I recently made an appointment to the doctor that had the earliest schedule in two weeks from now! Healthcare services are expensive for uninsured people, which have a higher probability to require special financial needs. Friends in other developing countries have told me that the most affordable COVID19 tests can take up to 12 days to be processed! With these and other additional examples, we believe the healthcare industry could use supporting tools based on tech and especially Machine Learning to speed the detection and treatment of several diseases and reduce the operational costs associated to making an accurate diagnosis.
What it does
We developed Doctor AId. It aims to serve as a support to speed-up the diagnosis of (for now) COVID-19, Diabetic Retinopathy and Respiratory Anomalies in the Lungs. It uses several Machine Learning methods, including CNN classifiers using VGG-19 and Res-Net-50. These predictions are served at scale using server-less functions. THe results are stored in a Cloud database do that Physicians can access the records and results for a specific patient and provide their verdict on the disease. Our platform can be used as a B2B (Product to Hospitals/Clinics) or B2C (Hospital to Customer/Patient).
The platform currently classifies for positive or negative results of COVID-19 based on chest x-rays, the severity of Diabetic Retinopathy from images of Retinas and provides word diagnosis of lungs anomalies from chest X-Rays using NLP through Visual Attention on a CNN-RNN structure.
How we built it
- The Back-End used primarily Node.js and Express.js among other libraries
- The server less APIs were built in Python 3 via Tensorflow and Keras and deployed in Google Cloud Functions. This also required the Google Cloud Build API
- Cloud Storage was used to store the information about the Classifiers and Machine Learning models to recreate the models in the cloud to make predictions.
- The database to store the records of the patients and their results was hosted in Firebase via Google Cloud Platform
- We started to work in hosting the application through Google Cloud App Engine.
Challenges we ran into
- Deploying the Google Cloud Functions effectively. Some libraries are not available in the cloud, so we had to do pip freeze and upload some of the dependencies as deployment packages.
- Connecting the Node.js application to Firebase
- Retrieving the data posted on Firebase
Accomplishments that we're proud of
- Connecting to the Node.js application to the Firebase database!
- Successfully deploying the COVID19 cloud function.
What we learned
- Many new ways to use Bootsrap 4
- Using Firebase with Real Time Database
- Hosting an application in Google Cloud App Engine
- How Hackathons work, this is the first Hackathon for Elise and the second for Alfonso.
What's next for Doctor AId
- Improving the detection algorithms to be more precise and accurate
- Improve the access to patients by the doctors and connect to other medical records from other platforms
- Add cybersecurity protocols
- Expand the Machine Learning detection algorithms to more diseases.
- Scale the platform to be accessible for developing countries and measure the impact on reducing the costs on the healthcare system.