Study: ML algorithm helps detect traumatic intracranial hemorrhage using preclinical data

A machine learning algorithm can accurately detect traumatic intracranial hemorrhage using information gathered before patients reach the hospital, according to a study published in JAMA network opened

Researchers built a preclinical triage system using data paramedics could provide, including patient age, gender, systolic blood pressure, heart rate, body temperature, respiratory rate, consciousness, pupillary abnormalities, post-traumatic seizures, vomiting, hemiplegia, clinical deterioration, whether head trauma occurred by excessive force or pressure, and whether the patient has sustained multiple injuries.

The study analyzed electronic medical records of 2,123 patients with head trauma who were transported to Tokyo Medical and Dental University Hospital from April 1, 2018 to March 31, 2021. The machine learning model detected traumatic intracranial hemorrhage with 74% sensitivity and 75% specificity using preclinical information.

In comparison, a prediction model using National Institute for Health and Care Excellence (NICE) guidelines, calculated after consultation with physicians, had a sensitivity of 72% and a specificity of 73%, which was not statistically different from the preclinical model. .

“While conventional screening tools must be examined by a physician, our proposed models only require patient information before transport, which can be easily obtained,” the study authors wrote.

“The results suggest that our proposed prediction models may be useful for constructing a triage system that can be used to assess the optimal setting to which a patient with a head injury should be transported. Further validation with prospective and multicenter datasets is needed.”


Researchers said assessing head trauma in the field could improve outcomes for patients. The current head trauma system requires paramedics to take patients to the hospital if they decide it’s necessary, where a doctor would assess whether a patient needs a CT scan. After a scan, the patient may need to be transported to another hospital.

By adding field triage, ambulances can get patients to the best place for care first, reducing time to treatment.

“Since the functional outcomes of patients with head injuries deteriorate when their transport is delayed, the transport time in step three should be shortened by building a reliable field triage instrument,” the researchers wrote.


As the use of artificial intelligence in healthcare increases, experts and studies have pointed to the importance of monitoring for bias, which could exacerbate existing health inequalities.

AI developers should also: perform thorough testing to ensure that the model works in all environments. Researchers in this lead trauma study noted that this is a limitation of their study, as it focused on a single site in Japan.

“Because this was a single-center study and included only patients who were hospitalized and underwent head CT, our dataset may not represent the general population of patients with head trauma,” they wrote.

“In addition, we suggest that our model may be underestimating high-risk patients, based on the calibration plot. To apply our model to clinical practice, we need to verify the predictive accuracy using a prospective external validation set and investigate the optimal cut-off value.” .”