Predicting with Uncertainty: Using Age Prediction Models to detect Brain Abnormalities
Benjamín Gutiérrez Becker (ML Research Partner from Technische Universität München)
Is aging absolutely inevitable and irreversible? Finding the ultimate elixir of eternal youth is a centuries-old quest, but with scientific discoveries developing at an unprecedented rate, it might not be entirely out of reach.
Scientists have been battling aging for decades and trying to find ways to stop it or at least slow down the process. The massive growth in scientific discoveries in biology, medicine, and disease diagnosis has incredibly boosted human longevity. We are now living longer than we ever did. Recent findings even point out that world population aged 60 or above is expected to more than double by 2050 and more than triple by 2100.
Studies have also highlighted that aging is not a homogenous process. Various factors can affect how we age, be it biological, genetic, psychological, environmental, and even dietary and lifestyle choices can significantly influence our aging process. For instance, phenomena like Super agers show how some individuals in their 70s and 80s can maintain robust physical and
cognitive abilities similar to that of much younger individuals by adopting a healthy lifestyle along with regular exercising and constant mental stimulation.
The advancement of artificial intelligence gives even more hope for the battle against aging. Progress in medical image segmentation and pattern recognition can help us identify aging and brain aging patterns more efficiently, and reach a more accurate and early diagnosis of diseases.
Can we estimate a person’s age based on a picture?
A machine learning algorithm can easily estimate a person’s age based only on a picture of their face. The apparent physical changes we associate with aging like grey hair and wrinkled skin are used to train these models.
Patterns related to brain aging, however, are more subtle and hard to discern, which makes the process of identifying brain aging patterns even more challenging.
The Subtlety of Brain Aging
The human brain with all its complexities remains a wonder; a three-pound intricate machine that performs a multitude of sophisticated functions; from intelligence, creativity, to emotion, reasoning, and memory. Due to the complexity of the brain structure and its inherent mechanisms, it is hard to grasp how the brain actually works and identify the patterns related to brain aging. Brain aging patterns are especially tricky to track because they are particular to each person and they depend on different factors such as gender, ethnicity, environment, psychological well-being, lifestyle choices, etc.
Similar to its impact on our physical features, aging has a tremendous effect on our brain structure. One of the noticeable patterns of brain aging is how different brain regions grow or shrink at different stages throughout our lifespan, which can lead to a decline in various cognitive abilities.
By tracing these changing patterns in the shape and size of different brain regions, machine learning models can estimate a person’s age based on MRI images of their brain. More importantly, new studies are leveraging current age prediction models to detect abnormal brain structures caused by a variety of conditions such as Alzheimer’s disease, schizophrenia, diabetes, etc.
Age Prediction Models: Strengths & Shortcomings
Age prediction models are an attractive tool to be exploited for measuring brain abnormality due to its simplicity and because models can be trained using only images of healthy controls without having to recruit patients suffering from a specific condition, especially that recruiting specific patient cohorts is a costly and often impossible process. The trained model, afterward, can predict the age of test subjects by measuring the prediction error between healthy controls and individuals diagnosed with a specific condition. Such models could identify brain anomalies as they begin to develop; enabling doctors with predictive insights to proceed with patient treatment before the actual symptoms are visible.
The problem, however, is how age prediction models strongly assume that brain changes caused by diseases are similar to the typical changing patterns related to brain aging. It is partially true that many diseases have similar degenerative patterns as aging for some brain structures, but diseases generally do not affect the whole brain in the same way that aging does.
An example of this is Alzheimer’s disease. Over a 50-year span, the volume of many brain structures, including the cerebellum (the area responsible for physical movement) and the hippocampus (which plays an essential role for long-term memory) consistently decrease. This decrease in the hippocampus volume is accelerated for individuals with Alzheimer’s disease; making the hippocampus of a 60-year old have a similar volume like that of a healthy 80-year old person. However, this accelerated aging does not impact the volume of the cerebellum. This means that 60-year-olds have a similar amount of cerebellum white matter whether they are diagnosed with Alzheimer’s disease or not (check the below boxplots to observe these patterns).
Uncertainty to measure Brain Abnormalities more efficiently
Diseases and aging affect the brain differently. So, how can we then use age prediction models to assess brain abnormalities more efficiently? In our recent paper: “Gaussian Process Uncertainty in Age Estimation as a Measure of Brain Abnormality,” we propose using prediction uncertainty instead of prediction error as a measure of abnormality.
What is prediction uncertainty?
Prediction uncertainty can be described as how confident a model is about a prediction. Generally, a model will have low uncertainty values when the test point is very similar to previously seen training examples; whereas it would have high uncertainty values for test points that are different from the training set.
In our case, we measure uncertainty using a Gaussian Process Regression (GPR) model. In a GPR model, the uncertainty of the predictions is measured by examining the density of samples of the training set falling in a specific region of the feature space. This means that a GPR model will be more confident about a prediction if it has already seen many similar cases before. On the other hand, if we ask for a prediction on a testing case that is entirely different from the training samples, we will still get a prediction, but with a high level of uncertainty.
Visualizing Our Uncertainty-based Measurements
Going back to our Alzheimer’s disease example, you can see in the video below an example of how our uncertainty-based framework operates. First, image-based features are collected from a set of healthy individuals (blue dots in the scatter plot). Given these training samples, we can then calculate an uncertainty map of the predictor at each point of the feature space (shaded area: blue corresponds to low uncertainty; red to high uncertainty). See how those areas with many samples correspond to areas with low levels of uncertainty? We assume that areas with low uncertainty levels correspond to areas with common-looking brains.
Finally, we perform age prediction on test subjects. Crosses correspond to healthy individuals and circles to individuals diagnosed with Alzheimer’s disease. In our experiments, we observed that predictions on healthy individuals tend to stay in areas with low uncertainty (the blueish area); while individuals with Alzheimer’s disease tend to drift towards areas with higher uncertainties (red areas). This happens because the features extracted from healthy individuals are generally similar to those in the training set; whereas individuals diagnosed with Alzheimer’s disease tend to have abnormal brain development, making their brains look different from those in the training set.
Predicting with Uncertainty: Why is it better?
How is prediction uncertainty different from the previously-mentioned approaches that use prediction error? While our approach still uses age prediction models; using uncertainty instead of prediction error as a tool to measure brain abnormalities proved to yield better results. Prediction uncertainty does not fall into the common pitfall of assuming that disease-based degenerative patterns are similar to that of accelerated aging.
Our experiments showed that prediction uncertainty could better identify significant differences between healthy controls and individuals with Alzheimer’s disease in a more consistent manner when compared to the standard approach using prediction error. We also found differences in prediction uncertainty between healthy controls and patients diagnosed with autism, which was not possible using the standard technique based on prediction error.
Our results pave the way towards more emphasis on the role of artificial intelligence and how it can transform modern medicine. By applying machine learning models, we can reach a better understanding of the intricacies of the aging process and how it affects our brain. More importantly, machine learning models such as the one we propose have the potential to detect different neurodegenerative diseases at very early stages; allowing physicians to adopt early treatments which can drastically improve the quality of life of patients.
Can artificial intelligence boost human longevity? The possibilities and applications of AI, particularly, in healthcare are yet to unravel, but it can certainly help us reach more groundbreaking discoveries that were — until recently — deemed unattainable.