PRISM

Inspiration

Perioperative myocardial infarction (MI) is one of the most significant yet elusive causes of morbidity and mortality in surgical care. They are especially difficult to detect in this setting due to the lack of symptoms in sedated or anesthetized patients, limited sensitivity and specificity of cardiac enzymes, and EKG signs that are subtle or transient. This issue worsens in hospitals in low-resource settings, in which there is a severe shortage of trained healthcare workers to identify complications. An automated mechanism of predicting MI in real-time would be an essential step towards closing the loop and providing timely access to intervention for this life-threatening emergency.

What it does

PRISM is a machine learning algorithm trained to rapidly detect heart attacks based on EKG data routinely collected at hospitals. Trained on hundreds of real patient EKGs, PRISM can differentiate between STEMI and healthy EKG data in real-time with high accuracy.

How we built it

We collected hundreds of myocardial infarction vs healthy EKG measurements, and generated new data from our sample using gaussian noise. We then trained an 18-layer convolutional ResNet on time series data from one lead. Our model learned better than we expected, achieving a high of 95% validation accuracy on a balanced dataset. To make our model accessible, we saved our model and are serving it over the web using Flask.

Challenges we ran into

The biggest challenge we encountered was finding enough publicly available EKG data to train the algorithm. We also had difficulties in selecting the most optimal framework for the algorithm and manipulating the data to properly input into the algorithms.

Accomplishments that we're proud of

Many of us had never worked with machine learning prior to coming to MedHacks this year, so we're very proud of the progress made on this project and how much we've learned in the process of successfully implementing hte algorithm

What we learned

Our group includes members with backgrounds from medicine, computer science, and engineering, so we took the opportunity to educate each other with our respective expertise. We have all learned about the practical implementation of machine learning algorithms and all the associated challenges, as well as challenges currently being faced by clinicians regarding the effective treatment of STEMI.