A deep learning algorithm has been designed and shown to have applications in early prediction of Alzheimer’s disease, as published in Radiology.
Prof Derek Hill, Professor of Medical Imaging, UCL, said:
“Diagnosis of Alzheimer’s disease is currently done by a medical professional, often in a memory clinic, taking account of the results of memory tests, conversations with the patients, and sometimes brain scans or tests of brain fluid. Diagnosis quality is often poor, but in many cases this is of relatively little consequence because there are few effective treatment options, and no drugs yet available to halt or slow down the disease.
“When effective new treatments become available, high quality diagnosis is likely to become more important as it will be used to identify which patients will benefit from treatments that are likely to be expensive and may have significant side effects.
“This paper proposes a method to diagnose based on a type of brain scan, FDG PET, that looks at brain tissue energy consumption. The experimental evaluation appears to show good diagnostic performance. However, the algorithm is not being trained or tested on a realistic real-world patient population. The ADNI dataset they use is quite un-typical of a memory clinic population, and already excludes many of the problem cases for diagnosis. Furthermore many experts would doubt that the type of brain scan used is sufficient for diagnosis as it doesn’t pick up the underlying molecular pathology (amyloid and tau) nor does it include any assessment of memory.
“This paper should be interpreted with caution. It is a good example of artificial intelligence applied to medical images, but is some way short of being a useful clinical tool that could replace diagnosis of Alzheimer’s by a healthcare professional.”
Prof John Hardy FMedSci, Professor of Neuroscience, UCL, said:
“This is a tiny data set, only looking at 40 people. It’s also a very selected data set and not representative of the whole population. So we can’t know yet whether this is relevant to most people. The limitations of the study are well described in the last page of the discussion in the paper.”
Dr Peter Bannister, Chair of the IET Healthcare Sector and Independent Consultant, Bannister Technologies, said:
“The press release clearly outlines the method and experimental protocol that was followed by the researchers. Techniques such as these are greatly needed to improve the diagnosis of diseases such as Alzheimer’s given the increasingly ageing society we live in today, and to enable the development of new therapies which can be offered earlier in the treatment pathway.
“While the initial results are positive, it is equally heartening to see the authors recognise the limited size of the testing carried out to date, and the need for further validation before this can be deployed in a clinical setting to support – but not replace – the neuroscientific experts. Ultimately, the greatest impact will arise from techniques which can apply clinical rigour much earlier in the diagnosis pathway, when imaging such as PET may not be routinely available.”
Dr Maryam Shoai, Postdoctoral Statistical Geneticist, Department of Neurodegenerative Disorders, UCL Institute of Neurology, said:
“This is a nice elegant study which is proof of concept that deep learning algorithms can be of immense use in predicting outcome of disease, however, the study cannot be taken as a predictive model in its current form given the small sample numbers. Furthermore, the test set and training set are essentially from the same ‘pot of samples’ and therefore can explain why the model looks so good, i.e. the model would not predict the outcome so well if the test set was truly independent.”
Dr David Llewellyn, Senior Research Fellow, University of Exeter and Fellow, Turing Institute:
“Artificial intelligence (AI) and machine learning have enormous potential to make sense of complex neuroimaging data and detect subtle early changes in the brain linked to dementia. However, the patients they used to train their algorithm were a highly selected group that do not reflect the diverse patients seen in routine clinical practice. For example, other types of dementia were excluded, making the job of classifying patients easier. Their algorithm therefore needs to be further refined, though in time it is likely that machine learning will help doctors to detect the earliest signs of dementia.”
Prof Duc Pham, Chance Professor of Engineering, University of Birmingham, said:
“This is another potentially very useful application of deep neural networks with the possibility to assist clinicians in diagnosing a debilitating disease at an early stage. The study is well balanced and the authors have recognised its pilot nature and limitations. As mentioned in the article, before the work is ready for clinical use, there is a need to increase the size of the independent test set to cover a more general patient population. I am confident that, if the required validation is done, clinicians could have a powerful tool at their disposal to help them prescribe preventative treatments in a timely manner.”
Prof Noel Sharkey, Emeritus Professor of Artificial Intelligence and Robotics, University of Sheffield, said:
“This is exactly the sort of task that deep learning is cut out for – finding high level patterns in data. Although the sample sizes and test sets were relatively small, the result are so promising that a much larger study would be worthwhile.”
Dr Carol Routledge, Director of Research at Alzheimer’s Research UK, said:
“The diseases that cause dementia begin in the brain up to 20 years before any symptoms start to show, presenting a vital window of opportunity for us to intervene before widespread damage occurs. This study highlights the potential of machine learning to assist with the early detection of diseases like Alzheimer’s, but the findings will need to be confirmed in much larger groups of people before we can properly assess the power of this approach.
“Currently in the UK, the use of PET scanning is mainly limited to research studies and clinical trials, to ensure that potential new medicines are tested in the right people. PET scans are a powerful tool, but they are expensive and require specialist facilities and expertise.
“Although recent advances in artificial intelligence offer a critical opportunity for tackling the challenge of diagnosing Alzheimer’s disease much earlier, this will require a major effort and significant investment. Alzheimer’s Research UK believes a technological approach, using big data and machine learning, could lead to huge benefits for people affected by dementia and their families.”
‘A Deep Learning Model to Predict a Diagnosis of Alzheimer Disease by Using 18F-FDG PET of the Brain’ by Yiming Ding et al. was published in Radiology at 15:00 UK time on Tuesday 6 November 2018.
Prof Derek Hill: “No conflicts.”
Prof Duc Pham: “I have no conflict of interest to declare.”
Dr David Llewellyn: “No interests to declare.”
Prof Noel Sharkey: “No conflicts of interest.”
Dr Carol Routledge: “No conflicts of interest.”
None others received.