervention arm (Figure 2). Simple urn-based randomization without replacement will occur independently in each batch. Participating centers within cities must have infrastructure for CXR interpretation (negatoscope or digital screen), reliable internet access, and commitment to study protocols.

All participating clusters will undergo a site initiation visit (SIV) prior to enrollment. All physician participants will complete the informed consent process and sign the informed consent form. During the SIV, local investigators will receive standardized training on study procedures, data entry, image capture protocols, and use of the study platform (LuAna). Training will include hands-on instruction to ensure the uniform application of inclusion criteria, proper anonymization of patient data, and completion of interpretation forms. Following the SIV, sites will begin patient recruitment and data collection. This study was registered on ClinicalTrials.gov (NCT06686251).

Radiograph interpretation will be performed by general practitioners or other nonradiologist physicians employed or otherwise engaged at participating trial sites. No restrictions will be imposed regarding age, sex, years since graduation, or previous training in imaging interpretation. All participating physicians will receive training using the study platform and must provide informed consent before contributing to interpretations. The training procedure emphasizes that AI outputs are advisory only; clinical responsibility remains fully with the interpreting physician, and AI suggestions never override clinical judgment. Afterward, they will complete a brief questionnaire assessing their baseline confidence in CXR interpretation, perceived challenges in reading CXRs, and initial impressions regarding the usability of the study platform (Multimedia Appendix 1).

LuAna is a cloud-based AI platform for CXR interpretation developed in partnership with the Brazilian Ministry of Health. The platform receives anonymized CXR with the accompanying physician report and generates 3 deep-learning outputs: (1) classification of images as normal or abnormal, (2) classification of specific radiologic findings (consolidation or pulmonary opacity, pneumothorax, atelectasis, pleural effusion, and cardiomegaly), and (3) probability scores for radiographic features suggestive of pulmonary TB. Operating fully in the cloud, LuAna provides decision support without requiring modifications to current imaging systems or clinical workflows.

Anonymized data will be gathered using the LuAna platform (accessible on desktop or mobile), which records CXRs and physician interpretations (Figure 3). Physicians will access the platform using secure login credentials. All data will be automatically logged to enable auditing, reproducibility, and safety monitoring.

Only demographic variables, including age group, sex, and race, will be collected in aggregated form to characterize the study population and allow for fairness analyses, ensuring equitable and unbiased AI performance. CXRs will be captured either by photographing printed films positioned on a lit negatoscope, by photographing digital images displayed on a picture archiving and communication system monitor, or by direct screen capture, all following strict protocols to guarantee full anonymization. Physicians will capture only the thoracic image, ensuring that all patient identifiers are systematically removed and never visible. A patient ID entered solely to return the radiologist’s report to the local health unit will be encrypted and remain inaccessible to the research team. Nondiagnostic images or studies failing predefined quality-control criteria will be excluded from analysis.

Physicians will complete a brief questionnaire of their first CXR interpretation based on the following predefined questions: (1) Is the CXR normal or abnormal? (2) Are the specified radiological findings present or absent? and (3) Does the CXR show features suggestive of pulmonary TB? In the intervention group, after the physician’s clinical management, AI output (including probability scores and heatmaps) will be made available. Physicians will then document their final interpretation and decision-making process and subsequently record any changes made after reviewing the AI results. All data will be securely stored in encrypted, deidentified cloud-based repositories.

The radiologists are responsible for producing the ground-truth (reference standard) for all CXR images during the study. The first step in ground-truth generation is to assign 2 thoracic radiologists to independently interpret the CXR images to set up the reference standard, classifying each finding as “normal” or “abnormal.” The correct classification matches the reference standard. For cases classified as abnormal, the presence or absence of the relevant radiological findings (consolidation, pulmonary opacity, pneumothorax, atelectasis, pleural effusion, and cardiomegaly) will be assessed, as well as the presence or absence of changes suggestive of TB. Based on these evaluations by the 2 thoracic radiologists, the ground truth is established. The radiologists will be blinded to both the physicians’ interpretations and the AI outputs.

