médical IA

Intraventricular haemorrhage: a "brainstem dorsal line" on the CT scan, read by a tabular foundation model, to predict six-month disability

Published on August 7, 2026 · 11 min read

A nine-hospital Chinese collaboration, coordinated by Southwest Hospital (Army Medical University), proposes the "Brainstem Dorsal Line" (BSDL), a semi-quantitative grade of brainstem compression by fourth-ventricle blood, readable on the admission CT scan. Combined with eight clinical and imaging variables and handed to TabICLv2 — a tabular foundation model that classifies "in context" without conventional retraining — it predicts a poor six-month functional outcome with an AUC of 0.90 on external validation and decent calibration. The signal is real and the marker is clever, but external validation rests on only 118 patients from a single country, the whole study is retrospective, the BSDL is still graded by hand, and the chosen model's edge over its rivals sits within statistical noise.

The context

Spontaneous intracerebral haemorrhage is among the deadliest strokes: it accounts for 10 to 20% of all strokes and remains the second leading cause of stroke death worldwide. When blood reaches the ventricular system — intraventricular haemorrhage (IVH), which complicates 20 to 60% of these bleeds — the prognosis darkens further: 30-day mortality reaches 40 to 60%, and only a quarter of survivors regain functional independence at six months. Neurological recovery plays out mostly in the first three to six months, hence the value of an early, reliable prognosis to guide the intensity of care and inform families.

Existing tools are crude. Clinicians rely on the Glasgow Coma Scale (GCS, scoring depth of coma from 3 to 15), haematoma volume, the presence of hydrocephalus, and the Graeb scores (original, oGS, and modified, mGS), which quantify the blood burden in the ventricles. These indicators come mostly from univariate analyses and poorly capture IVH's multifactorial pathophysiology. Above all, there is no prognostic model designed specifically for IVH — most work treats it as a mere binary variable — nor any standardised way to quantify a phenomenon that is nonetheless central: the mass effect of fourth-ventricle blood on the brainstem, a vital structure sitting just in front.

The method

The paper's first idea is an imaging marker. The Brainstem Dorsal Line is read on a sagittal reconstruction of the non-contrast admission CT: the dorsal contour of the brainstem is traced and the "shift distance" (SD) is measured — the maximum perpendicular gap between the compressed brainstem and its theoretical undeformed position. Three grades follow: grade 0, no or little blood in the fourth ventricle, a straight line; grade 1, blood present but no clear deformation, SD of 2 mm or less; grade 2, a bulky haematoma pushing the brainstem forward, SD above 2 mm. Two neuroradiologists blinded to outcome graded each scan, with a third resolving disagreements.

The BSDL then joins eight other variables kept by a three-step selection (univariate filtering, removal of overly correlated variables, then LASSO regression — a method that zeroes out the coefficients of the least useful variables). The nine final predictors are: BSDL grade, GCS, parenchymal haematoma volume, blood volumes in the lateral, third and fourth ventricles, acute hydrocephalus, oGS and mGS. The target is six-month functional outcome, measured on the modified Rankin Scale (mRS, from 0 for no symptoms to 6 for death) and dichotomised into favourable (mRS 0-3) or unfavourable (mRS 4-6).

The algorithmic core is TabICLv2 (Tabular In-Context Learning). It is a tabular foundation model: rather than being trained from scratch on the cohort, it was pre-trained on vast collections of synthetic tabular datasets, and it classifies a new patient by in-context learning — it is shown the training examples and predicts in a single pass, without a dedicated gradient descent. This family (which also includes TabPFN, tested here too) is reputed to be effective precisely on small clinical datasets. The authors compare it to seven other models: logistic regression, random forest, SVM, XGBoost, LightGBM, a neural network and TabPFN. Evaluation goes beyond AUC: it includes AUPRC (area under the precision-recall curve, more honest when positive cases are a minority), calibration (Brier score, Hosmer-Lemeshow test), clinical utility via decision-curve analysis, and interpretability via SHAP values.

The results

Development draws on 610 patients from seven centres, split into 488 for training and 122 for internal validation; external validation covers 118 patients from two other centres. Unfavourable outcomes make up about 28% of the development cohort and 26% of the external cohort. The new marker's signal is obvious from the description alone: BSDL grade 2 affects 47.1% of patients with an unfavourable outcome versus 4.5% of those with a favourable one.

On external validation, TabICLv2 reaches an AUC of 0.90 (95% CI 0.825-0.958), an AUPRC of 0.82, a Brier score of 0.099, a sensitivity of 77.4%, a specificity of 92.0% and an accuracy of 88.1%. Calibration holds: the Hosmer-Lemeshow tests are non-significant (p = 0.59 internally, p = 0.38 externally), a sign that the displayed probabilities roughly match observed frequencies — a quality too many studies neglect. SHAP analysis puts BSDL grade 2 at the top of the contributions, ahead of the GCS and parenchymal haematoma volume, with a dose-response relationship from grade 0 to grade 2.

