A paramedic gurney flies through the trauma bay carrying an unconscious elderly gentleman. He is already intubated and has a hive of doctors and nurses running alongside, placing intravenous lines and injecting medicine into his blood stream. He’s suffered a serious head injury in a car accident. With every passing minute, blood accumulates in the space above his brain, pressing on vital structures.
It was a cold winter afternoon in 2017, and the patient had been taken to a major regional hospital. When he arrived, the neurosurgeon on call had minutes to counsel the family on the man’s prognosis, and together they needed to decide whether to operate; surgery could save the patient’s life, but it could also commit him to a life dependent on a ventilator and a feeding tube, trapped in a coma or with limited brain function. Sometimes the quality of life matters more than just the presence of it. The challenge is how can doctors and family members make the right decision in these rushed and emotional moments.
It’s a dilemma that confronts neurosurgeons all too often. The decision to operate is complex, as they must weigh numerous factors such as the severity of the brain injury, prognosis, age, and other injuries. Doctors have long relied on relatively simple algorithms to guide their decision-making, but now researchers at a number of institutions around the world are designing and beginning to test artificial intelligence systems to more accurately predict likely outcomes and help surgeons decipher whether to operate on a patient with a traumatic brain injury.
AI is nearing clinical use in many arenas, from detecting cancer in lung scans and a stroke from brain CT images to providing early warning of sepsis or kidney failure in hospitalized patients. Traumatic brain injuries raise the stakes for these technologies: Can they be relied on when the need for speed is so great and the consequences of a wrong decision are so catastrophic? Turning to AI in these cases also raises ethical questions.
AI in medicine has been enabled by the explosion of patient data from electronic health records and more reliable and sophisticated patient monitoring and imaging. With artificial intelligence and machine learning methods, a computer is able to learn by recognizing patterns and details in a sample set of “training” data. It can then apply a new set of rules to subsequent patient data, empowering it to make more accurate diagnoses and treatment recommendations. Machine learning using advanced neural networks can distinguish findings on medical imaging that in some situations matches or even exceeds the ability of trained radiologists.
“In almost every instance, we find physicians can make better decisions when helped by machines,” said Dr. David Bates, chief of the division of general internal medicine and primary care at Brigham and Women’s Hospital, and co-author of a report on using electronic health record data to predict patient outcomes. He is optimistic about the potential of using AI in treating TBI patients: “Whenever there is a clinical situation with severe outcomes and a high proportion of patients with detailed imaging data, it is a good candidate for AI,” Bates said.
What makes TBI treatment decisions especially difficult is the lack of certainty. While a small subset of patients benefit from aggressive brain surgery, many die even with an operation. In some cases, operating on a patient saves his or her life, but he or she might become vegetative or so severely disabled that quality of life is completely lost. Sometimes it’s better to do nothing. And in many instances, patients are unable to express their own preferences due to their injury, leaving family members without clear guidance.
To complicate matters, a study presented by Dr. Theresa Williamson and colleagues at the 2018 American Association of Neurological Surgeons Annual Scientific Meeting demonstrates that doctors struggle in predicting outcomes for TBI patients and are extremely inconsistent when tested with sample patient cases. Another study, from 2010, similarly analyzed neurosurgeons’ decision-making with hypothetical TBI patient scenarios. It found there was wide variability among the tested surgeons about whether it would be appropriate to operate.
The researchers also found that neurosurgeons would be much less likely to desire surgeries for themselves than offer surgery for someone else in a number of TBI scenarios. Perhaps this is because patients and families often see the outcomes as a binary life-or-death situation, while neurosurgeons understand all too well the gray scale of potential outcomes following an operation, including vegetative and severely disabled states. AI might help alleviate some of these difficulties.
Currently, neurosurgeons are using statistical models without AI to help decide what to do. The most commonly used system is the International Mission for Prognosis and Analysis of Clinical Trials in TBI (IMPACT) model, which was derived from a study that evaluated outcomes of more than 8,000 patients with moderate or severe TBI. Neurosurgeons using the IMPACT model input key data points into the algorithm, such as age, pupil reactivity, and CT-scan findings, and the calculator then computes the probability of death or severe disability six months after the accident.
Studies demonstrate that when aided with this tool, surgeons are more consistent in their decision-making. AI has the potential to further improve the accuracy of outcomes predictions. One small study from the University of Vermont, based on data from 100 patients, showed that artificial neural networks have better predictive abilities for TBI patients when compared to seasoned neurosurgeons and residents.
In current trials, AI-capable computers are exposed to databases of TBI patients with up to hundreds of thousands of data point inputs. Advanced computational models then “learn” how to predict outcomes based on information from these large datasets. The data are often retrospective and fed to the machine manually by a data scientist, but in the near future, it’s possible that the training could become integrated with electronic medical records, so the process becomes real-time and automatic.
Experts caution that clinicians should proceed cautiously for now. There are no regulations governing the use of AI in emergency settings, and it’s uncertain who would be liable if something goes wrong. It’s also not clear how these algorithms can be vetted for accuracy, particularly given that some systems continuously learn from newly input data.
Biased data is another concern. Researchers need to ensure that the original “training” data set is representative of the whole population and can be applied broadly. Additionally, with systems that continuously learn, predictions can turn into a self-fulfilling prophecy. For example, if more TBI patients undergo operations based on suggestions from an AI program, the long-term outcomes of TBI patients at that hospital will be changed, and these outcomes may be used as inputs for the AI to learn as time passes. Thus, the model is affected by patterns of practice that are based on the model to begin with.
Another consideration is whether patients and families should be told when a prognosis is determined by machines. Thus far, there are no well-established guidelines.
Despite these challenges, there’s an emerging consensus that AI could be useful for determining a prognosis for TBI patients, and that it could aid clinicians, families, and patients in making a shared decision about the best course of treatment.
These tools “help guide the conversation,” said Dr. Stephen Honeybul, a consultant neurosurgeon at Sir Charles Gairdner Hospital in Australia and an expert on modeling outcomes for TBI patients. But doctors need to make sure to use AI technology as a guiding tool only, he added, and not as a replacement for old-fashioned clinical judgment.
Ultimately, the decision must be made with information that cannot be computed by the latest technology: patients’ and their family’s beliefs about what constitutes an acceptable quality of life.
The family of the elderly gentleman in the 2017 car accident decided to take aggressive measures, and he underwent an emergency operation to remove the blood from the top of his brain. Often this is the case, with family members telling the medical team, “Do everything you can.”
The man’s life was saved, but unfortunately, he had suffered severe injuries to his brain that could not be repaired. He spent weeks in the intensive care unit — tethered to uncomfortable breathing and feeding tubes, pricked with invasive IV lines. Finally, the decision was made to withdraw care and let him die.
The hope for AI is that it could help avoid such endings. Had the family known he was facing almost certain death, even after an operation, they might have chosen to forgo surgery and allow him to pass away quickly and painlessly on the day he arrived at the hospital.