How to change healthcare in less than 6 hours?
6 hours. This is the time which the participants of Lungs Decoded Challenge were given to use AI and technology to change how the diagnostics and treatment of lung diseases looks like. Many would say it is an impossible task to do in such a short time (and they probably would be right).
And yet, against all odds, 9 interdisciplinary teams delivered 9 prototypes to show us their vision on how they imagine the future of healthcare.
In this short article, we want to:
- give you a sneak peak into what Lungs Decoded Challenge is,
- what challenges the participants faced,
- and what solution was chosen as the most promising one.
We hope that this article will convince you to take a closer look into our project and data and see for yourself — would YOU be able to change healthcare in less than 6 hours?
What is Lungs Decoded Challenge?
Lungs Decoded Challenge was the hackathon-like event organised by MI2DataLab and CPPC in Warsaw, Poland. Our goal was to invite interdisciplinary teams to work on problems related to lung diseases based on data prepared in our lab (specifically in xLungs — Responsible Artificial Intelligence for Lung Diseases project). The data we shared with participants included x-ray images of lungs, anonymised tabular data about the patients and textual description of the case and the outcome of the diagnosis.
What we think was particularly unique in our event was the idea of interdisciplinary teams and mixed mentorship. Each team consisted of a mix of Data Scientists, ML Engineers, Doctors and UX Designers to make sure that various perspectives will be shown in the solutions.
Moreover, each team had the help of 2 types of mentors, static and moving. Static mentors worked with the team for the whole time, helping with leadership and project management. Moving mentors included medical experts and professional Data Scientists and ML Engineers who helped all of the teams in their problems.
Between the working sessions, participants also had a chance to listen to light talks related to lung disease diagnostics or visual communications. “As far as I know, this is the first hackathon-like event in Poland where medical doctors were explicitly invited to work together with developers on the solutions for healthcare. I’m extremely happy about that”, said radiologist Przemysław Bombiński, MD, during his presentation regarding challenges with lung data analysis.
What challenges did the participants face?
Each team had a chance to solve one of the tasks proposed by us. Tasks reflected various perspectives of different end users such as patients, doctors or even research teams working on similar types of data.
The first task was to create a prototype of the system for the doctors where they would be able to compare similar patients to one another. This task was particularly challenging due to the fact that data were multimodal. This means the participants had to find similar patients using not only tabular data, but also text and image associated with a particular patient.
What was also important from the perspective of the doctor was to create an interface which would be easy to read regardless of the fact how much time the doctor has for studying it (for example sometimes doctors have 5 minutes and sometimes they have 15 minutes to analyse the patient’s case).
Task number 2 was similar to the first one, but this time we wanted to face a perspective of the patient. The goal was to design the prototype of a system in which the patient would be able to receive their diagnosis in a safe and understandable way.
Although it may sound easy at a first glance, in this task the participants faced challenges like:
- How to translate the description of examination written by a doctor to a language understandable for patients in an automatic or semi-automatic way?
- How much information should the patient receive via the system and how much should receive from the doctor during their visit to keep the process safe?
- How to design the system to avoid information overflow?
Last but not least, in task number 3, we wanted to see if the participants are able to automatically translate unstructured, text data prepared by doctors during examination to structured, database-like systems.
Systems like that can be useful for various end users. For research teams, they can make their work faster and more efficient by encoding information into tables to make exploratory data analysis easier. For doctors, tabular data can be faster to analyse in comparison to text. Tabular data can also dictate the standard in which the doctors should input information regarding the patient into the system.
What was the most promising solution?
At the end of the day, each team presented their solution to all participants, mentors and organisers, who voted on the most promising project.
The winning team tackled task 3. They created a Streamlit app that used a GPT-3 based model to translate unstructured text to JSON format. Some of the results they showed were unbelievably impressive regarding the accuracy of the translation and the complexity of information encoded in JSON. To achieve that, they created a template of information that should be extracted and adjusted the model to look for that information in the text to fill the blanks.
Congratulations to the winning team: Dariusz Iwanow, Hanna Smach, Bartłomiej Krzepkowski and Jakub Stępka, as well as their mentor Maciej Chrabąszcz. And to you, our English speaking reader, if you managed to read through our often challenging Polish names. ;)
Conclusions
We strongly believe that Lungs Decoded Challenge proves it’s not all about the time we have, but the team, the engagement and vision of a better future. If you want to read about our solutions, stay tuned for the full report from the event where we will write about all of the brilliant ideas our teams had. You will find more information about it soon on the MI2DataLab LinkedIn page.
If you are interested in other posts about explainable, fair, and responsible ML, follow #ResponsibleML on Medium.
