The hierarchical medical system challenge. China implemented a hierarchical medical system (HMS) in 2015 to optimize health care resource allocation across tiered institutions. Despite this design, lower-tier primary and secondary hospitals remain chronically underutilized while tertiary hospitals face severe overcrowding -- outpatient and emergency visits to tertiary hospitals rose 27.7% from 2019 to 2023. Contributing factors include limited resources, staffing gaps, and public perception that lower-tier facilities provide suboptimal care.
AI as a solution for lower-tier institutions. AI-assisted diagnostic software functions as a second reader, delivering consistent and rapid CT image interpretation, reducing diagnostic errors, and minimizing care quality variability. However, clinical implementation in China has been concentrated almost exclusively in tertiary hospitals, driven by institutional capacity and IT infrastructure. This implementation pattern fails to leverage AI's potential precisely where it is most needed -- resource-constrained secondary hospitals handling population-level screening.
Alignment problem at tertiary hospitals. AI software designed for population-level screening is often misaligned with tertiary hospitals' core purpose of specialized complex case management. Studies have found that AI provides limited added benefit to senior physicians in complex cases, while its value is theoretically highest in primary and secondary settings where physician experience and diagnostic equipment are more limited.
Gap in implementation research. Most published studies have focused on AI algorithm performance metrics such as sensitivity and specificity, while real-world implementation strategies in lower-tier clinical environments remain underexplored. This study aimed to design, implement, and evaluate a tailored integration strategy for AI-assisted pulmonary nodule screening in a secondary hospital, using an established implementation science framework to assess what works and what barriers remain.
Barrier identification using CFIR. Key barriers to AI implementation were previously identified through Consolidated Framework for Implementation Research (CFIR) guided qualitative interviews and a 3-round modified Delphi process. The top 5 barriers selected were: unsatisfactory AI clinical performance at tertiary hospitals, misalignment between software functions and tertiary hospital goals, lack of collaborative networks between secondary and tertiary hospitals, lack of information security measures and certification, and lack of performance feedback and evaluation mechanisms.
ERIC strategy mapping. The Expert Recommendations for Implementing Change (ERIC) compilation -- 73 evidence-based implementation strategies developed through international Delphi consensus -- was systematically mapped to each CFIR barrier using the CFIR-ERIC mapping tool. Strategies with cumulative endorsement above 50% were designated high-priority candidates. A stakeholder consensus session with 27 participants (16 clinicians, 3 administrators, 4 informatics specialists, 4 AI vendors) then evaluated each strategy for contextual relevance, feasibility within existing workflows, and logical alignment with identified barriers.
Five integrated implementation strategies selected. Through consensus, five strategies were selected: (1) shifting AI deployment focus from tertiary to secondary hospitals through local consensus discussions; (2) developing a formal implementation blueprint aligned with secondary hospital clinical needs; (3) building a cross-institutional coalition to establish structured referral pathways; (4) collaborating with a certified data transfer platform to enable secure interhospital data sharing; and (5) establishing a bidirectional feedback mechanism returning tertiary hospital diagnoses to the referring secondary institution.
RE-AIM evaluation framework. The implementation was evaluated using the RE-AIM framework across four dimensions: Reach (patient and provider engagement), Effectiveness (impact on detection rates and workflow times), Adoption (uptake by intended users), and Implementation (feasibility, usability, and workflow integration). The Maintenance dimension was not assessed given the 3-month implementation window.
Secondary-tertiary hospital partnership. The implementation site was Fengtai Rehabilitation Hospital of Beijing (secondary, 211 beds, 300,000 annual outpatient visits) partnered with China Rehabilitation Research Center (tertiary, 1,100 beds, 470,000 annual outpatient visits). Both are publicly funded Beijing institutions. The secondary hospital served as the primary screening site; the tertiary hospital provided specialist review and final diagnosis for referred indeterminate cases.
AI software integrated with PACS. A commercially available AI-assisted pulmonary nodule screening system (approved by China's National Medical Products Administration in 2022, CE certified in 2020) was locally deployed at the secondary hospital and integrated directly with the hospital picture archiving and communication system (PACS). It automatically identified and segmented potential lesions in chest CT scans, providing malignancy probability scores from 0% to 100% for each nodule. The integration was designed to be seamless with existing diagnostic workflows without requiring additional steps for radiologists.
Cloud-based ESDR data transfer tool. An electronic source data repository (ESDR) cloud platform managed by the tertiary hospital enabled encrypted bidirectional data sharing between institutions. Using CT image IDs, the system automatically uploaded AI-generated nodule results and CT images from the secondary hospital's PACS to the cloud platform. Tertiary hospital physicians could then review imaging and diagnostics remotely, record final diagnoses and follow-up plans, and return recommendations to the secondary site.
Referral pathway and patient flow. Patients meeting referral criteria (malignancy risk 5% to 65%, dominant nodule above 8 mm, rapid growth, family history of lung cancer, high-risk factors including smoking and age) received structured referral documentation including demographics, symptoms, radiological findings, AI results, and insurance information. A free registration system was introduced at the tertiary hospital to remove financial barriers, and a facilitation team assisted with appointment scheduling and care navigation.
