Stealth Mode // Bootstrapped & Research-First

Unmasking Confounders
in Medical AI.

Atharva Shirahatti's 8-Gate auditing framework proves whether a diagnostic model predicts from true pathophysiology — or from a 21-year patient age shortcut.

“The AUC is not the contribution — the decomposition is.”

Request Access Meet the Founder

Core Pillars

Building the world's most rigorous diagnostic safety infrastructure.

Pre-Registered Integrity

Rigorous 8-gate verification preventing demographic shortcuts and scanner bias before clinical deployment.

Clinical Decision Safety

Built to bridge 562-patient multi-center research, hospital workflows, and regulatory compliance.

Reproducible Benchmarking

Strict adherence to IBSI, TRIPOD-AI, and STARD-AI global standardization frameworks.

Critical Discovery

A 21-Year Age Gap Was Driving AI Brain Tumor Diagnosis

Standard literature models reported AUC ≥ 0.94. CharakAstra's 8-Gate protocol revealed the model was not detecting tumor morphology — it was exploiting a covert 21-year demographic age disparity between diagnostic groups. True confounder-controlled AUC: 0.58.

8-Gate Protocol

Founder & Principal Investigator

Architected by Atharva Shirahatti

Pioneering confounder-controlled medical AI auditing to eliminate shortcut learning before clinical deployment.

Atharva Shirahatti

Founder & Lead AI Architect

CharakAstra Research Initiative

“The AUC is not the contribution — the decomposition is.”

— Atharva Shirahatti, on why standard evaluation metrics fail to detect clinical shortcut learning.

Leading the research and technical architecture of CharakAstra. Developed the 8-Gate diagnostic auditing protocol to systematically deconstruct medical AI models, proving whether predictive outputs represent true pathophysiological signatures or covert demographic and scanner shortcuts.

MULTI-CENTER

562 Patients

Validation Cohort

Multi-institutional retrospective and prospective validation across distinct imaging centers and scanner protocols.

IBSI COMPLIANT

544 Features

Stable Features (ICC ≥ 0.75)

Strict radiomics filtering discarding fragile markers sensitive to slice thickness and scanner noise.

ZERO LEAKAGE

100% Leakage-Free

5x5 Nested Cross-Validation

Strict patient-level data partitioning isolating feature selection, preprocessing, and model fitting.

UNMASKED BIAS

21-Year Age Gap

Confounder Discovery

Proved state-of-the-art brain tumor diagnostic models relied on patient age disparity rather than tumor imaging morphology.

The 8-Gate Protocol

Confounder-Controlled Diagnostic Auditing

Atharva Shirahatti's rigorous gate architecture: every medical diagnostic model must pass all 8 gates before clinical consideration.

Data Integrity

Acquisition Standardization & Scanner Invariance

Harmonizing multi-scanner variabilities, magnetic field strengths (1.5T vs 3.0T), and isotropic voxel dimensions.

Audit Benchmark

Harmonized voxel resampling & scanner-invariant intensity normalization.

Causal Disentanglement

Demographic & Confounder Balance

Detecting and isolating systemic age discrepancies (e.g. 21-year gap), biological sex, and referral skew.

Audit Benchmark

Zero demographic correlation with diagnostic target (r < 0.05).

Volume Integrity

Preprocessing & Mask Fidelity

Verifying rigid skull stripping, N4 bias-field correction, and contour delineation free of background marker leakage.

Audit Benchmark

Strict segmentation boundary audit with zero non-tissue contamination.

Radiomics Quality

Feature Stability & Reliability

Filtering radiomic markers under spatial perturbation to discard fragile, scanner-dependent noise artifacts.

Audit Benchmark

544 IBSI-compliant stable features with ICC >= 0.75.

Methodology

Leakage-Free Partitioning

Enforcing 100% patient-level disentanglement so no single subject spans both training and evaluation folds.

Audit Benchmark

5x5 Nested Cross-Validation with isolated preprocessing pipeline.

Model Interpretability

Causal Attribution & Shortcut Detection

Shapley and causal decomposition proving the model learns biological pathology rather than covert demographic cues.

Audit Benchmark

Performance maintained upon confounder neutralization (ΔAUC < 0.03).

Generalization

Multi-Center External Validation

Blinded, independent external evaluation across distinct hospitals and imaging protocols without retraining.

Audit Benchmark

Validated on 562-patient multi-institution external cohort.

Compliance

Regulatory & Clinical Audit Readiness

Comprehensive diagnostic quality dossiers formatted to international regulatory frameworks.

Audit Benchmark

100% adherence to TRIPOD-AI and STARD-AI validation protocols.

Interactive Diagnostic Simulator

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

Published Apparent AUC

--(Initial Report)

True Clinical AUC (Confounders Removed)

--Pending

8-Gate Audit Status

--Gates Verified

Clinical Deployment Verdict

— PENDING

Signal Decomposition Analysis

Atharva Shirahatti's Credo: “The AUC is not the contribution — the decomposition is.”

True Biology (--%) Demographic / Shortcut (--%)

G1: Scanner InvarianceQUEUED

Queued