Predicting Short-term MCI-to-AD Progression Using Imaging, CSF, Genetic Factors, Cognitive Resilience, and Demographics

In the Alzheimer’s disease (AD) continuum, the prodromal state of mild cognitive impairment (MCI) precedes AD dementia and identifying MCI individuals at risk of progression is important for clinical management. Our goal was to develop generalizable multivariate models that integrate high-dimensional data (multimodal neuroimaging and cerebrospinal fluid biomarkers, genetic factors, and measures of cognitive resilience) for identification of MCI individuals who progress to AD within 3 years. Our main findings were i) we were able to build generalizable models with clinically relevant accuracy (~93%) for identifying MCI individuals who progress to AD within 3 years; ii) markers of AD pathophysiology (amyloid, tau, neuronal injury) accounted for large shares of the variance in predicting progression; iii) our methodology allowed us to discover that expression of CR1 (complement receptor 1), an AD susceptibility gene involved in immune pathways, uniquely added independent predictive value. This work highlights the value of optimized machine learning approaches for analyzing multimodal patient information for making predictive assessments.

Cognitive resilience represents the ability of an individual to delay the deleterious effects of neurodegenerative pathologies on onset of cognitive symptoms 11 . While cognitive resilience has been widely used to explain the pathology-cognition disconnect in cognitively unimpaired individuals with AD pathology (and normal cognition), its relative influence specifically on MCI-to-AD progression has not been fully evaluated 12 . Furthermore, the complex relationships among the well-known AD biomarkers, cognitive resilience, and genetic factors remain largely unknown and estimating the independent predictive values provided by each of these factors may open the door to alternative strategies to delay or prevent the onset of dementia.
Several machine learning (ML) based approaches have been proposed for predicting MCI-to-AD progression 4,5,13,14 and clinical stage classification 15 utilizing high-dimensional clinical and biomarker data among the potential predictors. A limitation of the ML-based approaches that utilize high-dimensional data is the potential for overfitting, such that the classifiers are so optimally trained to fit the primary dataset that they perform poorly on previously unseen test data, ultimately limiting the generalizability and wider interpretability 15 of the predictive model. This is particularly important in this study because both a) an unbiased estimation of the contributions of cognitive resilience and genetic factors in predicting MCI-to-AD progression and b) the successful clinical translation of this technique, would require a generalizable model. However, prior studies on the predictability of MCI-to-AD progression have not performed satisfactory assessments of generalizability 4,5,13,14 . These approaches utilize non-linear classifications methods such as multi-kernel learning and artificial neural networks without establishing the need for non-linear models. Non-linear classification models are more susceptible to overfitting compared to linear classification models in problems involving high-dimensional data since they optimize relatively higher number of model parameters 16 . Whereas, a pitfall associated with linear models is that they provide suboptimal prediction performance when the data is not linearly separable. Therefore, our goals in this study with respect to the development of a predictive model are a) to understand whether linear classifiers are sufficient to provide clinically relevant accuracies in predicting MCI-to-AD progression based on a set of features derived from multiple modalities and b) to develop an analytical method that evaluates the generalizability of different classifiers and to aid model selection in clinical settings.
In this study, we used a machine learning-based approach and data from a well-characterized clinical cohort to identify individuals with MCI who rapidly progress to AD versus those with a more protracted course. Specifically, using a set of features derived from the ADNI multimodal biomarker and clinical dataset along with genetic factors and cognitive resilience measures, a linear model-based ML framework was developed for predicting MCI-to-AD progression. Based on a test of linear-separability 17 , we show that linear classification models perform just as good as non-linear models in this setting and a support vector machine classifier with a linear kernel provides the most generalizable performance with a test set area under ROC curve (AUC) value of 0.93. Using this framework, we evaluated the predictive values of previously underexplored factors of cognitive resilience and genetic factors and gauged their relative importance compared to commonly used CSF and imaging predictors. We found that, while the markers of AD pathophysiology (amyloid, tau, and neuronal injury) provided very high predictive values, the genetic factors and brain regions associated with cognitive resilience also displayed independent predictive values.

