Across Time and Space: Machine Learning Reveals Paths to Dementia
Researchers face a challenge in understanding the brain changes during the long course of Alzheimer’s disease. It’s not possible to track neurodegeneration continuously in individual people for up to 30 years, so instead scientists collect snapshots of the disease from different people in all stages of the disease. Now, using advanced computational approaches and a massive trove of MRI brain volume data, scientists have stitched together a series of these snapshots. This way, they identified disease subtypes with distinct progression patterns in people with Alzheimer’s disease or with mutations that cause frontotemporal dementia (FTD) and amyotrophic lateral sclerosis (ALS). They dubbed their method SuStaIn, for Subtype and Stage Inference.
- Machine learning identifies disease subtypes in AD, FTD, based on patterns of brain shrinkage over time.
- AD subtypes differ in their risk for progression to dementia.
- This approach could help stratify patients for clinical trials, personalized medicine.
In FTD/ALS, the subtypes correlated with causative mutations. In AD, three major subtypes and their trajectories ended on distinct patterns of neurodegeneration previously identified in postmortem studies. They also predicted the risk of clinical progression from mild cognitive impairment to dementia. The work, from the labs of Daniel Alexander, Jonathan Rohrer, and Jonathan Schott, all at University College London, appeared October 15 in Nature Communications.
Eric Oermann of Mount Sinai Health System in New York praised the work. “It is conceivable that such a model could help us improve our ability to diagnose patients with this disease and, most importantly, lead to better cohort selection for clinical trials. Many clinical trials of drugs for these diseases, particularly Alzheimer’s, suffer from poor phenotyping. SuStaIn is a step toward changing that,” he wrote to Alzforum.
Sorting Things Out. A computational approach assumes multiple disease phenotypes based on progressive brain atrophy (a), which can be reconstructed from cross-sectional brain imaging data (b, c), to produce a model for staging and subtyping patients based on a single MRI. [Courtesy of Young et al., Nature Communications.]
First author Alexandra Young set out to develop computational algorithms, including machine-learning techniques, that could transform cross-sectional imaging data to a longitudinal picture and reconstruct the sequence of disease progression. But she wanted to go a step further. Rather than have the computer assume everyone progressed along the same path, she asked if it could identify subgroups that followed different trajectories over time.
To start, Young drew on MRI regional brain volume data from 172 people in the Genetic FTD Initiative (GENFI) study. These volunteers all carried FTD- or ALS-causing mutations in progranulin (GRN), tau (MAPT), or C9ORF72. Participants spanned the clinical spectrum from asymptomatic to dementia. Young developed SuStaIn by combining clustering algorithms that identified similarities between scans with disease progression modeling to arrange the scans in logical order. By doing this, she created a new algorithm that simultaneously grouped people into disease subtypes and created a sequential staging of the topology of brain-volume loss in different brain regions. Without accounting for genotype, the algorithm identified four disease subtypes. Intriguingly, these roughly corresponded with genetic groups. Most GRN mutation carriers followed one trajectory characterized by asymmetric frontal lobe degeneration, while MAPT carriers displayed a mostly temporal lobe subtype. The C9ORF72 carriers broke into two distinct phenotypes, one dominated by frontotemporal lobe and the other by subcortical degeneration. It remains to be seen if the two trajectories correspond to the two clinical presentations of C9ORF72 mutations, namely behavioral variant FTD and ALS.
Just by knowing a person’s SuStaIn subtype, the researchers could accurately predict the genotype 86 percent of the time. If they relied only on a subtype model, where scans were classified into static groups without factoring in temporal progression, that accuracy fell to 69 percent. The authors pointed out that while they would not use this model to predict genotypes since those are easily identified, this exercise validates the algorithm’s ability to pick out true disease subtypes starting with only MRI data.
Dementia Three Ways. Atrophy in three distinct AD phenotypes (typical, cortical, and subcortical) starts and ends in characteristic brain regions. Sigma scores express decrease in brain volume compared with amyloid-negative, cognitively normal subjects. [From Young et al., Nature Communications.]