Translated into clinical terms, keep perspective. Out of 1,000 IVH patients meeting the study's criteria, about 260 will have a poor six-month outcome. At the chosen threshold, the model would correctly flag roughly 200, miss about 60 (called as recovering when they will fare badly), at the cost of roughly 60 false alarms (called unfavourable when they will recover). That is a real improvement over univariate scores, but a model missing nearly a quarter of poor outcomes is no oracle — and this figure describes discrimination, not proof that action changes because of it.

What is good

A concrete, free, physiologically plausible marker. The BSDL requires no extra imaging: it is read on the admission CT every patient already receives. It formalises a clinical intuition — brainstem compression by fourth-ventricle blood — that had lacked a standardised grid. The raw contrast (grade 2 in 47.1% of poor outcomes versus 4.5% of good ones) and its top rank in the SHAP analysis show this is not cosmetic refinement but the model's dominant predictor.

Methodological hygiene above the field's average. External validation on two centres not used for training, explicitly measured calibration (Brier, Hosmer-Lemeshow), decision curves for clinical utility, SHAP interpretability, and reporting compliant with the TRIPOD+AI framework. Many prognostic models stop at a flattering internal AUC; this one goes markedly further, which makes its claims verifiable.

An honest test bench for tabular foundation models. Pitting eight algorithms, including the recent TabPFN and TabICLv2, against each other on a small clinical dataset is useful in itself: it documents that in-context learning holds up on this kind of task, where deep learning is often thought to require large volumes. The preprint is open access (CC BY 4.0 licence), with no declared conflicts of interest, funded by public provincial grants, and both code and data are available on request.

What is less good

A population bias and a narrow external validation. The 728 patients come exclusively from nine Chinese centres, with tight inclusion criteria: age 18-80, ICU admission, parenchymal haematoma of 20 ml or less, near-perfect prior independence (mRS 0-1). The "external" validation stays Chinese and counts only 118 patients, hence wide confidence intervals (the AUC ranges from 0.83 to 0.96). Nothing guarantees transferability to other health systems, other populations, other neurosurgical practices — the classic population bias, here compounded by a modest test sample.

An edge for the chosen model that sits within noise, and a showcase metric. TabICLv2 crushes no one: on training, TabPFN does slightly better (AUC 0.820 versus 0.818); externally, the neural network posts a higher AUC (0.910) but a collapsed sensitivity (48%). The confidence intervals overlap heavily: presenting TabICLv2 as the winner is a trade-off choice, not a demonstrated superiority (a comparator to keep in perspective). And the AUC of 0.90, the most marketable figure, must be read alongside the AUPRC of 0.82 and above all the sensitivity of 77%: nearly one poor outcome in four slips under the radar (a misleading metric).

The risk of a self-fulfilling prophecy, and a still-artisanal marker. Predicting the outcome of a severe haemorrhage from severity markers raises the spectre of confounding by indication: if a heavily compressed brainstem (grade 2) leads to limiting care or declining surgery, then the "poor outcome" becomes partly a consequence of the decision, not just of the lesion. The authors admit to recording treatment strategies without analysing their interaction with the BSDL — an important blind spot for a score ultimately meant to guide intervention. Add that the BSDL is graded by hand by radiologists (subjectivity, inter-observer variability not quantified here), that no automatic segmentation yet exists, and that there is neither prospective validation nor regulatory clearance.

What it changes

For the research community, two contributions. First, a credible lead that tabular foundation models (TabPFN, TabICL) are competitive on small clinical cohorts, which could ease the data-volume constraint. Second, a marker, the BSDL, simple enough to be tested elsewhere, automated by segmentation, and confronted with non-Chinese cohorts. The next steps are self-evident: prospective validation, subgroup analysis, accounting for treatments received, and automating the grading.

For clinicians, nothing changes today: this is a research model, not deployed, with no prospective validation or clearance. The BSDL does, however, have immediate teaching value — it names and grades a phenomenon the neurosurgeon perceives intuitively — and could, if confirmed, sharpen how an admission CT is read. Caution bears on a double pitfall: not turning a score into a fixed prognosis, and not letting a pessimistic prediction alone decide the intensity of care.

For patients and the public, the promise is legible: to give, early and from an already-performed scan, a more accurate estimate of outcome, to inform better and allocate intensive-care resources better. But an estimate is not a sentence, it was learned on a particular population, and a tool of this kind should never be transplanted as-is to another context without fresh validation — nor, were it one day used to steer decisions, escape oversight as software as a medical device.