Conflict in the code: Machine learning technologies used in healthcare must be credible

Conflict in the code: Machine learning technologies used in healthcare must be credible

In the second article in our series, we explore the role of machine learning in healthcare today, and examine some common issues.

In medical practice and healthcare today, machine learning is already in use. It has led to exciting new developments that will continue to be refined, as well as redefining the analysis, detection, diagnosis and treatment of a range of acute and chronic diseases. There are variety of examples of the use of machine learning in healthcare, and here we examine some of them. 

Research and analysis of evidence

The volume of scientific literature produced each year is staggering; no human can keep abreast of all the research information available to them. This is one area where machine learning has already transformed our productivity. Using modern search engines, researchers need only cite a few key terms, and immediately find the most widely referenced papers on a given subject – or the newest, or the one from the most trusted source. What was previously a laborious task has been reduced to a few phrases and clicks. 

Machine learning is also being used by researchers to aggregate and make sense of the vast quantity of data available about individuals and populations. Linking and analysing large datasets, including genomic, social, image, and device data can be more efficient using machine learning, compared with traditional statistical techniques. 

Understanding and preventing disease

In order to improve biological understanding of disease, researchers deploy machine learning and predictive analytics to analyse these huge datasets. Machine learning algorithms have the potential to identify patients who are at risk, and suggest targeted preventative treatments in order to slow or even halt the development of disease. In New Zealand, Precision Driven Health and Vensa Health are working together to improve the both the efficiency and personalisation of laboratory test result processing

Screening and diagnosis

Many screening programmes, such as those for skin cancer, eye disease, bowel cancer and breast cancer, rely on visual inspection and interpretation of medical images. Machine learning systems have proven effective in screening for disease and, in New Zealand, have been used to assist in the diagnosis of breast cancer and skin cancer. The systems act as one stage in the screening process, escalating ‘non normal’ images to human specialists for further investigation.   

Personalised medicine

Machine learning approaches can help to personalise medicine for individual patients. Our medical practice usually treats all patients as similar, even though not everyone responds to treatment in the same way. Machine learning can help doctors to establish which particular treatment will work best for each particular patient. For example, in the field of radiotherapy dosing, researchers have used a ‘deep learning’ approach to help prescribe the ideal dose of radiation for individual lung cancer patients. This application is part of the burgeoning field of precision medicine. 

While these cutting-edge applications are undoubtedly exciting, there are some issues to consider. Building trust in our machine learning algorithms to assist in healthcare decision-making requires the techniques to be credible. 

Data quality, sample size and population source are key

Firstly, machine learning relies on the quality, scale and ‘representativeness’ of the data used to train the algorithm. As with any statistical model, if the data is poor quality, the sample size is small or drawn from a selective population, machine learning is unlikely to produce meaningful results that scale to population improvement. Good models are reproducible, and good medical decisions are generally explainable. Even if machine learning algorithms appear to be accurate on the data they used for training, credibility can be easily lost if that does not translate to where they are being used. 

The next article in our series will explore the issue of bias in more detail, and suggest some guidelines for conducting ethical machine learning.