Artificial Intelligence (AI) is increasingly being used in many sectors but its applicability in the medical industry is still in its nascent stages, with several research programmes under way to test the efficacy of machine learning in clinical settings. Lily Peng, Google Health product manager and a non-practising physician, spoke to Armaan Bhatnagar on the prospects of AI in medical imaging and biomedical diagnostics to help detect diseases like lung cancer, diabetic retinopathy and breast cancer metastasis:
How can AI be of use to the healthcare industry?
AI can help fill the gaps when you need to examine a lot of data and there is a shortage of human expertise. Mostly, the problems where AI can be useful are routine and common. Medical imaging is generally a good example since you have large datasets to work with and the information that is gathered is not always uniform.
Is data a prerequisite to building good AI models?
Yes, all good AI models depend on data that is representative of the task you are trying to reproduce. The rule of thumb is that AI is good for a task that you have done for about 10,000 times and don’t want to do for the 10,001st time – so you train a machine to take over the job. You will find many such examples of automation of routine tasks in daily lives.
Do you think doctors will trust AI-based technology in clinical settings?
It all depends on how the product is designed. Doctors are incredibly busy; but if you can help them do their work better and prove that the product is more effective for the patients, most of them will welcome the technology. If you implement technology for technology’s sake, and if you ask a doctor to spend more time doing something, they probably won’t do it. Also, it’s not just about how accurate the model is, but how easy and intuitive the product is to use.
Do you believe AI can work in conjunction with human specialists to help them arrive at an accurate diagnosis?
Human-computer interaction is an important part of our research. It is about how we communicate the model’s prediction in a way that is understandable to the doctors so that they can make use of that information. The output of an AI model is just one out of many pieces of information a doctor has to synthesise to make a decision for the patient.
In many ways, it is like a thermometer. Before thermometers, someone would just feel your forehead and say you are ‘kind of warm but not sure how warm’. But now you have a tool that tells you a precise reading of temperature. While you still may have to figure out other things, the information gives you a better way of quantitating what’s happening in a disease process in an individual. Similarly, in medical imaging, you don’t make a diagnosis based on the model’s prediction alone but that information can also help you arrive at a decision when presented with other factors.
What was the outcome of your research programme in India using machine learning system to detect diabetic retinopathy?
The first version of the model we trained was in partnership with a couple of hospitals in Tamil Nadu and a healthcare provider in the US. The model was as accurate as a general eye doctor. We improved on that and now the models are as accurate as a retina specialist. So, the accuracy level has been fairly high.
A lot of the challenges are more around understanding how to implement a system like this in a real-world setting than training the model itself. If you have the right kind of data and if you get doctors to agree on what the state of disease is, you can train a good model. The question is how you implement that system in a hospital.
Is there a scope for full-scale rollout of these models in the foreseeable future?
Besides regulatory approval, you need to figure out how patients are referred and managed following their screening. Currently, we are working live with an Indian eyecare hospital with a regulatory-approved model that is being used at a smaller scale. However, enrolling these programmes and then carefully rolling them out can take time, maybe to the order of years. It is still early stages, but we are looking forward to talking with our partners in India to see how we can make it work. But there is potential for full-scale implementation in the future.
Any other trend related to AI’s use in healthcare that excites you?
Currently, there are not that many screening programmes and I think they are an important way to address some of the biggest healthcare problems like cancer, heart disease and diabetes. So this is the area I personally like working on.
(The interviewer was in Tokyo on an invitation from Google)
DISCLAIMER : Views expressed above are the author’s own.