Domain Knowledge Inclusive Monotonic Neural Network Guides Patient-Specific Induction of General Anesthesia Dosing.

Area of Science:

• Anesthesiology
• Machine Learning
• Predictive Analytics

Background:

• Postinduction hypotension is a significant risk factor for adverse surgical outcomes.
• Anesthetic dosing relies on patient data and anesthesiologist expertise.
• Machine learning offers advanced methods for predicting hypotension, with neural networks showing promise.

Purpose of the Study:

• To develop machine learning models for predicting postinduction hypotension.
• To suggest generalized and patient-specific anesthetic doses for fentanyl and propofol.
• To incorporate clinical domain knowledge into predictive models.

Main Methods:

• Defined postinduction hypotension as mean arterial pressure <65 mm Hg within 10 minutes of induction.
• Utilized a dataset of 201,000 patient records with clinical data and medication history.
• Implemented and compared various classification algorithms, including a novel monotonic neural network (MNN).

Main Results:

• Gradient boosting and neural networks showed high performance but lacked domain knowledge integration.
• The MNN successfully incorporated the monotonic relationship between anesthetic dose and hypotension risk.
• Heatmaps visualized hypotension likelihood for average and specific patients, with MNN providing smoother predictions.

Conclusions:

• The MNN accurately predicted postinduction hypotension, comparable to existing methods.
• The model encodes clinically relevant monotonic relationships for enhanced interpretability.
• This tool aids anesthesiologists in optimizing patient-specific fentanyl and propofol dosages.