P506
Poster Presentations: P2
Table Agreement between neurochemical biomarkers of Alzheimer’s disease Monocentric Dataset
N Gender (Female %) Age, years CSF ^a-amyloid 1-42 CSF ^a-amyloid 1-42 values indicative of AD (%) CSF p-tau, CSF p-tau values indicative of AD (%) CSF t-tau CSF t-tau values indicative of AD (%)
Multicentric Dataset
HC
DC
AD
HC
AD
41 29.3 67.44 (10.62) 993.46 (325.20) ng/l 14.6
21 42.9 63.33 (9.53) 730.24 (237.23) ng/l 33.3
102 53.9 63.37 (3.99) 552.73 (216.24) ng/l 76.5
153 42.5 74.73 (5.55) 224.31 (63.06) pg/ml 32
74 36.5 75.42 (3.57) 144.49 (43.06) pg/ml 90.5
49.95 (16.77) ng/l 29.3 259.20 (106.62) ng/l 43.9
44.64 (13.39) ng/l 9.5 219.90 (77.33) ng/l 19
36.71 (67.97) ng/l 72.5 643.03 (427.59) ng/l 91.2
22.26 (11.31) pg/ml 32.7 69.34 (30.25) pg/ml 21.6
40.34 (20.51) pg/ml 81.1 123.33 (61.06) pg/ml 64.9
Data presented as mean (SD) or relative frequencies; HC: Healthy controls; DC: Diseased controls; AD: Dementia due to Alzheimer’s disease; CSF b-amyloid 1-42 values indicative of AD: b-amyloid 1-42 levels in cerebrospinal fluid (CSF) 642 ng/l or 192 pg/ml for the monocentric and multicentric dataset respectively; CSF p-tau values indicative of AD: tau phosphorylated at threonine 181 levels in cerebrospinal fluid (CSF) 61 ng/l or 192 pg/ml for the monocentric and multicentric dataset respectively; CSF t-tau values indicative of AD: total tau levels in cerebrospinal fluid (CSF) 252 ng/l or 94 pg/ml for the monocentric and multicentric dataset respectively. Neuroimaging Initiative (ADNI). The examination of the patients included laboratory screening and brain imaging. Diseased controls suffered from dysaesthesia, hypaesthesia or polyneuropathy. They had no subjective memory complaints and were independent in their activities of daily living. Their diagnostic workup with lumbar puncture and brain imaging did not reveal any abnormalities. The peptide concentrations were determined with commercially available ELISA. The biomarker values were dichotomized into indicative of AD or not according to published cut-offs. The degree of agreement between them was assessed with the Fleiss’ Kappa, which assesses the reliability of agreement between more than two raters in classification over that which would be expected by chance. Results: Demographic data and neurochemical biomarker profiles of the study groups are presented in figure. The Fleiss’ Kappa value was 0.429 in the monocentric dataset and 0.484 in the multicentric dataset. Both values indicate a moderate agreement. Conclusions: Our results somehow challenge the dominating traditional school of thought with regards to the neurochemical profile of AD dementia. They possibly point to the limits of the clinical utility of the established AD neurochemical biomarkers. Moreover, they imply that the model of AD pathogenesis, which the established neurochemical biomarkers reflect, is likely to be too simplistic. P2-096
COMBINING STRUCTURAL MRI, PROTEOMIC, AND GENETIC ADNI DATA FOR EARLY DETECTION OF ALZHEIMER’S DISEASE VIA RANDOM FOREST CLASSIFIERS 1
2
2
3
Ramon Casanova , Fang-Chi Hsu , Bryan J. Neth , Kaycee M. Sink , Stephen Rapp4, Jeff Williamson2, Mark A. Espeland2, Suzanne Craft1, 1 Wake Forest University School of Medicine, Winston Salem, North Carolina, United States; 2Wake Forest School of Medicine, Winston Salem, North Carolina, United States; 3Wake Forest School of Medicine, Winston-Salem, North Carolina, United States; 4Wake Forest University School of Medicine, Winston-Salem, North Carolina, United States. Contact e-mail:
[email protected] Background: Early detection of Alzheimer’s disease (AD) based on integrating information from different sources has become an intensive area of research. Machine learning techniques are in the forefront of this research due to their capability to deal with high dimensional data. Here we evaluate the joint performance of Random Forest classifiers estimated using proteomic, structural MRI (sMRI) and apolipoprotein E (ApoE) data for AD detection. Methods: ADNI structural MRI, ApoE and plasma proteomic data sets were used in this study. A total of 488 ADNI-1 participants from three diagnostic groups were studied: 49 cognitively normal (CN) adults, 330 adults with MCI, and 109 adults with AD. Baseline sMRI-based concatenated measures of cortical thickness and regions of interest volumes generated using FreeSurfer, 146 proteomic analytes concentrations and ApoE
SNPs rs 429358/rs7412 were analyzed. We evaluated RF classification performance when discriminating adults with MCI and AD from CN adults independently for each type of data and for two multimodal strategies: 1) Input space based on concatenation of the three types of data, and 2) Vote of RF classifiers independently estimated for each type of data. Mean classification accuracy was estimated using 30 random partitions of the data into training and testing datasets with sample sizes of 78 and 20 respectively. We matched numbers of participants in each cognitive status group. Results: When discriminating CN from AD, combining multiple types of information improved classification accuracy (Table 1). When discriminating CN from MCI proteomic data performed much better than sMRI and ApoE data, and their combinations led to worse results. Conclusions: Using all three types of data improved discrimination of CN versus AD patients, but not discrimination of CN versus MCI. These results suggest that classification, based on votes of classifiers estimated with different types of data, will improve when the individual classifiers produce similar accuracies. When one type of classifier clearly outperforms the rest, voting leads to losses in performance. In our study proteomic data performed best when discriminating CN from MCI suggesting that it may have greater value for detecting AD earlier. Table 1 Classification accuracy’s mean and SD are shown for each type of data and their combinations. Case
sMRI
ApoE
Proteomic Concatenation Vote
CN vs AD 83.7 (6.8) 79.8(8.4) 78.2 (9.6) 85.3 (7.2) CN vs MCI 66.5 (11.0) 64.0 (16.0) 85.7 (7.4) 84.8 (6.5)
P2-097
88.5 (6.7) 77.8 (10.0)
CEREBROSPINAL FLUID PROTEOMIC BIOMARKERS ASSOCIATED WITH NORMAL AGING AND ALZHEIMER’S DISEASE
Steven Lynham1, Vikram Mitra2, Malcolm Andrew Ward3, Abhay Moghekar4, Richard Obrien5, Madhav Thambisetty6, 1King’s College London, London, England, United Kingdom; 2Proteome Sciences Plc, Cobham, England, United Kingdom; 3King’s College London, London, England, United Kingdom; 4Johns Hopkins University School of Medicine, Baltimore, Maryland, United States; 5Johns Hopkins, Baltimore, Maryland, United States; 6National Institute on Aging, Baltimore, Maryland, United States. Contact e-mail:
[email protected] Background: Conventional approaches to biomarker discovery in Alzheimer’s disease (AD) compare changes in protein concentrations in patients relative to age-matched controls. This strategy ignores dynamic changes in protein levels during aging that may be relevant to early pathology in pre-symptomatic AD. Our goal was to identify proteome-based