NeoLume

Helping parents and healthcare providers detect neonatal jaundice early with AI-powered monitoring

Comment

NeoLume

Because Every Baby Deserves a Healthy Start

Pitch Deck: https://bit.ly/NeoLume

💡 Inspiration

Jaundice affects nearly 60% of full-term and 80% of preterm newborns globally, and can lead to irreversible brain damage or death if untreated. This issue is especially critical in under-resourced regions like Sub-Saharan Africa, where access to early diagnosis and treatment is limited. In some parts of Africa, severe neonatal jaundice occurs at 667.8 cases per 10,000 births, compared to just 3–4 per 10,000 in Europe and North America.

We were inspired to create NeoLume as a low-cost, scalable solution that bridges the medical infrastructure gap using existing mobile networks. Unlike expensive medical-grade bilirubinometers, our system empowers field volunteers and parents to detect early signs of jaundice using just a smartphone and SMS.

🧠 What does our solution do?

NeoLume is an AI-powered platform that detects jaundice in newborns through a photo sent via SMS. Here's how it works:

  • A parent or field healthcare worker uses our mobile-friendly web app to take a photo of a newborn.
  • The image is sent via Twilio SMS to our backend server.
  • A PyTorch-based deep learning model analyzes the image to detect signs of jaundice.
  • If a high likelihood of jaundice is detected, the user is notified via the app or SMS.

Meanwhile, healthcare organizations can access a secure web dashboard to:

  • Track cases geographically using interactive maps.
  • Monitor real-time case data, field worker activity, and medical tent coverage.
  • Gain insights and analytics for planning interventions and resource allocation.

🛠️ How we built it

NeoLume consists of three key components:

1. Frontend Web App (React.js)

  • Built for mobile-first usage.
  • Allows image capture and submission.
  • Integrated with Twilio to send/receive SMS-based image data.

2. Backend API (Flask + PyTorch)

  • Processes incoming images via SMS.
  • Applies pre-processing: brightness normalization, region-of-interest extraction.
  • Runs predictions using a CNN trained on neonatal jaundice datasets.
  • Responds back to users via Twilio and updates the case database.

🔗 Data source: https://zenodo.org/record/7825810
📄 Model reference: https://www.mdpi.com/2673-7426/3/3/37

3. Admin Dashboard (React + Node.js)

  • Provides NGO staff with a portal to track field workers and case clusters.
  • Visualized using Leaflet.js and Map APIs.
  • Displays charts, case counts, and analytics for planning.

✅ Validation

  • Model Accuracy: 86% across varying skin tones and lighting conditions (using 5-fold CV).
  • Latency: Twilio SMS pipeline tested under rural conditions, average latency <5s.
  • Mapping Accuracy: Simulated GPS coordinates successfully linked to cases.

🚧 Challenges we faced

  • Data diversity: Limited datasets across ethnicities and lighting conditions made training challenging. We used data augmentation techniques to address this.
  • GPS precision: SMS doesn’t always carry precise GPS, so we implemented fallback geolocation based on volunteer ID and area tagging.
  • Hardware limitations: Many users rely on older Android devices. Our app was optimized to work under low bandwidth and reduced memory conditions.

🏆 Accomplishments we’re proud of

  • Built a functional AI detection model deployable in low-resource environments.
  • Seamlessly integrated Twilio SMS to bridge the internet accessibility gap.
  • Created an intuitive real-time dashboard for NGOs and healthcare providers.
  • Designed for scalability, with thousands of users and cases in mind.

📚 What we learned

  • How to apply machine learning in healthcare ethically and responsibly.
  • How to build for low-connectivity environments using SMS and lightweight UIs.
  • Importance of cross-functional collaboration across backend, frontend, and AI teams.
  • Significance of privacy and data handling in medical use cases.

🚀 What’s next for NeoLume

🔧 Short-Term Improvements

  • Expand the training dataset with real-world, anonymized photos.
  • Add multilingual support for major West African languages.
  • Implement offline support for remote field workers with intermittent connectivity.

🌍 Long-Term Deployment

  • Partner with NGOs like Wellbeing Foundation Africa and Tiny Hearts Technology to conduct field trials and scale deployments across Nigeria and Ghana.
  • Integrate predictive analytics to detect regional trends and preempt outbreaks.
  • Continue aligning with SDG 3 (Good Health & Well-Being) to combat neonatal mortality globally.

🔗 Resources & Citations

NeoLume is not meant to replace physicians. It is an assistive tool designed to aid early detection in low-resource settings. Final diagnoses should always be confirmed by medical professionals.

Share this project:

Updates