Depth of Anaesthesia Index using Machine Learning
Hybrid Machine Learning Model
Machine Learning • Data Analysis • Model Design
Monitoring depth of anaesthesia during surgery is critical, but existing methods like the Bispectral Index (BIS) can be inconsistent across patients.
The goal of this project was to develop a more reliable index using EEG data, capable of accurately tracking anaesthetic depth over time.
Approach
Instead of relying on a single model, I explored both supervised and unsupervised learning approaches and combined them into a hybrid system.
A linear Support Vector Machine (SVM) was used to produce a continuous prediction aligned with BIS, while a K-means clustering model was used to identify underlying patterns in the data without labels.
These models were combined into a single index using a weighted approach, allowing the strengths of each method to complement one another.
Model Design
Feature selection played a key role in performance. Based on exploratory analysis, features with strong correlation to BIS (particularly x3 and x6) were prioritised, while redundant features were removed to avoid unnecessary complexity.
The SVM model was trained on selected features to produce a continuous DoA estimate, while the K-means model grouped the data into distinct anaesthetic states. Each cluster was mapped to a representative DoA value, allowing it to contribute meaningfully to the final index.
The final model combined both outputs using a weighting factor (α), which was tuned using a validation set to minimise prediction error.
Results
The final model showed strong agreement with the BIS reference, achieving:
- Correlation (r): 0.905
- RMSE: 8.06
- R²: 0.792
Compared to the original SVM-only model, the hybrid approach provided more consistent tracking, particularly during rapid transitions between anaesthetic states.
The final result is a hybrid DoA index that improves alignment with BIS while remaining interpretable and practical. This project reflects my approach to machine learning and how it can be used to solve real problems.