High AI software adoption rate. During the 3-month study period (December 2024 to March 2025), 1,375 patients underwent chest CT at the secondary hospital. Of 1,291 eligible patients (after excluding those under 18 or with prior lung surgery), 85.6% had their CT scans analyzed by the AI software. Patient demographics and nodule characteristics were not significantly different from the 1,591-patient historical control group collected in the same period the prior year.
Pulmonary nodule detection rate doubled. The overall pulmonary nodule detection rate in the intervention group was 65.2% -- significantly higher than the historical control rate of 32.4% (p less than 0.001). CT image assessment and reporting time improved from a mean of 30.4 minutes in the historical control period to 24.1 minutes in the intervention period (p less than 0.001), indicating workflow efficiency gains alongside the detection improvement.
High referral completion rate. Of 33 patients meeting predefined referral criteria, 25 (75.8%) provided written informed consent and were formally referred. Of those 25, 22 (88%) successfully completed the referral and visited the tertiary hospital. The median time from referral submission to tertiary hospital visit was 3 days. All 22 referred patients were accepted by the tertiary hospital, yielding a 100% referral acceptance rate.
CT image transmission mostly successful. Of 22 successfully referred patients, 20 (90.9%) had their CT images successfully transmitted via the ESDR platform to the tertiary hospital in advance of the appointment, enabling remote review by specialists. Two patients experienced PACS system failures that prevented upload, requiring them to carry physical CT films to the tertiary hospital instead.
Strong AI software adoption by physicians. Among 15 secondary hospital physicians (6 radiologists, 9 respiratory physicians), all 6 radiologists and 7/9 (77.8%) respiratory physicians reported using the AI software. Perceived alignment with institutional goals was high -- all 15 agreed the AI software supported the hospital's core objectives. Among 8 tertiary hospital physicians, 62.5% reported using the ESDR tool and 87.5% said its use aligned with strategic priorities.
AI software was perceived as easy to use. Among the 13 secondary hospital physicians who had used the AI software, 84.6% rated it as easy to use, and 93.3% reported it did not interfere with routine clinical practice. This suggests that the PACS-integrated deployment design successfully avoided the workflow disruptions that have been identified as a major barrier to AI adoption in prior studies.
ESDR tool created workflow burdens. In contrast to the AI software, only 3/5 (60%) of secondary hospital respiratory physicians and 4/5 (80%) of tertiary hospital physicians who had used the ESDR tool perceived it as user-friendly. The core problem was poor interoperability between the ESDR system and existing hospital information systems -- physicians had to manually access the platform via an external web link to upload or retrieve data, increasing workload and disrupting efficiency. This led 55.6% of secondary hospital respiratory physicians to report substantial burden from ESDR tool use.
Referral documentation completion was critically low. Despite 100% training coverage for both tools, only 28% (7/25) of referral records contained complete data entries. Critical fields including clinical symptoms, diagnostic conclusions, and insurance information were frequently missing, requiring manual supplementation. Additionally, only 44.4% of secondary hospital physicians reviewed the final diagnostic results returned through the ESDR tool, and only 50% of tertiary hospital physicians uploaded final diagnoses back to the platform -- severely limiting the bidirectional feedback loop that was a core strategy component.
Feasibility demonstrated but implementation gaps are actionable. The study demonstrated that deploying AI-assisted pulmonary nodule screening in a secondary hospital through a structured referral and data exchange framework is feasible and produces meaningful diagnostic improvements. The detection rate doubling and high referral completion are particularly significant given the secondary hospital's resource constraints. However, the data completeness gap (28% complete referral records) and low feedback review rates represent specific, addressable implementation failures rather than fundamental feasibility barriers.
Interoperability is the key technical barrier. The ESDR tool's requirement for manual external web access -- instead of integration within the hospital information system -- was identified as the primary driver of workflow disruption, data incompleteness, and low feedback adoption. This finding highlights that AI implementation success depends as much on the integration architecture of supporting data infrastructure as on the AI algorithm itself. Future deployments should prioritize native integration of data transfer tools within existing hospital systems.
Referral adherence determines overall value of AI screening. Consistent with prior research on AI-assisted screening for diabetic retinopathy, this study confirms that improved follow-up and referral adherence are essential for AI tools to deliver health value beyond detection improvement. High detection rates without effective referral pathways do not translate to patient benefit. The patient-centered design elements -- free registration, dedicated facilitation team, 3-day median referral time -- were critical to achieving the 88% referral completion rate.
Future research directions. Multicenter studies across diverse health care settings are needed to assess the generalizability, cost-effectiveness, long-term sustainability, and scalability of these implementation strategies. Future iterations should integrate the data transfer tool natively within hospital information systems, implement structured data entry forms with required fields to improve documentation completeness, and develop user-facing feedback interfaces that make reviewing tertiary diagnoses easier for referring physicians.