Bucket B: Interpretation

There is substantial overlap and a meaningful divergence, and getting the relationship right matters for how a funder should think about Bucket B's cost.

The overlap: foundation models, federated learning, causal inference methods, and uncertainty quantification all advance with general AI research. When the broader field produces better architectures, better training methods, or better uncertainty estimation, Bucket B inherits much of it. B4 (digital twins) and B5 (continuous learning) in particular ride general-AI and ML-infrastructure progress closely.

The divergence: medical AI faces constraints general AI does not. Strict privacy regulation shapes what data can be used and how. Regulatory validation demands interpretability and reproducibility that frontier general-AI systems often lack. Distribution shift for unusual patients, calibrated uncertainty for safety-critical decisions, and clinical accountability all push medical AI in directions general AI has no particular reason to go. Some of these constraints actively make the medical problem harder than the general one — a maximally capable but opaque model is useless if it cannot be validated for clinical use.

The consequence: Bucket B benefits from but does not fully ride general-AI progress. Dedicated investment in medical-AI-specific capabilities — calibration, validation methodology, causal reasoning under biological constraints — remains necessary. A funder should expect to fund both the medical-specific work and the general infrastructure it depends on, and should not assume that general-AI advances alone will deliver AIHS interpretation.

The bucket distinguishes predictive modelling — answering "what is likely to happen given observed data?" — from causal modelling — answering "what happens if we intervene?". The first improves with more data and better architectures, both of which the field has in abundance. The second requires inferring directed cause-and-effect relationships from data that does not, on its own, contain them: from observation alone one cannot in general distinguish "X causes Y" from "a hidden factor causes both," without intervention or strong assumptions.

This is why B3 is the central scientific gap in the entire study, and why it is classified very-high-risk on a 20-50 year horizon — far longer than the other Bucket B advances. The B3 page itself treats the formal frameworks (Pearl's structural causal models and do-calculus, the Neyman-Rubin potential-outcomes framework) and why biology denies them their preconditions, in depth.

At the bucket level, the key point is that a breakthrough is not guaranteed to arrive on the AlphaFold template, because causality is not present in observational data the way 3D structure is present in protein sequences. A breakthrough might come from perturbation biology scaled by orders of magnitude (so interventional data becomes large enough to learn causal structure directly), from hybrid mechanistic-learned models that extrapolate correctly outside their training distribution, or from genuinely new causal-inference theory suited to feedback-rich biological systems. We cannot currently predict which — if any — will deliver. That irreducible uncertainty is exactly what the Phase II decision gate is built to assess.

Even without comprehensive causal modelling, AIHS interpretation can cover the diseases for which causal mechanisms are already well understood from prior science — and that is most of common medicine: diabetes, hypertension, the common cancers, most infections, many chronic diseases. For these, mechanistic models combined with patient-specific parameterisation produce credible intervention predictions without needing a B3 breakthrough.

So the realistic ceiling without B3 looks like this: AIHS for well-characterised conditions operates at near-full capability — continuous monitoring, integrated interpretation, and validated intervention selection. AIHS for rare, complex, or poorly-understood conditions remains advisory rather than autonomous: the system provides data integration and pattern recognition, flags its own uncertainty, and routes the decision to a clinician who directs it.

This is still transformative relative to current practice — a system that handles the common cases autonomously and supports clinicians intelligently on the hard ones would change medicine substantially. It simply is not the universal capability the full AIHS vision implies. The distinction matters for honest communication: the study's baseline expectation is this partial-capability outcome, with universal capability as the conditional upside if B3 (and rare-variant interpretation in B1) yield to research. The Phase II decision gate is structured precisely to determine which scenario the field is in and recalibrate ambition to evidence.

Because several Bucket B capabilities are physically impossible without Bucket A diagnostic advances — the dependency is not a matter of convenience but of what can exist at all.

The clearest case is B2 (real-time epigenomic state inference), classified high-risk and explicitly dependent on Bucket A. Inferring chromatin accessibility, methylation, and histone modifications in a living patient without destroying samples requires the in-vivo readout capability of A1 — there is no way to infer epigenomic state in vivo without the means to read cellular state in vivo. B2 cannot move faster than the substrate technology A1 provides.

The same logic runs throughout. B3 causal modelling needs the rich, continuous, multi-parameter data that A3 and A4 produce — causal inference is starved without dense longitudinal observation. B4 digital twins need patient-specific data from across the diagnostic stack to parameterise against. Interpretation is, almost by definition, interpretation of diagnostic data; the quality of the interpretation is bounded by the quality and richness of the data feeding it.

Pursued in separate silos, the two buckets would repeatedly stall waiting on each other — Bucket B teams building interpretation methods for data that does not yet exist, Bucket A teams producing data with no integrated consumer. Tightly coupled, each pulls the other forward: diagnostic advances open new interpretation possibilities, and interpretation needs sharpen the requirements for diagnostic development. This is why the study recommends co-funding and shared programme structure across A and B rather than parallel independent efforts.