Age prediction on the basis of brain anatomical measures

INTRODUCTION

It is a well-known fact that the brain changes during aging. An often-reported finding is that particular brain areas change more than others [Fjell and Walhovd, 2010; Fjell et al., 2009b]. However, this issue is far from clear since reported age-related brain changes and differences differ across studies and methods used. Combining anatomical data from six different samples, Fjell and colleagues [Fjell et al., 2009b] revealed that the frontal cortex (especially the superior, middle, and inferior frontal gyri), and some parts of the temporal (e.g., the middle temporal gyri) and parietal cortex (e.g., the precuneus, the inferior and superior parietal cortices, and the temporo-parietal junction) are subject to age-related cortical thinning. The strongest effects were seen in the superior and inferior frontal gyri, as well as in the superior parts of the temporal lobe. In contrast, the inferior temporal lobe and anterior cingulate cortices were relatively less affected by age [Fjell et al., 2009b].

Most studies examining age-effects on brain tissue, however, have focused on cortical measures, with only a few studies focusing on possible age-related differences with respect to subcortical volumes (e.g., thalamus, caudate, amygdala, hippocampus, putamen, pallidum, and accumbens) [Fjell and Walhovd, 2010; Jäncke et al., 2015; Li et al., 2014; Tang et al., 2013]. Similar to studies investigating cortical regions, these studies have demonstrated substantial subcortical volume losses during the course of healthy aging.

Although studies using between-group statistics may indicate the typical age-related anatomical differences, it has been argued that between-group analyses might be less useful for automatic diagnosis of group differences and age-related diseases [Seidman et al., 2004]. The wide utilization of modern machine learning techniques in the neuroimaging community has made it possible for researchers to discover biomarkers of aging and to develop automatic classification systems. In addition, it may prove helpful to delineate a typical pattern of anatomical measures for a particular age in order to more precisely identify deviations from normal aging.

In this context, many efforts have been made to use structural and functional magnetic resonance imaging (sMRI and fMRI, respectively) to classify different patient groups and to statistically separate them from healthy control subjects. For example, several studies have tried to classify Alzheimer's and schizophrenia patients using anatomical measures [Fan et al., 2008; Hinrichs et al., 2009; Hinrichs et al., 2011]. In addition, several attempts have been made to predict age on the basis of anatomical data. The underlying idea of these studies is to detect possible discrepancies between predicted and chronological age, which might help identifying pathological structural changes. Only a handful of studies have been published so far conducting age estimation based on MRI scans. Lao et al. [2004] used support vector machine (SVM) techniques to classify elderly subjects into one of four age groups on the basis of anatomical measures and reached an accuracy rate of 90%. Ashburner [2007] applied an algorithm for “diffeomorphic” image registration to estimate the age of healthy subjects and reported good age prediction accuracy. Neeb and colleagues [Neeb et al., 2006] used brain water maps (which is a new method for the absolute and quantitative mapping of water content in vivo) to predict age and gender in 44 healthy volunteers aged 23–74 years. They applied a linear discriminant analysis (LDA) with jackknife cross-validation for age prediction and obtained good classification performance. Age predictions have also been conducted in the so-called BrainAge project from the Jena group [Franke et al., 2015; Franke et al., 2012; Franke et al., 2014; Franke et al., 2010; Gaser et al., 2013]. These studies predicted age in the context of voxel-based morphometry (VBM) methods using a relevance vector machine (RVM) technique. Using this technique, they obtained a correlation between predicted and true age of r = 0.92.

Several classification techniques have been used so far, ranging from “classical” LDA [Kasparek et al., 2011; Leonard et al., 1999; Nakamura et al., 2004; Takayanagi et al., 2011], and linear and non-linear regression analysis to various variants of SVM approaches [Ashburner, 2007; Castro et al., 2014; Dai et al., 2012; Fan et al., 2008; Hinrichs et al., 2011; Klöppel et al., 2012; Sato et al., 2012; Schnack et al., 2014; Wachinger et al., 2015; Wachinger et al., 2014]. Which of these techniques are best suited for age prediction is an open question that needs further investigation.

