Thus, predicted age tended to underestimate the age of older participants and overestimate age of younger individuals. Such results align with the large rank bias reported in Tables 1 , 2. We did not observe a significant association of prediction error with sex or site Table 4. Table 4. From BLUP models, we estimated the total association between age and the brain features.
In absence of reported SE in the PAC results, we cannot conclude whether the different prediction accuracies are statistically different from each other.
It is important to keep in mind that ranking of prediction accuracy may be highly dependent on the metric chosen as well as on the test data Statistical testing can provide a confident ranking of algorithms, and inclusion of other datasets is needed to conclude about the generalizability and performance of the prediction scores on samples with other demographics, MRI machines, or patient groups for instance. We present analyses that we performed prior to the challenge closing that informed our method , as well as post-hoc analyses in which we explored new avenues.
Lastly, we detail our approach for the second PAC challenge minimize MAE, while controlling bias though in much less detail as we came sixth out of six entries with a MAE almost 2 years greater than the winner. Ensemble learning with weights estimated via linear regression led to a significant reduction of MAE of about 0. Score combination using random forest also outperformed the algorithm with minimal MAE Inception V1 , but the result was somewhat dependent on the folds considered.
The difference between linear model and random forest was too small to conclude about a significant difference Table 2. The weights given to each algorithm via linear regression were highly dependent on the folds and iterations, which might be an artifact of the large correlations between the scores. Nevertheless, few weights were consistently set to 0 across all folds and iterations , suggesting that all seven algorithms contributed to the ensemble learning Supplementary Figures 1, 2.
Our results align with previous publications that highlighted the benefits of ensemble learning, which combines different models 56 or different data Each processing stream allows multiple user-defined options e. Importantly, the image processing maximizing age prediction may not be the best suited to predict another phenotype e.
Lastly, the good performance of BLUP-median on gray matter maps raises the question of cost-efficiency and updatability of prediction, considering that deep learning models require about 24 h of computing on a GPU, while BLUP only takes a few minutes on a single CPU. In addition, we found very similar performance of ensemble prediction from our seven different algorithms compared with that of seven independently trained Inception V1 scores.
We conclude that using a variety of algorithms may not offer an advantage over using several well-performing ones. Due to limited computing resources, we did not investigate whether increasing the number of Inception V1 algorithms further reduced the MAE, though our age prediction did not progress when combining the 21 models estimated throughout the analysis.
Finally, our predictions showed a large age bias: overestimating age on younger participants and underestimating it on older participants. We also identified older individuals as main contributors of the MAE, suggesting much is to be gained by improving the performance on this sub-population. Our attempt to re-train part of the network on adults above 40 years of age age specialized six-layer CNN was not conclusive in improving the age prediction accuracy.
Other avenues for research include enriching the training sample in specific age groups or demographics that show a lower performance. We did not find error or absolute error to be associated with sex or site, despite differences in global head size, or site differences in term of scanners, demographics, and image qualities.
An investigation on a larger dataset may be more powered in detecting subgroups with larger MAE. To finish on bias, we found that rescaling the scores using the median and median absolute deviation per site could reduce drastically the bias but resulted in an increase in MAE Table 1 and Supplementary Tables 10, Low bias age predictors avoid subsequent association analyses e.
We did not systematically investigate the use of white matter maps to improve prediction accuracy. Only the six-layer CNN was trained on both gray matter and white matter maps, and it did not outperform the other algorithms.
In addition, our split design allowed for well-powered statistical testing and weighted estimation for ensemble learning; however, it may not be the optimal split to minimize the MAE. This suggests that the prediction accuracy we report here might be close to the theoretical maximum achievable from linear predictors, even though this claim is weakened by the fact that prediction R 2 is not a sufficient statistic here as age was not normally distributed thus, it might be inflated.
Importantly, the high prediction accuracy we report does not ensure that PAD best discriminates cases from controls in a clinical sample More generally, prediction accuracy is not a linear function of training sample size [see 58 ], and we can expect further significant improvement in age prediction to require much larger sample sizes.
We would also like to point out that reducing the MAE below 1 year is unlikely, when training algorithms on rounded age, which was the case here. Finally, PAC participants were described as healthy individuals, though screening of all brain related disorders is impossible, which raises the question of unknown diagnosis for participants with large prediction error. In conclusion, we achieved a MAE of 3. We identified several contributors to prediction accuracy: algorithm choices, image processing options, and ensemble learning.
Publicly available datasets were analyzed in this study. The ensemble of datasets presented in this article are not readily available because they are held and distributed by the PAC team. Requests to access the datasets should be directed to Pr. Hahn hahnt wwu. Cole james. Ethical review and approval was not required for the study on human participants in accordance with the local legislation and institutional requirements. Acquisition, analysis, or interpretation of data was performed by all authors.
