Poznan, Poland, June 2019
A blog by Sally Vernon
The European Society for Artificial Intelligence in Medicine (AIME) was established in 1986, and has two principal aims – to foster fundamental and applied research in the application of artificial intelligence (AI) techniques to medical care and medical research, and to provide a forum for discussing the progress made. AIME holds a conference every two years, and I was delighted to be able to attend the 2019 event in Poznan, Poland.
Only 53 papers out of 135 submissions were accepted, so I was very pleased to give a short presentation on our paper “Identification of Patient Prescribing Predicting Cancer Diagnosis Using Boosted Decision Trees”. This work, led by Dr Jo French, was a partnership project between Public Health England and Health Data Insight, funded by a Pioneer award from Cancer Research UK.
Distilling the main ideas into five minutes meant I had to be surprisingly ruthless, but it was worth it when I received compliments on a “clear, interesting and enthusiastic talk”. Despite it being the last coffee break on the final day, so many people came to ask questions afterwards I never made it to the coffee!
The conference touched on many themes. Some of the main ideas that have stayed with me include:
- Machine learning classifiers can solve such a wide variety of medical problems – taking an image, and learning to identify pathological features; taking an electronic patient record, and learning to predict likely outcomes of interventions; taking a broad disease type, and learning to identify specific subtypes to optimise personalised medicine; taking a geographical trace of a person’s day, and learning to identify daily activities like ‘coffee with friends’ or ‘swimming’, and thus their likely health status.
- Diversity of data, and how big data can also be very small data. For example, data collected from wearable devices, sound recordings and video imaging can be many gigabytes, but only relate to a very small number of patients.
- Topological data analysis – the idea that data has shape and structure, and you can understand your data better by looking at simplified topological signatures, such as persistence diagrams and mapper graphs.
- Tension between performance and explainability in machine learning algorithms – how the classifiers with the best performance are often from deep neural networks, where we do not understand why the machine makes the decisions it makes. There was interesting discussion around what that means for medical AI – whether it is ethical to make a clinical decision with no understanding of why the decision has been made, and whether there could be biases learned by the algorithm from the training data that could lead to unfair discrimination or treatment. There was also interesting discussion about the subtle differences between an explainable algorithm, an interpretable algorithm, and a transparent algorithm, and what properties an algorithm needs to be a trusted algorithm.
- How the world changes, and how we build trust in new systems. How new technology can be met with a mix of suspicion and curiosity. How we cascade out trust in a human chain – computer scientists inspiring key early adopters in the medical community, and those early adopters showcasing ideas to the wider clinical community. How our legal and regulatory frameworks change more slowly than the development of our technology, and how we need to balance between responding rapidly to a changing world, but doing due diligence and maintaining our ethical foundations.
The conference was mostly attended by academic computer scientists, with people from industry, medicine and ethics also contributing to the discussions. All presentations were given in plenary, and the programme ran from 09.00 until after 18.00 on both days. There were three types of talk – invited speakers, who spoke for an hour, long papers, scheduled with twenty minutes to talk and five minutes for questions, and short papers, with five minutes to talk, and questions in the coffee break by their poster. The short papers meant that the programme was varied and exciting, and a wide range of work could be presented – although lots of the research was technical and complex, and presenting it in five minutes was sometimes challenging!
I was left with the impression that artificial intelligence has a huge amount of potential, but that there is still a long way to go to get from theoretical models to their reliable applications in medical environments. However, the impact of those applications when we do get there has the potential to be huge, fundamentally altering the way our health care systems make life-changing decisions.
AIME 2019 was absorbing, motivating and very educational. I would recommend it to anyone with a basic understanding of machine learning techniques looking for inspiration for how to apply these techniques to new problems in the medical field. The next event will be held in Porto in 2021.