Materials and Methods
study participants. This report utilized data from the Alzheimer's Disease Neuroimaging Initiative (ADNI), a multisite longitudinal study of older adults representing clinical stages along the continuum from normal aging to AD 18 . All study participants provided written informed consent, and study protocols were approved by each local site's institutional review board. Further information about ADNI, including full study protocols, complete inclusion and exclusion criteria, and data collection and availability can be found at http://www.adni-info.org/. All methods as stated on the website were performed with the relevant guidelines and regulations. Since all the analyses were performed on de-identified ADNI data which is publically available for download, IRB Review was not required. In addition, all methods were carried out in accordance with the approved guidelines.
In choosing ADNI participants to study, we used these inclusion criteria: i) the person had at least three years of follow-up; ii) the person had all the data modalities of interest (as specified below); and iii) the person was diagnosed with MCI at the baseline evaluation. We identified 135 participants who met those criteria. A total of 39 of the 135 progressed to AD within three years of the MCI diagnosis (and are referred to as the MCI-P group); the remaining 96 progressed to AD after three years or remained in MCI until their last follow-up after three years (and are referred to as the MCI-NP group). predictive factors. Figure 1 illustrates potential factors that may impact progression from MCI to AD.
Below we describe the specifics of the individual biomarkers utilized in this work.
All the biomarkers utilized in this work were downloaded from the LONI data archive (https://ida.loni.usc. edu/) where the pre-processed ADNI data are hosted.
Cerebrospinal fluid (CSF) biomarkers. CSF levels of amyloid beta (Aβ) and total (T-tau) and phosphorylated (P-tau) tau proteins were assayed by the ADNI Biomarker Core as previously described 19 .
Magnetic resonance imaging biomarkers. Structural MRI (SMRI) scans at baseline were downloaded and processed as described previously 20 . FreeSurfer v5.1 was used to obtain volume and thickness measures for standard regions of interest (ROIs) as surrogates for cerebral atrophy. We scaled the volumes by total intracranial volume. In addition, we also included volumetric measurements of hippocampal subfields.
Positron emission tomography (PET) biomarkers. Fluorodeoxyglucose (FDG) and F-florbetapir PET imaging from the baseline visit were analyzed as surrogates for neuronal injury and amyloid pathology, respectively 18  FDG-PET, AD-specific ROIs representing the temporal, angular, and posterior cingulate gyri were utilized 21 . For F-florbetapir PET, we included regional amyloid deposition assessed by standardized uptake value ratio (SUVR) in the temporal, parietal, and cingulate cortex, as well as a composite global measure of multiple regions 22 .
Cognition. Scores on the Mini-Mental State Examination (MMSE) 23 at baseline were utilized as measures of cognitive performance.
Cognitive resilience. The number of errors on the American National Adult Reading Test (ANART) (which is an estimate of pre-morbid verbal IQ) and years of education were utilized as surrogate measures of cognitive resilience.
Genetic factors. Genotype and gene expression data from peripheral blood samples in ADNI were obtained as previously described 24 . In this study, we specifically analyzed APOE ε4 allele status (carrier vs. non-carrier) as well as expression data for the top genes with validated associations to AD (APOE, BIN1, CLU, ABCA7, CR1, PICALM, MS4A6A, CD33, and CD2AP) as listed in the AlzGene database 25 .
Demographics. Gender and age at the baseline visit were utilized in the predictive model.

Data aggregation and ML preprocessing.
A total of 94 potential predictive factors were included for analysis. A matrix was generated with 135 rows (representing study participants) and 94 columns (representing the potential predictive factors for MCI-to-AD progression). Prior to further analysis, we centered and standardized all data on a feature-by-feature basis by subtracting the mean and then dividing by the standard deviation. Figure 2 shows a flow diagram of the prediction framework developed for this study. The framework consists of four major steps: information-theoretic feature selection, classifiers and hyper-parameter optimization, goodness-of-fit evaluation, and generalized performance evaluation. They are explained in the following paragraphs. A unique aspect of this workflow is the ability to specify the number of parameters optimized in the classification model by using an information-theoretic feature selection method. This is particularly useful in assessing the generalizability of a classifier as a function of the number of model parameters.

