From Risk Scores to Agents: What "AI" Really Means in Nephrology

This presentation reframes AI as a practical toolbox rather than a buzzword. It moves from classic statistical models and bedside risk calculators toward machine-learning approaches for clinical prediction, prognostics, and decision support. The emphasis is on what makes models useful in practice: adequate data scale and diversity, interpretability, and validation that accounts for workflow and resource burden rather than metrics alone.

The middle section covers deep learning as a method for medical imaging, physiologic signals, and phenotype discovery. It then introduces large language models as tools for structured extraction from free text, reporting support, educational use, and guideline-grounded retrieval. The limitations are treated seriously, especially hallucination and the need for explicit benchmarking before deployment.

The closing section focuses on what agents add beyond chat: language models connected to tools, workflows, and knowledge bases that can automate multi-step tasks while preserving safety, transparency, and real-world utility.