I can’t get the fog to lift: Deterioration progresses differently in people with different levels of memory problems
Dementia research, which explores everything from risk factors, to diagnosis, to long–term treatment outcomes in patients, aims to address the growing health crisis of people suffering from dementia; a set of diseases characterised by memory and other cognitive issues. Worldwide, there are approximately 55 million people suffering from some form of dementia, such as Alzheimer’s, which is the most common type of dementia. Currently, we spend 1% of global GDP (Gross Domestic Product, the output of a country through manufacturing and services) caring for those with dementia and with the expected increases in global life expectancy, the number of those who need to be cared for in the long term, because of dementia, is only going to increase.
However, none of these oft-quoted statistics captures the experience of those living with a dementia-type illness.
I am a Data Scientist and a second-year PhD candidate at Goldsmiths, University of London. My PhD looks at the possibility of using artificial intelligence and machine learning; whereby computers can learn and make predictions, in order to predict those who go on to develop dementia in general, and in particular Alzheimer’s Disease. As a computer scientist and statistician, my work focuses less on a specific area of dementia research, but rather on the techniques used to produce robust and validated results. As a result, I have conducted research on healthy individuals and their risk of predicting Alzheimer’s, as well as those with a clinical diagnosis of Alzheimer’s disease, attempting to predict their long-term outcomes. I am also, somewhat frantically, a full-time data science contractor with my own business.
When I was volunteering at a dementia activities group, we would see many older people, all of whom would be suffering from some form of memory complaint or other cognitive issues. What defined this group, besides receiving some form of dementia diagnosis, was just how different they all were and how the disease progressed differently in each of them. They came from various backgrounds with different, often fascinating, life stories to tell. One man had hiked the coast-to-coast trail whilst facing backwards to raise money for charity, another was one of the first postmistresses, at a time when the roles were predominantly for men. Another woman was in the Women’s Auxiliary Air Force during World War 2 and a man with early-onset Alzheimer’s had been a vet before his sudden and dramatic decline.
It is decline specifically that our latest paper addresses. The paper is called A Machine Learning Approach for Predicting Deterioration in Alzheimer’s Disease. During my volunteer days, I would see some people I had grown to care about dramatically and tragically decline. However, some would persist in what seemed like a state of mild forgetfulness for a long time, sometimes even years. This paper sought to explore the problem of decline within disparate groups, and in the end, we discovered that the task of predicting deterioration in healthy individuals is quantifiably different to predicting deterioration within those who already had mild cognitive issues.
One of the datasets I use to try and predict dementia is the Alzheimer’s Disease Neuroimaging Dataset or ADNI. This is a database consisting of different data types that were already collected from the same participants. This is so researchers could use this to create impactful research into Alzheimer’s Disease. Participants who attended the study did so at various points from 2004 onwards. The sample included those who had no cognitive issues (called ‘Cognitively Normal’ in our paper) and those who had been diagnosed with Mild Cognitive Impairment. This is a formal diagnosis that is used to assess individuals that are experiencing some form of memory complaint or other cognitive issues, but these issues are not yet serious enough to justify a diagnosis of Alzheimer’s Disease or another type of dementia.
The data collected in ADNI included brain imaging data such as images collected from MRI or PET (these are different types of equipment that are used for brain imaging), and fluid biomarker data collected via the spinal cord using an injection which draws out cerebral spinal fluid from the spine. A biomarker is used in science as an indicator for a biological pathway that may be happening in the body. These biomarkers are known to reveal certain characteristics that indicate the possibility of dementia such as the build-up of amyloid plaque and tau. Both of these are specific types of protein that sometimes build up in the brain and are indicative of dementia pathology. The ADNI data also contains neuropsychological tests, such as the Mini-Mental State Exam, which are administered by a GP or a frontline health professional, as well as data on genetics, and participant demographics.
Our paper used machine learning in order to predict deterioration in those who had been diagnosed as cognitively normal and those who had been diagnosed as having mild cognitive impairment at the start of the study. We defined deterioration as the participant receiving a worse diagnosis upon their final visit to the study. For example, if a participant received a diagnosis of cognitively normal at the start, but a diagnosis of either mild cognitive impairment or Alzheimer’s disease at their final visit, they were defined as having deteriorated. The same applied to participants who were diagnosed with mild cognitive impairment at the start of the study, except their definition of deterioration was having received a diagnosis of Alzheimer’s disease.
We separated these two types of participants (cognitively normal and mild cognitive impairment) into two groups and applied machine learning models to those groups separately. As our paper specifically wanted to compare the difference in predicting deterioration between the two groups, it made sense to treat them separately so that the differences would hopefully be apparent.
We applied six different machine learning models to each of these groups in turn. Machine learning uses statistics to build models that learn from the data and make predictions about future data based on those learnings. In essence, the models learn how to make decisions without a human telling them how to make those decisions. These models ranged in complexity from standard regression models that would be taught in any undergraduate statistics course, to much more complicated models such as Gradient Boosting Machine and Support Vectors Machines. The variety of models used is important because it allows for the possibility that one of our groups would produce better results in making predictions on a specific type of machine learning technique.
As it turned out, we were able to clearly see several differences between the two groups, not only in terms of how well we were able to predict deterioration, but also in how the model was built to make those predictions.
We found that, in general, all models were better at predicting deterioration in the cognitively normal group than in the group with mild cognitive impairment. The list of important predictors for the best models for each group was different as well. For the cognitively normal group, the top 6 most important predictors for their best model were all neuropsychological tests, such as the Mini-Mental State Exam. In comparison, the top predictors from the mild cognitive impairment group were more of a mixed bag, with imaging, demographics, neuropsychological testing, and even the effect of time being cited in the top 6 most important predictors for the best model.
Another interesting finding was the type of models that tended to do best in the two groups, with the cognitively normal group favouring more complicated models, and the mild cognitive impairment group favouring less complicated modelling based on simple calculations to separate those who had deteriorated from those who remained stable.
So what does this tell us?
Ultimately the strongest conclusion one can draw from this work is that the task of predicting deterioration is more complicated than we might at first think. In particular, the task of predicting which cognitively normal person will deteriorate is different from predicting deterioration in those already suffering some form of cognitive issue.
From a clinical perspective, it certainly gives pause for thought. We want to be able to predict deterioration so that we can either implement strategies to try and delay that deterioration or otherwise allow patients an informed understanding of their likely disease progression. However, as we tackle this problem of predicting deterioration, this paper has shown that perhaps the problem is in fact not one problem but two: Firstly, predicting a decline in healthy people and secondly, predicting a decline in those already struggling with cognitive complaints. Each of these challenges should be taken in isolation, according to this paper, and being able to predict each one will provide us with different opportunities in managing and improving a patient’s long-term outcome. However, given that this is a relatively small study, performed on a clinical dataset, an important next step is to validate these results on larger, community-based data, to see if the results can be replicated.