Confounder-Controlled
Diagnostic AI Auditing
Designed by Atharva Shirahatti, the 8-Gate Audit Protocol is a systematic, gate-locked verification pipeline that every medical diagnostic model must pass before clinical consideration.
“The AUC is not the contribution — the decomposition is.”
Why a gate-locked system?
Sequential Accountability
Each gate is a prerequisite for the next. A model cannot achieve an honest G6 causal attribution score without first passing G5's leakage-free partitioning. The order is not arbitrary — it mirrors the causal chain of how shortcuts propagate.
Binary Verdicts, Zero Ambiguity
Every gate produces a pass or fail verdict against a quantitative benchmark. There is no partial compliance. A model that fails G2's demographic balance gate is not deployable regardless of its AUC.
Regulatory Pre-Alignment
The 8-Gate output is formatted as a regulatory-ready dossier under TRIPOD-AI and STARD-AI reporting standards, providing the documentation regulators require for AI-assisted diagnostics clearance.
Confounder-Controlled Diagnostic Auditing
Atharva Shirahatti's rigorous gate architecture: every medical diagnostic model must pass all 8 gates before clinical consideration.
Acquisition Standardization & Scanner Invariance
Harmonizing multi-scanner variabilities, magnetic field strengths (1.5T vs 3.0T), and isotropic voxel dimensions.
Harmonized voxel resampling & scanner-invariant intensity normalization.
Demographic & Confounder Balance
Detecting and isolating systemic age discrepancies (e.g. 21-year gap), biological sex, and referral skew.
Zero demographic correlation with diagnostic target (r < 0.05).
Preprocessing & Mask Fidelity
Verifying rigid skull stripping, N4 bias-field correction, and contour delineation free of background marker leakage.
Strict segmentation boundary audit with zero non-tissue contamination.
Feature Stability & Reliability
Filtering radiomic markers under spatial perturbation to discard fragile, scanner-dependent noise artifacts.
544 IBSI-compliant stable features with ICC >= 0.75.
Leakage-Free Partitioning
Enforcing 100% patient-level disentanglement so no single subject spans both training and evaluation folds.
5x5 Nested Cross-Validation with isolated preprocessing pipeline.
Causal Attribution & Shortcut Detection
Shapley and causal decomposition proving the model learns biological pathology rather than covert demographic cues.
Performance maintained upon confounder neutralization (ΔAUC < 0.03).
Multi-Center External Validation
Blinded, independent external evaluation across distinct hospitals and imaging protocols without retraining.
Validated on 562-patient multi-institution external cohort.
Regulatory & Clinical Audit Readiness
Comprehensive diagnostic quality dossiers formatted to international regulatory frameworks.
100% adherence to TRIPOD-AI and STARD-AI validation protocols.
Real-Time 8-Gate Audit Verification
Experience how Atharva Shirahatti's 8-Gate protocol unmasks demographic shortcuts in standard high-AUC models.
Awaiting automated verification run
Signal Decomposition Analysis
Atharva Shirahatti's Credo: “The AUC is not the contribution — the decomposition is.”
Queued
Awaiting automated gate verdict.
Queued
Awaiting automated gate verdict.
Queued
Awaiting automated gate verdict.
Queued
Awaiting automated gate verdict.
Queued
Awaiting automated gate verdict.
Queued
Awaiting automated gate verdict.
Queued
Awaiting automated gate verdict.
Queued
Awaiting automated gate verdict.
Request an audit
Ready to validate your diagnostic AI?
Partner with CharakAstra to run the full 8-Gate protocol on your model. We work with clinical research groups and MedTech organisations under NDA.
Request Collaboration