What is fMRI?
Ever wonder how researchers get those colorful blobs on brain images? It's usually through a technique called functional neuroimaging. Specifically, many functional neuroimaging studies use functional MRI (fMRI) as the main tool for data collection.
Our brains have an intricate set up of blood vessels that supply different regions. When a region is engaged in activity, there is a higher need for oxygen at that region. So what happens? Hemoglobin in red blood cells delivers oxygen to those spots through increased blood flow. Because of this, fMRI is fully called “BOLD fMRI,” meaning blood oxygenation level dependent fMRI.
An MRI machine is like a giant, electromagnetic camera (swipe to see what it looks like). The magnetic strength of commonly used machines is about 50,000 times greater than the Earth's magnetic field. In fMRI, the magnet is
used to track blood flow across the brain.
Two main kinds of pictures are structural (i.e., what does your brain look like) and functional (i.e., where in your brain is activity happening). The structural images are usually "high resolution," meaning you can see brain structures clearly.
Functional images are not so clear and are actually purely gray scale in their raw data form (swipe for picture). These gray scaled images are in 4D, meaning they are a 3D representation of the brain over time (the 4th dimension). The gray scale images show signal changes as a result of blood flow over time.
After data collection is complete, we put raw images through image processing steps that improve the signal we’re interested in over messy background noise. Part of this processing is to align the structural MRI high resolution brain picture to the functional images.
Finally, we align each person’s data to a standard brain template, meaning some parts of the brain have to get slightly stretched or smushed to fit in this template. Once every participant’s data is aligned, we can finally make comparisons across our sample using statistical tests. Results of these statistical tests show up as the colorful blobs that you see!
Neuroimaging Biomarkers of Subjective Experience
Pain, sadness, anxiety, hallucinations, fatigue, and nausea. What do all of these symptoms have in common? They are subjective experiences, meaning only the individual perceiving the symptom knows what it feels like to him/her. You might tell others that you are in pain, and they can draw from their person experiences with pain to try to understand your experience. However, they will never truly know what your pain feels like. In other words, it is intangible because you can’t directly see or feel someone else’s subjective experience. The person needs to tell you about it for you to understand.
The question of where subjective experiences originate in the brain has been a focus of cognitive neuroscience for a long time without conclusive answers. More recently, this question has been applied to people with
distressing subjective experiences that seek clinical treatment. Specifically, functional neuroimaging researchers have wondered whether we can use brain scans to provide evidence for a patient’s subjective experience by identifying objective markers.
Objectivity is the opposite of subjectivity because it refers to tangible, measurable data. Often, subjective and objective data are combined in clinical diagnosis. For example, a person telling a healthcare provider that (s)he is experiencing increased thirst and need to urinate (subjective report) who shows very high levels on A1C blood tests (objective marker) is ultimately diagnosed with Type I Diabetes. A blood test is an example of a “biomarker,” or an objectively measured characteristic that acts as an indicator of normal or abnormal biological functions and helps a healthcare provider track the body’s responses to treatments.
So the big question is: can a functional brain scan act as an objective biomarker of your subjective symptoms? Although some researchers are hoping that this will be the case one day, there are important philosophical, methodological, and ethical factors to consider before we use fMRI to help diagnose conditions that are predominately based in subjective experience, like chronic pain, chronic fatigue, major depressive disorder, bipolar disorder, etc.
Before continuing, I’ll emphasize that I think neuroimaging is a wonderful tool to help researchers understand what might be going on in our brain during a specific mental process. I use it in my research regularly. But, there are important limitations that don’t often get discussed in the media, and there is often an over-promise of neuroimaging’s state of the science. These limitations don’t matter as much when we are looking at data from the average of a group of individuals to speculate about brain function. These limitations DO matter strongly when we think about how we might use neuroimaging to make healthcare decisions for individual people that can strongly impact their quality of life.
In some studies proposing biomarkers of subjective experience, justifications have sounded something like “[Pain] is primarily assessed by means of self-report, an imperfect measure of subjective experience” . Some have even gone so far as to say, “Neuroimaging has the potential to become an objective measure of pain and replace subjective report” . These examples come from the field of pain neuroimaging, but can also be found across the research areas of depression, schizophrenia, etc.
These statements bring up the problem of reverse inference. In neuroimaging, reverse inference happens when we assume that the presence of brain activation (i.e., the blobs on an analyzed brain scan) necessarily means that a person is engaged in a specific mental task. For example, activity in a region called the amygdala is often associated with fear. So if a provider sees a brain blob in the amygdala on a patient’s scan, the reverse inference would be assuming that (s)he is experiencing anxiety. But, how will we ever know if the person really was in a state of anxiety? By asking them! So, we can’t get rid of the patient’s self-report if we want to confirm the patient’s experience.
Others have suggested that it’s not so much the presence of one blob, but the pattern of several blobs that will make for a useful biomarker. These blob patterns are generated from machine learning techniques, often by asking people who have the clinical condition of interest (e.g., major depression) and people who don’t have that condition to be scanned. Then, we apply machine learning to all of the data and see how well the machine learning algorithm separates between people with and without the clinical condition. But, this approach relies on researchers’ ability to distinguish between patients and non-patients. How do we do that for conditions that rely on subjective experience? We rely on the person’s self-report of symptoms in a clinical assessment!
