Pipeline & Model Selection

Configure BP prediction pipeline and classification layer for all scans

Currently Active: hybrid

All new scans will use this configuration for BP prediction and classification

BP Prediction Pipeline

Select which pipeline produces SBP/DBP values. Click to activate — changes take effect immediately.

Classification Layer (Post-V4)

Select which classification model assigns the 3-category label (Low-Normal / Normal-PreHTN / Hypertension). These layers run on top of V4 pipeline outputs using 61 meta-features.

V7.1 Meta-Correction is a cascaded pipeline: V6 predicts first → GBC Detector checks if V6 is likely wrong → if flagged, GBC Corrector (trained on misclassified samples only) reroutes to the correct class. Walk-forward accuracy: 91.0% (vs V6's 71.9%, +19.2% improvement). Corrector now differs from V6 on 100% of flagged cases. See for threshold tuning.

User-Facing Display Model

Select which model's output is shown to end users in the BP result card. This controls the category badge, BP reading, and model label displayed after a scan. All models run in shadow mode regardless of this setting.

Users see: V14 Hybrid output

How it works: All models (V4→V6→V8→V9→V14) always run in shadow mode on every scan. This setting only controls which model's output is displayed to the user in the BP result card. The badge will show the selected model name (e.g., "V14 Hybrid", "V9 KNN", or "V4 Pipeline").

Full Pipeline Comparison

Feature	V2	V3	V4	V6	V7
Type	BP + Cat	BP + Cat	BP + Cat	Cat Only	Cat Only
Architecture	Sequential Stacking	Cascaded-Parallel	Routed Ensemble	5-Model Stacking	V6 → Detect → Correct
Category System	4-Cat	4-Cat	3-Cat	3-Cat	3-Cat
In-Sample Acc	79.7%	78.6%	91.2%	95.5%	97.2%
Walk-Forward Acc	—	—	—	71.9%	90.8%
SBP MAE	6.84	6.99	1.35	N/A	N/A
Error Correction	—	L9	V4	—	Meta

V4 Pipeline uses a unified SBP-only 3-category system. Thresholds: Low-Normal (SBP ≤115), Normal-PreHTN (116-137), Hypertension (≥138).

V6 & V7 are classification-only layers that run on top of V4 outputs. They use 61 meta-features (model predictions, probabilities, residuals) but do not produce BP values — BP comes from V4.

Layer 9 Error Correction Backfill

Apply Layer 9 error correction to existing scans that have V3 data but no L9 correction.

V4 Category-Aware Routed Ensemble Backfill

Apply V4 pipeline (Classifier → Specialist Routing → Per-Category Correction) to existing scans.

V7 Meta-Correction Backfill (v7.1)

Re-run V7 meta-correction on existing scans with the latest v7.1 models. The corrector is now trained on misclassified-only samples for better error detection and correction.