a-t-em/asthma-prediction

Binary classification model for predicting diagnosis of asthma

This is a submission of the final project for DataTalksClub's MLOps Zoomcamp 2024. The purpose of the project is to demonstrate the ability to create a machine learning pipeline for a self-selected problem that adheres to MLOps best practices as taught during the course (e.g. with reusable code functions, monitoring of training metrics, testing, accessible prediction endpoint etc.)

Background

Asthma is a chronic lung disease that affects the airways, causing them to narrow and swell. Common symptoms include coughing, wheezing, chest tightness, and shortness of breath.

Machine learning methods can help identify early signs of asthma in patients by analyzing features like wheezing and exposure to various environmental allergens and irritants.

Dataset - asthma_disease_data.csv

Source: Rabie El Kharoua. (2024). 🌬️ Asthma Disease Dataset 🌬️ [Data set]. Kaggle. https://doi.org/10.34740/KAGGLE/DSV/8669080

Brief Description: This is a small, curated dataset containing anonymized health data of 2,392 patients, a minority of which have been diagnosed with Asthma Disease.

Model - model.pkl

The model is a simple Random Forest binary classification model that has been fine-tuned using hyperopt over a limited range of hyperparameters.

Since it is a decision-tree based model and the original dataset is well-curated with no missing values, no preprocessing of data is necessary. However, since the dataset is highly imbalanced, with asthma cases being a minority, minority oversampling techniques were used to create a balanced dataset for training.

Model inputs: select features with relatively higher correlation with the target.

Model outputs: an array containing the value 0 for a negative diagnosis (no asthma) and the value 1 for a positive diagnosis (asthma).

Files

exploration.ipynb -- a notebook initially published on Kaggle that conducts exploratory data analysis on the dataset and examines the efficacy of the default Random Forest Classifier. The F1 score of 0.93 yielded by the model trained on a handful of select features on the withheld 20% of the dataset suggests that a decision tree based model is well-suited for the problem.

build_model.py -- a script for loading and preprocessing the dataset from the CSV file and using hyperopt to fine-tune a Random Forest Classifier on it. The best performing model from 10 hyperopt trials is saved as model.pkl, and it is used to make predictions on the test dataset. The predictions are then saved together with the corresponding features in a CSV file under the data folder.

scoring_script.py -- a script for launching a Flask app (web service) locally that allows for posting a data record input in JSON format to a prediction endpoint that scores the input using model.pkl. The app returns a prediction between 0 and 1 (inclusive) within a JSON object.

testing.py -- a script for testing the Flask app from scoring_script.py to ensure that it returns a prediction between 0 and 1 (inclusive). Enter pytest testing.py into the terminal to run it.

evidently-metrics.ipynb -- a notebook creating monitoring reports and dashboard comparing the training dataset and the test dataset (with predictions included) using Evidently. To view the dashboard in the browser locally, enter evidently ui in the terminal after running the entire notebook to create the workspace and the associated dashboard panels and reports.

Usage

Clone the repository and run make setup in the terminal (which requires Make to be installed first). This will launch a virtual environment and install all dependencies from requirements.txt. To launch mlflow server with SQL backend locally and train the model, enter make build_model into the terminal. You can then view the monitoring metrics in your browser at localhost:5000.

To deploy the model as a prediction endpoint, you can launch the Flask app with make launch_app. It would then be available on port 8080 for posting JSON requests. You may test it with custom test cases using make test command that will run the test case in testing.py.

Alternatively, you can run make all to set up the environment, build the model, and launch the Flask app successively; or if the environment is already ready, you can enter make run to only do the latter two.

Dependencies - requirements.txt

To replicate the development environment, clone the respository onto your local machine and run the command pip install -r requirements.txt in your terminal.

The main libaries utilized are listed below:

io
os
pickle
json
requests
pandas
numpy
sklearn
imbalanced-learn
seaborn
matplotlib
mlflow
pytest
hyperopt
flask
evidently
pylint