Traumatic brain injury is not well understood in the research laboratory, nor the clinical setting. With increasing attention and research grants—thanks to mainstream media coverage of sports-related concussions—scientists have been making strides in the laboratory to better identify the characteristics of brain injuries. However, in recent years, there has not been as much success in hospitals and other clinical settings.
Now, Rahul Raj and his team of researchers from Helsinki University Hospital (Finland) have applied a 21st-century solution to the problem—an AI-based algorithm that can predict the probability of a patient dying with 80 to 85 percent accuracy.
Patients brought into the ICU with a severe TBI are unconscious, making it especially challenging to accurately diagnose and monitor the patient during care. Many variables are monitored in the ICU, with intracranial pressure (ICP) and cerebral perfusion pressure (CPP) being the cornerstones of TBI intensive care. Despite that, current prognostic models do not include these factors, but rather are static and based upon simple variables that are assessed upon admission.
Raj and his team developed two AI-based algorithms to predict 30-day patient mortality—the ICP-MAP-CPP (intracranial pressure, mean arterial pressure, cerebral perfusion pressure) and the ICP-MAP-CPP-GCS (Glasgow Coma Scale). The ICP-MAP-CPP algorithm predicted 30-day mortality with a discrimination of 67 percent on day one to 81 percent on day five (new predictions given in 8 h intervals after the first 24 h). Meanwhile, the ICP-MAP-CPP-GCS algorithm predicted 30-day mortality with a discrimination of 72 percent on day one to 84 percent on day five.
“The real-time predictions that are based on the dynamic algorithms could be used to alert the physician about subtle neuroworsening and to quantify the effect of different medical and surgical interventions on prognosis,” wrote the study co-authors in the paper recently published in Scientific Reports.
The millions of data points collected per patient in the ICU on a daily basis is overwhelming for any human. While AI can naturally sequence much more data, the data still needs to be interpreted by a human brain. For this reason, Raj stressed the importance of developing an algorithm simple enough to provide actionable big data in a clinical setting. A more complex algorithm would risk overfitting and generalizability, but the researchers do acknowledge that a larger sample size and more robust machine learning techniques could develop the models further to suggest the most optimal course of treatment.
The dynamic algorithms Raj and his team developed are open-source and free to be used for further development.
“With additional multicenter studies, these predictive algorithms are likely to be improved. We believe that an internationally validated algorithm that could capture dynamic changes in prognosis during intensive care could aid clinicians to make more data-driven treatment decisions, potentially improving quality of care,” the researchers concluded.