Authors (including presenting author) :
Leung KY(1), Lui CT(1), Lau CL(1), Hung CB(2), Kwok KY(3), Chan Victor(3), Hau LM(4)
Affiliation :
(1)Department of Accident and Emergency, NTWC
(2)Department of Family and Medicine, NTWC
(3)Department of Radiology, NTWC
(4)Department of Quality and Safety, NTWC
Introduction :
Hospital Authority (HA) developed an Artificial-Intelligence Chest-X-Ray (AICXR) to alert physicians on abnormal lung nodules. Meanwhile, commercial AICXR supporting multiple pathologies detection were available, with sophisticated annotation and reporting features. Psychometric evaluation of clinicians’ attitudes regarding AICXR adoption were lacking.
Objectives :
Evaluate trust and perception by radiologists, emergency and family physicians, regarding accuracy, annotation, reporting features and accessibility of single-pathology HA AICXR versus a multi-pathology commercial AICXR.
Methodology :
A trial was conducted in New-Territories-West-Cluster (NTWC) from April to August 2023. HA AICXR focusing on nodules and a commercial AICXR covering 10 pathologies were simultaneously implemented in Department of Accident and Emergency (A&E), Family Medicine (FM) and Radiology. In prior validation, both models demonstrated comparable accuracy on detecting nodules. HA AICXR was integrated into Clinical-Management-System (CMS), while access to commercial AICXR required separate webportal logon.
A voluntary, anonymous survey was collected in August 2023, focusing on model performance, perceived practical impact and accessibility. Opinions were collected at likert-scale, with results presented in percentages.
Result & Outcome :
84 responses were collected. Majority of respondents were satisfied with commercial AICXR (A&E:59%, FM:71%, Radiology:93%), 56% respondents had more than 5 years of work experience.
Commercial AICXR demonstrated satisfactory diagnostic accuracy. 57% radiologists expressed confidence while the rest were neutral with no negative feedback. FM demonstrated higher confidence regarding accuracy than A&E (A&E:47%, FM:63%). Most respondents agreed commercial AICXR help assisting diagnosis (A&E:54%, FM:74%, Radiology:57%).
Access to commercial AICXR among frontline physicians was infrequent (A&E:26%, FM:45%), mainly upon encountering uncertainties. In contrast, Radiologists demonstrated high utilization (86%), predominantly for assisting reporting. Most frontline physicians identified separate webportal logon as a barrier (A&E:82%, FM:75%, Radiology:28%).
Comparing both models, frontline physicians found commercial AICXR more accurate (A&E:65%, FM:56%), with clearer annotation (A&E:83%, FM:78%), and better interpretation with abnormality likelihood score (A&E:66%, FM:63%).
In conclusion, commercial AICXR covering multiple pathologies with clearer annotation appeared better accepted by clinicians, especially if integrated to CMS. The observed heterogeneity among clinicians of different specialties might be explained by different CXR interpretation training. The findings shed light on future development and implementation of AICXR in healthcare system.
