🧠 Hematoma Expansion Prediction — 20 Years In, Still No Clinical Standard
After two decades of research, predicting hematoma expansion (HE) in ICH remains one of neurocritical care’s most critical unsolved problems.
The evidence shows:
- 📊 NCCT markers (blend sign, black hole sign, hypodensities) → High specificity (~~92–95%) but low sensitivity (~~28–32%) → ~70% of expanders still missed
- • 📈 Composite & CTA-based tools → BAT score: C-statistic ~0.77 → CTA spot sign: pooled OR 8.49 (29 studies) → Yet no routine clinical adoption
- • ⚠️ Definition inconsistency → “Expander” rates vary from 13% to 44% depending on criteria → IVH-inclusive definitions improve outcome prediction, but challenge biomarker consistency
- • 🔬 Latest validation (Zhu et al., Neuroradiology, Mar 2026, n=685) → Combined models maintain AUC 0.82–0.87 across definitions
- • 🤖 AI / Radiomics → Single-center: AUC up to 0.90+ → Multicenter reality: 0.76–0.81 → Persistent overfitting gap
The field needs to go to:
- 🧩 Multimodal integration — imaging + CTA + clinical data in one architecture
- • 📉 Beyond AUC — calibration + decision-curve validation
- • 🏥 Tiered deployment — from NCCT-only settings to comprehensive stroke centers
- At PurpleAI, our HE prediction research is built on this exact framework:
- ➡️ Automated ➡️ Multimodal ➡️ Multicenter-validated
The gap is no longer in model development — it’s in bridging published performance to real clinical utility.
