Data-driven categorization of postoperative delirium symptoms using unsupervised machine learning.

Area of Science:

• Medical research
• Data science
• Oncology

Background:

• Postoperative delirium (POD) lacks comprehensive phenotyping, hindering understanding of its mechanisms and management.
• Hypothesis-free symptom categorization is proposed as a method to uncover underlying delirium mechanisms.
• Previous evidence for data-driven phenotyping of POD is limited.

Purpose of the Study:

• To explore postoperative delirium phenotypes in patients undergoing invasive cancer surgery.
• To utilize a data-driven, unsupervised machine learning approach with minimal prior assumptions.
• To identify distinct symptom clusters and patient subgroups associated with delirium.

Main Methods:

• Recruited 286 patients undergoing elective invasive cancer resection.
• Assessed delirium daily for 5 days post-surgery using the Delirium Rating Scale-Revised-98 (DRS-R-98).
• Applied K-means clustering to DRS-R-98 scores to derive symptom features and classify patient subgroups.

Main Results:

• Identified four key symptom features: mixed motor, cognitive/higher-order thinking with perceptual disturbance, acute/temporal response, and sleep-wake cycle disturbance.
• Classified 91 delirious patients into seven distinct subgroups based on symptom profiles.
• Derived three delirium clusters, two subsyndromal delirium clusters, and an insomnia cluster using machine learning.

Conclusions:

• Unsupervised machine learning successfully delineated patients into distinct delirium and subsyndromal delirium clusters post-cancer surgery.
• The identified clusters include delirium, subsyndromal delirium, and insomnia phenotypes.
• Further validation and research into the pathophysiology of these clusters are crucial for understanding POD mechanisms.