ML-based prediction framework.
Feature reduction using joint mutual information (JMI). Mutual information between two random variables quantifies the amount of information shared between them. Mutual information is a more comprehensive measure of the relationships between random variables than statistical correlation-based approaches, which measure linear relationships only. Mathematically, mutual information is denoted by I Q R ( ; ) and defined for discrete random variables Q and R as shown in Eq. 1, where Q and  denote the alphabets of Q and R, respectively. Demographic factors like age, gender, and disease risk factors can also influence the progression. Indicated using solid arrows are factors that influence MCI-to-AD progression and broken lines indicate measurements that were used to measure those factors. Factors and measurements highlighted in red are those that have not been studied in previous MCI-to-AD progression studies in conjunction with the rest. When F k is one of the attributes in a set of attributes and Y is an outcome or class that can be predicted by the attributes, mutual-information-based approaches can be used to select the most predictive attributes. One such approach is to treat the attributes as independent random variables, rank them in descending order based on their mutual information with respect to the outcome Y, and select the top n number of attributes. One limitation of that approach is that useful and parsimonious sets of features should be both (i) individually relevant, and (ii) not highly correlated with each other. Joint mutual information shared between and Y is defined as shown in Eq. 2, where F k and Y denote the alphabets of F k and Y, respectively.

Figure 2.
A flow diagram illustrating the prediction framework. The framework uses a machine learning-based approach to learn a classifier using 80% of the full dataset and to test its performance on the remaining 20% of the data. Specific details of each step in the framework are as follows. (A) Stratified data partitioning: After the order of the subjects is randomized, the MCI-NP and MCI-P groups are separately partitioned with an 80%-training/20%-testing split. The respective training and testing sets from MCI-NP and MCI-P groups are combined to form the overall training and testing data for a single cross-validation. (B) Feature select loop: The top n out of 94 features that best jointly correlate with the class labels (P-MCI or MCI-NP) are selected using the joint mutual information (JMI) criterion. (C) Inner CV loop: A combination of hyper-parameters is selected for each classifier based on a tenfold cross-validation. (D) Goodness-of-fit metrics: The classifier learned in the previous step is tested on the testing dataset and measured on its performance. (E) Outer CV loop: A fivefold cross-validation is utilized to produce generalized performance metrics accounting for non-uniformly distributed data. A JMI-based feature selection method starts with an empty set of attributes and iteratively adds F i s that, when added, provide the maximum increase in the joint mutual information shared between the set of attributes and the outcome 26 . It has been shown to be the most stable and flexible feature selection method 27 among all the information-theoretic feature selection methods developed to date.

