Applying Machine Learning to the ANZELA-QI Database to Predict Adverse Outcomes for Patients Undergoing Emergency
Dafydd Jones1, Joshua Blum1, Catherine Cartwright1
1Department of General Surgery, Royal Hobart Hospital, Hobart, Australia.
Background:
Emergency laparotomy is associated with high rates of morbidity and mortality. Accurate, individualised risk prediction models can be used to improve shared decision-making, discharge planning and enhance patient flow. This study used the ANZELA-QI database to apply novel machine learning models to stratify the risk of adverse outcomes in patients undergoing emergency laparotomy.
Methods:
Data were extracted from the ANZELA-QI database. Three machine learning techniques were employed: logistic regression, XGBoost and random forest. Selected clinical and demographic predictor variables were used to train and test the machine learning models in the prediction of mortality, post-operative ICU admission, non-return home and prolonged hospitalisation.
Results:
A total of 8615 cases from 35 hospitals was available from the ANZELA-QI database. Complete data were available in 5195 cases for mortality, ICU admission and non-return home outcomes, and 4469 cases for length of stay. In this cohort 2175 (42%) were admitted to ICU, 601 (12%) died, 1483 (29%) did not return to usual place of residence and 2983 (67%) were admitted for over 1-week post-operatively. Machine learning models demonstrated the greatest accuracy in the prediction of ICU admission and length of stay. The sensitivity and specificity for ICU admission were 0.7 and 0.74, respectively. For admission longer than one week, the overall accuracy was 75%.
Conclusion:
This study applied novel machine learning programs to the ANZELA-QI database to develop risk stratification models for adverse outcomes in patients undergoing emergency laparotomy. The results showed high accuracy for the prediction of prolonged length of stay and post-operative ICU admission.