A chest radiograph will be labeled as suggestive of pulmonary TB if at least 1 radiographic feature considered compatible with pulmonary TB is present, as determined by the adjudicated thoracic radiologist reference standard. Examples include cavitation, upper-lobe–predominant consolidation or opacities, fibronodular changes or scarring with upper lobe volume loss, or a miliary pattern. If none of these features is present, the examination will be labeled as “not suggestive of pulmonary TB.”

For physician performance analyses, an examination will be considered correctly classified when the physician’s binary classification (normal vs abnormal; suggestive vs not suggestive of pulmonary TB; presence or absence of specific radiographic patterns) matches the adjudicated thoracic radiologist reference standard. In cases of any disagreement between the 2 experts in any of the evaluated items, a third radiologist, more senior, will adjudicate.

All findings will be registered in the CARPL, with abnormalities annotated to allow direct comparison with AI outputs. This will serve as the reference value for model validation and performance analysis (Figure 4).

Patient identifiers will be used solely to return reports of critical imaging findings to clinical sites, when present. These identifiers are cryptographically isolated from the research dataset and are never accessible to the research team (Figure 4). If the radiologist identifies a potentially critical finding, such as tension pneumothorax, a large obstructive mass, or near-complete lung consolidation, an alert will be sent to the originating site via email and/or phone (Figure 4). Patient data will remain encrypted; only the automated system will link the alert to the corresponding patients, ensuring privacy and prompt clinical follow-up. An institutional lifetime license ensures consistent labeling with the same software version throughout the study.

At the end of the study, all participating physicians and radiologists will be invited to complete a standardized questionnaire assessing the usability of the LuAna platform and its integration into routine clinical workflows. Responses will be used to address implementation challenges and opportunities for improvement (Multimedia Appendix 1).

An independent Data and Safety Monitoring Board, comprising external experts in clinical trials, radiology, statistics, and ethics with no conflicts of interest, will oversee the study in accordance with international guidelines and good clinical practice. The board will be responsible for safeguarding patient safety and ensuring study integrity.

The sample size was calculated to detect a minimum difference of 10% in the primary outcome—a binary indicator of whether the physician correctly labeled a CXR—between the AI-assisted and control groups. The calculation accounted for the stepped-wedge cluster-randomized design with batched implementation using the following assumptions: 9 clusters randomized in 3 batches to cross over from control to intervention, 10 time periods, a baseline detection rate of 67% in the control group, an expected increase to 77% in the intervention group, and an intracluster correlation coefficient ranging from 0.001 to 0.003.

The design effect was adjusted for the correlation structure induced by repeated measures within clusters over time using methods described by Hussey and Hughes [28] and extended to account for batched randomization, as described by Hemming et al [35,36]. The power calculation targeted 80% power at a 2-sided significance level of 5% and accounted for potential variability in cluster sizes. The final average cluster size was 37, and the overall sample size was inflated by 10% to account for anticipated exclusions due to image quality issues or missing data (dropout). This approach ensures adequate power to detect a meaningful effect of the intervention while maintaining the integrity of the stepped-wedge design with staggered implementation across batches.

Primary analysis will follow the modified intention-to-treat principle. All eligible CXRs (the unit of analysis in this study) will be analyzed according to the study group (control or intervention), to which their corresponding cluster was allocated at the time of image acquisition, regardless of physician adherence to the intervention or exposure to AI output. Images of poor quality, as defined by our radiologists, will be excluded postrandomization.

The primary outcome analysis will use a generalized linear mixed model with a binomial distribution and logit link function to estimate the effect of the AI-assisted intervention on detection probability. The model will include fixed effects for the intervention group (AI-assisted vs control), calendar time (to adjust for secular trends), time since intervention rollout (to account for learning or adaptation effects), and batch effects. In addition, random intercepts for clusters will be specified to account for intracluster correlation.

