Limited incremental prognostic value of an AI-derived hematoma expansion probability beyond admission clinical and CT features for 30-day functional outcome after spontaneous intracerebral hemorrhage: a retrospective cohort study
This retrospective cohort study of 117 patients with spontaneous intracerebral hemorrhage found that adding an AI-derived hematoma expansion probability to conventional admission clinical and CT features provided no significant incremental value in predicting 30-day functional outcomes compared to standard predictors alone.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
When a sudden bleed occurs inside the brain, known as a spontaneous intracerebral hemorrhage, the clock starts ticking immediately. Doctors rely on a quick brain scan, a computed tomography (CT) image, to see the size of the bleed and the pressure it is putting on the brain. They also look at the patient's level of consciousness and other basic signs to guess how the person will recover over the next month. For years, these traditional clues have been the standard for making those predictions. However, a new wave of technology promises to do more. Artificial intelligence can now scan those same brain images and calculate a specific number: the probability that the bleed will get bigger. The big question for the medical world is whether this new, high-tech number actually helps doctors make better predictions than the old, simple clues they already have.
A team of researchers at Puyang Oilfield General Hospital in China set out to answer this question by looking back at the records of 117 adults who had been treated for this type of brain bleed. They wanted to see if adding the artificial intelligence's prediction of bleeding expansion to the standard list of patient details would improve the accuracy of predicting a poor recovery. A poor outcome in this context means the patient is left with significant disability or passes away within 30 days. The researchers built two different prediction tools. The first tool used only the information doctors gather when a patient first arrives: their age, gender, the size of the bleed, how much the brain is shifted out of place, their level of consciousness, and whether blood had spread into the fluid-filled spaces of the brain. The second tool took that same list and added the artificial intelligence's calculated probability that the bleed would expand.
The researchers tested these tools using a rigorous method to ensure the results were reliable and not just a lucky guess. They found that the standard tool, using only the basic clinical and scan details, was already quite good at predicting who would struggle to recover. It achieved an AUC of 0.81. When they added the artificial intelligence's expansion probability to the mix, the performance did not improve. The new tool predicted outcomes with almost the exact same accuracy as the old one. In fact, the difference was so tiny it was statistically meaningless, suggesting that the artificial intelligence was not providing any new, useful information that the doctors couldn't already see from the initial scan and the patient's condition. The study also compared their standard tool against a well-known scoring system called the ICH Score, which is widely used in hospitals. Their tool performed similarly to this established score, showing no clear advantage for either approach in this specific group of patients.
The findings suggest that for predicting the 30-day fate of a patient with a brain bleed, the complex artificial intelligence calculation of bleeding expansion does not add value beyond what a doctor can determine from a standard admission assessment. The traditional factors, such as the size of the bleed and the patient's level of alertness, appear to capture almost all the necessary information needed for a short-term prognosis. The researchers noted that their study was limited to a single hospital and a relatively small number of patients, so the results are a strong signal but not a final verdict. They concluded that while the artificial intelligence technology is impressive, it did not prove itself to be a necessary addition to the doctor's toolkit for this specific task in this group of people. To know for sure if this technology can help in a wider setting, larger studies involving different hospitals and more diverse groups of patients will be needed.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.