NeuroFusion — multimodal neuroimaging + biomarker fusion platform for TBI outcome prediction. Public methodology, model cards, and integration reference for research partners.
Active research development. NeuroFusion Phase 1 targets CENTER-TBI and TBICare dataset integration.
methodology/— Multimodal fusion architecture (CT + MRI + biomarkers + EHR + time-series)model-cards/— Model cards for each published NeuroFusion model (TRIPOD+AI compliant)reproducibility/— Reproducibility scripts and synthetic data examplesconfigs/— Example multiverse analysis configs (BC-000 to BC-127 biomarker combinations)
NeuroFusion integrates six clinical data modalities for TBI outcome prediction:
- Structural neuroimaging (CT, MRI) — processed via MONAI; Marshall + Rotterdam scale extraction
- Blood biomarkers (GFAP, UCH-L1, S100B, NSE, NfL, tau, IL-6) — 128 combinatorial combos (BC-000 to BC-127)
- Clinical trajectory (GCS serial, pupillary response, vital signs) — 6-hour interval time-series
- Structured EHR (demographics, mechanism, comorbidities)
- ICP monitoring (TIL score, daily ICP burden — T2/T3 facilities only)
- Functional outcome (GOSE at 3, 6, 12 months)
The multiverse analysis engine (evidenceos-research/evidenceos-multiverse) runs ~25,000 valid analytical cells across 8 axes to characterize how model performance varies with analytical choices — generating Coverage Vibration of Effects (CVoE) as a novel model stability metric.
- CVoE (Coverage Vibration of Effects): first metric to quantify conformal coverage instability across analytical multiverse
- SAFE Set: Rashomon(ε) ∩ Coverage-Valid(δ) — the clinically deployable intersection of near-optimal and coverage-guaranteed models
- Assay harmonization axis: explicit correction for Simoa vs Abbott i-STAT biomarker platform differences
Research contributions welcome from neuroimaging and biomarker scientists. All contributions must pass TRIPOD+AI reporting standards and include a model card. See CONTRIBUTING.md.
- Methodology documentation: CC-BY-4.0
- Reproducibility scripts: MIT
- Model weights: model-specific (see individual model cards)
EvidenceOS Research Lab — research@evidenceos.com
multiverse-analysis-toolkit— Multiverse analysis frameworkcbim-framework— Patient classification ontologyevidence-capsules— Output Evidence Capsule schema