Classification methods.
We evaluated a number of classifiers in this study to understand the concepts of linear separability and generalizability in the context of predicting MCI-to-AD progression. Support vector machine (SVM), multiple kernel learning (MKL), and generalized linear models (GLM) with elastic-net regularization are the classifiers used in this study. SVM classifier, because it allows the transformation of features using linear and non-linear kernels, provides us the ability to evaluate the suitability of linear and non-linear classifiers for this problem. On the other hand, MKL allows the application of different kernel transformations for features from different modalities while optimizing more model-parameters than an SVM classifier. Although MKL and SVM are similar classification paradigms, MKL facilitates the integration of multiple modalities at the expense of potential overfitting. Finally, GLM classification with elastic-net regularization is an extension of the commonly known logistic regression classifier with additional regularization to minimize overfitting. In this study, we evaluated the overall generalizability of these classifiers that have distinct optimization objectives. A support vector machine is a binary classifier that finds the maximum margin hyperplane that separates the two classes in the data 28 . Suppose that the data being classified are denoted by  ∈ × X N P (N subjects and P features), and that the data from subject i are denoted by . Let us also use ∈ − Y { 1, 1} N to denote the class labels for all the subjects (where −1 and 1 are numerical labels for the two classes), and to denote the class label for subject i. The optimization problem to find the optimal hyperplane (described by weights opt opt has been found, the predicted class label for subject i is obtained as the sign of . This formulation assumes that the data are fully linearly separable between the two classes. When that is not the case (but there is still a linear classification task), slack variables and a tolerance parameter (box-constraint) can be introduced to obtain separating hyperplanes that tolerate small misclassification errors.
Dual formulation of SVM has received considerable interest due to its ability to use different kernel transformations of the original feature space without altering the optimization task, and due to its advantages in complexity when the data are high-dimensional 29,30 . The dual form of the above optimization problem can be written as shown in Eq. 4, where the operation . denotes an inner product between two vectors.
Notably, the dual formulation can be expressed simply in terms of the inner product between X i ( ) and X j ( ) (which is a scalar value). Therefore, any transformation of the features can replace the original features, i.e., X X , where φ is a feature mapping. Furthermore, for any such feature transformation φ, we can define a kernel function K such that With the kernel transformation, the cases where the original data are not linearly separable, may be solved as transforming the data to higher dimensions may introduce linear separation in the transformed domain. Linear, radial basis function (RBF), and polynomial kernels are widely used kernels in this context.
Multiple kernel learning is a classification technique that builds upon the above property of SVM and introduces additional variables in order to weight the kernel transformation of 31 individual features. It does so by replacing the kernel function K X X ( , ) , where K m is the kernel function of the m th feature and β m is its weight. This technique is particularly useful when the data contain different modalities of features that need to be weighted differently. In a standard kernel-SVM, either the kernel transformation is uniformly weighted, or the weights are determined manually. It has been shown that the same dual formulation of SVM can be utilized to find the optimal weightings of the kernel transformation under the constraint that Logistic regression is a subclass of generalized linear models (GLM) and is well-suited for binary classification tasks 32 . It models the response variable as a binomially distributed random variable whose parameters are described by predictor variables and model parameters. Regularization of the model parameters is utilized to avoid effects related to overfitting. The objective function of a regularized GLM model is where h x ( ) is the logistic function defined as x and β α P ( ) is the regularization term with the elasticity parameter α. The regularization term in general might contain both L 1 -norm terms and L 2 -norm terms, a situation that is commonly referred to as elastic net regularization with β β α β = + α α − P ( ) 1 2 2 2 1 33 . The elasticity parameter α ranges in (0, 1] with values of α → 0 approaching ridge regression and α → 1 approaching LASSO regression.
Hyper-parameter optimization. The hyper-parameters of the three classifiers are SVM's kernel, kernel scale σ, and box-constraint C, MKL's kernel, and GLM's penalty term λ and elastic-net parameter α. The optimal values of hyper-parameters are chosen by performing a grid-search of parameter values with a k-fold cross-validation within the training dataset.
Goodness-of-fit evaluation and generalized performance metrics. Goodness of fit of a classifier is evaluated by i) predicting the classes of the test dataset by using the classifier that was trained on the training dataset, and ii) comparing the predictions against the true class labels of the test dataset. The comparison is performed using standard performance metrics such as receiver operating characteristics (ROC) curve analysis, area under ROC curve (AUC), sensitivity, specificity, accuracy, precision, recall, and F1-score.
Because of heterogeneity in the data, choosing one partition of training and test datasets is not sufficient to credibly evaluate the performance of a classifier. A common practice to obviate the effect of heterogeneity in the data is to perform an k-fold training-testing cross-validation of the dataset. One run of this procedure is carried out by randomly choosing a subset of the dataset as training data, and testing on the rest of the dataset. That is repeated n different times, and the performance metrics of all n evaluations are averaged. In addition to this, the proportion of the different classes in the randomly chosen training dataset was kept constant in our approach via a stratified data-partitioning to eliminate any variability in the performance of the classifiers due to class imbalances in the data 16 . We performed those extra steps to obtain a generalizable set of performance metrics for the analyzed classifier.

Implementation of the prediction framework. The prediction framework was implemented in MATLAB
version R2017b and the code is publicly made available at https://gitlab.engr.illinois.edu/varatha2/adni_mci2ad_prediction. We utilized a standard fivefold cross-validation with 80% training data and 20% testing data to evaluate the performance of the classifiers. Training and testing data selection was performed using the stratified data-partitioning approach explained previously. The JMI criterion was used to identify ∈ … n {5, 10, , 90, 94} features based on the training dataset and corresponding labels. Using the reduced dataset, we trained SVM with a linear kernel, SVM with an RBF kernel, MKL with a linear kernel, LR with elastic net regularization, and RF. The highest average AUC obtained with a tenfold cross-validation within the training set was used as the selection criterion for the optimal hyper-parameters. We generated performance metrics for each classifier with the best identified hyper-parameters by predicting the labels of the test set and comparing them against the true class labels of the test set. The performance metrics of all five of the (outer) cross-validation runs were averaged to generate generalized performance metrics. test for linear separability. Linear separability of a dataset with two classes signifies that a linear classifier (or a hyperplane) can separate the two classes in the data as well as nonlinear classifiers can. We show that by using a slight modification of the histogram of projections 17 method, which is considered a test for class separability using SVM classification. The histogram of projections method 17 , is as follows. An SVM classifier is trained using the training data and a chosen kernel. Then, the test samples are projected on the 1-dimensional line perpendicular to the maximum margin hyperplane learned by the SVM model. Histograms of these projections for the two different classes are compared to evaluate the degree of separation achieved by the classifier. In our approach, we applied a sigmoid transformation to the projections of the test samples to obtain likelihood probabilities, and then, based on the likelihood probabilities, we created histograms of the two classes. This modification was performed in order to eliminate the effects of scaling when different kernel transformations are applied in SVM classification and to maintain the range of histograms within a closed set [0, 1].