Results will be reported as adjusted odds ratios with 95% CIs and 2-sided P values. Model fit and assumptions will be evaluated using the Residual Diagnostics for Hierarchical (Multilevel/Mixed) Regression Models (DHARMa) [37]. In case of convergence issues or model instability, generalized estimating equations with robust SEs will be used as an alternative to estimate population-averaged effects. Independent models will also be fitted for each radiological finding, as well as for the pulmonary abnormality and TB algorithm.

Model performance metrics will be reported with 95% CIs [38,39]. Responses to questions regarding the mobile interface will be summarized as frequencies. Chi-square test and ordinal logistical regression will be used to investigate associations between these responses and physician characteristics (eg, years in practice and biological sex). Additionally, generalized linear mixed models will be fitted with the algorithm’s output as the dependent variable, adjusted for image characteristics, to investigate their effect on AI classification.

Missing data are expected to be minimal due to app-enforced completeness checks. Incomplete records will be described, and sensitivity analyses using multiple imputation will be performed if the proportion of missing data exceeds 10%. All analyses will be performed in R (R Foundation for Statistical Computing), and code scripts will be version-controlled and archived to ensure reproducibility.

This study was reviewed and approved by the local Ethics Committee of Hospital Israelita Albert Einstein (CEP Einstein), São Paulo, Brazil, registered under IRB00005041, in accordance with applicable national and international regulations (CAAE number 82494524.0.1001.0071). A waiver of written informed consent for patients was granted, as the study is strictly observational and involves no intervention in routine clinical care, no modification of clinical management, and no access to patients’ medical records. Only chest radiographs obtained as part of standard care were used, and all images were fully anonymized prior to storage and analysis, with no directly identifiable information included in the research dataset. However, written informed consent was obtained from the physicians responsible for enrolling patients in the LuAna trial, as they are considered direct research participants.

All data are handled under strict confidentiality procedures. Chest radiograph images were collected via the LuAna app and stored in an anonymized format on a secure cloud-based server.

The analytical database contains exclusively deidentified data. Access to identifiable information at the study sites, when applicable, is restricted to authorized local investigators. Study data will be securely stored at the hosting institution for a minimum of 5 years, in compliance with institutional and regulatory requirements.

Posttrial access pathways for the tool will be discussed with the Brazilian Ministry of Health to ensure continuity if the intervention proves beneficial. Technical details of the algorithm, including its development, training, validation, and performance characteristics, will be described in accordance with the CLAIM (Checklist for Artificial Intelligence in Medical Imaging) reporting guidelines [40], complementing the CONSORT-AI and SPIRIT-AI standards applied to the clinical trial components [32,33].

This trial is a pivotal effort to evaluate how AI can enhance diagnostic accuracy, clinical workflows, and physician performance in interpreting chest radiographs.

By using a batched stepped-wedge cluster-randomized design across a nationally representative sample of health care settings, we aim to rigorously assess the real-world utility of LuAna, an AI-powered platform for CXR interpretation developed in the public sector to assist general practitioners.

Despite ongoing debate about the advantages of SW-CRTs, including concerns about equipoise and delayed access to interventions [41], this design offers distinct benefits for implementation research. Governed by the same ethical principles as all clinical research [42], SW-CRTs are particularly valuable when evaluating innovations that are expected to be beneficial and for which withholding access may be contentious [42]. In the SUS context, this approach ensures that all centers ultimately gain access to AI support, promotes fairness and acceptability, and provides a pragmatic pathway to sustainable integration of the intervention into routine care.

The central goal is to determine whether LuAna improves clinicians’ ability to detect radiographic abnormalities, particularly in environments with limited access to radiology expertise. Comparisons against adjudicated, double-blinded radiologist interpretations will enable the quantification of diagnostic agreement, while stratified analyses by site type and physician characteristics will offer insights into contextual factors that may influence AI performance.

Importantly, the trial also investigates the clinical impact beyond diagnostic accuracy. By evaluating patient management decisions and turnaround times, we aim to capture how AI affects treatment pathways and efficiency in routine care. Usability assessments will provide critical feedback for refining and scaling the platform, thereby ensuring seamless integration into diverse clinical workflows across both mobile and desktop interfaces.

