Artificial Neural Network Subgrouping of Corticobasal Syndrome Patients
Morgan Clinton+, Jim DeLeo+, Carl Leonard+ and Jordan Grafman++ National Institutes of Health Bethesda, Maryland
+ Scientific Computing Section, NIHCC ++ Cognitive Neuroscience Section, NINDS
OBJECTIVE
To explore using artificial neural network methodology to differentiate corticobasal syndrome (CBS) patient subgroups with a longer range goal of developing practical diagnostic and treatment selection comparative effectiveness tools that would improve patient outcome.
INTRODUCTION
Corticobasal Syndrome (CBS) is a neurodegenerative disorder that has several associated major subgroups. A computerized clinical decision support system that could use clinical data to provide improved subtyping of this disorder could potentially improve diagnoses and outcome for CBS patients. We set out to explore the use of artificial neural networks to build such a system.
Corticobasal Syndrome
Corticobasal Syndrome (CBS) is a degenerative disorder of the brain in which nerve cells die over time, causing a progressive decline in the ability of hand usage (usually more pronounced on one side). Other symptoms include arm or leg stiffness, tremor, gait unsteadiness, and speech difficulty. Some patients also have some decline in thinking ability, such as loss of skilled activities, poor problem solving abilities, poor concentration, problems with language, and forgetfulness [1, 2, 3].
CBS Subtypes
• Alzheimer’s Disease • Corticobasal Degeneration • Pick’s Disease • Progressive Supranuclear Palsy • Frontotemporal Dementia
• Terminal neurodegenerative disease. • Associated with brain plaques and tangles. • Most common cause of dementia. • Occurrence more typical after age 65.
Alzheimer's Disease
• Gradual nerve cell loss in cerebral cortical
• •
areas and basal ganglia. Causes problems with mobility. Markedly asymmetrical.
Corticobasal Degeneration
FDG PET images showing patterns of metabolic activity that are characteristic of patients with Alzheimer’s disease, Pick’s disease (frontotemporal dementia) and elderly individuals with no dementia.
Red indicates high FDG uptake
Figure 1.
Blue indicates low FDG uptake
Pick’s Disease
• Degeneration of brain nerve cells. • Tau proteins accumulate into “Pick bodies.” • Associated with shrinking of the frontal
•
and temporal anterior lobes of the brain. Symptoms may include: (1) changes in behavior or (2) problems with language acquisition.
Progressive Supranuclear Palsy (PSP)
• Rare neurodegenerative disease.
•
• Affects both neurons and glial cells. • Neurons display tangles and clumps of tau
protein. Symptoms include loss of balance, personality changes, vision problems, and later dementia and slurring of speech.
Frontotemporal Dementia (FTD)
• •
• • •
• •
Umbrella term for a diverse group of rare disorders, including Pick’s Disease. Predominantly affects the frontal and temporal lobes – portions of these lobes atrophy, or shrink. (Figure 2.) These brain areas are generally associated with personality and behavior. Signs and symptoms vary, depending upon the portion of the brain affected. Some people undergo dramatic changes in their personality and become socially inappropriate, impulsive or emotionally blunted. Others lose the ability to use and understand language. This disease is often misdiagnosed as Alzheimer’s.
Left Hemisphere
Right Hemisphere
Frontal lobes
Temporal lobes
Red regions correspond to degeneration of neurons resulting in lower brain volume
Figure 2.
Right lobe appears normal - Initially FTD affects the brain unilaterally
Artificial Neural Networks
Artificial neural networks (ANNs) are a broad class of general purpose computational tools inspired by biological neural networks. They are well suited for classification and clustering applications. Classification ANNs are used to predict a specific outcome like a diagnosis or to answer a question like “Will this patient do well on this treatment plan?” The NIH Clinical Center Scientific Computing Section has experience using different ANN paradigms in biomedical clustering and classification applications [4, 5]. In our work here we developed a new ANN paradigm called HubuHTM
Clustering ANNS
In the context of our work here “Clustering ANNs” refers to ANNs that group patients with similar characteristics. We developed a new clustering algorithm called HubuHTM. It uses test values described below to group the patients in our study. The HubuHTM algorithm finds the closest point to every point in the data set and counts the number of times each point is a closest point. The points with the highest counts are considered hubs or archetypes. Figure 3 shows an example of a HubuHTM graphic output from a study in which only two variables were used, viz. the “Beck Depression Index II” and the “Boston Naming Test Score.” Once the hubs are identified by the program, every point is assigned to the hub closest to it. Since points represent patients here, our computer program implementation of HubuHTM produces a report that lists patients in each cluster. (See Table 2.) The report is reviewed to see if it reveals any new knowledge about CBS.
Data Acquisition
Our data was provided by Jordan Grafman, Ph.D., a cognitive psychologist in the National Institute of Neurological Disorders and Stroke (NINDS). This data was associated with 107 CBS patients and came from one of his previous studies. The sample included patients without a distinct pathology for a specific subgroup disease. In Dr. Grafman’s previous work with these patients, cognitive psychological tests were performed. There were 40 parameters in the original data set. We selected 17 of them for our work here. To illustrate our new clustering methodology we chose only two of these, viz. the “Beck Depression Index II” and the “Boston Naming Test Score.” We selected the 61 patients with no missing data. (See Table 1.)