Results
Our results are organized as follows. Table 1 represents the overall summary statistics of the dataset analyzed in this study with some important demographic factors. First, we present the results of our evaluation of whether linear models are sufficient to classify this dataset. Second, we report the analytic method we developed to evaluate generalizability and the results on the generalizability of the analyzed linear classifiers. Third, we report the relative importance of the biomarkers and clinical variables utilized in this study evaluated using the most generalizable classifier selected based on the previous result. Finally, we report the analysis based on LASSO regression 34 and a set of identified features that provide independent information in predicting MCI-to-AD progression.

MCI-P (n = 39)
MCI-NP (n = 96)  Linear separability. First, we evaluated the linear separability of the data in order to understand whether linear classifiers, which are less susceptible to overfitting, were sufficient to perform prediction of MCI-to-AD progression. We applied the modified histogram of projections method to our dataset by training two SVM classifiers with linear and RBF kernels and by plotting histograms of sigmoid transformations of the projections of test samples. Figure 3A shows a generic illustration of our approach based on the histogram of projections method. Figure 3B,C show the histograms obtained using our approach for linear and RBF kernels, respectively. The two histograms have similar shapes, and the misclassification errors obtained by choosing an appropriate threshold are also close (20.52% for the linear kernel and 19.78% for the RBF kernel). Figure 3D shows a grouped scatter plot of the probabilities obtained using linear versus RBF kernels for the MCI-P and MCI-NP classes. There is a significantly high correlation between the probabilities obtained using the two kernels (ρ = 0.99, p < 1e-6), indicating that the classification boundaries learned using the linear and nonlinear kernels are similar. This experiment suggests that the dataset being analyzed is linearly separable. Hence, we limited our analyses to linear classification techniques for the rest of this paper.

Generalizability of classifiers.
Second, we evaluated the overfitting characteristics of the linear classifiers in order to identify models that were likely to provide highly generalizable performance on this dataset. Three linear classifiers-multiple kernel learning (MKL) with linear kernels, support vector machine (SVM) with linear kernel, and generalized linear model (GLM) with elastic-net regularization-were trained using 80% of the whole dataset as the training set but using only a subset of all the features. We selected the subset (out of a total of 94 features) using the joint mutual information (JMI) criterion, by varying the number of features used each time as {5, 10, 15, …, 85, 90, 94}. We utilized tenfold stratified cross-validation to obtain average training and testing area under ROC curve (AUC) metrics and their respective standard deviations on each occasion. Figure 4 shows plots of the cross-validated AUCs with their standard deviations against the ratio number of features used in training number of training samples for both is not small (it is >0.14 in this case). Those results suggest that both SVM with a linear kernel and GLM with elastic-net regularization have good generalizability (with a consistent AUC value of approximately 0.9) regardless of the ratio of the number of features to the number of training samples, and MKL had good generalizable performance only when this ratio was small. Further, particularly for this dataset, all linear classifiers showed a reasonable test set predictability when the number of features was appropriately chosen during the training phase.

