Novel Preoperative Risk Stratification Using Digital Phenotyping Applying a Scalable Machine-Learning Approach.

Area of Science:

• Computational biology and bioinformatics
• Medical informatics
• Surgical outcomes research

Background:

• Perioperative risk classification is crucial for patient care and decision-making.
• Electronic health records (EHR) and machine learning can identify digital phenotypes for personalized care.
• Digital phenotypes may predict postoperative adverse events.

Purpose of the Study:

• To hypothesize that preoperative digital phenotypes are associated with postoperative adverse events.
• To develop and validate digital phenotypes for predicting surgical outcomes.
• To compare the predictive performance of digital phenotypes with the American Society of Anesthesiologists (ASA) score.

Main Methods:

• Retrospective analysis of 7251 patients undergoing laminectomy, colectomy, or thoracic surgery over 9 years.
• Extraction of 77 preoperative features from EHR data.
• K-means clustering to identify 3 distinct phenotypes per surgery, followed by random forest model training and validation.

Main Results:

• Three distinct digital phenotypes were identified for each surgery, with varying risk profiles.
• Adverse outcomes, including mortality and prolonged hospital length of stay (LOS), increased progressively across phenotypes (low-risk alpha to high-risk gamma).
• Digital phenotypes combined with the ASA score demonstrated superior predictive accuracy (AUROC) for hospital mortality and prolonged hospitalization compared to the ASA score alone.

Conclusions:

• Three digital phenotypes were successfully identified for three common surgical procedures.
• These phenotypes provide a basis for anticipating adverse postoperative events.
• The integration of digital phenotypes with existing risk scores enhances perioperative risk assessment.