Turning to AD, the investigators pulled 3T MRI data from 793 ADNI participants, spanning the range from cognitively normal to dementia. The machine-learning technique identified three predominant subtypes of AD. About one-third of the subjects had what the authors called typical disease, where atrophy showed up first in the hippocampus and amygdala then spread through the temporal lobe, and into the parietal, frontal, and finally the occipital cortex. Another third displayed a cortical subtype, involving early changes in the insula and cingulate cortices followed by occipital/parietal and finally widespread cortical pathology. The final group showed a subcortical profile, starting in the pallidum, putamen, and caudate before progressing to the temporal lobe and further cortical areas much later in disease. In each subtype, the final stages of progression resembled distinct patterns of postmortem neuropathology previously described.
Young got similar results when she analyzed a second, largely independent set of 1.5T MRI scans from 576 ADNI participants, of whom 59 had also had the 3T scan. In this analysis she also identified an additional, minor subtype. Accounting for only 4 percent of scans, it appeared to coincide with the rare posterior cortical atrophy variant of AD.
The model drew distinctions between the subtypes early in disease: Even in people with MCI, 37 percent of subjects could be strongly assigned to a subtype based on a single MRI. The stratification became more pronounced later, at dementia stages, where 78 percent of people were strongly assigned to one of the three subtypes. This means that at least some early stage people could, theoretically, be stratified by subtype, Alexander told Alzforum, although at present the majority cannot.
Triangulating Trajectories. Probability plots show how classification of each person by clinical group works better with clinical progression. Dots nearest corners have the highest probability of belonging to that subtype, while those in the middle show no strong inclination toward a specific phenotype. [Courtesy of Young et al., Nature Communications.]
The classification offered useful prognostic information, where the risk of progressing from MCI to dementia varied depending on the subtype and stage. Members of the typical group had the highest likelihood of moving from MCI to dementia, while subcortical had the lowest. As seen in the FTD data, using both subtype and stage information gave better predictive power than either alone.
“The study is well-done and mathematically rigorous,” said Oermann. “Some of the assumptions seem a bit generous, and the feature set seems a little sparse. They only consider volumes from a few major areas, and we already know that those volumes do correspond to the disease. That said, it is an ambitious attempt to integrate our knowledge of one of the few well-recognized phenotypes (cerebral atrophy) with a comprehensive chronological model of disease progression.”
The study convincingly shows that SuStaIn performs better than previous machine-learning techniques to classify genetic subtypes of FTD in the GENFI study, and to predict conversion from MCI to AD in ADNI, wrote Gil Rabinovici of the University of California, San Francisco. He nevertheless was less convinced of SuStaIn’s utility for classifying people in early disease stages. Still, Rabinovici thinks this may reflect a limitation of MRI, which he said is inherently a later-stage disease biomarker, rather than of the model itself. “It will be interesting to see how the model performs compared to true longitudinal data in inferring stage-related atrophy,” he added. He noted that these data should be available in ADNI and in GENFI—and for other neuroimaging modalities as well.
“The technique is novel and potentially very relevant for clinical use,” Dragan Gamberger, a computational biologist at the Ruđer Bošković Institute, Zagreb, Croatia, wrote to Alzforum. With Murali Doraiswamy, Gamberger previously used machine-learning methods to identify a cluster of people in ADNI with markedly faster brain atrophy and clinical decline, who would be suitable subjects for clinical trials (Gamberger et al., 2017). While Gamberger cannot tell if Young identified the same cohort, the proof of the technique will be its clinical application, he wrote. “For all clustering approaches, there is no objective measure to validate them, and there is no exact way to determine if one methodology produces better results than another. The ultimate measure of the quality is the usefulness of the results for medical practice,” he wrote.
Senior author Alexander stressed the need for continuing the technique’s validation. “At the end of the day, it is a machine-learning algorithm and we all know those can go wrong,” he said. He’d like to use SuStaIn to classify participants in recently completed clinical trials to see if it would identify subgroups with different responses to treatment.
Going forward, Young said she wants to understand more about what it means to be assigned to a subtype and how that affects progression. “At the moment, we are just looking at atrophy patterns but ultimately we’ll include a larger range of markers, including cerebrospinal fluid measures, amyloid PET, tau PET, and FDG to capture the full progression.” That should give more power to discern differences at the earliest disease stages, she said.
