Perioperative risk classification, planning, and risk reduction are essential due to the variability of operations and the range of comorbidities of the patients receiving surgery in an emergency environment. In this sense, machine learning has the capacity to provide data-driven forecasts based on the multivariate interactions of tens of thousands of examples. For a study, researchers sought to develop and evaluate cross-validatable interpretable models for estimating post-operative mortality following any emergency surgery on elderly patients.

The study was a secondary analysis of the FRAILESEL study, a multi-center (N = 29 emergency care units), national observational prospective study that collected data between June 2017 and June 2018. The study looked at the perioperative outcomes of elderly patients (age ≥65 years) who underwent emergency surgery. The primary outcome was set at the 30-day mortality, and the following data were gathered: demographic and clinical information, medical and surgical history, preoperative risk factors, frailty, biochemical blood examination, vital parameters, and operative information.

Over 238 (9.26%) of the 2,570 patients that were included (50.66% men, median age 77 [IQR = 13] years) belonged to the non-survivors category. With a test accuracy of 94.9% (sensitivity = 92.0%, specificity = 95.2%), MultiLayer Perceptron was the solution that performed the best. Model explanations demonstrated how non-chronic cardiac-related comorbidities affected everyday living activities and increased the chance of postoperative mortality by lowering consciousness, raising creatinine levels, and decreasing saturation.

A rigorously cross-validated model outperformed current tools and ratings from the literature in the prospective observational analysis regarding prediction accuracy. The model provided critical information for shared decision-making processes necessary for risk reduction by using preoperative data and providing patient-specific explanations.