The Problem
The American Heart Association published the first Scientific Statement on Acute Myocardial Infarction in Women. They discovered that:
- Cardiovascular disease is the #1 leading cause of death for women in the United States
- Women have "atypical" symptoms more often than men and are 7 times more likely to be misdiagnosed than men
- High risk abnormalities in ECGs are missed by physicians in Emergency Medicine
The Objective
How can we design a product that will effectively reduce the likelihood of misdiagnosing heart disease in women
- Develop a website and mobile app that enables users to upload their ECG signals and determine whether their results are normal or abnormal.
- Design a digital experience that is very simple, intuitive and unobtrusive.
What is HERA?
The accuracy of a physician's diagnosis improves with their experience in reading
ECG
signals. Artificial Intelligence offers a solution to the problem of misdiagnosis,
particularly
when equipped with a substantial dataset of ECG signals.
Hera is a web-based product that leverages deep learning to predict the likelihood
of a
patient having heart disease based on their ECG results.
Users can upload a photo of their ECG results via our web or mobile application, and
our system meticulously
analyzes these results using a deep learning model to estimate the probability of
heart-related issues.
Core Functionalities of Product
The Backend
The deep learning model is constructed using a Python-based CNN algorithm, comprising
four
convolutional layers and two fully connected layers. To initiate the model training
process,
we ensure that the input data adheres to the required format.
We obtained the ECG dataset from the PhysioNet website and performed preprocessing to
render
it compatible with the CNN in the Keras library.
Following this, we partitioned the dataset into separate sets for training and
testing,
enabling us to evaluate accuracy.
In the case of the Android app, we utilized Xamarin and Firebase for development, while
the
web application was created using a CherryPy server.
The Design
HERA takes its name from the Greek goddess of women's health
Initial Questions
Collecting user data was just as important as dislaying results. Questions like:
- Are you experiencing any chest pains? If yes, for how long?
- Have you been feeling light-headed?
- Are you experiencing unusual fatigue?
were important in determing how the patient might be feeling.
User Interface
It was very important that we kept the interface very simple since most of our users
would be
well in their late 30s
and older. With the project duration being short, I couldn't go over multiple iterations
as
there were other parts of the project to get to.
The product sport a user-friendly, straightforward user interface and include an image
uploading feature.
Our strategy involves storing these images on the server, with the intention of
leveraging
the collected data to enhance our model.
Due to the time constraint, we were unable to implement the functionality of displaying
prompts to the
patients. The final interface ended up being a page to upload your ECG result image.
Landing page of website
Accomplishments
- Winner: Most Efficient Hack
- Winner: Rockwell's Diversity Prize
Going forward
- A login feature to store and secure each patient's data
- A large dataset enhances the precision of deep learning models in predicting abnormalities. Hera will store all the ECG results on our server for the purpose of improving our model's accuracy.
- Implement asking patients for symptoms before collecting ECG result
- Have the appropriate prompt to ensure the upload is in the right format
- Find good dataset that can be used to train the model.