The group is also starting to analyze how genetics and other factors relate to the subtypes. Genome-wide association analysis gives them preliminary evidence that one subtype of AD associates with a genetic predisposition to diabetes, for example. “By dividing people into groups, you can reveal hidden associations with genetics and lifestyle that you wouldn’t see if you consider everyone all together,” Alexander said.
In FTD, the researchers are intrigued by their finding of two subtypes in a single genetic group, the C9ORF72 carriers. Young said they want to follow up on that result to see if the two subtypes represent two ends of a disease spectrum or two distinct groups, and how they relate to genotype and the clinical phenotypes of ALS versus FTD. She said they also find a hint of a second subset in the MAPT carriers, which might stem from a select few mutations, but they need a larger study to confirm this. Fortunately, GENFI has now doubled its number of participants, and they are using the additional power to validate and further explore the phenotypes.—Pat McCaffrey
- Gamberger D, Lavrač N, Srivatsa S, Tanzi RE, Doraiswamy PM. Identification of clusters of rapid and slow decliners among subjects at risk for Alzheimer's disease. Sci Rep. 2017 Jul 28;7(1):6763. PubMed.
No Available Further Reading
- Young AL, Marinescu RV, Oxtoby NP, Bocchetta M, Yong K, Firth NC, Cash DM, Thomas DL, Dick KM, Cardoso J, van Swieten J, Borroni B, Galimberti D, Masellis M, Tartaglia MC, Rowe JB, Graff C, Tagliavini F, Frisoni GB, Laforce R Jr, Finger E, de Mendonça A, Sorbi S, Warren JD, Crutch S, Fox NC, Ourselin S, Schott JM, Rohrer JD, Alexander DC, Genetic FTD Initiative (GENFI), Alzheimer’s Disease Neuroimaging Initiative (ADNI). Uncovering the heterogeneity and temporal complexity of neurodegenerative diseases with Subtype and Stage Inference. Nat Commun. 2018 Oct 15;9(1):4273. PubMed.
To make an annotation you must Login or Register.
This study describes SuStaIn, a very interesting new machine-learning approach that incorporates both temporal (i.e., disease stage) and phenotypic (i.e., disease topography) dimensions toward modeling MRI data in genetic FTD and sporadic AD. The main advance is the inclusion of both temporal and phenotypic dimensions in a single model, overcoming a limitation of previous approaches which considered only one of these dimensions at a time.
The study convincingly shows that SuStaIn performs better than traditional machine learning in classifying genetic subtypes of FTD in the GENFI study, and in predicting conversion from MCI to AD in ADNI. I find the data a bit less convincing in terms of the utility of SuStaIn for individual classification at the earliest disease stages—it seems like there the classification probabilities are centered in an “uncertain zone” in unaffected FTD mutation carriers and patients with MCI (Figure 5). This may represent a limitation of MRI, which is inherently a later-stage disease biomarker, rather than a limitation of the model itself. It will be interesting to see how the model performs compared with true longitudinal data in inferring stage-related progression of atrophy—these data should be available in ADNI and GENFI—as well as with other neuroimaging modalities.
Finally, as acknowledged by the authors, disease heterogeneity is more complex than just topography and stage. It will be exciting to see how the boundaries of machine learning can be pushed even further to capture the complexity of these biological phenomena.
Rudjer Boskovic Institute
This is a very good paper. I completely agree with the main finding that segmentation of the AD population is a relevant first step for better analysis of this disease. All our previous experiments on the ADNI dataset demonstrated that there exist strong differences among MCI and AD subpopulations (see references below). From the paper I cannot conclude if subpopulations we have managed to detect are well in agreement with the subpopulations that have been detected in here by Young et al.