Prediction performance of linear classifiers.
Based on the previous experiment, we selected the subset of features that provided the most generalizable predictive performance for each linear classifier. Table 2 lists all the cross-validated goodness-of-fit metrics (AUC, sensitivity, specificity, accuracy, precision, recall, and F1-score) obtained using MKL with linear kernels, SVM with a linear kernel and GLM with elastic-net regularization, with their respective subsets of features. All linear classifiers produced comparable results, while SVM with a linear kernel provided the best result. predictive ability of individual modalities. To understand the predictive ability of individual modalities in a statistically impartial manner, we restricted our analysis in this subsection to only one linear classifier. Based on Table 2, we chose the SVM classifier with a linear kernel, as it provided the best performance in terms of the AUC metric. Then, we repeated our classification procedure using a stratified tenfold cross-validation for each of the modalities (using all the variables in the respective modality each time). Figure 5A shows a bar-chart of the average AUCs obtained by individual modalities. CSF proteomic markers (including Aβ, total-Tau, and phosphorylated-Tau) provided the best individual prediction capability, followed by the imaging markers in the order amyloid PET, FDG-PET, and SMRI. Gene expression, clinical, cognitive resilience, and demographic markers showed lower predictive ability than did CSF and imaging-based markers. Figure 5B shows a different bar-chart of the average AUCs obtained by iteratively removing the respective features of the individual modalities in a descending order based on their individual predictive abilities seen in Fig. 5A. A gradual decline in the AUC is observed, with the largest drop occurring when SMRI features were removed.  Table 2. Cross-validated goodness-of-fit metrics for linear classifiers. Here we list the goodness-of-fit metrics (AUC, sensitivity, specificity, accuracy, precision, recall, and F1-score) obtained for the test dataset (20% of the whole dataset), using the subset of features that provided the most generalizable result, as shown in Fig. 4. Their average values and standard deviations were computed using a tenfold stratified cross-validation.
Scientific RepoRts | (2019) 9:2235 | https://doi.org/10.1038/s41598-019-38793-3 Best predictors of MCI-to-AD progression. Next, we sought to identify a minimal set of predictors required to model MCI-to-AD progression. This was motivated by an observation about Fig. 5: although CSF features provided the best individual AUC (Fig. 5A), the removal of CSF features from the data (Fig. 5B) did not result in a substantial drop in the AUC (the reduction in AUC is within the error limits). That suggests that other modalities provide overlapping information that may be correlated with the CSF markers, and also indicates that all the modalities might have a notable level of correlation with each other. We used a generalized linear model (GLM) classification method with L1 regularization (known as LASSO regression) to identify a sparse set of features with minimal within-correlation and maximal prediction potential of clinical progression 34 . We chose to use LASSO regression instead of SVM classification for this task because the feature weights assigned by the SVM classifier with a linear kernel do not represent the independent predictive values of individual features when the features themselves are correlated with each other. However, the L1-norm-based regularization of LASSO regression may enable identification of the smallest possible set of predictors, since it penalizes the classifiers that utilize a large number of features in the resulting predictive function. Table 3 shows the results from this approach, identifying the factors that provided the best spread of independent information to predict MCI-to-AD progression. This model included PET, MRI, and CSF variables in addition to age and expression of CR1 (complement receptor 1) and was able to predict MCI-to-AD progression with an AUC of 0.92.

Discussion
In this work, we developed an accurate and generalizable machine learning-based methodology for predicting short-term progression of MCI to AD dementia in a well-characterized clinical cohort. A unique aspect of this work was our evaluation of important genetic markers and cognitive resilience markers in addition to the neuroimaging biomarkers and clinical examination. The results suggest that a combination of selected neuroimaging, blood and CSF biomarkers, and demographic traits reflect the underlying pathophysiology and factors that drive clinical progression in the AD spectrum. The methodology allowed us to discover that expression of CR1 and AD-pattern neuropathology and neurodegeneration (MRI measures in the frontal lobes) in brain regions associated with cognitive resilience added independent predictive values in predicting MCI-to-AD progression. We also evaluated the relative importance of individual predictors and identified a minimal set of predictors that are important for predictive modeling of MCI progression to AD.
predictors of MCI progression to AD. The broad pathogenesis of AD has been well-described conceptually, from initial alterations in molecular and cellular pathology to neurodegeneration and eventually to clinical impairment sufficient to cause dementia 2 . However, it is not yet fully understood what specific factors "shift the curve" to either promote or inhibit the development of dementia. As a result, we initially approached our study in a relatively unbiased fashion, casting a broad net for potential predictive factors out of multidimensional clinical, neuroimaging, and other biomarker data. Through our ML approach, we found that measures of AD neuropathology (CSF amyloid and tau and cerebral amyloid assessed by PET) and neuronal injury (assessed by FDG-PET and structural MRI) explained the most variance in separating fast versus slow progression from MCI to AD dementia. The importance of biomarkers in predicting progression has been studied in the past, but the approach we took allowed us to rank various predictors (including the biomarkers) as well as identify the minimal predictor set that are key to the overall models. When the biomarkers were excluded, predictors such as cognitive performance, cognitive resilience, and expression of selected genes individually explained less of the outcome variance and collectively appeared to offer less new information to the predictive model. Our results support the hypothesis that CSF biomarkers and imaging can be used as surrogates for neuropathology and brain health and can serve as key indicators of future MCI prognosis.

