With a slew of FDA nods this year alone, artificial intelligence can now be widely used in hospital settings to quickly detect the presence of bleeding within the skull or brain. But a new AI tool is aiming to spot signs of internal bleeding much earlier, before a person with head trauma has even arrived at the hospital.
It’s not quite ready for its primetime debut—having only been tested so far in retrospective analyses—but those early results are already proving the tool’s potential to vastly trim down the typical process of examining and triaging a patient with traumatic intracranial hemorrhage.
That process can waste precious treatment time, as patients undergo an initial analysis in the ambulance and then, once they’ve arrived at a hospital, a CT scan. The results of that scan can in turn require another transfer if the initial facility isn’t equipped to treat those findings.
The new machine learning model, which was developed by researchers from Tokyo Medical and Dental University in Japan, aims to cut out those transfers, potentially enabling patients to go straight from the ambulance to appropriate treatment.
The researchers described their work in a study published this month in the journal JAMA Network Open.
To build the predictive AI model, they compiled data from the electronic health records of more than 2,100 patients who were admitted to Tokyo Medical and Dental University Hospital with head trauma. They trained the algorithms to detect links between certain data points and patients who were ultimately found to have intracranial hemorrhages.
The AI was trained to focus specifically on data that are typically collected by ambulance staff as a head trauma patient is transported to a hospital: age, sex, blood pressure, heart rate, body temperature, respiratory rate, state of consciousness, pupil abnormalities, the type of head trauma and any related scarring, the presence of other bodily traumas, post-traumatic seizures, loss of consciousness, vomiting, partial paralysis and clinical deterioration.
After being trained on one set of the total patient group, the machine learning algorithms were put to the test, with a goal of sorting through the remaining 421 cases to predict which ones were ultimately diagnosed with intracranial hemorrhage.
In that analysis, the AI achieved sensitivity of 74% and specificity of 75%, edging out those of a screening method commonly used by hospitals to determine whether a patient admitted with head trauma should undergo a CT scan. Another simpler algorithm that took into account only five criteria—disorientation, state of consciousness, severity of the head trauma, scarring and pupil abnormalities—scored a sensitivity of 74% and specificity of 70%, comparable to the standard screening method.
According to the study’s authors, those results indicate that an AI model could one day be used by ambulance crews at the scene of a trauma to immediately determine the best possible medical facility for a patient, possibly reducing further transfers and time-consuming evaluations.
Once the algorithms have been validated in a prospective study of a larger, multicenter cohort, the researchers said the AI would ideally live as an app on the computers or smartphones of ambulance crews for easy access.