Routine clinical data are sufficient for intracerebral haemorrhage prognostication; haematoma radiomics adds no detectable incremental value: a three-tier validation study

Technical Summary: Routine Clinical Data vs. Haematoma Radiomics in ICH Prognostication

Problem Statement
Spontaneous intracerebral haemorrhage (ICH) is associated with high early mortality and limited functional recovery. While existing prognostic tools like the ICH Score and FUNC Score are widely used, they rely on specialist neurological assessments (GCS, NIHSS) that may not be feasible in all clinical settings. Conversely, CT radiomics has been proposed as an automated alternative to capture quantitative imaging data beyond visual inspection. However, previous studies claiming high predictive performance (AUCs 0.88–0.89) often lack rigorous validation, specifically failing to separate feature selection from model evaluation, leading to potentially optimistic and misleading results. This study addresses the critical question of whether 851 haematoma-core CT radiomics features provide incremental prognostic value over routine admission clinical data when evaluated under a strict, bias-controlled framework.

Methodology
The study employed a single-centre retrospective design involving 570 acute ICH episodes (556 patients) from 2020 to 2024.

• Data Processing: Automated haematoma segmentation was performed using nnU-Net v2, achieving a Dice coefficient of 0.945. From the segmented haematoma cores, 851 IBSI-compliant radiomics features were extracted using PyRadiomics v3.0.1, including first-order statistics, shape features, and wavelet textures.
• Model Development: Two primary logistic regression models were constructed:
  1. Clinical-only: Utilizing 16 pre-specified candidate variables (demographics, labs, and clinical scores) reduced via LASSO.
  2. Clinical + Radiomics: Combining the clinical variables with the 851 radiomics features.
• Validation Framework: To prevent data leakage and overfitting, the study utilized a rigorous three-tier validation approach:
  1. Grouped Nested Cross-Validation (5×5): Feature selection (LASSO) was performed independently within each inner fold, ensuring patient-level separation (GroupKFold) so that recurrent episodes from the same patient never appeared in both training and testing sets.
  2. DeLong Paired ROC Comparison: Used to statistically compare the Area Under the Curve (AUC) of the clinical model against the combined model and established benchmarks (ICH Score, FUNC Score).
  3. Bootstrap Out-of-Bag Estimation: Used to assess feature selection stability and quantify optimism.
• Outcomes: The primary outcomes were 90-day unfavourable functional outcome (modified Rankin Scale >2) and all-cause mortality.

Key Results

• Clinical Model Performance: The clinical-only model, utilizing 10 distinct variables across two outcome-specific models, achieved robust performance with grouped nested-CV AUCs of 0.866 for functional outcome and 0.795 for mortality. This significantly outperformed the ICH Score (ΔAUC = 0.181 and 0.088, respectively) and the FUNC Score for functional outcome (ΔAUC = 0.079).
• Radiomics Incremental Value: Adding the 851 haematoma-core radiomics features to the clinical model resulted in no detectable incremental value.
  • For functional outcome, the combined model AUC was 0.872 (ΔAUC = +0.005, p = 0.805).
  • For mortality, the combined model AUC was 0.771 (ΔAUC = −0.024, p = 0.439).
• Radiomics-Only Performance: A model relying solely on radiomics features performed poorly, yielding AUCs of 0.558 for functional outcome and 0.616 for mortality.
• Feature Stability: Bootstrap analysis revealed that core clinical variables (e.g., NIHSS, age) were selected in 100% of iterations, whereas no radiomics feature exceeded a 90% selection frequency, indicating instability and a lack of consistent signal in the high-dimensional imaging data.
• Overfitting Analysis: The difference between in-sample apparent AUC and nested-CV AUC was minimal for the clinical model (0.026) but larger for the combined model (0.037), suggesting that apparent improvements in the combined model were driven by feature-selection overfitting rather than genuine signal.

Significance and Claims
The authors conclude that for ICH prognostication in settings where CT is the primary imaging modality, routine admission clinical and laboratory data provide reliable risk stratification without the need for complex radiomics pipelines. The study demonstrates that a compact clinical model can outperform established scoring systems (ICH and FUNC scores) and that, under rigorous validation conditions, haematoma-core radiomics features do not confer significant additional predictive power.

The paper emphasizes that the lack of incremental value from radiomics may be attributed to:

1. The high prognostic signal already captured by clinical severity markers (NIHSS, GCS, haematoma volume).
2. The homogeneous nature of acute haematoma cores on non-contrast CT, which limits textural heterogeneity.
3. The exclusion of peri-haematomal oedema (PHE) texture, which has shown promise in other studies.

Crucially, the authors assert that separating feature selection from model evaluation is essential for honest radiomics assessment. They position their work as a corrective to previous literature that may have reported inflated performance due to methodological flaws, advocating for a pragmatic approach where clinical data remains the cornerstone of ICH prognostication unless radiomics can demonstrate robust, validated incremental value in multi-centre settings.