Dataset Information

Ensemble Learning of Convolutional Neural Network, Support Vector Machine, and Best Linear Unbiased Predictor for Brain Age Prediction: ARAMIS Contribution to the Predictive Analytics Competition 2019 Challenge.

ABSTRACT: We ranked third in the Predictive Analytics Competition (PAC) 2019 challenge by achieving a mean absolute error (MAE) of 3.33 years in predicting age from T1-weighted MRI brain images. Our approach combined seven algorithms that allow generating predictions when the number of features exceeds the number of observations, in particular, two versions of best linear unbiased predictor (BLUP), support vector machine (SVM), two shallow convolutional neural networks (CNNs), and the famous ResNet and Inception V1. Ensemble learning was derived from estimating weights via linear regression in a hold-out subset of the training sample. We further evaluated and identified factors that could influence prediction accuracy: choice of algorithm, ensemble learning, and features used as input/MRI image processing. Our prediction error was correlated with age, and absolute error was greater for older participants, suggesting to increase the training sample for this subgroup. Our results may be used to guide researchers to build age predictors on healthy individuals, which can be used in research and in the clinics as non-specific predictors of disease status.

SUBMITTER: Couvy-Duchesne B

PROVIDER: S-EPMC7770104 | biostudies-literature | 2020

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

Ensemble Learning of Convolutional Neural Network, Support Vector Machine, and Best Linear Unbiased Predictor for Brain Age Prediction: ARAMIS Contribution to the Predictive Analytics Competition 2019 Challenge.

Couvy-Duchesne Baptiste B Faouzi Johann J Martin Benoît B Thibeau-Sutre Elina E Wild Adam A Ansart Manon M Durrleman Stanley S Dormont Didier D Burgos Ninon N Colliot Olivier O

Frontiers in psychiatry 20201215

We ranked third in the Predictive Analytics Competition (PAC) 2019 challenge by achieving a mean absolute error (MAE) of 3.33 years in predicting age from T1-weighted MRI brain images. Our approach combined seven algorithms that allow generating predictions when the number of features exceeds the number of observations, in particular, two versions of best linear unbiased predictor (BLUP), support vector machine (SVM), two shallow convolutional neural networks (CNNs), and the famous ResNet and In ...[more]

PMID: 33384629

Similar Datasets

Project description:Despite important advances from Genome Wide Association Studies (GWAS), for most complex human traits and diseases, a sizable proportion of genetic variance remains unexplained and prediction accuracy (PA) is usually low. Evidence suggests that PA can be improved using Whole-Genome Regression (WGR) models where phenotypes are regressed on hundreds of thousands of variants simultaneously. The Genomic Best Linear Unbiased Prediction (G-BLUP, a ridge-regression type method) is a commonly used WGR method and has shown good predictive performance when applied to plant and animal breeding populations. However, breeding and human populations differ greatly in a number of factors that can affect the predictive performance of G-BLUP. Using theory, simulations, and real data analysis, we study the performance of G-BLUP when applied to data from related and unrelated human subjects. Under perfect linkage disequilibrium (LD) between markers and QTL, the prediction R-squared (R(2)) of G-BLUP reaches trait-heritability, asymptotically. However, under imperfect LD between markers and QTL, prediction R(2) based on G-BLUP has a much lower upper bound. We show that the minimum decrease in prediction accuracy caused by imperfect LD between markers and QTL is given by (1-b)(2), where b is the regression of marker-derived genomic relationships on those realized at causal loci. For pairs of related individuals, due to within-family disequilibrium, the patterns of realized genomic similarity are similar across the genome; therefore b is close to one inducing small decrease in R(2). However, with distantly related individuals b reaches very low values imposing a very low upper bound on prediction R(2). Our simulations suggest that for the analysis of data from unrelated individuals, the asymptotic upper bound on R(2) may be of the order of 20% of the trait heritability. We show how PA can be enhanced with use of variable selection or differential shrinkage of estimates of marker effects.

Project description:The objectives of this study were to develop an efficient algorithm for calculating prediction error variances (PEVs) for genomic best linear unbiased prediction (GBLUP) models using the Algorithm for Proven and Young (APY), extend it to single-step GBLUP (ssGBLUP), and apply this algorithm for approximating the theoretical reliabilities for single- and multiple-trait models in ssGBLUP. The PEV with APY was calculated by block sparse inversion, efficiently exploiting the sparse structure of the inverse of the genomic relationship matrix with APY. Single-step GBLUP reliabilities were approximated by combining reliabilities with and without genomic information in terms of effective record contributions. Multi-trait reliabilities relied on single-trait results adjusted using the genetic and residual covariance matrices among traits. Tests involved two datasets provided by the American Angus Association. A small dataset (Data1) was used for comparing the approximated reliabilities with the reliabilities obtained by the inversion of the left-hand side of the mixed model equations. A large dataset (Data2) was used for evaluating the computational performance of the algorithm. Analyses with both datasets used single-trait and three-trait models. The number of animals in the pedigree ranged from 167,951 in Data1 to 10,213,401 in Data2, with 50,000 and 20,000 genotyped animals for single-trait and multiple-trait analysis, respectively, in Data1 and 335,325 in Data2. Correlations between estimated and exact reliabilities obtained by inversion ranged from 0.97 to 0.99, whereas the intercept and slope of the regression of the exact on the approximated reliabilities ranged from 0.00 to 0.04 and from 0.93 to 1.05, respectively. For the three-trait model with the largest dataset (Data2), the elapsed time for the reliability estimation was 11 min. The computational complexity of the proposed algorithm increased linearly with the number of genotyped animals and with the number of traits in the model. This algorithm can efficiently approximate the theoretical reliability of genomic estimated breeding values in ssGBLUP with APY for large numbers of genotyped animals at a low cost.

