A team of data scientists from the University of Pittsburgh School of Medicine in the US, and neurotrauma surgeons from the University of Pittsburgh Medical Centre, has developed the first automated brain scans and machine-learning techniques to inform outcomes for patients who have severe traumatic brain injuries.
The advanced machine-learning algorithm can analyse vast volumes of data from brain scans and relevant clinical data from patients. The researchers found that the algorithm was able to quickly and accurately produce a prognosis up to six months after injury. The sheer amount of data examined and the speed with which it is analysed is simply not possible for a human clinician, the researchers say.
More on machine learning in medicine: Are machine-learning tools the future of healthcare?
Publishing their results this week in Radiology, the scientists’ new predictive algorithm has been validated across two independent patient cohorts.
Co-senior author of the paper Shandong Wu, associate professor of radiology, bioengineering and biomedical informatics at University of Pittsburgh in the US, is an expert at using machine learning in medicine. The researchers used “a hybrid model machine-learning framework using deep learning and ‘traditional’ machine learning, processing CT imaging data and clinical non-imaging data for severe traumatic brain injury patient outcome prediction,” he tells Cosmos.
Wu says the team used data from the University of Pittsburgh Medical Center (UPMC) and another 18 institutions from around the US. “By using the machine learning model when the patient is admitted early in the emergency room, we’re able to build a model that can automatically predict favourable or unfavourable outcome or the mortality or the other recovery potential,” he says.
“We find our model maintains prediction performance, which shows our model is capturing some critical information to be able to provide that kind of prediction.”
Co-senior author Dr David Okonkwo, a professor of neurological surgery at the University of Pittsburgh and a practising neurosurgeon, also spoke with Cosmos. After presenting the same data to a small group of neurosurgeons, Okonkwo says “the machine learning model significantly outperformed human judgment and experience”.
The success of the first model, based on specific data sets from within the first few hours of the injury, is “extremely encouraging and telling us that we’re on the right path here to build tools that can complement human clinical judgment to make the best decisions for patients,” says Okonkwo. But the researchers believe it can be made more powerful and accurate.
“The first three-day window is very critical for better or for worse for patients with severe traumatic brain injuries. The most common reason for someone to die in the hospital after a traumatic brain injury is because of withdrawal of life-sustaining therapy, and this most commonly happens within the first 72 hours,” Okonkwo says.
“If we can build a model that is based off of that first three days’ worth of information, we think that we can put clinicians in a better place to identify the patients that have a chance at a meaningful recovery.”
The study is one of many using machine learning in different areas of medicine, says Wu. “There are tons of new leading research in the past couple of years, using all kinds of imaging or clinical data and machine learning or deep learning to address many other medical issues, diseases or conditions,” he says.
“Our study as on top of that, another strong study showing, you know, critical care and severe trauma and brain injury population, how our techniques or how deep learning can provide more information, or additional tools to help physicians like David here to provide improved care to patients.” Okonkwo says machine-learning tools are intended “not to replace human clinical or human judgment, but to complement human clinical decision making”.