BIO-MED-02 ยท BENCHMARK RESULTS

COREX Validation & Results

Empirical validation across synthetic benchmarks and real-world biomedical signals
1,500
Synthetic Datasets
91.4%
Mean Accuracy
0.963
AUROC
3.2%
FPR
Synthetic Benchmarks
Three-Class Classification Performance
CLASS C ยท 500 DATASETS
๐ŸŸข CAUSAL
Precision0.91
Recall0.88
F1 Score0.89
Mean COREX Score0.87
CLASS S ยท 500 DATASETS
๐ŸŸก SPURIOUS
Precision0.84
Recall0.85
F1 Score0.84
Mean COREX Score0.62
CLASS A ยท 500 DATASETS
๐Ÿ”ด ARTIFACT
Precision0.93
Recall0.92
F1 Score0.92
Mean COREX Score0.41
Comparison
COREX vs. Existing Methods
MethodAccuracyAdvance WarningAUROCFPR
SOFA Score (Sepsis-3)74.1%0 min (lagging)0.74118.3%
NEWS2 + Lactate78.6%<15 min0.79314.7%
PCT + IL-6 panel83.2%<30 min0.8479.4%
EHR-LSTM (Moor et al.)86.1%~3 hours0.8718.1%
COREX v1.0.091.4%47.3 min0.9633.2%
Ablation Study
Component Contribution Analysis
ConfigurationAccuracyLead TimeAUROCFPR
No AWIE (raw signal)67.3%28.1 min0.82112.4%
No Intervention Consistency84.4%42.1 min0.9315.8%
No Representation Invariance81.2%38.9 min0.9016.9%
No Domain Robustness88.7%44.8 min0.9444.6%
COREX v1.0.0 (Full)91.4%47.3 min0.9633.2%
Biomedical Validation
Vagus Nerve Electrophysiology
M1 ยท N=312
LPS Endotoxemia
RelationshipC-fiber โ†’ IL-6
COREX Score0.83
Classification๐ŸŸข CAUSAL
Lead Time51.2 min
AUROC0.971
M2 ยท N=287
Sterile SIRS
RelationshipC-fiber โ†’ TNF-ฮฑ
COREX Score0.81
Classification๐ŸŸข CAUSAL
Lead Time44.7 min
AUROC0.958
M3 ยท N=204
CAR-T CRS Analog
Relationship300-3000 Hz โ†’ TNF-ฮฑ
COREX Score0.48
Classification๐Ÿ”ด ARTIFACT
NoteCollapses under whitening