A new machine learning model predicts 30-day mortality for brain bleed patients using routine clinical data instead of expensive brain scans, challenging the assumption that advanced imaging is required for accurate prognosis.

When a patient suffers an intracerebral hemorrhage, clinicians face a high-stakes race against time. The standard playbook demands rapid, expensive head scans to guide prognosis. But what if the most valuable predictive data is already sitting in basic, routine lab work?

This study challenges the assumption that we need advanced imaging to predict early mortality. By relying strictly on clinical data collected within the first 24 hours, the researchers show that simpler data can yield highly accurate survival predictions. This shifts the focus from high-tech imaging to smarter, faster analysis of bedside vitals.

Stripping away the scans

The researchers built their models using data from the Medical Information Mart for Intensive Care (MIMIC) database. They trained and tested the algorithms on 1,478 patients from the MIMIC-IV dataset, splitting them into 1,034 for training and 444 for internal testing. To see if the tool could hold up over time, they ran a temporal validation on a separate group of 339 patients from the older MIMIC-III CareVue database.

Out of nine machine learning setups, a LightGBM model performed best. It did not need any imaging data to achieve strong results. This approach builds on previous efforts to model intensive care outcomes, such as a deep learning framework designed to predict ICU readmission after brain bleeds.

How the model performed

• An internal test area under the curve (AUC) of 0.859, showing strong accuracy in distinguishing survivors from non-survivors.
• A temporal validation AUC of 0.811, proving the model remains stable when applied to older patient cohorts.
• An internal calibration slope of 1.017 and a Brier score of 0.132, indicating highly reliable risk probabilities.
• A temporal validation Brier score of 0.174, showing only a minor drop in accuracy over time.

The limits of one hospital

The real value here is the use of SHapley Additive exPlanations (SHAP) to make the model interpretable. Instead of a black box that spits out a risk score, this tool shows clinicians exactly which vital signs or lab values drove the prediction. This transparency is crucial for bedside adoption, as doctors will not trust a tool they cannot audit.

However, we must view these results with healthy skepticism. Because all the data came from a single medical center in Boston, we do not know how this model will perform in rural clinics or different hospital systems. Without true geographic validation, this tool remains a promising prototype rather than a ready-to-use clinical standard. Clinicians should view it as a proof of concept that routine data holds untapped prognostic power.

This analysis is based on research published in Scientific Reports.