All studies that have predicted age on the basis of anatomical data have worked with cross-sectional datasets. Thus, cohort effects could possibly have influenced the morphological features. For example, nutritional status, education, health, and social interactions have substantially changed within the last 40–60 years. Since there is ample evidence available demonstrating that these factors influence brain anatomy and body size [de Bruin et al., 2005; Pannacciulli et al., 2006; Taki et al., 2004; Taki et al., 2006], it is necessary to control for brain/head size. Here we used intracranial volume (ICV). ICV more closely indicates head size rather than brain size, since it comprises brain tissue, meninges, cerebrospinal fluid, and interstitial volumes. ICV does not correlate perfectly with forebrain volume (R² = 0.58) [Jäncke et al., 2015], demonstrating that more than 40 percent of ICV variance is unrelated to brain tissue. In addition, it has been shown that small brains in terms of ICV exploit the space within the skull more efficiently than larger brains [Jäncke et al., 2015]. Conversely, there is disproportionately more space available for brain tissue in larger brains. To account for these non-brain-tissue influences, it would be necessary to control for ICV influences when predicting age on the basis of anatomical measures.

A major problem with attempts to predict age from anatomical measures is the problem of over-fitting, which is present even when using relatively small sample sizes. In this study, we will use brain anatomical measures from a large sample in freely available databases, in addition to samples from our own laboratory. This amounts to a sample of 3,144 subjects spanning an age range from 8 to 96 years. Secondly, in contrast with the aforementioned studies, we will use cortical thickness, volume and area measures from 148 regions of interest obtained using the FreeSurfer analysis tool. In addition, we will use subcortical volume measures and anatomical measures of larger brain regions (mean cortical thickness, volume, and area, as well as mean subcortical volume) in order to examine whether these larger ROIs will provide similar or even better classification results than the collection of small ROIs. Using this collection of anatomical measures, we will explicitly answer the following questions: (1) Is it possible to predict age on the basis of a combination of anatomical measures? (2) Which combination of anatomical measures is most important in predicting age? (3) Do different classification techniques (linear discriminant analysis, SVMs, ridge regression, random forest, or neural network classification) substantially differ in terms of their classification accuracy?

METHODS

RESULTS

Table 2 presents the results of the prediction analyses for the nine datasets and the six prediction techniques (MLR, RR, NN, KNN, SVM, and RF). The table shows the R² values computed for the entire sample (R² total), the training sample (R² training), and the test sample (R² test). In addition, the shrinking factor (R² shrinkage: difference between R²'s obtained for the training and the test sample) is shown. The R² values are rank ordered according to R² test.

Focusing on the R² values obtained for the test sample (R² test in Table 2), one can see that the R² values range between R² = 0.40 (for the AREA dataset, using KNN) and R² = 0.84 (for NN and SVM applied to the ALL dataset). The R² values for R² test being larger than 0.8 are grey shaded (in dark grey) and printed in bold and italic letters. The R² values for R² test larger than 0.7 are light grey shaded. The best predictions were obtained for NN and SVM for the ALL dataset. Interestingly, the NN method applied to the small 11 LBR dataset revealed similar good results with a R² = 0.83. In addition, the frequently and easy to apply MLR method revealed good results for the ALL and ALL+ datasets with R² values of 0.82 and 0.81, respectively.

In Figure 1 all R² values are shown as a radar plot in order to demonstrate the relationship between the four R² values in more detail and much clearer. In Table 3 the mean, minimum, and maximum R² values obtained for the six methods are listed in rank order. As one can see from this Table, NN reveals the largest R² values while KNN turns out to be the least precise method. MLR is the second precise method with a mean R² value of 0.77.

Figure 2 shows the absolute prediction errors (absolute difference between predicted and chronological age) broken down for age, the different methods, and datasets. As one can see from these figures the absolute prediction error increases with increasing age.

Table 4 shows mean prediction errors broken down for three age groups (8–18, 18–65, and 65–96 years). This Table reveals that depending on the used prediction method the absolute prediction error varies considerably. There are relatively small prediction errors for ALL and ALL+ feature set with an error of 1.23 years for the young subjects (8–18 years) obtained for the NN technique. This error increases to 4.23 and 5.23 years for the middle-aged (18–65 years) and older subjects (65–96 years), respectively. The prediction errors obtained with the other techniques are a bit higher.

DISCUSSION

With the current study we aimed at predicting chronological age on the basis of a combination of brain anatomical measures. Linked to this overall rationale, we were interested to answer different sub-questions. First, we wanted to clarify which combination of anatomical measures is most suitable to predict age. Secondly, we explored whether a successful age prediction relies on a fine-grained extraction of anatomical measures from many small brain regions, or whether it is sufficient to use brain measures from larger brain regions. Thirdly, we compared different statistical techniques with respect to their prediction accuracy. In the following paragraphs, we will discuss our findings and relate them to the current literature.

The main finding of our study is that age can be predicted with a high accuracy on the basis of anatomical measures. We identified 33 predictions with accuracies being larger than R² = 0.70 with the highest accuracy of R² =0.84. Thus, the prediction accuracy is in the same range (or even better) as has been demonstrated by previous studies using different anatomical measures and statistical prediction methods (R²s ranging between 0.49 and 0.91) [Ashburner, 2007; Franke et al., 2013; Franke et al., 2012; Franke et al., 2014; Franke et al., 2010; Lao et al., 2004; Neeb et al., 2006]. However, different from the afore-mentioned studies, we used a much larger sample (n > 3,400) spanning a large age range (7–96 years), and we used different sets of anatomical measures (thickness, volume, surface area).

Most astonishing was the finding that very good prediction accuracy was achieved even when using a small set of anatomical measures comprising only 11 larger brain regions with averaged anatomical characteristics across larger brain regions (e.g., grey and white matter volume, and cortical thickness, volume and area: 11 LBR). With this set of anatomical measures, we achieved a good prediction accuracy of R² = 0.83 using the NN technique. The lowest prediction accuracy for this set of anatomical measures was obtained using the SVM technique (R² = 0.61). Linear regression and RR revealed also reasonable results (R² = 0.73 and R² = 0.70, respectively).

The best prediction accuracy obtained with this reduced feature set only marginally differed from the best prediction accuracy achieved with a more fine-grained feature set (R² = 0.84 for the ALL dataset using NN or SVM, and R² = 0.83 for the 11 LBR dataset using NN). Thus, it is clearly not necessary to use a large and sophisticated set of anatomical measures comprising very small brain areas to achieve accurate age prediction. The brain regions used for the 11 LBR dataset comprised relatively large brain compartments taken from the cortex, brainstem, and the cerebellum (cortical volume, cortical thickness, cortical area, cortical gray matter volume, cortical white matter volume, cerebellar gray matter volume, cerebellar white matter volume, subcortical volume, brainstem volume, corpus callosum, white matter hypo-intensities). Thus, a combination of these different cortical, sub-cortical and cerebellar brain measures are at least as effective for predicting age than larger datasets comprising relatively small cortical measures.

That the anatomical dataset with the large cortical, subcortical and cerebellar volume LBRs are so useful for predicting age might result from the fact that age effects on brain anatomy are mostly distributed and not focal. Several recent studies have shown that older subjects demonstrate smaller volumes in many distributed cortical, subcortical and cerebellar regions [Fjell et al., 2009a; Fjell et al., 2009b; Jäncke et al., 2015; Tang et al., 2013; Ziegler et al., 2012]. As cortical volumes become smaller with increasing age, cerebrospinal fluid and white matter T1-weighted hypo-intensity (T2-weighted hyper-intensity) volumes increase. Thus, combining volume measures from these different brain regions seems to be a good and valid strategy for capturing age-related anatomical differences more accurately and efficiently.