Information added by genetic factors and cognitive resilience. An important aspect of this work
is the incorporation of genetic factors and cognitive resilience into our predictive model. Most cases of AD are thought to be genetically complex, with multiple factors presumed to contribute to susceptibility and protection, with the largest known factor being the APOE ε4 allele 35 . Understanding of the genetic architecture of AD has greatly expanded over the last decade as a result of genome-wide and rare-variant association studies, among other approaches 7 . However, the specific genetic factors that influence progression at various clinical stages of AD are still not well-characterized. Our ML approach identified expression of CR1 as a key factor in predicting fast versus slow progression to AD dementia, which is a novel finding of this study. Interestingly, our final minimal set model included CR1 expression but did not include APOE ε4 allele status, suggesting that the former provided unique information for clinical course prediction while the latter was already represented by surrogate biomarkers, specifically amyloid deposition 36 . Polymorphisms (genotype variations) in CR1 have been associated with AD status and endophenotypes in numerous large-scale studies [37][38][39][40][41][42] . CR1 encodes a receptor involved in complement activation, a major immune mechanism with a wide array of functions; it has been proposed that it impacts the clearance of amyloid in AD 43 . Previous findings on CR1 have been illuminated by a more recent and extensive body of work highlighting immune system pathways as potential cruxes in AD pathophysiology [44][45][46][47][48] .
Our new findings relating CR1 expression to progression from MCI to AD dementia provide further validation of that previous work and argue for greater focus on CR1-and on genetic variation in MCI-to-AD progression more broadly 49 -to enable better understanding of the mechanisms underlying AD and its clinical trajectories. It is not unexpected that in our ML-based prediction model for MCI to AD progression, CR1 gene expression contributed relatively less to explaining variance than biomarkers of neuropathology and neuronal injury. Surrogates of pathology (fluid biomarker and neuroimaging) reflect the extent of disease related changes and are likely more proximal to clinical manifestations of disease compared to gene expression which may be upstream of these changes and thus possibly modifiable by concomitant forces over time. An example is the relationship between APOE ɛ4 and amyloid: while APOE ɛ4 is a key driver of amyloidosis, the measured effect of amyloid load on cognition is significantly stronger than the impact of APOE ɛ4 on cognition 36,50 . Complementary methods to incorporate collective effects of multiple genes, including pathway analyses and polygenic risk scoring [51][52][53][54] , could be incorporated into our ML approach in future. In addition, for this study we analyzed microarray gene expression (rather than genotype) data, which offers the conceptual advantage of being a dynamic (rather than static) marker of late-life conditions but which has the disadvantage of representing a downstream effect of influences earlier in life with the potential to be modified by interactions with other heritable and non-heritable factors. Finally, for this study we limited the focus to the AlzGene Top 10 list, but other genes with less well-known population-level effects on case-control status may have larger contributions to late-life clinical progression which could be discovered with an unbiased genome-wide approach. Despite these constraints, our approach serves as a proof of concept that incorporating genetic data can add value to ML-based clinical prediction models and highlights CR1 for further study on its potential effects at the inflection of MCI to AD progression.
Although cognitive resilience has been shown to contribute significantly to delaying the onset of clinical impairment, neither education nor verbal IQ was identified as a key predictor of MCI progression. There are two possible explanations for those results. (i) Preservation of structure or lack of atrophy, as observed via MRI, may be a better surrogate of resilience than are estimates of pre-morbid IQ or educational attainment. It has been shown that cognitive resilience is captured well by greater volume and metabolism, especially in the frontal and cingulate regions 55 . (ii) Cognitive resilience may be more relevant before the onset of cognitive symptoms or impairment 12 , and thus may be of less importance in cognitively impaired individuals, who are the focus of this study.  strengths of the computational approach. We presented a machine learning-based prediction framework for predicting MCI-to-AD progression using state-of-the-art classification techniques under a widely accepted cross-validation setup that accounts for sample bias and unbiased hyper-parameter optimization.
