Language-Based Process Phase Detection in the Trauma Resuscitation.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Surgical Process Modeling

Background:

• Surgical process modeling (SPM) traditionally relies on video and sensor data for tracking.
• Spoken language contains valuable semantic information for understanding process progression.
• Trauma resuscitation involves complex communication critical for patient care.

Purpose of the Study:

• To develop and evaluate a deep learning model for predicting trauma resuscitation phases using verbal communication.
• To compare the performance of a language-based model against existing visual-audio systems.
• To explore multimodal approaches incorporating audio data and identify challenges in automatic speech recognition for this domain.

Main Methods:

• Utilized a long-short term memory (LSTM) deep learning architecture to process verbal communication logs.
• Employed a two-stage LSTM process: first for sentence representation, then for sequential group meaning extraction within a time window.
• Trained the model on 24 manually-transcribed trauma resuscitation cases and tested on 6 cases.

Main Results:

• Achieved an accuracy of 79.12% in predicting trauma resuscitation phases.
• Demonstrated superior performance compared to existing visual-audio systems, especially during critical phases.
• Identified significant challenges associated with using automatic speech recognition (ASR) to replace manual transcription.

Conclusions:

• Verbal communication, processed via deep learning, is a viable and effective method for trauma resuscitation phase prediction.
• The proposed LSTM model shows promise for enhancing surgical process modeling, outperforming current multimodal systems in specific areas.
• Further research is needed to overcome ASR limitations for seamless integration into clinical workflows.