Project: Predict survival of coronary artery disease patients for preemptive medical treatment

Name: Michael Ng

Folder structure

• data folder
  • place survive.db here
• output folder
  • sample folder
• src folder
  • config.ini
  • pipeline_Classifier.py
  • readConfig_loadData.py
• eda.ipynb
• README.md
• requirements.txt
• run.sh

Install

This project requires Python 3 and the following packages installed:

• configparser == 5.2.0
• pandas == 1.3.5
• numpy == 1.22.0
• matplotlib == 3.5.1
• sklearn == 0.0
• scikit-learn == 1.0.2
• seaborn == 0.11.2

Instructions

---

1. Open the config.ini file in your desired editor. It has these contents and further instructions on permissible values:

   ---

   [DB]
   DBNAME = survive.db
   TABLENAME = survive
   [PARAM]
   ALGO = 1
   TESTSIZE = 0.30
   SEED = 42
   FEATURES = Gender,Smoke,Diabetes,Age,Ejection Fraction,Sodium,Creatinine,Pletelets,Creatinine phosphokinase,Blood Pressure,Hemoglobin,Height,Weight,Favorite color

   ---

   Brief description:

   • DBNAME - [string] fullname of SQL database file to be placed in data folder
   • TABLENAME - [string] name of table in database
   • ALGO - [integer] classification algorithm
   • TESTSIZE - [float] percentage of data to reserve as test set
   • SEED - [integer] random state seed for train-test split
   • FEATURES - [string] concatenated string of selected features for training the model on

   ---

2. At your shell terminal where you have activated your python environment, type "sh run.sh" and enter. Give it a full minute to complete.

   ---

3. It will run the run.py python script in the src subfolder, which performs this series of instructions:

   a. Read config.ini and assign variables (with readConfig_loadData.py module)
   b. Load SQL data
   c. Load to dataframe
   d. Clean data
   e. Run pipeline* (with pipeline_Classifier.py module)
   f. Save plot as JPG in output folder

---

Sample JPG output

---

* Pipeline flow

1. Separate features and labels
2. Split data into training set and test set
3. Select algorithm
4. Generate string list for Categorical Features from p_ls_features
5. Generate string list for Numeric Features from the difference
6. Generate a dynamic dictionary that updates the values when different sets of features are selected
7. Generate integer list for Numeric Features (needed for scaling pre-processing)
8. Generate integer list for Categorical Features (needed for one hot encoding pre-processing)
9. Define preprocessing for numeric columns (make them on the same scale)
10. Define preprocessing for categorical features (encode them)
11. Combine preprocessing steps
12. Create preprocessing and training pipeline
13. Fit the pipeline to train a logistic regression model on the training set
14. Get predictions from test data
15. Calculate ROC curve
16. Format plots arrangement
17. Text display of selected Features shown in 1st subplot
18. Print input parameters in 1st subplot
19. Plot ROC curve in 2nd subplot
20. Plot Confusion Matrix in 3rd subplot
21. Print metrics in 4th subplot

---

Key findings from the EDA

• We want a high Recall value as the cost of a false negative is high. We do not want to miss identifying a patient who truly needs preemptive medical attention (False Negative). Precision is less important as the costs of further tests is not disproportionately costly in risk, time or funds. Accuracy is not ideal too as the population of non-Survivors and Survivors is not symmetrical.
• The most impactful features on predictive ability are 'Ejection Fraction', 'Pletelets', 'Blood Pressure', and 'Hemoglobin'.
• The Random Forest Classifier generates the best predictions, followed by K-Nearest Neighbour.

---

Algorithms available for experimentation

• Classification
  • Logistic Regression
  • Random Forest Classifier
  • Support Vector Machine
  • K-Nearest Neighbour

---

Metrics available

• Accuracy
• Recall (use this)
• Precision
• F1 Score
• AUC (use this)

---

Other considerations

• It would be useful to test the models on data from another hospital to minimise the impact on data entry/regime errors.

---

Updates

• [eda.ipynb] plot_distribution method renamed to histo_boxplot and shifted to new plotter package, in submodule plot_distribution.py
• [plotter package] in __init__.py, import histo_boxplot to save typing submodule plot_distribution
• [plotter package] in submodule plot_distribution.py, reformatted histo_boxplot method to be compliant with PEP-8 using pycodestyle check
• [plotter package] in submodule plot_distribution.py, added docstrings to histo_boxplot method
• [plotter package] in submodule plot_distribution.py, converted histo_boxplot method into Distribution class (retained histo_boxplot method for reference) - benefits are that metrics can now be accessed as instance objects of Distribution class
• [plotter package] in submodule plot_distribution.py, added docstrings to Distribution class
• [eda.ipynb] in cell with "examining the impact of individual features on the predictions", made loop for generating individual feature's output plot
• [eda.ipynb] in cell with features_dict, add comments for all steps
• [readConfig_loadData.py] Changed from using ConfigParser() methods of get(), getint(), getfloat(), to calling its self[key] (see https://docs.python.org/3/reference/datamodel.html#object.__getitem__)
• [plot_distribution.py]: in Distribution class, removed redundant attributes and non-public methods
• [readConfig_loadData.py]: ConfigParser() self[key] returns strings, need to explicitly typecast to other types when assigning to variables
• [readConfig_loadData.py]: removed redundant SQL connection object code replicated in run.py
• [readConfig_loadData.py]: renamed to read_config.py as SQL data loading is not within and to comply with PEP-8 naming convention; updated run.py import of this module
• [run.py]: 'Load SQL data' and 'Load to dataframe' section moved to submodule read_SQL.py as df_from_SQL method
• [clean_data.py]: 'Clean data' section in run.py made into a method and shifted to submodule clean_data.py; updated run.py import of this module
• [read_SQL.py, clean_data.py]: added docstrings
• [pipeline_classifier.py]: split pipeline_classifier method's output plotting and JPG generation into separate plot_model_output method; this permits pipeline_classifier method to simply output the trained model for other uses, and also allows plot_model_output to be used for other models
• [pipeline_classifier.py]: added option of JPG generation into plot_model_output method
• [run.py, eda.ipynb]: updated calling of pipeline_classifier method
• [pipeline_classifier.py]: reformatted methods to be compliant with PEP-8 using pycodestyle check
• [read_config.py]: removed assignment to global variables and made into pop_config_values method, allowing clear visibility of where they are used in run.py
• [run.py, read_config.py, read_SQL.py, clean_data.py, pipeline_classifier.py]: all modules final compliance with PEP-8 using pycodestyle check
• [read_config.py] 'ALGO' and 'TESTSIZE' config.ini errorneous entries will raise ValueError
• [read_SQL.py] 'FEATURES' config.ini errorneous entries will raise ValueError