Authors (including presenting author) :
Lui CT (1), Hau LM (2), Tsoi LK (2), Tang PF (2), Wu SN (2), Lam MS (3), Wong CW (4), Yuen WH (4), Kwok WY (4), Lau MC (4), Wong MH (4), Yuen HY (5), Ma KC (5), Yee YC (5), Lee WS (5), Chung HY (5), Lo Edwin (6), Chan Peter (6), Palmerston JB (7), Lui Kenneth (7), Lau Kenny (7), Lam YF (7), So Byron (7), Chung Amy (7), Lee Dennis (7), Ng Owen (7), Chan KYJ (7)
Affiliation :
(1) Department of Accident and Emergency, NTWC Department
(2) Quality and Safety Department, NTWC
(3) Nursing Services Department, NTWC
(4) Department of Medicine & Geriatrics, Pok Oi Hospital
(5) Department of Surgery, Pok Oi Hospital
(6) Information Technology Department, NTWC
(7) Head Office Information Technology & Health Informatics, Hospital Authority
Introduction :
Worldwide, patient deterioration occurs in wards with patient morbidities. There is room for early identification, or prediction, for patients at-risk of deterioration for early intervention. In many cases, early intervention can save lives. Numerous early warning systems based on vital signs have been developed in the past decades, including England’s NHS National Early Warning Score (NEWS) and an updated Modified Early Warning Score (MEWS). However, the impact has been heterogeneous due to limited on model accuracy and workflow integration.
Objectives :
To implement an artificial intelligence model in wards to predict otherwise unexpected deterioration events and enact timely clinical interventions to improve patient outcomes.
Methodology :
An AI model utilizing a fully-automated data pipeline with demographics, vital-signs and laboratory data as model inputs was used to predict the risk of unexpected patient deterioration. The model was previously trained on Australian datasets (n=258,732) and has been extensively validated with HA retrospective datasets in 2022 (n=43,188) with 90% accuracy. Similar accuracy was demonstrated in a prospective silent trial in 2023 in POH (n=2,777). After these rigorous “offline” evaluations, the AI model was integrated to HA CMS with workflow reengineering for live pilot testing in wards. Frontline ward nurses are notified whenever there is a newly identified at-risk patient, through the HA Chat messaging application on working mobiles, as well as on Smart-Care-Centre dashboards. A standardized clinical response and escalation methodology, the Smart-CARES bundle, was designed for nurses and doctors to enact timely interventions for deteriorating patients when an alert is activated by the AI model. The simplified A-to-G approach is a workable approach for ward nurses and junior doctors. A series of training sessions, training kits and reference cue-cards were implemented prior to pilot testing of the system. The working bundle has been piloted in two wards in POH since 11-Sep 2024 and rolled out to 5 more wards in December.
Result & Outcome :
A 5-domained evaluation has been conducted to assess the system during the first two months of the pilot trial in POH on model performance, technical pipeline implementation, clinical compliance and escalation, patient outcome, and staff acceptance. The model performance post-hoc evaluation noted similar accuracy as anticipated and notably greater accuracy than MEWS/NEWS. There were an average of 0-3 and 4-6 alerts per ward per day in SUR and M&G wards, respectively. A majority of the adverse event cases were flagged up 36-48 hours prior to an adverse event (ICU consultation, CPR, or mortality). The chart review on the nursing, intern and physician assessment and intervention identified proactive nursing intervention according to Smart-CARES nurse bundle in 29% of cases. Intensified monitoring and organ support by interns and physicians were satisfactory, while room for improvement was found in root cause workup and targeted treatment in the intern group. Staff survey noted good acceptance and trust by frontline staff. Preliminary assessment noted positive impacts, reducing incidence of adverse events and hospital length-of-stay.
