Healthcare AI should reflect
true biology, not data noise.
Founded by Atharva Shirahatti — CharakAstra exists to build the definitive quality-assurance and bias-auditing standard for medical AI, rooted in pre-registered scientific evidence.
“The AUC is not the contribution — the decomposition is.”
The Problem We Solve
Modern diagnostic AI studies routinely report AUC values above 0.90, creating the impression of clinical-grade accuracy. But these metrics reveal nothing about what the model actually learned.
Atharva Shirahatti's research discovered that leading brain tumor AI models achieved high AUC not by learning to distinguish tumor subtypes from imaging — but by exploiting a 21-year average patient age gap between diagnostic groups embedded in the training data.
When this confounder is neutralized, the model's AUC crashes from 0.94 to 0.58 — barely above chance. The model never learned medicine. It learned demographics. CharakAstra exists to catch and eliminate this.
Core Commitments
What we stand for
Unmasking Shortcuts
Diagnostic AI that achieves AUC ≥ 0.94 may do so by exploiting a 21-year patient age disparity rather than actual tumor morphology. We build the tools to detect and eliminate this before clinical contact.
The Decomposition Imperative
AUC alone tells you whether the model ranked correctly — not why. Our 8-Gate decomposition separates true biological signal from demographic, scanner, and site confounders.
Rigorous Method, Zero Hype
562-patient multi-center validation. 544 IBSI-compliant stable features (ICC ≥ 0.75). 5×5 leakage-free nested cross-validation. Pre-registration before analysis begins. Evidence first, claims second.
Multi-Center at Core
A model that generalises within one institution but fails across hospitals is clinically useless. Every CharakAstra gate is designed to detect this failure before deployment.
Regulatory Standards (TRIPOD-AI & STARD-AI)
We align every audit report with TRIPOD-AI and STARD-AI international reporting frameworks, producing regulatory-ready dossiers from the start.
Patient-First Architecture
If a model cannot pass all 8 audit gates, it has no business touching a clinical decision. Our infrastructure exists to enforce this standard across the AI medical research community.
Join the initiative
Partner with Atharva Shirahatti's research team
Clinicians, research groups, and MedTech leaders: work with us to validate the 8-Gate protocol across centers, modalities, and imaging systems.