Comparative Analysis on Machine Learning and Deep Learning to Predict Post-Induction Hypotension.

Area of Science:

• Anesthesiology
• Medical Informatics
• Machine Learning

Background:

• Hypotension during anesthesia induction poses significant post-operative risks.
• Early prediction of hypotension is vital for patient safety and improved surgical outcomes.

Purpose of the Study:

• To predict hypotension occurring between tracheal intubation and incision one minute in advance.
• To evaluate the efficacy of machine learning and deep learning models for hypotension prediction.

Main Methods:

• Utilized vital and electronic health records (EHR) from 282 patients undergoing laparoscopic cholecystectomy.
• Trained meta-learning models (Random Forest, XGBoost) and deep learning models (CNN, DNN).
• Compared performance using raw vital records versus feature-engineered data.

Main Results:

• On raw data, CNN achieved 72.63% accuracy, outperforming Random Forest (70.32%) and XGBoost (64.6%).
• With feature engineering, Random Forest with feature selection yielded the highest accuracy (74.89%).
• CNN performance decreased to 68.95% with feature engineering.

Conclusions:

• Random Forest, enhanced by feature selection, demonstrated superior predictive performance for hypotension.
• While CNN showed promise, feature engineering significantly impacted model efficacy.
• The data examination period is a critical factor for improving prediction accuracy.