The methodology they use is novel and potentially very relevant. It is a statistical approach to clustering (segmentation) of data sequences in a multiview setting. A known fact about all clustering approaches is that there is no objective measure to validate them and there is no exact way to determine if a result of one methodology is better than a result produced by another methodology. The ultimate measure of the quality is only usefulness of the results for the medical practice. We have previously shown how computational AD subtyping can be validated using longitudinal clinical outcomes. Our unbiased multilayer clustering algorithm identified, in the same ADNI data set, a cluster of people with markedly greater brain atrophy and clinical decline rates. Such a cluster is highly suitable for clinical trials (Gamberger et al., 2017).
Gamberger D, Lavrač N, Srivatsa S, Tanzi RE, Doraiswamy PM. Identification of clusters of rapid and slow decliners among subjects at risk for Alzheimer's disease. Sci Rep. 2017 Jul 28;7(1):6763. PubMed.
Gamberger D, Ženko B, Mitelpunkt A, Lavrač N, Alzheimer’s Disease Neuroimaging Initiative. Homogeneous clusters of Alzheimer's disease patient population. Biomed Eng Online. 2016 Jul 15;15 Suppl 1:78. PubMed.
Gamberger D, Ženko B, Mitelpunkt A, Shachar N, Lavrač N. Clusters of male and female Alzheimer's disease patients in the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. Brain Inform. 2016 Sep;3(3):169-179. Epub 2016 Mar 30 PubMed.
Mount Sinai Medical Center
It is a curious fact that most radiologists, and even most neurologists, when presented with an MRI scan of a patient suspected to have a neurodegenerative disease, rarely pay much attention to the morphology of the brain, other than to comment on hippocampal and/or gross lobar atrophy. The study by Alexandra Young and colleagues informs interested neuroscientists everything they would want to know about patterns of regional brain atrophy, but were afraid to ask. The study is an exhaustingly detailed report describing the heterogeneity of patterns of regional brain atrophy in two major neurodegenerative diseases and their subtypes, namely frontotemporal dementia and Alzheimer’s disease.
It is not, however, enough to merely recognize patterns of regional atrophy, which are characteristic of the subtypes of these two groups of diseases. The patterns of atrophy change as the disease progresses, thus providing diagnostic information not only about the disease subtype, but also about the stage of the disease. The authors have used a machine-learning technique on MRI volumetric data from a large number of cases derived from the Genetic FTD Initiative (GENFI) and the Alzheimer’s Disease Neuroimaging Initiative (ADNI) to disentangle patterns of atrophy that are identifying characteristics of each subtype in these two disease groups and of the disease stage, hence “Subtype and Stage Inference” or “SuStaIn.”
One might ask how this interesting technique has utility beyond providing information about subtype and stage. It turns out that the some of the genetic heterogeneity of the frontotemporal dementias can be identified by recognition of the phenotypic heterogeneity. Within each genetic subtype, an even more fine-grained analysis using this method allows discovery of further phenotypic heterogeneity.
In Alzheimer’s disease, by recognizing patterns of atrophy for each of three major subtypes of disease (limbic predominant, hippocampal sparing, and typical, which is a combination of the first two subtypes), it is possible to identify the stage of the disease, and most likely to predict expected rate and pattern of progression of the disease. This information could be useful to the clinician who is often asked by the patient to predict the rate of progression of the disease. Further, accounting for this heterogeneity in the expected individual rates of cognitive and functional progression could be very useful in analyzing effectiveness of an intervention in clinical research trials, much as knowledge of the APOE genotype has made a large impact in analysis of Alzheimer clinical trials.
Mount Sinai Health System
The study is well-done and mathematically rigorous. Some of the assumptions seem a bit generous—it likely isn’t true that biomarkers have a discrete and stepwise progression, uniform priors isn’t really accurate, etc., and the feature set seems a little sparse—they only consider volumes from a few major areas and we already know that those volumes do correspond to the disease. That said, it is an ambitious attempt to integrate our knowledge of one of the few well-recognized phenotypes (cerebral atrophy) with a comprehensive chronological model of disease progression. It is conceivable that such a model could help us improve our ability to diagnose patients with this disease and, most importantly, lead to better cohort selection for clinical trials. Many clinical trials of drugs for these diseases, particularly Alzheimer’s, suffer from poor phenotyping. SuStaIn is a step toward changing that.
Make a Comment
To make a comment you must login or register.