Project description:BackgroundRecently, there has been a growing interest in the genetic improvement of body measurement traits in farm animals. They are widely used as predictors of performance, longevity, and production traits, and it is worthwhile to investigate the prediction accuracies of genomic selection for these traits. In genomic prediction, the single-step genomic best linear unbiased prediction (ssGBLUP) method allows the inclusion of information from genotyped and non-genotyped relatives in the analysis. Hence, we aimed to compare the prediction accuracy obtained from a pedigree-based BLUP only on genotyped animals (PBLUP-G), a traditional pedigree-based BLUP (PBLUP), a genomic BLUP (GBLUP), and a single-step genomic BLUP (ssGBLUP) method for the following 10 body measurement traits at yearling age of Hanwoo cattle: body height (BH), body length (BL), chest depth (CD), chest girth (CG), chest width (CW), hip height (HH), hip width (HW), rump length (RL), rump width (RW), and thurl width (TW). The data set comprised 13,067 phenotypic records for body measurement traits and 1523 genotyped animals with 34,460 single-nucleotide polymorphisms. The accuracy for each trait and model was estimated only for genotyped animals using five-fold cross-validations.ResultsThe accuracies ranged from 0.02 to 0.19, 0.22 to 0.42, 0.21 to 0.44, and from 0.36 to 0.55 as assessed using the PBLUP-G, PBLUP, GBLUP, and ssGBLUP methods, respectively. The average predictive accuracies across traits were 0.13 for PBLUP-G, 0.34 for PBLUP, 0.33 for GBLUP, and 0.45 for ssGBLUP methods. Our results demonstrated that averaged across all traits, ssGBLUP outperformed PBLUP and GBLUP by 33 and 43%, respectively, in terms of prediction accuracy. Moreover, the least root of mean square error was obtained by ssGBLUP method.ConclusionsOur findings suggest that considering the ssGBLUP model may be a promising way to ensure acceptable accuracy of predictions for body measurement traits, especially for improving the prediction accuracy of selection candidates in ongoing Hanwoo breeding programs.

Project description:Across the majority livestock species, routinely collected genomic and pedigree information has been incorporated into evaluations using single-step methods. As a result, strategies that reduce genotyping costs without reducing the response to selection are important as they could have substantial economic impacts on breeding programs. Therefore, the objective of the current study was to investigate the impact of selectively genotyping selection candidates on the selection response using simulation. Populations were simulated to mimic the genome and population structure of a swine and cattle population undergoing selection on an index comprised of the estimated breeding values (EBV) for 2 genetically correlated quantitative traits. Ten generations were generated and genotyping began generation 7. Two phenotyping scenarios were simulated that assumed the first trait was recorded early in life on all individuals and the second trait was recorded on all versus a random subset of the individuals. The EBV were generated from a bivariate animal model. Multiple genotyping scenarios were generated that ranged from not genotyping any selection candidates, a proportion of the selection candidates based on either their index value or chosen at random, and genotyping all selection candidates. An interim index value was utilized to decide who to genotype for the selective genotype strategy. The interim value assumed only the first trait was observed and the only genotypic information available was on animals in previous generations. Within each genotyping scenario 25 replicates were generated. Within each genotyping scenario the mean response per generation and the degree to which EBV were inflated/deflated was calculated. Across both species and phenotyping strategies, the plateau of diminishing returns was observed when 60% of the selection candidates with the largest index values were genotyped. When randomly genotyping selection candidates, either 80 or 100% of the selection candidates needed to be genotyped for there not to be a reduction in the index response. Across both populations, no differences in the degree that EBV were inflated/deflated for either trait 1 or 2 were observed between nongenotyped and genotyped animals. The current study has shown that animals can be selectively genotyped in order to optimize the response to selection as a function of the cost to conduct a breeding program using single-step genomic best linear unbiased prediction.

Dataset Information

Ensemble Learning of Convolutional Neural Network, Support Vector Machine, and Best Linear Unbiased Predictor for Brain Age Prediction: ARAMIS Contribution to the Predictive Analytics Competition 2019 Challenge.

Publications

Ensemble Learning of Convolutional Neural Network, Support Vector Machine, and Best Linear Unbiased Predictor for Brain Age Prediction: ARAMIS Contribution to the Predictive Analytics Competition 2019 Challenge.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets