Atharva Shirahatti
Atharva Shirahatti is the founder of CharakAstra, building the definitive quality-assurance and bias-auditing standard for medical AI models — unmasking demographic shortcuts (e.g. 21-year age gaps and tumor grade skews) in diagnostic machine learning before clinical deployment.
About
Atharva Shirahatti is a researcher and founder buildingCharakAstra — a bootstrapped, stealth-mode research initiative focused on confounder-controlled diagnostic AI auditing.
His core thesis is simple but critical: "The AUC is not the contribution — the decomposition is."Standard medical AI evaluation metrics like AUC can be dangerously misleading — models can achieve AUC ≥ 0.94 by exploiting covert demographic shortcuts (like a 21-year patient age gap between diagnostic groups) rather than learning true pathophysiology.
CharakAstra's 8-Gate Audit Protocol systematically deconstructs medical AI models to prove whether predictive outputs represent true pathophysiological signatures or covert demographic and scanner shortcuts — before they touch a clinical decision.
562 Patients
Multi-center validation cohort across distinct imaging centers and scanner protocols.
544 Features
Stable features with ICC ≥ 0.75. Strict radiomics filtering discarding fragile markers sensitive to slice thickness and scanner noise.
100% Leakage-Free
5×5 nested cross-validation with strict patient-level data partitioning isolating feature selection, preprocessing, and model fitting.
Unmasked: The 21-Year Age Gap
Atharva Shirahatti's research proved that state-of-the-art brain tumor diagnostic models achieved high AUC not by learning tumor morphology — but by exploiting a 21-year average patient age gap between diagnostic groups embedded in training data. When this confounder is neutralized, the model's AUC crashes from0.94 to 0.58 — barely above chance.