Alberto Paderno

Presentation
AI in healthcare is shifting from model-centric achievements to outcome-oriented, bedside decisions. This talk surveys the evolving landscape—from classic predictive models to multimodal and generative systems—and asks a single practical question: when does AI measurably improve patient care? Rather than centering on any one technology, we connect algorithmic advances to concrete clinical actions across imaging, pathology, endoscopy, and longitudinal EHR use.

First, we review endoscopy AI for detection, segmentation, and quality metrics, and how real-time assistance can change actions (biopsy, resection, surveillance) rather than just scores.

Second, we show how VLMs enable generalizable perception and reasoning across modalities, powering retrieval, report drafting, and interactive decision support while preserving clinician oversight.

Third, we examine pathology foundation models trained on whole-slide images: their strengths in representation learning, few-shot adaptation, and uncertainty estimation; and how slide-level outputs translate into case-level risk, treatment triage, and workflow acceleration.

We outline design patterns that make models usable: human-in-the-loop loops, uncertainty-aware UX, data lineage and versioning, and change-controlled updates. We align with current regulatory expectations for high-risk AI/IVDR, emphasizing intended use, performance studies, and post-market surveillance.

The audience will leave with a practical roadmap to deployable decision support: start with a clear clinical question and baseline, validate externally, quantify net benefit at actionable thresholds, integrate into the workflow with guardrails, and monitor in the wild.

Bio
Assistant Professor at Humanitas University, Milan, Italy