Manuscript drafting or manuscript revision for important intellectual content was performed by all authors. Approval of final version of submitted manuscript was done by all authors.
Study supervision was carried out by NB and OC. The sponsors had no role in study design, data analysis or interpretation, writing, or decision to submit the report for publication. The members from his laboratory have co-supervised a Ph. OC's spouse is an employee of myBrainTechnologies —present.
The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. Heterogeneity in healthy aging. J Gerontol Seri. Mol Psychiatr. Horvath S, Raj K. DNA methylation-based biomarkers and the epigenetic clock theory of ageing.
Nat Rev Genet. Sajedi H, Pardakhti N. Age prediction based on brain MRI image: a survey. J Med Syst. Biomarkers of aging. Exp Gerontol. Brain age predicts mortality. Biological age predictors. Franke K, Gaser C. Longitudinal changes in individual brainAGE in healthy aging, mild cognitive impairment, alzheimer's disease. BrainAGE in mild cognitive impaired patients: predicting the conversion to alzheimer's disease. Gray matter age prediction as a biomarker for risk of dementia.
Accelerated brain aging in schizophrenia and beyond: a neuroanatomical marker of psychiatric disorders. Schizophrenia Bull. Quantitative neurobiological evidence for accelerated brain aging in alcohol dependence. Transl Psychiatr. Predicting brain-age from multimodal imaging data captures cognitive impairment. Brain Behav. A Nonlinear simulation framework supports adjusting for age when analyzing brainAGE.
Front Aging Neurosci. Estimation of brain age delta from brain imaging. Predicting brain age with deep learning from raw imaging data results in a reliable and heritable biomarker.
Genetics of brain age suggest an overlap with common brain disorders. Brain age prediction using deep learning uncovers associated sequence variants. Nat Commun. Brain aging comprises many modes of structural and functional change with distinct genetic and biophysical associations. Horvath S. DNA methylation age of human tissues and cell types. Genome Biol. Cole JH, Franke K. Predicting age using neuroimaging: innovative brain ageing biomarkers.
Trends Neurosci. MRI signatures of brain age and disease over the lifespan based on a deep brain network and 14 individuals worldwide. Ashburner J. A fast diffeomorphic image registration algorithm.
Fischl B. A family of fast spherical registration algorithms for cortical shapes. Multimodal Brain Image Analysis. Springer International Publishing. Release Year. E-Everyone Items PEGI 3 9 Items 9. G 1 Items 1. Manual Included 90 Items Online Playability 13 Items Multiplayer 12 Items Complete Edition 11 Items Demo 1 Items 1. Special Edition 1 Items 1. Wii Motion Plus Compatible 1 Items 1.
Brand New 27 Items Like New 53 Items Very Good Items Good Items Acceptable 14 Items Please provide a valid price range. Buying Format. All Listings. Accepts Offers. Buy It Now. Item Location. Canada Only. Some are hard but others are easy depending on your strong points.
Over all it is a pretty interesting game that you can play for a few minutes or for as long as you want. I bought it to sharpen my mind and I am happy to say that it has done just tha t. It is a good game so if you want it and have the money you may very well find yourself pleased with it as well.
Read full review. Brain Age 2's subtitle tells you almost everything you need to know about Nintendo's new mind-flexing sequel. It's "more training in minutes a day. This means it's perfect for people who played the first game and are looking for more. The games in Brain Age 2 are more advanced than the previous game, which had you doing basic math problems half the time. Now, those math problems have twists, such as one game where you continually add numbers, but one number is removed from the screen after a second or so, which forces you to use your memory to remember numbers and solve the problem.
Another puts the keys of a piano on the touch scree n and has you follow along with some sheet music to play a tune. Another has voices say two or three words at the same time, and you need to pick apart the garbled speech then write down the words you hear. All in all, the different games are satisfyingly different from what was in the first game and make for a solid companion piece.
I love the different games it gives you in a scattered way, so you can't just play all of them the first day you get it and then be done with it. I love that Sudoku is always available, and that I can train with the games at any time. Your scores are recorded once a day, however, so when you do them for the training part, make sure you've practiced! I did and my scores leapt and I've only had the game for about a week!
It's great for skills that you don't think about using, but you DO use. Believe it or not, this game would be great for those summer months when kids are out of school - it'll help them keep at least the bottom line basics so they aren't completely at a loss when school starts again! Teens have a tendency to divide and conquer.. If you never played the first one, this one has a tutorial that helps you get started.
What's really cool is it tells you what part of your brain is being sharpened, and explains what that part of the brain does. It also gives you tips on how to get better scores.
As you do more activities on more than one day you unlock other areas of the game like a relaxation and unwinding game. I do not like holding the game console on it's side to play the game. And sometimes it misunderstands what you write for an answer so you get it wrong and get a bad score if you don't write exactly how the game wants you to. But that just made me work harder on having better writing skills.
0コメント