I'm not sure your comment about it being easy to overly diagnosis rare conditions captures the reality of practising medicine for anyone who's been a clinician long enough. Part of the training is to think common first before rare, and to sense when there are peculiarities. Rarity stands out, so inexperienced clinicians/medical students will seek this out in patients.
As paviva mentions diagnosis is based on imperfect data with a high noise-to-signal ratio. Which I think is actually a testament to human's ability to navigate all of that and still be able to care and treat illness.
I think the trends in technology with healthcare will more likely be to support and supplement humans rather than taking them out of the equation. I do agree that there will probably be a shift in how we practice as in how much we need to retain as doctors. Doctors in the next 5-10 years will certainly need a much greater understanding how to critique ML used to make recommendations. I doubt clinicians will need to understand the inner depths of how an ML model is working at a code level though. There are already several guidance papers on how to evaluate ML models.
But clinical reasoning will still be a human endeavour in my opinion. The adjuncts to diagnosis is already happening in more visual concentrated specialities like radiology,pathoogy and ophthalmology, where ML algorithms I think have a good fit. No doubt this is where we will start to see more reliance on the ML models. But again, that uncertainty and ability to take an imperfect data set and make a decision with conscious and subconscious understanding will be aided by these models but not necessarily determined by them. To be able to capture all of the other inputs that are required for diagnostic reasoning will still escape ML inspired technology. Blood tests and imaging alone do not make a diagnosis, its the whole context from symptom to investigations together.
Another area of technology is the use of hand held devices for diagnostics, portable US for example. US is known to have better accuracy than chest X-rays for particular conditions. It's not quite there yet in terms of the taking that on a ward round, but that feels like a more realistic proposition in the next 5-10 years.
I also expect that we will be looking far far more about patterns in this noise of data. Looking at 100,000's of medical records to try to understand the patterns of disease we as humans just can't perceive. Looking at predictive modelling for things like cancer. Cancer being a large bucket of atleast 200 types of diseases, most of them are rare and no one clinician has seen enough of them to really get predictive or understand the nuance of them to really diagnose with confidence. Additionally clinical trials are poor datasets for sample sizing and in application to real life scenarios that we end up having to rely on.
So being able to get through millions of notes and finding all these cases to make predictions seems to me to be a fruitful area of investigation.
Interesting project! Wondering if it’s open source ?
Also I’d be interested to know a little more about the translation from notes to testing. To me it seems when I’ve used Anki that it takes great care and time to make a well formed question. A poorly formed question has a massive detriment in knowledge recall. Just wondering how you’ve managed to bridge that dilemma to automate the process.
Very interested in contributing, I’m a programmer, clinical educator and love spaced repetition.
Sorry, this is a commercial project, not open source. In fact, if this fails, we will starve.
Anyway, it is the core of this app that automatically connects the naturally written note-taking into a form that can use Spaced Repetition. It's hard to explain in more detail, but I hope you enjoy the experience of being able to use Spaced Repetition for the content by simply taking notes in outliner form. Anyway, thanks for the compliment.
“Commercial” is not incompatible with “open source”. If you have a policy of “release last year's version under AGPLv3” (and then buy indefinite proprietary licenses for the best community patches off their authors), we'll all benefit from open source without you having competitors leeching off your work.
In fact, some companies can do it with no time delay, like https://plausible.io – but, of course, it depends on your business model and your field.
Agree, would much rather this was an open source project like Anki and then charge for extra services like cloud sync or whatever to generate revenue.
I feel very uncomfortable contributing huge amounts of content to a product that I have no control over. With no Linux version I can only submit all my data to the cloud. It does appear you can export to JSON or Anki deck though so you could always do regular exports to ensure you have a copy of your data.
I will offer that 'Where does money come from' By Josh Ryan-Collins, Tony Greenham, Richard Werner, Andrew Jackson. Its a fantastic short book on how money actually works in this system.
The follow up to this is Richard Werners study on central banking.
A lost century in economics: Three theories of banking and the conclusive evidence
As paviva mentions diagnosis is based on imperfect data with a high noise-to-signal ratio. Which I think is actually a testament to human's ability to navigate all of that and still be able to care and treat illness.
I think the trends in technology with healthcare will more likely be to support and supplement humans rather than taking them out of the equation. I do agree that there will probably be a shift in how we practice as in how much we need to retain as doctors. Doctors in the next 5-10 years will certainly need a much greater understanding how to critique ML used to make recommendations. I doubt clinicians will need to understand the inner depths of how an ML model is working at a code level though. There are already several guidance papers on how to evaluate ML models.
But clinical reasoning will still be a human endeavour in my opinion. The adjuncts to diagnosis is already happening in more visual concentrated specialities like radiology,pathoogy and ophthalmology, where ML algorithms I think have a good fit. No doubt this is where we will start to see more reliance on the ML models. But again, that uncertainty and ability to take an imperfect data set and make a decision with conscious and subconscious understanding will be aided by these models but not necessarily determined by them. To be able to capture all of the other inputs that are required for diagnostic reasoning will still escape ML inspired technology. Blood tests and imaging alone do not make a diagnosis, its the whole context from symptom to investigations together.
Another area of technology is the use of hand held devices for diagnostics, portable US for example. US is known to have better accuracy than chest X-rays for particular conditions. It's not quite there yet in terms of the taking that on a ward round, but that feels like a more realistic proposition in the next 5-10 years.
I also expect that we will be looking far far more about patterns in this noise of data. Looking at 100,000's of medical records to try to understand the patterns of disease we as humans just can't perceive. Looking at predictive modelling for things like cancer. Cancer being a large bucket of atleast 200 types of diseases, most of them are rare and no one clinician has seen enough of them to really get predictive or understand the nuance of them to really diagnose with confidence. Additionally clinical trials are poor datasets for sample sizing and in application to real life scenarios that we end up having to rely on.
So being able to get through millions of notes and finding all these cases to make predictions seems to me to be a fruitful area of investigation.