On the 14th February 2018, I went to the Machine Learning in Healthcare workshop in Moorgate. This was run by the team over at Re•Work (more about their work can be found here). The event was billed as a training and professional development event for those interested in moving healthcare from reactive to proactive and predictive care using aspects of machine learning.
We kicked off the workshop with an overview of machine learning from Dr Nophar Geifman (University of Manchester) who a very clear overview of the main methods in Machine Learning. We learned about how neural networks can be developed that allow for many more connections between input and output than just a simple model, for example regression. As someone who has spent most of the last 3 months building a logistic regression models for a stroke cohort study, this was a revelation! I could immediately see that application to the field of epidemiology and how we could get many more insights particularly in hypothesis generation. The strength seems to lie in the not having assumptions (e.g. from a researcher) when we approach a dataset. We still need the quantification and exploration from a causal, epidemiological point of view but the field of ML for hypothesis generation is very exciting field that can be allied to epidemiology.
From a clinical point of view, machine learning may one day provide us with many more tools at our disposal. For example, Dr Geifman gave the example of a machine learning tool that learns to identify skin cancer using a deep neural network. The tool “learns” a test dataset of biopsy-proven cancer images and achieved a performance that was on par with consultant dermatologist. As a GP trainee, I often see skin lesions which don’t always fit into neat categories. We tend to have a very low bar for referral to a specialist, and whilst we have experimented with using e-clinics where you can take a photo and send it to the local dermatology team, this has not always been practical. This is mainly due to the limitations in referral system, finding the practice camera (and cable!), uploading over unreliable wifi, time (as it takes about a week to get any feedback) and the paperwork ensued because of consent forms. If there was a way to transfer this machine learning algorithm into an app form that stores data on the phone or in another encrypted manner, this can only be a good thing.
What is clear to me is that we don’t need to fear machine learning taking our jobs as doctors, but we need to get involved to make sure it is clinically useful. It turns out machine learning is very effective at predicting certain things but only under certain conditions. For example, it is startlingly good at predicting in-hospital mortality from electronic health records (see more here). However clearly this requires there to be a good set up in capturing this information and deployment of this information into the clinical environment. For this to happen there has to be buy-in both from the managers and politicians, but also the clinical staff. We need to be able to explain “predictions” from a machine learning algorithm to patients and colleagues when it is not always clear what has happened under the hood e.g. in deep learning. I think part of what needs to happen is for us to train more clinicians the concepts behind machine learning and large data-driven tools. This brings us on nicely the part of the afternoon that we spent thinking about the ethics and legal implications of artificial intelligence with a talk from Andreas Theodorou, who is working in this area at the University of Bath. As he pointed out a key part of developing AI tools is the buy in and co-operation from humans. This is not by any means a given. Stephen Hawkings described AI potentially creating a dystopic future (see here) and on a more humorous front, I read only last week that testing of self-driving cars in California has been plighted by humans attacking them (see here). Clearly there is some strength of feeling in this area and that will only be increased when we are talking about our health and the personal information contained in our health records.
The workshop was excellent and covered a lot more ground that I can talk about within a blog post. If you get the opportunity to attend one of these events in the future, I thought it was well worth it. I’ve just bought tickets to a 2 day event in September on Deep Learning in Healthcare (see more here). Hopefully I will see you there to continue this conversation!
Photo by: unsplash-logo贝莉儿 NG