PrAViC: real-time video classification with early, confident decisions presented at ECAI 2025

At ECAI 2025 in Bologna — Europe’s leading conference dedicated to artificial intelligence — we presented PrAViC, a probabilistic framework enabling fast and confident real-time video classification. The work was showcased in the main conference track by Dr. Eng. Arkadiusz Lewicki, CTO of Prometheus MedTech.AI.

PrAViC (Probabilistic Adaptation Framework for Real-Time Video Classification) addresses one of the most critical challenges in modern perception systems: making reliable decisions without waiting for full video sequences. Traditional 3D CNNs require entire clips to be processed, making them impractical for time-sensitive environments. PrAViC shifts this paradigm through a probabilistic formulation that enables early, confident predictions — often after just a few frames.

Our experiments demonstrate up to 2.6× higher efficiency compared with conventional approaches, while maintaining 91–93% accuracy and operating fully in real time. Importantly, PrAViC requires no architectural changes or retraining, allowing seamless integration into existing 3D CNN models.

The potential applications are broad — from driver monitoring and human–robot interaction to real-time medical diagnostics, where early and reliable recognition in echocardiography can directly impact patient outcomes.

Prometheus MedTech.AI was also represented by Magdalena Trędowicz, whose research focuses on multiple instance learning and model interpretability at Jagiellonian University.

ECAI 2025 brought together leading researchers and industry practitioners shaping the future of artificial intelligence. We are proud that our contribution highlighted a key direction for the field: smarter, faster and more trustworthy decision-making, rather than merely scaling model size.