title	emoji	colorFrom	colorTo	sdk	python_version	app_file	pinned
ML Stroke Predictor	🧪	pink	indigo	docker	3.10.2	app.py	true

ML Stroke Predictor

This predictor (SVM model) was trained on health data pertaining to strokes. Given certain data, it is inteded to predict whether that data incidates if a stroke is likely to happen or not. The demo was created using gradio.

Demo

Link to Demo on Huggingface

This predictor was made for educational purposes only. It is not intended to and cannot be used to make or predict medical diagnoses.

Description:

This is a supervised learning, classification task, using a Support Vector Machine (SVM). The data used to train this model contains 10 features. The final model utilizes 8 features:

• age
• BMI
• gender
• has heart disease
• has hypertension
• marriage status
• residence type (urban or rural)
• smoking status

I used pandas to preprocess the data and scikit learn to train and test the model. “Stroke” examples represents 5% of the full dataset, and the model is accurate ~94% of the time.

How the demo works

Body mass index (BMI) is one of the features of this predictor. A user is first asked to submit a weight and height to calculate BMI.

The user is then asked to submit the following:

• Age
• Gender
• Marital status
• Smoking status
• Resdience type
• If they have hypertension (high blood pressure)
• If they have heart disease

Finally, a prediction is made of whether the entered data is likely to result in a stroke or not.

Using this model

To run locally:

Clone this repository.

$ git clone https://github.com/jaded-gloryy/ml-stroke-predictor.git

Run the following command. A link to the demo will be generated as output.

$ python app.py

Challenges:

Abstractions

This was my first time implementing abstractions throughout the programming process. Typically I get function happy, only to come back later and implement classes. Throughout this project I was able to see and experience the benefits of building abstractions right into the process through the encoder classes I incorporated.

UI Interface

This was my first time using gradio (or any user facing interface for a ml model) along with balancing error handling from user inputs vs choosing interfaces with the proper constraints built in. Figuring out where to call out or raise an error was also a new challenge due to the different abstractions I implemented. I learned how to build the interface for my model separate and independent of using the model itself.

Integrations

Integrating my GitHub repo with a HuggingFace "Space" was a bit challenging. I wasn't initially familiar with the process but I was able to find documentation on how to sync my repo with a HuggingFace repo using GitHub actions. Now everytime I push code to my "main" branch it will perform an upload to the "space" on hugging face. After getting those to work seamlessly, and adding the necessary items for integration (yml in the ReadMe and an app.py file), I did some troubleshooting to ensure that proper versions of all my dependencies were installed and there were no conflicts.

Future work

Going forward I’d like to test a few different models to compare their performance to this SVM. I’m particualry interested in producing confidence intervals around these predictions. I would also like to look deeper into my error rate, ie. what percent of predictions are false positives (a stroke is predicted when no stroke occurred) and also negatives (no stroke is predicted when a stroke has occurred).

File summary

Dockerfile: Contains logic to create a Docker Image.

demo.ipynb: Contains logic for the interface of this project.

api.py: Contains api logic for this demo, including requesting BMI caluclation, requesting a prediction, and supplementary validation functions.

classes: Contains logic for encoding categorical input data and formatting.

stroke-prediction.ipynb: Contains logic for preprocessing health data, training the model, and saving a trained model.

model.txt: Contains the trained, serialized model used in the demo.

data: Contains the health data the model is trained on as well as a BMI table used to calculate BMI percentiles for children and teens under 20. (For children and teens, BMI is age- and sex-specific and is often referred to as BMI-for-age )