SEPTEC

Keras ML framework accuracy

Inspiration

Sepsis is life threatening condition involving organ dysfunction caused by a dysregulated host response to infection. Onset of sepsis heralds septic shock and subsequently death. The World Health Organisation reports an estimated 6 million deaths per year from sepsis. According to Centre for Disease Control, it is a contributing diagnosis in one-third of all hospital deaths. Early detection and initiation of antibiotics are critical for improving sepsis outcomes. Each hour of delayed treatment is associated with a 4-8% mortality.

A 2016 task force from the Society of Critical Care Medicine and European Society of Intensive Care Medicine (SCCM/ESICM) defined sepsis as an increase of two or more points in the sequential (sepsis-related) organ failure assessment (SOFA) score. The management strategy focuses on early recognition and prompt initiation of antibiotics. Time to antibiotics is a crucial factor in determining outcomes.

Vital signs are monitored periodically for every patient in the hospital. While the clinician sees these as a snapshots, computers can analyse more subtle trends. We wanted to harness the power of machine learning to detect patterns in vital signs to predict development of sepsis and and alert the clinician to turn on their spidey-sense.

What it does

Septec is a machine learning algorithm that is designed to alert the clinician to the possibility of impending sepsis, thereby decreasing time to diagnosis and antibiotics. It tracks patients' vital signs (Heart Rate, Temperature, Systolic and Diastolic Blood Pressure and Oxygen saturation) and can predict development of sepsis 12 hours before onset with 75% accuracy.

How I built it

The model is currently trained on open source data from the PhysioNet trial and grows more robust as it trains on increasing volumes of data. Scikit-learn data analytics tools are used to preprocess the data. Then, Keras API and tensorflow is used to implement the binary classification machine learning technique. We trained our model on open source data that had been subject to parsing. Then this model was run on a testing data set to assess the accuracy and error rate of our classification. We built a full stack application using React, NodeJS and Express to represent this information. Future steps are to integrate the keras API in nodejs and our full stack app to complete the functionality of the app.

Challenges I ran into

Open source data is rarely well organized or high quality. We had to perform extensive data cleaning to generate meaningful variables and outcomes. Team dynamics tend to get tricky in high stakes situations and navigating them required tact.

Accomplishments that I'm proud of

Choosing to work with extremely talented and insanely self-motivated people right off the bat
Tackling a problem that is clinically relevant with potential to prevent millions of deaths
Over the course of these 48 hours, we explored uncharted softwares and mastered them overnight, running on coffee and camraderie
Successfully navigating the murky waters of open source data

What I learned

Clear communication is crucial. Task delegation with clear deadlines go hand in hand with impromptu cafe visits. Put together, my team members have explored and now built fully functional application in softwares they were previously alien to.

What's next for SEPTEC

The applications for Septec are wide ranging. Notably in the post operative realm where infections are the leading cause of morbidity and mortality, this tool can be used to risk stratify the patient and aid the clinician in early diagnosis.