VocalScan

Inspiration

We got the idea from seeing how tough Parkinson’s can be for people we know; especially how hard it is to catch early and how little day-to-day info doctors actually get. We wondered: what if everyday things like your voice or a quick spiral drawing could give clues about Parkinson’s? VocalScan tries to make that possible, using just a phone and a browser. It’s not meant to just diagnose it also helps track treatment and progress to support both patients and clinicians, while keeping privacy at the forefront.

What it does

With VocalScan, you can:

• Record a short voice sample and draw a spiral in your browser.
• Enter basic info (like age/gender) to make results more accurate.
• Get an estimated Parkinson’s risk level (LOW/MEDIUM/HIGH) with confidence scores.
• See your past results on a simple dashboard combining voice and spiral data.
• Get personalized tips, insights, and treatment care based on your results.
• Communicate with an AI coach and multiple app functionalities for daily care

How we built it

• Backend: Flask APIs with Firebase Auth for login, Firestore for storing data, and Cloud Storage for audio and images.
• Machine learning: Combined:
  • Voice features (using tools like praat-parselmouth for analysis).
  • Spiral drawings (processed with a CNN model).
  • Basic demographic info.
  • Fused everything together with a Fusion model for a final prediction.
• Frontend: Flask templates + vanilla JavaScript for recording, drawing, and showing results.
• Models: Trained in PyTorch with export-ready versions for deployment.

Challenges

• Combining different types of data (voice, images, demographics) into one model.
• Making the system fast and responsive while running ML models.
• Designing an assessment flow that’s clear and user-friendly.
• Keeping authentication and file uploads secure.
• Finding viable datasets for Parkinson's detection

What we’re proud of

• A full pipeline: from voice and spiral input to final results, all in one place.
• A system that still works even if one input is missing.
• A clean, simple UI that guides users through the process.
• Proper authentication and secure encrypted handling of health-related data, all stored in a database
• A modular ML setup that’s easy to update and improve.

What we learned

• Combining multiple data types leads to better, more stable results.
• User experience matters as much as accuracy; if it’s not easy to use, people won’t use it.
• Firebase is powerful for real-time apps but needs careful planning for complex data.
• Even with deep learning, traditional feature engineering still adds value.
• Designing for missing or incomplete data is essential in real-world use.

What’s next

• Faster model deployment (ONNX), bigger and better training data, and more advanced spiral analysis.
• New features like detailed trends, medication timing insights, and smarter coaching.
• Tools for doctors: dashboards, automated reports, and clinical validation.
• Scaling up with mobile apps, offline use, and privacy-preserving updates.
• Long-term studies to validate accuracy and fairness across different groups.
• Healthcare Provider API Integration

Built With

• firebase
• flask
• javascript
• python
• pytorch
• tailwind
• vanilla