The Clinical Challenge of Spontaneous Intracerebral Haemorrhage

Managing patients with spontaneous intracerebral haemorrhage remains a formidable challenge in acute neurology. Historically, acute stroke care has focused on early survival and haematoma expansion. However, clinicians recognize that long-term disability, secondary vascular events, and late mortality impose immense burdens on survivors. Therefore, accurate long-term risk stratification is essential for optimizing outpatient therapy and guiding secondary prevention strategies.

Traditional prognostic tools often fail because they rely solely on initial clinical examinations and cranial computed tomography scans. Consequently, these scores neglect progressive underlying microvascular pathologies such as cerebral amyloid angiopathy. Magnetic resonance imaging reliably identifies subtle markers of cerebral small vessel disease. These imaging features provide critical insights into vascular vulnerability. Unfortunately, conventional regression tools lack the flexibility to synthesize complex multidimensional imaging and laboratory data.

To address this gap, modern prognostic workflows must integrate advanced predictive algorithms. Furthermore, clinicians require transparent algorithms rather than opaque black-box systems. Explainable artificial intelligence frameworks now allow physicians to evaluate both cohort trends and individual risk contributions. This shift empowers medical teams to tailor blood pressure goals, surveillance intervals, and antithrombotic decisions with exceptional precision.

Cohort Architecture and Explainable Modeling Methods

To evaluate long-term outcomes effectively, investigators analyzed a registry of 842 acute stroke patients admitted to The Second Affiliated Hospital of Zhejiang University between November 2016 and April 2023. The research team collected baseline clinical profiles, laboratory parameters, and brain magnetic resonance imaging scans. Specifically, neuroimaging analyses captured cerebral small vessel disease features, including white matter hyperintensities, enlarged perivascular spaces, cerebral microbleeds, and cortical superficial siderosis.

Researchers applied the least absolute shrinkage and selection operator alongside stepwise multivariable Cox regression to isolate influential prognostic variables. Subsequently, they constructed multivariable Cox proportional hazards models designed to forecast all-cause death and secondary haemorrhagic recurrence. Model discrimination and calibration underwent validation through concordance indexes, integrated Brier scores, and time-dependent receiver operating characteristic curves.

In addition, the authors eliminated black-box opacity by deploying modern interpretability techniques. They utilized SurvSHAP(t) to capture global feature importance over continuous survival timeframes. Meanwhile, they implemented SurvLIME to generate patient-specific explanations for individual clinical encounters. Consequently, these interpretive methods bridge computational data science and bedside neurology, allowing clinicians to trace how each variable shifts individual survival curves.

Predictive Accuracy for Recurrence and All-Cause Death

During a median follow-up of 36 months (interquartile range: 12 to 51 months), investigators documented 86 deaths (9.1%) and 62 recurrent haemorrhages (6.6%). The developed Cox models demonstrated strong discriminatory power across multiple time points. Specifically, the all-cause mortality model achieved a robust concordance index of 0.841, accompanied by an integrated Brier score of 0.079. These findings confirm high accuracy in identifying patients at risk of late death.

Similarly, the recurrence prediction model delivered excellent prognostic value, attaining a concordance index of 0.759 and an integrated Brier score of 0.063. Predicting recurrent haemorrhage is difficult because complex biological and haemodynamic interactions govern vascular rupture. Nevertheless, this model retained reliability by incorporating neuroimaging biomarkers alongside clinical risk factors. Furthermore, time-dependent area under the curve measurements confirmed sustained predictive stability throughout the multi-year observation window.

Therefore, the statistical findings validate this explainable methodology as an exceptional risk stratification instrument. The models maintain high precision while avoiding overfitting across extended survival periods. Ultimately, these prognostic tools equip clinicians with dependable probabilities, facilitating proactive intervention before devastating clinical events manifest.

SurvSHAP and SurvLIME Insights into Key Risk Factors

The implementation of SurvSHAP(t) and SurvLIME provided clear insights into the determinants of patient survival. For all-cause mortality, global interpretability analyses identified patient age, haemorrhage aetiology, and admission haemoglobin concentrations as dominant predictive drivers. Older patients exhibited reduced survival trajectories, particularly when acute bleeding stemmed from structural or systemic vascular disease. Additionally, low baseline haemoglobin significantly increased late mortality risks, illustrating how systemic frailty compromises physiological resilience.

