Authors (including presenting author) :
Wei Y(1), Lai PBS(2), Cheung SYS(3), Lam TYT(4), Hung CT(1), Yeoh EK(1), Chong KC(1)
Affiliation :
(1)Centre for Health Systems and Policy Research, School of Public Health and Primary Care, The Chinese University of Hong Kong, (2)Department Surgery, Faculty of Medicine, The Chinese University of Hong Kong, (3)Department of Surgery, North District Hospital, (4)School of Nursing, The Chinese University of Hong Kong
Introduction :
Traditional risk assessment methods are still the major tools for postoperative outcome prediction in current clinical practice. However, the prediction performance varies by population and the accuracy remains questionable.
Objectives :
With using 12 years of surgical data (both preoperative and intraoperative) from all 17 public surgical hospitals, we aim to establish a surgical risk calculator for post-operative mortality in Hong Kong
Methodology :
This is a machine learning based model development study using 12-year surgical data from 17 public hospitals based on the Surgical Outcomes Monitoring and Improvement Program in Hong Kong. A total of 124 patient-level variables will be collected from approximately 288 thousand patient records for model training and validation. The primary outcome is the 30-day mortality, defined as the occurrence of death within 30 days after a surgery regardless of the cause. To predict the study outcomes, we employed seven statistical and machine learning models. Prediction performance will be evaluated using the area under the receiver operating characteristic curve (AUROC).
Result & Outcome :
We have conducted a pilot analysis to preliminarily assess the prediction performance of our suggested models for the prediction of 30-day mortality after emergency surgeries. All the 7 prediction models were well-calibrated in predicting the 30-day mortality of emergency surgeries with the AUROC being at least 0.88 (Figure 1). In the model validation, four of the models had AUROC > 0.85 with random forest having the highest prediction power. In the elective surgeries, 4 out of the 7 models had a validated AUROC > 0.85 with logistic regression and multilayer perceptron neural network having the highest prediction power. In conclusion, our establishment of a convenient surgical risk calculator in Hong Kong is timely and helpful to assist surgeons for a risk assessment and provide reliable information for counselling and communication with patients.
