Applying AI in the medical field is becoming more prevalent and its applications widespread. Nevertheless, there is a long road ahead for any AI system to gain trust by hospitals and for its doctors to use the predictions or insights generated without reserve.
In 2024, a prominent university hospital approached GoFlek for a pilot project to help understand its Cardiac Arrest Emergency Registry (CAER) data which is gathered from trauma cases admitted to hospital emergency. The sample CAER data given had a single target variable (patient survival) and 30 input variables ranging from patient condition at admission, vital signs as well as treatment given over a two days’ period.
The hospital had two main objectives from the pilot project:
- Comparing its overall current practice to an AI system that can “mechanically” predict outcome or automatically find insights like associations or influencers in the data.
- Running WHAT-IF simulation cases and comparing doctors’ recommendation or predictions with those of the system under different scenarios.
The sample two years' data was anonymized during implementation and analysis.
Hospital ... tight spot
The hospital had been for a few years looking at the AI field and searching for a ML system that allow users to run exploration, discovery and prediction tasks, all at the same time and without the need to build different models or retrain them for various scenarios or activities. In other word, one model fits all.
Most importantly, they wanted an “open box” like system that can trace back the probabilistic computations, and where the doctors or analysts can peek and understand how a prediction or a discovery came about. After all, the whole reason for the pilot exercise is to increase trust in AI, gain firsthand experience in the technology, as well as to assess technical complexities for future projects.
GoFlek ... to rescue
For GoFlek this was a great pilot project to run its novel Flek Machine in the medical field. Moreover, it was an opportunity to put to the test its FlekML Engine that was especially designed for this sort of multi–use case: running predictions tasks while at the same time allowing users to collaborate and interact with the same model for exploration or discovery purposes.
Flek Analytics … results
Overall, the pilot project was a success and the probabilistic model built by Flek was able to generate insights or make predictions that confirmed what the doctors expected for about 85% of the cases.
For example, Flek discovered that there is a strong association between the death of a cardiac arrest patient and the number of shocks given as well as whether his heart failed due to asystole. Additionally, being a male patients and not given a shock while presumed cardiac had a great influence on whether the patient died. All these insights and others, were remarkable from the hospital’s perspective because our ML engine did not have any prior medical knowledge or “experience” in the medical field.
Moreover, and from the hospital's view point, the project was also a success because the data analysts and the doctors could easily interact with the system using a variety of WHAT-IF scenarios. Chiefly, checking patient survival under different conditions such as: if the patient was admitted with a known heart condition or if she was transported via an ambulance or if given a specific treatment in the first hour of admission. These parameters and many others were tweaked and used to run different simulation cases.
The entire project took about 3 weeks to complete, including data cleaning and transformation, and then 1 week for WHAT-IF sessions and another for insight validation with the doctors at the hospital.