In 2022, Miami Dolphins quarterback Tua Tagovailoa returned to a game against the Buffalo Bills after sustaining a head injury that the NFL later acknowledged should have been classified as a concussion.
Team medical staff and an independent consultant incorrectly attributed his visible physical instability to a previously reported back injury rather than a neurological issue, despite the NFL and players' association later stating that concussion protocol had been followed.
While rules have since been revised to prevent such a situation from occurring again, the fact remains: diagnosing a concussion is neither quick nor easy.
Research out of FIU aims to change that.
Hear vs. see
It is estimated that more than 50% of concussions in the United States go undiagnosed, and that roughly 70% occur in sports settings. Identifying a concussion during a football or soccer match, professional or otherwise, can be challenging. On the sidelines, evaluation often relies on brief self-reported symptom checks, such as asking whether an athlete has a headache or feels dizzy, before a decision is made about returning to play.
Classified as mild traumatic brain injury, concussions can impair cognitive function but do not appear on CT scans or other imaging. Diagnosis may rely on assessments of vision, eye movements, reflexes and balance.
"The problem is that these tests are not the most accurate," says Christian Poellabauer, a professor in the Knight Foundation School of Computing and Information Sciences.
While severe concussions can be more obvious – such as the reflexive upper-limb "fencing response" exhibited by Tagovailoa on the field during two other episodes – "It's very difficult to find those more subtle cases," Poellabauer says.
"The concern is if you have too many concussions in a row, or if you keep playing and you get another one, that's going to have long-term effects on your health."
Although most concussions typically cannot be seen, Poellabauer and his team have made an important discovery: They can be heard.
Better AI, better results
Poellabauer began studying the correlation between "speech biosignatures" and the most common form of traumatic brain injury some 10 years ago.
"We were looking for ways of measuring or detecting concussions that are a little bit more foolproof," recalls the computer scientist who focuses on developing novel healthcare solutions.
He landed on speech biosignatures, which capture subtle acoustic, phonetic or linguistic biomarkers unique to an individual. While sometimes likened to fingerprints for their uniqueness, speech biosignatures can show changes over time – depending upon the presence of illness or injury or even intoxication – and so do not remain fixed in the way fingerprints do.
Poellabauer's research group took several voice samples from each of hundreds of high school and college athletes, including those involved in boxing, tackle football, lacrosse, rugby and cheerleading, both before and during their respective seasons. The information collected reflected that of test subjects who eventually experienced confirmed concussions and control subjects who did not.
Comparisons between individuals' two sets of data showed differences on measures such as amplitude, frequency (pitch) and vibration for those who had documented brain trauma. Imperceptible to the human ear, the variations were captured by artificial intelligence.
With the rapid improvement of machine learning, Poellabauer's earlier research has accelerated in recent years. The tool is currently able to correlate changes in voice to cases of brain injury with greater than 90% accuracy.
Doctoral candidate Rahmina Rubaiat is picking up the thread by working to identify a single word or sound that could best be used for both baseline and diagnostic samples. Participants in the original study had been recorded verbalizing eight different words, phrases and sounds.
Simplifying the test would allow athletic trainers or other personnel to take a sample easily from each member of a team in the pre-season and store it, on a tablet using an app, for comparison with the same simple test repeated in the event of an incident, Rubaiat explains. Results would indicate the degree of injury – mild, moderate, severe – to inform a rest and recovery period. Retesting could continue until measures return to baseline.
No concussion should ever be ignored, Rubaiat says. "A severe concussion or mild back-to-back concussions, after a few years, may lead to experiencing some other neurological impairments," she says in an echo of Poellabauer's concerns.
Voice: the next frontier?
Pollabauer has expanded his research to examine how a voice-based test might diagnose the onset of neurological diseases such as Parkinson's and Alzheimer's, both of which show changes in acoustic properties of the voice with progression. His work might even help to distinguish when someone with a neurodegenerative disorder might have the added trauma of a concussion from, for example, a fall.
Outside interest has focused on the possible application of his findings to the assessment of neurological impairment caused on the job in fields with inherent physical dangers such as law enforcement, fire service, the military and construction.