Aside from these philosophical points, there are methodological concerns in how we might identify and use neuroimaging biomarkers of subjective experience. Two fundamental considerations in this category are reliability and validity (see link for information about why those metrics are important). Briefly, we want to make sure that our neuroimaging biomarker is reliable (i.e., if we scan someone with a clinical condition at two separate times, will his/her brain blobs show the same pattern?) and valid (i.e., if we want to measure anxiety, are we actually measuring anxiety or do the blobs reflect excitement? Both experiences can result in a brain blob on the amygdala).
Second, some have expressed concerns about whether fMRI is even the right tool to produce neuroimaging biomarkers. These authors have noted that this technology doesn’t measure neurotransmitters specifically and is too far removed from neuron function to provide meaningful diagnostic information. With fMRI, we measure signal changes that happen with changes in blood flow over time. This blood flow is thought to be a signal of brain activity that generates from changes in neurochemistry and neuron function, but we do not directly measure those processes.
Third, reports looking across various studies have demonstrated that sample sizes matter in how well machine learning algorithms perform. Larger samples tend to show weaker accuracy for determining which participants do or do not have the condition in question. This finding is thought to result from a greater variety of symptoms in larger samples and better uniformity of symptoms in a smaller sample. Additionally, these algorithms are often applied to assume that 50% of people will have the symptom of interest and 50% of people will not. This approach does not reflect how often a symptom is likely to be present in a real world clinic (i.e., higher %) or general population (i.e., lower %). By using a 50/50 split, machine learning algorithms become biased in their ability to identify who does and does not have the symptom.
Finally, there is an over representation of positive, published findings. There is a general bias in scientific publishing, so that mostly positive results get published, and negative results are not reported. This bias in publishing is important to consider because the way some researchers generate their machine learning algorithms is based on what we know about brain regions associated with a symptom/clinical condition. If we only look at the published studies and ignore the unpublished, negative findings, we might be biased in how we choose brain regions to include for our machine learning algorithm that ultimately creates the neuroimaging biomarker.
In a perfect world, objective biomarkers of subjective experiences would be highly reproducible, appropriately capture the symptoms they aim to measure, remain free from sampling bias, and undergo development without reverse inference. Let’s say all of those factors are in place. What happens though, if the biomarker says a person does not have the symptom, but a person is adamant that (s)he does have the symptom? In the research world, we call this Type II error, or the likelihood of a false negative.
As a healthcare provider, do you change your belief in their subjective report and treatment of this person? What if you are the patient who has a biomarker telling a provider you do not have the symptom, but you know you are experiencing the symptom? Is there anything you can do now to convince your provider of your symptom and get the treatment that you need? These are just some of the ethical questions associated with this line of research.
 Logothetis, N. K. (2008). What we can do and what we cannot do with fMRI. Nature, 453(7197), 869.
 Strother, S., La Conte, S., Hansen, L. K., Anderson, J., Zhang, J., Pulapura, S., & Rottenberg, D. (2004). Optimizing the fMRI data-processing pipeline using prediction and reproducibility performance metrics: I. A preliminary group analysis. Neuroimage, 23, S196-S207.
 Biomarkers Definitions Working Group, Atkinson Jr, A. J., Colburn, W. A., DeGruttola, V. G., DeMets, D. L., Downing, G. J., ... & Spilker, B. A. (2001). Biomarkers and surrogate endpoints: preferred definitions and conceptual framework. Clinical Pharmacology & Therapeutics, 69(3), 89-95.
 Schnack, H. G., & Kahn, R. S. (2016). Detecting neuroimaging biomarkers for psychiatric disorders: sample size matters. Frontiers in psychiatry, 7, 50.
 Robinson, M., Boissoneault, J., Sevel, L., Letzen, J., & Staud, R. (2016). The effect of base rate on the predictive value of brain biomarkers. The Journal of Pain, 17(6), 637-641.
 Robinson, M. E., Staud, R., & Price, D. D. (2013). Pain measurement and brain activity: will neuroimages replace pain ratings?. The Journal of Pain, 14(4), 323-327.
 Wortolowska K. How neuroimaging can help us to visualise and quantify pain? Eur J Pain Suppl. 2011;5(S2):323–7.
 Davis, K. D., Racine, E., & Collett, B. (2012). Neuroethical issues related to the use of brain imaging: can we and should we use brain imaging as a biomarker to diagnose chronic pain?. Pain, 153(8), 1555-1559.
 Woo, C. W., & Wager, T. D. (2015). Neuroimaging-based biomarker discovery and validation. Pain, 156(8), 1379.
 Poldrack, R. A. (2006). Can cognitive processes be inferred from neuroimaging data?. Trends in cognitive sciences, 10(2), 59-63.
 Bonnet, L., Comte, A., Tatu, L., Millot, J. L., Moulin, T., & Medeiros de Bustos, E. (2015). The role of the amygdala in the perception of positive emotions: an “intensity detector”. Frontiers in behavioral neuroscience, 9, 178.
 Davis, K. D., Flor, H., Greely, H. T., Iannetti, G. D., Mackey, S., Ploner, M., ... & Wager, T. D. (2017). Brain imaging tests for chronic pain: medical, legal and ethical issues and recommendations. Nature Reviews Neurology, 13(10), 624.
 Wager, T. D., Atlas, L. Y., Lindquist, M. A., Roy, M., Woo, C. W., & Kross, E. (2013). An fMRI-based neurologic signature of physical pain. New England Journal of Medicine, 368(15), 1388-1397.