Machine Learning Prediction of Postoperative Mortality in Older Emergency Surgery Patients

Area of Science:

• Surgical Outcomes Research
• Machine Learning in Medicine
• Geriatric Surgery

Background:

• Older adults comprise 35% of emergency general surgery (EGS) admissions.
• This demographic faces higher mortality risks compared to younger patients.
• Accurate prediction of postoperative mortality is crucial for this population.

Purpose of the Study:

• To develop and evaluate a deep mixture of neural networks (DMNN) model for predicting postoperative mortality in older adults undergoing EGS.
• To compare the DMNN model's performance against the established American College of Surgeons National Surgical Quality Improvement Program (NSQIP) calculator.

Main Methods:

• Utilized the 2023 NSQIP database, including approximately one million patient records.
• Included patients aged 55 years and older undergoing EGS.
• Developed a DMNN model using clinical data and biomarkers; assessed performance via AUC, confusion matrices, and Shapley values.

Main Results:

• The DMNN model achieved an AUC of 0.910 in the test set (n=1,000).
• Compared to the NSQIP calculator (AUC 0.936), the DMNN demonstrated significantly higher sensitivity (0.542 vs. 0.119) at a 0.5 probability threshold, with slightly lower specificity.
• The DMNN identified fewer missed deaths but had more false positives; performance varied by risk threshold.

Conclusions:

• The DMNN model shows potential as an adjunct tool for perioperative mortality risk prediction in older EGS patients.
• Further training and utilization of this model could enhance the accuracy of adverse outcome prediction.
• The DMNN model performed adjunctively with the NSQIP risk calculator, offering specific advantages in sensitivity at higher risk thresholds.