Authors (including presenting author) :
Cheng HL(1), Chan JKY(1), Lui LMY(1), Chung APM(1), Cheng ITH(1), Lee DPK(1), Ho ESL(1), Pang JYW(1), Lansley P(1), Cheung NT(1)
Affiliation :
(1) Information Technology and Health Informatics Division, Hospital Authority Head Office
Introduction :
Chest X-ray (CXR) is the most common imaging request. All CXR examinations requested by General Out-Patient Clinic (GOPC) would have reporting. However, the report turnaround time (TAT) was about 20-40 days in which about 8% of these examinations have findings that warrant early clinical attention. Besides, CXR examinations in Accident and Emergency (A&E) Department, In-patient setting and Specialty Out-Patient Clinic (SOPC) are not reported routinely by radiologists. Human perceptual error might occur. Hospital Authority (HA)’s Clinical Management System (CMS) contains 30 years of big data for Artificial Intelligence (AI) model development. AI is more stable in performance and perceptual error can be avoided. Numerous healthcare data was retrieved from various CMS modules and used for AI model training and validation. Local CXR AI models were developed to shorten the report TAT and provide real time decision support.
Objectives :
By using the comprehensive medical record system data to develop the local CXR AI model to: - shorten the report TAT for CXR requested by GOPCs - provide real time decision support for mass and nodule detection in CXR - support clinical workflow by providing a CXR AI flagged list for second screening
Methodology :
Various criteria, such as hospitals, diagnosis and image modalities were retrieved via Clinical Data Analysis and Reporting System. CXR images and reports would be retrieved from Radiology Information System for CXR labelling and report classification. To shorten the report TAT, the CXR AI model was developed by using over 4,000 images for threshold tuning while nearly 5,000 images were used for validation. CXR required early attention would be classified as “High probability” and Radiologists would then provide CXR report with priority. To support the report prioritization, an AI Radiology portal was developed. To provide real time decision support for mass and nodule detection, over 1,300 images were used for threshold tuning while nearly 15,000 images were used for validation. After the AI model was developed, a relatively sensitive threshold was selected for picking up lung mass or nodule. AI results would be shown by AI indicator which would be displayed at CMS. Clinicians could access the CXR image directly via the AI indicator. Besides, clinicians could retrieve suspicious cases through AI Result Patient List for more efficient second screening.
Result & Outcome :
CXR AI model for GOPC report prioritization was fully implemented in all GOPCS. An interim evaluation was done in HKWC, NTWC and PWH. Report TAT in 6 consecutive months was reviewed. The average TAT in these 3 clinical areas were 10-15 days. The CXR AI model for mass and nodule detection was fully implemented in all HA A&Es. The model was further extended to SOPC and In-patient area since November 2021. An impact analysis of TSWH A&E CXR AI pilot was done from 12 March 2021 to 11 April 2021. 530 out of 2,759 CXR images were tagged with suspicious mass and nodule. Flag rate was around 20%. Individual case review was conducted from 1 April 2021 to 10 April 2021. 174 out of 920 CXR were tagged with suspicious mass and nodule. 8 images showed new lung shadows whereas 4 cases were likely to be carcinoma. Besides, the average length of documentation was increased in AI tagged cases when compared with those cases without AI tag. It implied that the AI indicator improved clinician vigilance on AI tagged cases. “Reporting request” rate and SOPC referral were reviewed. The rate of “Reporting request” between 12 March 2021 to 11 April 2021 was 5.8% after CXR AI implementation. This was similar to that in the previous 3 months (5.3%-6.1%). The SOPC referral rate did not have significant increase. The SOPC referral rate of AI tagged cases was 6.1% while that of cases without AI tag was 9.4%. When comparing the overall SOPC referral to the previous year, the SOPC referral rate was 9.1% in April 2021 while that was 8.8% in April 2020.