In this study, we used six different statistical techniques for age prediction. It turned out that the accuracy of these methods is mostly depended on the dataset. However, the KNN method had the weakest results overall, with an average accuracy of R² = 0.62. Each of the techniques used provide different advantages and disadvantages, which should be carefully taken into consideration when planning to predict age on the basis of anatomical measures.

The MLR approach is a standard statistical technique frequently used in psychology and neuroscience. It is relatively easy to program and can be run quickly even with a large number of predictor variables. However, a problematic aspect is the possibility of correlated predictors, which can distort the prediction accuracy substantially (e.g., suppression effects) [Pedhazur, 1997]. In addition, in order to achieve reasonably good prediction accuracies, it is necessary to use large samples for the training, a prerequisite that was easily met by our sample with 3,144 participants. In our study, the MLR technique revealed very good results for nearly all datasets, with the exception of AREA and THICKNESS.

Similar to the MLR technique, RR is simple to program, runs quickly (even a bit faster than MLR) and can be used in so-called ill-defined mathematical conditions and when too many predictors are used. For example, when it is impossible to calculate an inverse matrix (a mathematical operation, which is necessary to calculate linear regressions), RR is a suitable solution [Hoerl & Kennard, 1970]. Nevertheless, possible statistical dependencies between the predictors can distort the prediction results, similar to the MLR technique. For our dataset, RR revealed similar results to the MLR technique.

The NN approach is a bit more complicated to implement in terms of programming effort. It is a time- and resource-consuming technique, and requires large sample sizes. It can, however, handle many correlated predictors as well as linear and non-linear relationships without any a priori knowledge about the relationships. In our study, the NN approach provided the best results with small R² shrinkages, although it must be kept in mind that the differences to the other techniques are quite small. Nevertheless, out of our nine datasets, NN provided the best results for eight datasets.

The KNN approach is a time-consuming technique with some similarities to the NN technique. The precision of this technique is highly dependent on the predictors used, as has been shown by our study. KNN revealed mixed results in our study, with only one good result for the 11 LBR dataset that was in the same range as the NN results. For all other datasets, KNN had moderate or weak predictive power.

The SVM technique is a bit more demanding to implement, but is relatively fast and provides stable results. Similar to the KNN, the results strongly depend on the selection of predictors. The results in our study were roughly similar to those from the NN technique. SVM provided good results only when large datasets were used (ALL and ALL+).

The RF technique is a very time-consuming technique that can handle correlated predictors, requires only small learning sets and is less affected by correlated predictors. Similar to the KNN and NN approaches, the results strongly depend on the selection of predictors used. For our data, RF revealed similar results to the MLR technique.

To summarize our findings with respect to the precision of the techniques used, NN provided the best results across all datasets, irrespective of whether they comprised only a few or many anatomical measures. However, the relatively simple and widely used MLR technique also provided good results. Since RR is nearly as good as MLR, it can be used in mathematically problematic situations. RF, although providing reasonable results, is too time consuming and it offers no added gains.

The absolute prediction errors (difference between predicted and chronological age) we obtained in our analyses are quite similar to those reported in previous studies [Brown et al., 2012; Cherubini et al., 2016; Franke et al., 2010]. However, it is worth mentioning that our study cannot be compared entirely with these studies, either because they only studied relatively young subjects [Brown et al., 2012] or they employed a much smaller sample than we used in our study [Cherubini et al., 2016]. However, the results are astonishingly similar in showing that the prediction error increases with increasing age [Brown et al., 2012; Franke et al., 2010]. In addition, the prediction errors in our study are about in same range as it was reported in these studies, at least for the methods and datasets for which we achieved good and very good prediction results.

CONCLUSION

Taken together, the results of the present study demonstrate that chronological age can be predicted quite well on the basis of anatomical measures. The NN approach revealed to be the approach with the best prediction accuracy. In addition, it was evident that good prediction accuracies can be achieved using a reduced, but nevertheless age-representative dataset (e.g., 11 LBR).