babyBoss_reinforcement

Inspiration

Throughout this journey, the driving force behind our project was the stark healthcare inaccessibility observed in remote and underprivileged communities. People living in these communities may require a longer time to travel to see a doctor due to poor transportation infrastructure. When prescribing medical treatments, doctors may also find it difficult to prescribe these treatments specific to each individual, given their medical history, and how they have previously responded to previously prescribed medicine. Inspired by stories of individuals struggling to access even the most basic healthcare services, we set out to harness the power of machine learning to bridge this gap, particularly focusing on the management of HIV. This aligns with the 3rd goal of the United Nation's Sustainable Development Goals, which strives to reduce inequalities in healthcare access.

What it does

Our project utilizes an advanced reinforcement learning model to analyze user-provided health data—such as age, gender, ethnicity, CD4 count, and viral load—to generate personalized HIV treatment recommendations. This solution aims to support patients and healthcare providers by offering accessible, data-driven advice for managing HIV effectively.

How we built it

We built our project by first delving into reinforcement learning principles, focusing on creating an environment and agent that could learn from medical data. We utilized Python for the development, incorporating libraries such as Pandas for data management, and Streamlit for creating a user-friendly web application. The integration of these technologies allowed us to process input data, run it through our trained model, and output treatment recommendations directly to users.

Challenges we ran into

One of the main challenges we faced was how to secure the data and how to make use of the data we obtained to design the reinforcement learning algorithm. Given the nature of healthcare data, we are unable to play out a large number of scenarios where the agent makes interventions to learn the optimal policy. It is necessary to be able to learn from observational historical data, which makes designing the algorithm a lot trickier. Moreover, in medicine, we are almost never able to observe every change and possibility in treatment. The model has to estimate the state of conditions of patients without all the data present. Finally, one of the main challenges we faced was designing a good reward function, which balances the short term improvement with overall long term success. We also do not want to have just one reward given based on survival or death of the patients. Hence, it was challenging to determine which actions resulted in the reward or penalty, and how to design the amount of reward or penalty given.

Accomplishments that we're proud of

We are particularly proud of developing a solution that can make a real difference in people's lives by providing them with personalized healthcare advice. By creating our own reward system, the agent is able to determine what actions, or treatments, are more suitable given a patient's state. We also managed to implement more than one algorithm, despite picking up reinforcement learning only on day 1. While we had theoretical knowledge of the subject, we lacked in the technical aspect of it. As such, overcoming technical and ethical challenges associated with this project has been a significant achievement for our team.

What we learned

Throughout this project, we gained a deeper understanding of reinforcement learning and its potential applications in healthcare. We also learned the importance of interdisciplinary collaboration, combining insights from computer science, medicine, and ethics to address complex health challenges. Moreover, we developed a greater appreciation for the intricacies of data privacy and security in healthcare applications.

What's next for babyBoss_reinforcement

Moving forward, we plan to further refine our model by incorporating more comprehensive datasets, potentially including genomic data and drug resistance information, to enhance the accuracy of our treatment recommendations. We can look into short-term and long-term rewards for the agent. We also aim to expand our solution to other diseases and conditions, broadening our impact. Collaborating with healthcare professionals and organizations will be crucial as we seek to validate our tool clinically and explore its integration into existing healthcare systems.

Built With

• python
• streamlit