Authors: (including presenting author): :
Ling GY (1), Kwok KY (2), Chan KYJ (3), Chan LW (4), Chau CW (5), Cheung KH (6), Chung PMA (7), Kung CMA (8), Kung SW (9), Lee SF (10), Leung MWM (11), Liu HCJ (7), Lo YH (6), Lui CT (1), Ng WLO (12), Siu YCA (13), Tse CF (1), Wong MK (12), Wong OF (14), Chan K (15), Woo PSP (15)
Affiliation: :
(1)Accidents and Emergency Department, Tuen Mun Hospital, (2)Radiology Department, New Territory West Cluster, (3)Information Technology and Health Informatics Division, HAHO, (4)Accident and Emergency Department, Pamela Youde Nethersole Hospital, (5)Accident and Emergency Department, Queen Elizabeth Hospital, (6)Accident and Emergency Department, Prince of Wales Hospital, (7)Head Office IT, Hospital Authority, (8)Department of Infection Emergency and Contingency, Head Office Major Incident Control Center, HAHO, (9)Accident and Emergency Department, Tseung Kwan O Hospital, (10)Diagnostic Radiology Department, Alice Ho Miu Ling Nethersole Hospital, (11)Accident and Emergency Department, Queen Mary Hospital, (12)Information Technology and Health Informatics Division, Hospital Authority, (13)Accident and Emergency Department, Ruttonjee Hospital, (14)Accident and Emergency Department, North Lantau Hospital, (15)Statistics and Data Science Department, Strategy and Planning Division, HAHO
Keyword 1: :
artificial intelligence
Keyword 2: :
deep learning model
Keyword 3: :
non-contrast computed tomography of the brain
Keyword 4: :
intracranial haemorrhage
Introduction: :
Non-contrast computed tomography of the brain (NCCTB) has been widely used to detect intracranial haemorrhage (ICH) in symptomatic patients. However, its interpretation is subject to perceptual and cognitive errors. A deep learning model (DLM) for detecting ICH and midline shift was deployed across 17 public emergency departments in Hong Kong. This study evaluated its impact on both patient outcomes and emergency physicians’ perceptions before and after implementation.
Objectives: :
The implementation analysis compared demographics, baseline characteristics and clinical outcomes of patients of ICH before and after DLM implementation. A cross-sectional survey of emergency physicians was conducted between March and June 2025 to assess knowledge, trust, and perceived efficiency related to DLM use.
Methodology: :
This is a mixed-method study consisted of pre- and post-implementation analysis and a questionnaire study on emergency physicians.
Result & Outcome: :
Survey response rate was 42.4% (n=236), with 93.2% agreed the DLM was trustworthy. Higher DLM knowledge was associated with shorter documentation and decision-making times, higher diagnostic accuracy, greater confidence in patient discharge, and improved efficiency (p< 0.01). Experience in EDs was not significantly associated with usage patterns, efficiency or familiarity. In the implementation analysis, thirty-day all-cause mortality among ICH patients decreased from 20.4% to 18.3% (p=0.042) after DLM implementation. No significant differences were observed in ED length of stay (p = 0.733) or door-to-neurosurgery time (p = 0.947). Post-implementation reattendance or readmission rates at 7 days were reduced from 1.9% to 1.5% (p< 0.001). Post-implementation evaluation was independently associated with lower mortality (aOR 0.856, 95% CI 0.746–0.981, p=0.0258) in logistic regression. Implementation of DLM was associated with improved patient outcomes without delaying care, alongside high clinician trust and perceived efficiency gains. Workforce artificial intelligence literacy and model familiarization are critical to maximizing clinical impact.
