There has been an abundance of studies demonstrating biological changes related to ME/CFS, but a diagnostic biomarker has yet to be established. Diagnostic biomarkers are measurable clues within the body that give information about the presence or progression of a disease.
As currently the diagnosis of ME/CFS is largely based on self-report measures and the exclusion of other conditions, many people with ME/CFS face a lack of validation and friction when accessing healthcare and social services due to scepticism. Hence finding a diagnostic biomarker is so important as this will provide unquestionable ‘laboratory proof’ which will help dispel scepticism over the disease and aid faster diagnosis for a person presenting to their GP.
Therefore, the latest contribution by Dr Karl Morten and colleagues to this concerted effort is very much welcome. The research group combined Raman spectroscopy – an analytical tool which uses light – and artificial intelligence (AI) to study white blood cells of individuals with ME/CFS in comparison to individuals with multiple sclerosis (MS), and healthy controls. The lead author of this paper was Raman spectroscopy specialist, Dr Jiabao Xu.
Using these techniques, the researchers state they were able to distinguish with high accuracy (91%) between healthy controls, individuals with MS, and individuals with ME/CFS, and further differentiate (with an accuracy of 84%) between mild, moderate and severe ME/CFS. These findings are indeed promising. However, it is important to note it was a small study which needs be replicated on a larger scale, and under varied conditions, to validate the findings.
A deeper look
What did the study do?
Morten and colleagues set out with the aim of developing a diagnostic method for distinguishing ME/CFS from healthy controls and other diseases which cause fatigue, such as MS.
Peripheral blood mononuclear cells (PBMCs) – a subset of white blood cells – were obtained from the UK ME/CFS Biobank which stores frozen blood samples from individuals with ME/CFS (fulfilling the Canadian Consensus Criteria and/or Fukuda criteria), and from individuals with MS, in addition to healthy controls. The study involved 98 samples in total: 61 ME/CFS, 21 MS and 16 healthy controls. Using their clinical profiles, the ME/CFS cohort was further divided based on symptom severity (mild, moderate or severe). The implied rationale for including individuals with MS in this study was that their clinical symptoms overlap significantly with people with ME/CFS, therefore it would be useful to know if it is possible to differentiate these conditions using laboratory tools.
The researchers initially attempted to assess mitochondrial respiration (the process by which cells generate energy) in PBMCs in 41 of the 98 individuals, but this approach did not effectively differentiate between disease cohorts. Changing tack, they employed Raman spectroscopy which is a technique that enables the analysis of the molecular and chemical composition of cells by measuring the scattering of laser light and translating this into Raman spectra (graphs). Whilst there were only 98 subjects, a huge amount of data (over 2,000 Raman spectra) was produced from their PBMCs.
The Raman data were then analysed using linear discriminant analysis (LDA). In simple terms, LDA is a classification model – a smart tool that helps sort items into different groups/categories based on their characteristics. Hence, the intention was that it could be trained to differentiate between those with ME/CFS, MS and healthy controls by identifying patterns and characteristics within the molecular/chemical profiles obtained through Raman spectroscopy. The researchers set aside a test dataset for which they already knew the diagnoses, which allowed them to assess whether the LDA model was correctly trained to distinguish the Raman data based on diagnoses.
Additionally, LDA was used to identify the most significant peaks in the Raman spectra which related to certain important molecules, primarily amino acids – tryptophan, tyrosine and phenylalanine. All of these molecules have been previously implicated in theories relating to the development of ME/CFS, including in a pilot study by the team suggesting phenylalanine as a potential biomarker.
LDA is only one type of classification model. The researchers employed seven more classification models and combined the outputs to create a more powerful model. This technique is known as ensemble machine learning, which is essentially artificial intelligence which takes into account the predictions made by individual models to create a more precise prediction. Therefore, Raman data from an individual’s blood sample can be fed into an ensemble ‘machine’ which should theoretically be able to provide a diagnosis based on all of its training data.
What did they find?
Findings from initial analysis of Raman data
Using LDA alone, the researchers were able to differentiate the data from Raman spectroscopy into separate groups representing each of the cohorts – ME/CFS, MS and healthy controls. However, the researchers had some difficulty additionally differentiating ME/CFS disease severities.
This is visualised below in a simplified version of the figures in the paper. As seen in figure 1, LDA could separate the Raman data into three clusters (ME/CFS, MS and healthy controls). However, when tackling the additional task of differentiating between ME/CFS disease severities, LDA revealed distinctive clusters for severe ME and healthy controls, but an overlap between mild and moderate ME/CFS and MS (figure 2). Whilst further manipulating the data allowed for better differentiation, the LDA model was only accurate at classifying the data into the five groups (mild, moderate and severe ME/CFS; MS; and healthy controls) around half the time.
Note: LD1 and LD2 in the figures above simply represent two theoretical ‘measuring sticks’ which LDA uses to gain different perspectives on the data.
Additionally, LDA demonstrated a universal increase in the amino acids, tryptophan and tyrosine, across all subgroups of ME/CFS, and MS. On the other hand, relative to controls, phenylalanine was significantly reduced in moderate and severe ME/CFS, but elevated in mild ME/CFS and MS. The researchers suggested these mixed findings could be due to ‘metabolic subtypes’ existing in individuals with ME/CFS. Mixed findings related to levels of lipids was also observed.
Findings from advanced analysis of Raman data
As mentioned previously, the researchers were aiming for a greater accuracy and so employed ensemble machine learning. They were not disappointed: the ensemble model demonstrated an ‘enhanced predictive power’ and showed a much higher capacity in differentiating between ME/CFS patients, MS patients and healthy controls, and also between subgroups of ME/CFS based on symptom severity:
- The overall accuracy of the ensemble model in classifying Raman data as either ME/CFS, MS or healthy control was 91%.
- The overall accuracy of the ensemble model in classifying Raman data into five groups (mild, moderate and severe ME/CFS; MS; and healthy controls) was 84%.
Whilst there is still some room for improvement, these are highly promising findings.
Discussion
According to the researchers, “this is the first research using Raman spectroscopy and advanced machine learning techniques to discriminate subgroups of ME/CFS patients based on the symptoms severity, achieved with high accuracy, sensitivity, and specificity”. The study findings open the possibility of a simple and minimally invasive diagnostic method for ME/CFS that only requires a small sample of blood for analysis. Although they do point out a barrier to adoption could be that the Raman spectroscopic approach is “not yet readily available in certified diagnostic laboratories”.
Whilst this is a pioneering study, the researchers recognise that it is still the early stages and the approach needs to be validated in larger study populations, and further optimised. Additionally, as the blood samples were frozen, as they were obtained from the UK ME/CFS biobank, the researchers would also like to test the approach on freshly fixed (non-frozen) samples for comparison.
Read more about what makes an effective diagnostic biomarker