Robustness will be evaluated through sensitivity analyses on image quality, artifacts, and medical devices, thereby validating the model’s generalizability under real-world conditions. Additionally, longitudinal analysis of physician interpretations will explore whether repeated AI exposure improves independent performance, highlighting AI’s potential educational role.

Beyond its methodological rigor, the LuAna trial advances an area in which robust evidence remains scarce: the real-world, prospective clinical evaluation of AI tools. While algorithmic performance is well documented in retrospective studies, few investigations have examined how AI behaves once deployed within the operational complexity of large, heterogeneous public health systems [40,43,44]. This gap has been repeatedly highlighted by international frameworks, including the DECIDE-AI (Developmental and Exploratory Clinical Investigation of Decision Support Systems Driven-Artificial Intelligence) guidelines, which emphasize the need for pragmatic, early-stage clinical evaluation before widescale deployment. Indeed, studies show positive primary outcomes for AI in clinical practice, but concerns remain due to mostly single-center trials, limited demographic data, and inconsistent reports on operational efficiency, which affect generalizability and practical use [45,46].

Evaluating AI-based decision-support systems involves several challenges: understanding their nature as complex interventions; considering user variability and the biases it introduces; integrating human factors into the collaboration between clinicians and AI; assessing both real patients and their data; dealing with continuous system changes; preventing the reproduction of health inequalities; ensuring that results generalize to different contexts; and guaranteeing reproducibility in the real world. By implementing LuAna across a nationally diverse range of health care settings, this study not only provides valuable empirical data on AI performance in real-world situations but also offers a unique opportunity to examine issues related to fairness [47-49].

Variability in patient demographics, imaging equipment, staffing, and workflow conditions across Brazil’s public health system creates a natural stress test for equity: the trial can assess whether AI-assisted interpretation benefits all subpopulations and settings equally or whether performance disparities emerge [50,51]. In doing so, it addresses 2 critical gaps in the global AI literature—real-world clinical validation and algorithmic fairness—offering evidence that is seldom available yet essential for responsible, scalable, and equitable integration of AI into routine care.

Grounded in operational realities and health care equity, this study contributes to a broader movement toward responsible, evidence-based AI integration in clinical practice. It aligns with international recommendations for validation frameworks and addresses a global need for tools that are both technically robust and contextually appropriate for low- and middle-income health systems.

Although the stepped-wedge design enhances real-world applicability by including diverse health care settings, the findings may still be influenced by context-specific factors, such as infrastructure, physician training, and patient demographics. Variation in image acquisition, whether from printed films, computer displays, or mobile capture, may introduce heterogeneity that affects both physician interpretation and algorithm performance. In addition, while the trial evaluates diagnostic accuracy, turnaround times, usability, and model robustness, it does not directly measure patient-level outcomes such as morbidity or mortality. Equity analyses are further limited by the narrow range of demographic variables collected, which may not fully capture socioeconomic or clinical determinants of performance. Finally, the lack of detailed clinical data, including comorbidities, treatment history, and laboratory findings, restricts the ability to contextualize radiographic interpretations and fully assess the AI integration into patient care pathways.

Future work should prioritize expanding datasets to include a broader spectrum of acquisition standards, additional radiographic projections (eg, lateral views), and complementary imaging modalities such as CT. Scaling to additional centers across Brazil—capturing greater geographic, demographic, and service-level diversity—will further strengthen dataset representativeness and enhance algorithm robustness. The adoption of standardized annotation protocols will support higher-quality data curation and reduce uncertainty in the reference standard. In parallel, strategies such as transfer learning and active learning should be explored to optimize performance in data-limited environments.

Ultimately, LuAna seeks to demonstrate that AI-powered diagnostic support can be safely and effectively integrated into public health care, empowering clinicians and improving outcomes for underserved populations. This study will lay the groundwork for broader AI adoption in medical imaging, grounded in evidence, equity, and clinical impact.

If proven effective, LuAna could help bridge diagnostic gaps by supporting nonspecialist clinicians, accelerating decision-making, and strengthening health care delivery in underserved regions, offering a scalable, evidence-based model for equitable AI integration in medical imaging.