Not clinical decision support.

Human progress in clinical medicine is stunted by the speed of evidence generation. Randomised trials guard against mistaking association for treatment benefit, but they are slow, expensive and cover only a fraction of clinical decisions. A 2013 Oxford hospital audit found trial evidence supporting about one in five primary treatments; many others had convincing non-experimental support1. One UK exercise estimated £770,000 for specialist trial support2; a 2018 analysis estimated a median US$19m for trials supporting new drug approvals3.

Conventional trials alone cannot close this gap.

We need a capability that distinguishes when evidence supports treatment from when uncertainty warrants an experiment.

This inverts clinical decision support: rather than only recommending care, we would identify repeated decisions where we could instead safely learn. Randomising without equipoise, where benefit is established in one direction or another, is unethical. But decisions without learning are also unethical.

Our challenge is to map efficiently map a decision into a searchable specification of patients, treatments and outcomes. We need to identify relevant evidence from the scientific literature and local evidence (digital twins). We need to combine this with metrics of search coverage, causal validity and quantify relevance to the particular decision.

Done well, we’d rapidly identify decisions without evidence creating the ethical justification to randomise. Each randomisation event would create learning, and shrink the uncertainty, and improve future decision making.

Offline reinforcement learning illustrates the opportunity: learn treatment strategies from past records, then test them prospectively. AlphaGo could learn through self-play without harming anyone. Medicine cannot explore so freely. By bounding exploration risk we could allow retrospective policy learning to progress into controlled clinical experimentation.

  1. Chapman et al. Lancet, 2013. https://doi.org/10.1016/S0140-6736(13)62286-2 ↩︎

  2. Comparative costs and activity from a sample of UK clinical trials units. Trials, 2017. https://pmc.ncbi.nlm.nih.gov/articles/PMC5414193/ ↩︎

  3. How much do phase III trials cost? Nature Reviews Drug Discovery, 2018. https://www.nature.com/articles/nrd.2018.198 ↩︎