Our app, StrokeGuard, is designed to empower individuals with the tools and knowledge they need to prevent strokes. By combining the latest in medical research with user-friendly technology, StrokeGuard offers personalized health assessments. Users can easily track their blood pressure, BMI, glucose levels, and other vital health metrics through the app. The app will include educational resources to help users understand the signs of a stroke and what actions to take in an emergency, promoting awareness and proactive health management.
Welcome to our data science project! We're diving into a world of numbers and information to help an insurance company in India. They want to know who might be interested in buying travel insurance. To figure this out, we're using a bunch of data they've collected over the years, like who they've talked to and what they've bought. Our goal is to build an ML model that can predict who's most likely to want travel insurance. By doing this, we hope to give the company useful advice on how to reach more people in India who might be interested in buying insurance for their travels.
We will be taking into account that we will be selling the insurance vy 5000₹ (Rupee) this will help to have a forecast as a benchmark and understand better the base case for the project.
In this summary, we will outline the key points extracted from the analysis. We have identified the following:
We observed initial indications suggesting that age is correlated with the likelihood of having a stroke, with individuals above 40 showing increased probabilities. Additionally, we did not find strong correlations between the dependent variable stroke and other categories. However, there is a notable suggestion that being married may correlate with higher stroke risk, likely due to the fact that a significant portion of individuals above 40 are married, thus implying this connection.
Next, we can locate the workbook where everything was developed.
We apply a chi-square test for the categorical variables, and we found that there is no association between hypertension, heart_disease, ever_married, work_type and smoking_status meaning that this feature could not be helpful at the time to predict the target variable. but because this variables make sense we will leave them just because our intuition But for gender and Residence type we could see a significant dependence between the target variable stoke, in this context residence type was a surprising result!
After that we apply a test to verifying the numerical categories, and we found that both values in the both circumstances could have a different distribution when we compared BMI, age and glucose levels; to if they had a stroke, so we could expect that we will be using this features for the predictive models
Here are the conclusions of the Inference analysis
Inference and Modeling results
Here tou can find the workbook of the analysis
We will stick with this 3 metrics for the choosing model
- F1 score
- Recall
- ROC curve.
And we finally get to the results from the models and this are the models choose for the deployment.
| Model Name | Recall | F1 | ROC AUC |
|---|---|---|---|
| GNB_001 | 0.8125 | 0.135417 | 0.748808 |
| ensemble_model_001 | 0.8750 | 0.15053 | 0.756982 |
had the best performance if we compare them with the other models.
Here you had the workbook were we can see the work
We see that some features, like Age, have significant predictive power in these models, which makes sense but not in all cases. We could optimize this variable by applying different transformations. Additionally, features like heart disease and heart attack have evident connections with having a stroke. On the other hand, some features, such as residence type and gender, do not show a strong connection with the target variable and do not contribute much to the prediction capabilities.
Here you had the workbook were we can see the work
Finally, we deploy the model using Streamlit infrastructure, because was easily to manage and deploy, so we could see the app here:
-
Our next step involves requesting additional data. While our current dataset is sufficient, acquiring more data is expected to enhance the models' performance.
-
Furthermore, establishing an infrastructure that allows us to provide weekly updates to the model could further improve its performance over time.
-
Additionally, developing a dedicated in-app chatbot could assist patients by addressing questions related to the topic effectively.
MIT licensed This project is licensed under the [License Name]. See the License.md file for details.