Development and Validation of a Novel Machine Learning Model to Predict the 30-day Mortality of Emergency Laparotomy in the Hong Kong Population

Emergency laparotomy (EL) is a commonly performed procedure and high risk surgery that is known to have a high mortality and morbidity rate. Despite various audits and studies to identify the risk factors and introduce protocols aimed at improving surgical outcomes, the short term mortality after EL remains high. Worldwide data demonstrates that short term (30-day) mortality ranges between 5.3-21.8%, and long term (1-year) mortality rates ranges between 15-47% (Ref 1). Older patients have been identified as the subgroup suffering from highest mortality rates, and efforts implemented in older patients undergoing EL including: the use of risk calculators for mortality prediction, increased peri-operative input from geriatrician and critical care, higher consultant surgeon and anesthetist presence in the operating theatre, and introduction of enhanced care pathways. Apart from age and specialist input, other risk factors for mortality after EL include: frailty, surgical duration, cancer-related surgery, stoma care, patient selection, pre-operative sepsis and physiological parameters, pre-existing comorbidities, ASA status (Ref 2).

Mortality prediction models currently in clinical use for EL include the Portsmouth-Physiologic and Operative Severity Score for the enumeration of Mortality and morbidity (P-POSSUM), Acute Physiology and Chronic Health Evaluation II (APACHE-II), American College of Surgeons National Surgical Quality Improvement (ACS-NSQIP), and the most recent addition of the (NELA) risk calculator. The National Emergency Laparotomy Audit (NELA) performed in the UK since 2012 has been a paradigm shift in evidence-based improvement for patients undergoing EL, demonstrating a reduction in national 30-day mortality rate (11.8% vs.8.7% in 2012 vs. 2012) after identification and implementation of specific recommendations (Ref 3).

Using the data from the large NELA UK cohort between 2014-2016, the NELA risk calculation tool was developed to estimate 30-day mortality, and takes into account patient demographics, ASA status, physiological parameters, vital signs, and details regarding severity and nature of surgical intervention. Multiple studies in the UK, Australia, Singapore have shown the NELA risk calculator is comparable, if not superior, to P-POSSUM for mortality prediction and risk stratification to differentiate between lowand high-risk patients undergoing EL (Ref 5, 6, 7). However, no risk scoring is perfect. The NELA risk model was shown to underpredict, and P-POSSUM to over-predict observed mortality (Ref 8). Since its introduction, NELA has been a pioneer in developing evidence-based interventions and guiding directions for future research in patients undergoing EL, but its implementation in Hong Kong has been limited by lack of validation of accuracy in our patient population.

Frailty is defined as: an objective measure of increased vulnerability and decreased physiological reserve, resulting in accumulation of physiological deficits in multiple systems, and can occur in patients of all ages, but occurs most commonly in older patients. Frailty is a well known risk factor for poor surgical outcomes in EL (Ref 9, 10), but has yet to be incorporated into commonly used risk calculators. There are many risk scoring and surrogate indices for frailty, sarcopenia and osteopenia. Clinical frailty score (CFS) is the most commonly used index for frailty, and CFS alone has been shown to provide prognostic information for patients undergoing EL, but still underperforming compared to NELA. Interestingly, addition of CFS to NELA did not increase the accuracy of the risk model prediction (Ref 11).

The application of deep learning and machine learning is gaining traction, and has been used to develop various risk prediction models and future event prediction (Ref 4). Accumulation of vast datasets from anesthetic records can prove to be a treasure trove for data scientists to uncover new trends and predictions which would previously be overlooked. Risk calculators are helpful tools for clinicians to aid in clinical decision making, but the accuracy and validation of these risk calculators have not been done in this vicinity. Using machine learning algorithms and incorporation of frailty into risk calculators, we hope to develop a novel algorithm with high accuracy and generalizability, to be introduced into clinical use.