Causal Sensitivity Scores: A Deeper Dive into AI's Clinical Responsiveness

clinical AI systems, appearances can be deceiving. Two systems might score similarly on traditional metrics yet diverge dramatically in behavior when faced with dynamic patient data. This stark contrast in performance metrics raises a critical question: How do we measure an AI's true clinical responsiveness?

Introducing the Causal Sensitivity Score

The Causal Sensitivity Score (CSS) steps in where traditional metrics fall short. This interventional metric mutates oncology cases across five key dimensions, biomarker shifts, treatment failures, and more, to see if AI recommendations adapt correctly. It assigns scores on a crisp scale of 0, 0.5, to 1, providing a clear picture of how models respond to patient changes.

In a striking revelation, six frontier models from three labs showed almost inverted rankings when evaluated by CSS compared to the Consensus Match Score (CMS), a typical coverage metric. The model previously deemed worst by CMS catapulted to the top under CSS, while a mid-ranker plummeted to last place. This flip exposes a flaw in coverage-focused evaluations: they miss the responsiveness that CSS captures.

The Blind Spot in AI Models

Every model studied failed when it came to surgery-status interventions, achieving only a dismal 17.2% on CSS for this category. This glaring blind spot, invisible to CMS metrics, highlights a universal safety concern in clinical AI systems.

in experiments mimicking tool use, most models improved their CSS scores. Yet, the lowest-scoring model remained stubbornly unresponsive, retrieving the same chart sections without updating its recommendations. This reveals a deeper structural issue, a responsiveness deficit that only a counterfactual evaluation like CSS can unmask.

Why CSS Matters

So why should we care? Because CSS offers a more nuanced understanding of AI performance in clinical settings. It surfaces hidden deficits that coverage metrics gloss over. If the AI can hold a wallet, who writes the risk model? The question isn't just academic. It’s about patient safety and how we ensure AI systems are truly fit for purpose.

As AI systems become more agentic, metrics like CSS won't just be useful, they'll be essential. They'll provide a dense reward signal for reinforcement learning, pushing AI towards better real-world performance. Slapping a model on a GPU rental isn't a convergence thesis. We need to go deeper, and CSS is a step in that direction.