How Machine Learning Predicted Post-Cardiothoracic Surgery Complications

Samara Rosenfeld
DECEMBER 06, 2018


Recurrent neural networks (RNN) provided significantly better accuracy levels than the clinical reference tool in predicting severe complications during critical care after cardiothoracic surgery, a new study found.
 
Alexander Meyer, M.D., department of cardiothoracic and vascular surgery at German Heart Center Berlin, and his team used deep learning methods to predict several severe complications — mortality, renal failure with a need for renal replacement therapy and postoperative bleeding leading to operative revision — in post-cardiosurgical care in real time.

“For all tasks, the RNN approach provided significantly better accuracy levels than the respective clinical reference tool,” the researchers wrote.

>> READ: Subtle Medical's AI Imaging Tool Earns FDA Clearance

Mortality was the most accurately predicted, scoring a 90 percent positive predictive value (PPV) and an 85 percent sensitivity score. Renal failure had an 87 percent PPV and 94 percent sensitivity score.

The deep machine learning method also showed area under the curve scores that surpassed clinical reference tools, especially soon after admission.
 
Of the data studied, postoperative bleeding was the most difficult method to predict, due to how accurate the predictions were for mortality and renal failure. Postoperative bleeding had a PPV of 87 percent and sensitivity of 74 percent.
 
The team studied electronic health record (EHR) data from 11,492 adults over the age of 18 years old who had undergone major open-heart surgery from January 2000 through December 2016 in a German tertiary care center for cardiovascular diseases.
 
Patients’ data sets were studied for the 24 hours after the initial study, and if any complication occurred, patients were labeled accordingly.
 
Researchers measured the accuracy and timeliness of the deep learning model’s forecasts and compared predictive quality to established standard-of-care clinical reference tools.
 
Meyer told Healthcare Analytics News™ that one of the major findings of this study was that the system developed outperformed all three pre-existing benchmarks. He added that it is possible to work on a real-time uncurated clinical data stream.

With this information, physicians in emergency care units can perform interventions immediately if a patient is experience complications.
 
“Health systems should openly embrace this technology and ideally try to make use of it,” Meyer said.
 
At the very least, health systems can try to get regulations and developments so that this technology can be used.
 
In a clinical setting, technology like this is difficult to implement and generally demands a financial incentive.
 
Hospitals can work with researchers and companies to push this technology forward and gain support from politicians to help provide financial means and ways to attain these tools.  

Get the best insights in healthcare analytics directly to your inbox.

Related
Mobile Machine-Learning Program Predicts Autism with High Accuracy
AI Struggles Across Health Systems, Needs Wide-Range Testing​​​​​​​
Blockchain Technology: A Foundation for the Future of Healthcare?

SHARE THIS SHARE THIS
13
Become a contributor