The Data
ID 2 3 5 7 9 10 13 14 16 20 21 22 23 24 26 X1 13 36 11 10 4 12 31 47 5 4 2 21 11 8 12 X2 48 33 50 19 48 26 44 16 41 54 51 49 50 25 33 ID 29 32 33 36 38 43 44 45 46 48 50 51 53 54 56 X1 4 19 17 23 38 9 8 6 13 19 20 10 26 14 22 X2 60 52 56 47 50 42 31 49 38 44 12 49 56 54 37 ID 57 59 60 61 63 64 65 66 67 68 70 71 74 75 77 78 79 80 X1 10 6 4 19 25 20 8 17 12 17 16 8 6 13 24 5 11 6 X2 56 24 49 59 50 40 50 53 57 41 33 48 59 39 49 56 51 48
ID 81 83 84 85 86 89 90 93 95 96 97 98 104 105
X1 8 14 6 37 10 12 24 4 2 1 14 6 31 27
X2 53 38 51 55 49 36 52 51 26 55 50 58 48 55
Table 1. Data used in this study.
P-67
P-48
P-24
Figure 3.
HubuHTM REPORT
HUB P-24 P-48 X1 8 13 X2 25 38 Patients 7, 10, 14, 24, 44, 50, 59, 95. 3, 9, 16, 26, 43, 46, 48, 56, 64, 68, 70, 75, 80, 83, 89.
P-67
17
53
2, 5, 13, 20, 21, 22, 23, 29, 32, 33, 36, 38, 45, 51, 53, 54, 57, 60, 61, 63, 65, 66, 67, 71, 77, 78, 79, 81, 84, 85, 86, 90, 93, 96, 97, 98, 104,105.
Table 2. Hub clustering's of CBS patients.
HubuHTM REPORT ANALYSIS
P-24 (8 patients)
• The 2 deceased patients in this group had a • •
confirmed autopsy diagnosis of Alzheimer’s. The hub’s average Boston Naming Test (BNT) total correct score for the 8 patients was 23.4, lower than group P-48 (39.4), and group P-67 (52.0). Beck Deviation Index was greater than average and had the highest deviation from the patient mean. (See Table 3.) P-67 (38 patients)
• The average BNT Score was higher than the other
•
two groups. (See Table 3.) 4 out of the 8 deceased patients from this hub were diagnosed with CBD.
HUB
Boston Naming Test (Group Average)
Beck Depression Index –II (Group Average)
P-24
P-48
23.4
39.4
15.4
14.5
P-67
52.0
14.8
All Patients
44.9
14.8
Table 3. Psychological Test Scores
RESULTS
We have developed and demonstrated HubuHTM, a new clustering algorithm that shows promise for discovering new knowledge related to Corticobasal Syndrome. Although we used only two parameters for simple demonstration purposes, the HubuHTM methodology is easily applied to any number of parameters. In our demonstration here the report was human produced but with a little extra work could be easily generated by the computer. The graphical and tabular results produced by HubuHTM data visualization and text reports are to be understood as qualitative impressions that may lead to new knowledge, new hypotheses, or at least novel views of the data that may spark new insights. Preliminary experience using this algorithm with two-dimensional data seems to have already suggested interesting new findings.
CONCLUSIONS
We have developed and demonstrated a new artificial neural network paradigm called HubuHTM and applied it to clinical data on corticobasal syndrome patients. HubuHTM clusters patients into hubs and qualitative inferences are drawn from studying these clusters. We are enthusiastic about the initial work we have done here and believe that with a little more work HubuHTM can become an important new computational tool for gaining deeper insights into CBS and other diseases. We believe this methodology will facilitate the discovery of new knowledge - knowledge that could enhance diagnostic and treatment selection comparative effectiveness, and knowledge to more rapidly convey clinical findings to bedside practice – the ultimate objective of translational medicine.
REFERENCES
1. Litvan I, Hauw JJ, Bartko JJ, Lantos PL, Daniel SE, Horoupian DC, Et al. Validity and reliability of the preliminary NINDS neuropathologic criteria for progressive supranuclear palsy and related disorders. J Neuropathol Exp Neurol 1996; 55:97-105. 2. Chand P, Grafman J, Dickson D, Ishizawa K, Litvan I. Alzheimer disease presenting as a corticobasal syndrome. Mov Disord, 2006;21(11):2018-2022 4. 3. Boeve BF, Lang AE, Litvan I. Corticobasal degeneration and its relationship to progressive supranuclear palsy and frontotemporal dementia. Ann Neurol 2003: 54(Suppl.5)S15–S19. 4. Litvan I, DeLeo JM, Hauw JJ, Daniel SE, Jellinger K, McKee A, et al. 1996, June). What can artificial neural networks teach us about neurodegenerative disorders with extrapyramidal features? Brain,119(3), 831-839. 5. J.E. Dayhoff and J.M. DeLeo, Artificial neural networks: Opening the black box, Cancer 91 (2001) (8 Suppl.), pp. 1615-1635.
ACKNOWLEDGEMENTS
I am very thankful to the Department of Clinical Research Informatics (DCRI) in the NIH Clinical Center for providing me with an internship and the opportunity to work at the NIH this summer. It has been both incredibly enjoyable and educationally stimulating. I would like to thank Jim DeLeo and Carl Leonard for all of their help and mentoring along the way. They have provided wonderful support on this project, as well as on several other tasks. A special thanks to fellow summer students James Woo, Simone Campbell, and Rishi Gharpuray for their hard work and commitment as a team. I would also like to acknowledge Dr. Grafman and Michael Tierney in the NINDS, for their continued support, guidance and communication throughout this project. - Morgan Clinton
Corticobasal Syndrome Patient Subtyping with HubuHTM – a New Clustering Algorithm Morgan Clinton Jim DeLeo Jordan Grafman
H
UBU
HTM
Assign all entities to closest hubs.
Construct “no missing data” view for next m feature combination. no n ≥ nmin? yes Find hubs.
m: no. features nmin: min no. entities amin: min no. attractors rmax: max no. reports
Generate report.
Save report if it is meaningful.
rmax met? no
yes
amin criteria met?
no
yes
yes
More combinations?
no
End
Figure 3. HubuHTM in data mining mode.
END