Well Hospital is a fast-growing medical Center in Toowoomba. In the past years, Well Hospital has served Toowoomba local community well. Aiming to increase the safety of surgery, Well Hospital is going to develop a Depth of Anesthesia (DoA)

Assessment Item 1

Background

Well Hospital is a fast-growing medical Center in Toowoomba. In the past years, Well Hospital has served Toowoomba local community well. Aiming to increase the safety of surgery, Well Hospital is going to develop a Depth of Anesthesia (DoA) index based on the features of raw electroencephalograph (EEG) data obtained from the bispectral index (BIS) monitor. Against a group of talented programmers, you really want to win the contract to develop the system. But first, you need to develop a new DoA index based on machine learning techniques and the data provided by Well Hospital.

Data Description

At the beginning of the project, 12 cases of patient data are given by Well Hospital as attachments. Project data set 1 (for reports 1 and 3) includes 8 cases of training data (Train_1 to Train_8) and 4 cases of testing data (Test_1 to Test_4). In each case of training data, there are 7 data sets which include one BIS index (BIS) and 6 feature sets (x1, x2, x3, x4, x5 and x6). In each case of testing data, there are also 7 data sets. But the BIS index in testing sets cannot be used for feature selection or model building. The BIS data is obtained from the BIS monitor, and the 6 feature sets are calculated from the raw EEG data using different feature extraction methods. 

• For each case, all the data sets (BIS values, x1, x2, x3, x4, x5 and x6) have the same number of data points. 
• All the data sets are stored chronologically. For example, BIS(1) is the BIS value of the first second. x2(4) is the x2 feature value calculated from the fourth second EEG data. Each feature value corresponds to its BIS value in the time series. 
• All the feature sets x1 from different cases are calculated by the same feature-extracting methods. So do x2, x3, x4, x5 and x6.

A data set with 4950 segments is given in Project data set 2 (for report 2 working), in which each segment is the data collected from different patients (not in chronological order). For all the 4950 segments, there are 7 data sets (p1, p2, p3, p4, p5, p6 and p7) that can be considered feature sets. The feature sets are calculated from the corresponding segments of raw EEG data using 7 different feature extraction methods respectively (which are different with x1 to x6). However, all the feature data p1 from different segments are calculated by the same feature extraction method. So do p2, p3, p4, p5, p6 and p7.

The labels of 4950 segments were given in Project data set 3 (for report 2 testing). However, you can only use the labels to evaluate your classification results. Don’t use them for feature selection and model building. In this project, you need to design a DoA index based on the training data set using the machine learning techniques you learnt in CSC6003, and then assess the performance of your index based on the testing data set.

Supervised learning Report

In this report, the following contents should be covered:

• Do research for machine learning applications on the DoA assessment. Please summarize other researchers’ outcomes and find out the limitations of different machine learning methods. Please focus on the supervised machine learning application on feature selections and model building for DoA assessment. No need to explain detailed background information on the DoA assessment or EEG feature extraction in your report. 
• Analyse the data sets given in Project data set 1, what are the statistics information of the data and what are the relationship among the data sets? 
• According to the data sets and findings during your research, discuss which machine learning methods you will use in this project and show your reasons.

You need to select two methods from the following options: 

1. Decision tree or random forest 
2. Linear model or non-linear model 
3. K-NN 4. Neutral networks

• According to the findings during your research, discuss which DoA evaluation methods you will use to compare the new index with the BIS index and show your reasons. For example, R square, Pearson coefficient, Bland-Altman plot diagram and scatter plot diagram and so on. 
• Use the two supervised machine learning methods you selected to process the training data sets separately. The feature selection methods and results need to be present in your report logically (the selected feature sets can be either the same or different for the model building of two supervised machine learning methods). 
• Two DoA index design (model building) based on two supervised machine learning methods must be presented clearly and logically in your report. You need to present your methods and results with flow charts, key equation(s), figure(s) or table(s). The programming code(s) and supporting figure(s) or Excel data should be presented in the appendix of the report. The label (BIS value) cannot be any part of your new index. It means the new index is calculated by an equation including feature sets (x1, x2, x3, x4, x5 or x6), not a BIS value. The new index may be similar to: New index = 4*x1+5*(x2)^2

If you use a neural network or support vector machine, you may not provide the equation of the new index as above. For different methods, the modelling presentations are different. Please present the feature selection, parameter setting or kernel function selection clearly.

Assessment Requirements Summary

Objective:
The main goal of this assessment is to develop a Depth of Anesthesia (DoA) index using machine learning techniques based on EEG-derived features, and evaluate its performance against the BIS index using the provided training and testing datasets.

Key Pointers to Cover:

1. Conduct research on supervised machine learning applications for DoA assessment, including limitations and previous outcomes.
2. Analyze the training dataset (Project dataset 1): compute statistics, explore relationships among features (x1–x6), and identify relevant patterns.
3. Select two supervised machine learning methods from the options provided (Decision Tree/Random Forest, Linear/Non-linear model, K-NN, Neural Networks).
4. Perform feature selection for each method and justify choices.
5. Build two separate DoA indices using the selected methods, with clear presentation of models, flowcharts, key equations, and results.
6. Evaluate the performance of the new indices using appropriate comparison methods (R⊃2;, Pearson correlation, Bland-Altman plots, scatter plots, etc.).
7. Provide supporting code, figures, and Excel data in the appendix.
8. Ensure the BIS index is not used in constructing the new index.