Autonomous detection, grading, and reporting of postoperative complications using natural language processing.

Area of Science:

• Medical Informatics
• Computational Linguistics
• Health Services Research

Background:

• Surgical complications like venous thromboembolism (VTE) are challenging to identify.
• Natural language processing (NLP) is underutilized for detecting VTE in clinical documentation.

Purpose of the Study:

• To develop and validate an NLP algorithm for identifying and grading the severity of deep venous thrombosis (DVT) and pulmonary embolism (PE).

Main Methods:

• Utilized a "bag of words" NLP approach on surgical patient reports (2011-2014).
• Classified VTE presence and severity using electronic health record data.
• Trained and validated the algorithm on American College of Surgeons National Surgical Quality Improvement Project data.

Main Results:

• Achieved high sensitivity and specificity for DVT (85.1%, 94.6%) and PE (90%, 98.7%).
• Identified imaging documentation as a key factor in discordances.
• Determined that most identified VTE cases required therapeutic anticoagulation.

Conclusions:

• NLP reliably detects postoperative VTE.
• Expansion of NLP for identifying other conditions from clinical notes is recommended.