Summary
A retrospective preprint describes a machine-learning model that combines CT-derived radiomics with clinical information to predict outcomes after spontaneous intracerebral haemorrhage. The authors report improved performance over clinical and conventional imaging variables alone.
A retrospective multicentre preprint describes a machine-learning system that combines features from admission CT scans with clinical information to estimate outcomes after spontaneous intracerebral haemorrhage, a type of bleeding within brain tissue. The authors report that the combined models performed better than models using clinical and conventional imaging information alone. The study was posted on medRxiv on October 2, 2026.
How the system uses CT scans
The researchers used non-contrast computed tomography (NCCT), a scan taken without injecting contrast dye. Deep-learning software automatically segmented the haematoma—the area of bleeding—from each scan. The team extracted 1,690 radiomics features, numerical measurements intended to capture image patterns such as texture and heterogeneity. After screening for reproducibility and selecting features, they combined 42 of them into a score called the Rad-score.
The Rad-score was added to demographic, clinical and conventional CT variables in two machine-learning models: one for functional-outcome prediction and another for prognostic risk stratification. In this approach, radiomics is intended to quantify image characteristics that may not be captured by measuring haematoma volume and broad anatomical location alone.
The abstract reports 2,680 patients, with 1,876 assigned to training and 804 to internal validation. It also reports an independent external validation group of 196 patients. For functional-outcome prediction, the integrated model’s R² was 0.93, compared with 0.35 for the conventional model, while mean squared error fell from 387 to 61. For risk stratification, the area under the receiver operating characteristic curve (AUC) rose from 0.59 for a clinical-only model to 0.95 for the integrated model. The reported AUC in external validation was 0.86.
What the results can tell clinicians
The findings suggest that quantitative features extracted from routine CT images may add prognostic information when combined with clinical variables. If further validated, such a system could support clinicians assessing risk after an intracerebral haemorrhage. The reported statistics describe model performance; they are not evidence that using the system changes treatment or improves patient outcomes.
This is a retrospective study reported as a preprint, and the authors say prospective validation is warranted. The abstract does not specify the functional-outcome measure or the thresholds used to define risk groups, which makes the clinical meaning of the reported scores harder to assess. Its cohort accounting is also ambiguous: the training and internal-validation groups already total 2,680, while the abstract separately lists 196 external-validation patients without clarifying whether they are included in that overall count.