Conversely, the model identified distinct pathological drivers for haemorrhage recurrence. Here, cortical superficial siderosis and prior intracranial bleeding emerged as the most critical determinants. Cortical superficial siderosis reflects repetitive subarachnoid bleeding episodes, signifying advanced cerebral amyloid angiopathy and severe vascular fragility. Consequently, patients exhibiting these neuroimaging features experienced substantially higher recurrence rates regardless of other clinical comorbidities.

Furthermore, patient-level SurvLIME explanations showed that identical clinical factors contribute variably across distinct clinical profiles. For example, severe microvascular lesions elevated recurrence risks in normotensive patients, whereas persistent hypertension dominated risk trajectories in younger cohorts. Thus, explainable artificial intelligence clarifies the heterogeneous biology underlying stroke recovery, replacing uniform prognostic generalizations with tailored mechanistic insights.

Translating Neuroimaging Markers into Clinical Decision-Making

These empirical findings present vital practical lessons for stroke teams, neurologists, and intensive care physicians. Currently, clinicians frequently struggle to balance ischemic risks against recurrent bleeding when considering antithrombotic therapies. However, by detecting cortical superficial siderosis and quantifying small vessel burden on magnetic resonance imaging, physicians can identify high-risk subsets requiring cautious management.

Moreover, identifying low haemoglobin as a major mortality predictor underscores the importance of addressing systemic medical conditions after hospital discharge. Physicians must treat nutritional deficiencies, gastrointestinal bleeding sources, and chronic anaemia aggressively rather than viewing them as minor laboratory abnormalities. Concurrently, strict long-term blood pressure control remains essential, especially for individuals whose predictive profiles highlight hypertensive arteriolosclerosis.

Finally, integrating explainable machine learning into routine clinical workflows enhances patient counseling and informed consent. When multidisciplinary teams can explain algorithmic predictions to families, shared decision-making becomes far more effective. Therefore, adopting interpretable predictive tools provides a reliable path toward personalized neurovascular management, ensuring that therapeutic intensity matches individual long-term biological risks.

What makes explainable models superior to traditional prognostic scores?

Explainable models surpass conventional scoring systems by integrating multidimensional clinical data and neuroimaging markers through advanced survival algorithms. Moreover, techniques like SurvSHAP and SurvLIME illuminate specific risk drivers for individual patients. Consequently, clinicians gain transparent, personalized prognostic probabilities rather than rigid, generalized cohort estimates, which markedly enhances clinical decision-making.

How do cerebral small vessel disease markers influence haemorrhage recurrence?

Neuroimaging markers of cerebral small vessel disease, particularly cortical superficial siderosis and cerebral microbleeds, indicate severe underlying microvascular fragility from cerebral amyloid angiopathy or arteriolosclerosis. Consequently, patients exhibiting widespread small vessel lesions possess vulnerable vessel walls prone to repeated mechanical failure, thereby experiencing substantially elevated risks of recurrent intracranial bleeding.

Why is low admission haemoglobin linked to higher long-term mortality?

Low baseline haemoglobin often reflects underlying physiological frailty, systemic inflammation, chronic occult blood loss, or severe malnutrition. Furthermore, reduced oxygen-carrying capacity exacerbates secondary ischaemic brain injury and impairs neurological recovery after intracranial bleeding. Therefore, systemic anaemia compromises long-term physiological resilience, resulting in elevated late all-cause mortality among stroke survivors.

Disclaimer: This content is for informational and educational purposes only... Refer to the latest local and national guidelines for clinical practice.

References

1. Yang KC et al. Explainable models for predicting long-term outcomes in patients with spontaneous intracerebral haemorrhage: a retrospective cohort study. Stroke Vasc Neurol. 2025 Oct 27. doi: 10.1136/svn-2024-003864. PMID: 40000031.


2. Greenberg SM, Ziai WC, Cordonnier C, et al. 2022 Guideline for the Management of Patients With Spontaneous Intracerebral Hemorrhage: A Guideline From the American Heart Association/American Stroke Association. Stroke. 2022;53(7):e282-e361.


3. Charidimou A, Boulouis G, Frosch MP, et al. The Boston criteria version 2.0 for cerebral amyloid angiopathy: a multicentre, retrospective, MRI-neuropathology diagnostic accuracy study. Lancet Neurol. 2022;21(8):714-725.