Histogramm-Derivative Risikokalibrierung für Screening-Scores der Künstlichen Intelligenz

Details

Veröffentlichungs-Nr.

EP4800703

Anmeldetag

27. Februar 2026

Veröffentlichung

2. September 2026

Rechtsraum

EP

Abstract

In an embodiment of the present invention, a method converts continuous scores from an AI disease classifier into clinically meaningful risk estimates. Historical score-outcome data are binned into percentiles, and the cumulative disease fraction is fitted with a differentiable curve for example a two-Beta mixture. The curve's derivative yields an instantaneous risk ratio that, combined with baseline prevalence, provides an absolute disease malady probability. Risk ratio and absolute probability ensures comparability across model versions, improving interpretability. The invention includes associated systems, software, and computer-readable media.

Anmelder

Firmen

Westerhoff, Malte
Visage Imaging GmbH

Land

🇩🇪 Deutschland