Due to health problems faced throughout life, such as Vinícius who survived cancer in adolescence (who, by the way, suffered from errors in the decision making of doctors), we decided to create a technological solution to improve the assertiveness of decisions made in hospitals.

What it does

We use simple, cheap and widely accessible information, such as symptoms, vital signs and exams as a simple CBC, to predict the possible outcomes of the patient's condition.

How I built it

Through databases provided by Brazilian hospitals, we were able to apply Machine Learning algorithms to extract a classifier. At first, we made a binary classification ('hospitalization' or 'discharge from hospital'). Then, we made it a little more complex, to predict - in case of hospitalization - exactly which sector the patient would probably be hospitalized ('ward', 'semi-intensive' or 'ICU'). The next step now is to test other modeling approaches such as Autoencoders and bipartite network link prediction. Our intention is to make a temporal analysis, which accompanies the patient throughout the journey, always calculating the probable prognosis to guide the doctor in decision making and help him to assess whether the plan of action outlined is having an effect.

Challenges I ran into

Hospital data is missing, and is EXTREMELY unbalanced. These were two very big challenges we had when building the application.

Accomplishments that I'm proud of

Building such a complex application in such a short time was really an achievement. We built a model to predict if the suspect patient of Covid-19 will need hospitalization and a model that makes this prediction with 85% accuracy.

What I learned

I learned about Metrics for Imbalanced Classification, learned about different AI approaches. I learned about different approaches to AI to solve problems that are usually viewed differently. For example, one of the attempts of rating, I used the same system that Netflix uses to recommend movies based on user preference. Thinking outside the box in terms of algorithms brought a very big gain.

What's next for trIAge

The next steps of trIAge are now:

  • join the developed model with the prototyped front-end (not yet developed)
  • increase the database for other applications the part of Covid-19
  • co-create the product in partnership with hospitals and health operators, bringing greater assertiveness to the creations.
  • Predict not only what sector of hospitalization the patient will need, but also the entire temporal journey of the patient BETWEEN the sectors, generating greater predictability of resources for the hospitals.
Share this project: