Very excited to have 2 new papers in press today in Lancet Digital Health, alongside an editorial from the journal highlighting our work.
I am immensely proud of the work we have done here and honestly think this is the most important work I have been involved in to date 🥳
1/7
#Medical#AI has a problem. Preclinical testing, including regulatory testing, does not accurately predict the risks that AI models pose once they are deployed in clinics.
I've written about this before in my blog, for example in:
1) describe a step by step method for algorithmic auditing in health, building on the 🔥 work by Raji et al
2) audit a high accuracy model we developed @theAIML for hip fracture dx, ID-ing several serious risks that were not detected by standard testing.
3/7
The high performance hip fracture model (AUC 0.994 vs 0.969 for radiologists) fails unexpectedly on an extremely obvious fracture and produces a cluster of errors in cases with abnormal bones (Paget's disease etc).
These findings (and risks) were only detected via audit.
4/7
We are excited that this work is impacting policy. Professional orgs such as @RANZCRcollege are incorporating audit into their practice standards (ie ranzcr.com/college/docume…) and we are talking with regulators and governance groups on how audit can make AI systems safer.
5/7
I'll leave it at that for now (although expect a blog post in the near future 😂), so I'll just leave the links here:
The audit paper is my first senior (last) author publication (co-seniored with the amazing @Denniston_Ophth), and both papers have been published under my new name!
They are also the final papers of my PhD, which is now completed!
🥳🥳🥳
8/7
Urrgh I'm sorry, I don't know how I didn't link @rajiinio's profile here. If you don't know her, check her out, she does incredible work!
Also the other authors on the algorithmic audit paper are awesome too!
• • •
Missing some Tweet in this thread? You can try to
force a refresh
We've put out a preprint reporting concerning findings. AI can do something humans can't: recognise the self-reported race of patients on x-rays. This gives AI a path to produce health disparities.
This is a big deal, so we wanted to do it right. We did dozens of experiments, replication at multiple labs, on numerous datasets and tasks.
We are releasing all the code, as well as new labels to identify racial identity for multiple public datasets.
2/8
Humans can't detect race better than chance, but AI performs absurdly well on the task. As you can see here, AUC scores are in the high 90s, and are maintained on external validation on completely distinct datasets and across multiple different imaging tasks.
Docs are ROCs: A simple fix for a methodologically indefensible practice in medical AI studies.
Widely used methods to compare doctors to #AI models systematically underestimate doctors, making the AI look better than it is! We propose a solution.
The most common method to estimate average human performance in #medical AI is to average sensitivity and specificity as if they are independent. They aren't though - they are inversely correlated on a curve.
The average points will *always* be inside the curve.
2/7
The only solution currently is to force doctors to rate images using confidence scores. While this works well in the few tasks where these scales are used in clinical practice, what does it mean to say you are 6/10 confident that there is a lung nodule?
Alright, let's do this once last time. Predictions vs probabilities. What should we give doctors when we use #AI / #ML models for decision making or decision support?
This discussion was getting long, so I thought I'd lay out my thoughts on a common argument: should models produce probabilities or decisions? Ie 32% chance of cancer vs "do a biopsy".
I favour the latter, because IMO it is both more useful and... more honest. IMO:
I personally suspect the biggest problem is automation bias, which is where the human over-relies on the model output.
Similar to self driving cars where jumping to complete automation appears to be safer than partial automation.
But interestingly (and perhaps counter-intuitively) this could also mean that "blind" ensembling (where the human gets no AI input, and the human and AI opinions are combined algorithmically) might be better than showing the doctor what the AI thinks.
@weina_jin The weird thing about CV in AI is that you don't actually end up with a single model. You end up with k different models and sets of hyperparameters.
It allows an estimate of generalisation for a *group* of models, but that is still a step removed from a deployable system.