Aura | Devpost

Doctor SOAP (Subjective, Objective, Assessment, and Plan) Note
Model Results
Forums Page
Body Map of Symptoms
Patient Vitals
Patient Diagnosis with Explanation
Home Page
Justification of Diagnosis
Presentation Front Screen

Inspiration

One of our members, Ilan, lives with Crohn’s disease, a chronic autoimmune condition that causes severe pain, inflammation, and fatigue. For years, he made frequent doctor visits with no clear answers while his health steadily declined. It wasn’t until he was hospitalized that he finally received a diagnosis, at the cost of thousands of dollars in medical bills.

Ilan’s experience revealed a painful truth: millions of patients suffer for years before getting answers, often paying a heavy emotional and financial price.

That’s why we built Aura.

Our mission is to protect the health, mental wellbeing, and financial stability of patients by helping them reach answers earlier, at a fraction of the cost and in a fraction of the time.

What it does

With Aura, we can use the same information doctors source, but train statistical models to identify autoimmune diseases years before medical professionals, at a fraction of the cost, and with 96% accuracy to help people like Ilan live happier and healthier lives.

We admit a machine learning model shouldn't be the only frontier for diagnosis. Instead, Aura analyzes and builds doctor-ready reports with citations from peer-reviewed journals so that medical professionals can determine the best next steps.

How we built it

We built an ETL pipeline that generates meaningful features before training a two-stage XGBoost model. This model is pretrained with public data from ~100k patients and runs predictions each time a patient wants to evaluate their medical records in search of potential immune diseases which doctors can't diagnose. Additionally, we fine-tuned an 8-billion parameter quantized model on general medical data and patient summaries. By masking the doctors' final diagnoses and training strictly on preliminary data, this model successfully predicted the presence of a disease with 80% accuracy.

Challenges we ran into

Patient data is held very tightly by hospitals as-per HIPAA regulations, thus making sourcing data a very difficult task. Additionally, we had to experiment frequently with different types of models until we found XGBoost, which gave us great discriminatory accuracy on unseen data.

Accomplishments that we're proud of

We are proud of our frontend website because we felt the its easy-to-use interface and welcoming design helped patients understand their health better, while providing a bright design for people to enjoy in a time of agony.

What we learned

As non-medical students, our team learnt (and still needs learn) so much about medical data. A few interesting facts we learnt were:

4-7 years: Average time to autoimmune diagnosis
10%: Global population affected
80%: Patients are women
4+ doctors: Average before correct diagnosis
Irreversible damage: Delayed diagnosis costs organs
$17.5B - $35B: Potential saves in unnecessary visits

What's next for Aura

Aura has proven to be pivotal in early diagnosis and to help patients gain a sense of control over their health. With some additional research and polishing, we believe that Aura could have a strong business model in the healthcare industry.