Our results indicate that linear classifiers are sufficient to identify patients who have the potential to progress from MCI to AD within 3 years, via the use of multimodal measurements including imaging, CSF, genetic, and clinical data. The reason for this might be that all the features were engineered prior to application of machine learning in such a way that a dichotomy within a feature correlates with disease progression. Linear classifiers are less susceptible to overfitting and thus can be appropriate for clinical translation. In addition, we developed an analytic method for evaluating the generalizability of three different linear classification approaches, namely GLM-ElasticNet, multi-kernel SVM with linear kernels, and standard SVM with a linear kernel. Multi-kernel SVM showed a slightly greater tendency to overfit than the other linear classifiers in our experiments. Because multi-kernel SVM optimizes twice as many model parameters as standard SVM or GLM, it seems that the number of samples used in this study was not enough to train multi-kernel SVM in a generalizable manner. However, as indicated in the experiments, when the number of features used in the classification was small, multi-kernel SVM seemed to provide a more generalizable performance compared to when more features were used in the classification, since there were fewer optimized model parameters in the former scenario. Even so, standard SVM and GLM provided better generalizability and achieved higher classification accuracies with more features. In comparison with previously published predictors of MCI-to-AD progression, our approach achieves superior predictive performance, with a 0.93 area under ROC curve providing a 6% improvement over the current best predictor 4 , which utilized only CSF and imaging modalities for the prediction of short-term MCI-to-AD progression. It is our firm belief that AUCs of test sets are the most reliable measure of a classifier's general performance, and hence that our evaluation presents a fair assessment of the proposed ML-based predictor.
Limitations and future directions. Our study has several limitations, which may be addressed by future work. We used publicly available data from the ADNI and ultimately analyzed only a modest sample, because we could only use data from patients for whom we had complete data for all the assessed predictive variables. Replication of our methods in similar and larger cohorts would help validate our approach. In addition, application of our final model to an independent dataset could extend our findings by assessing their success in predicting clinical progression from known baseline data. Further, use of alternative surrogate variables-such as verbal episodic memory performance (rather than MMSE scores) as a reflection of cognitive impairment, and lifetime intellectual enrichment (rather than, or in addition to, the AMNART or educational attainment) as a reflection of cognitive resilience-as well as newer biomarkers such as Tau-PET might produce a more optimal model. We also chose to focus on differentiation of faster versus slower progression from MCI. However, it is not yet known whether the factors that are important at that clinical stage are more broadly generalizable to earlier stages of AD at which clinical intervention might have greater impact. Finally, a broader investigation of genetic factors, including genome-wide genotype and expression data, might increase the accuracy of our model and raise additional loci for follow-up.
In conclusion, we used a machine learning approach to integrate data from multiple sources to predict progression from MCI to AD dementia. By analyzing the linear-separability of the dataset, we established that linear classifiers are sufficient for predicting MCI-to-AD progression in the proposed setting. Furthermore, we also developed an analytical method to assess the generalizability of classifiers based on the number of model parameters. We also showed that, the markers of AD pathophysiology (amyloid, tau, and neuronal injury) provided very high predictive values, the genetic factors and brain regions associated with cognitive resilience also displayed independent predictive values. Our findings, particularly in relation to the impact of genetic factors and cognitive resilience in the setting of biomarkers and clinical tests, warrant further investigation using larger datasets and may have downstream benefits leading to improved prognostication and risk stratification in AD.

Data Availability
Data used in preparation of this article were obtained from the Alzheimer's disease Neuroimaging Initiative (ADNI) database (http://adni.loni.usc.edu). Thus, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data, but did not participate in this analysis or the writing of this report. A complete listing of ADNI investigators can be found at http://adni.loni.usc.edu/wpcontent/uploads/ how_to_apply/ADNI_Acknowledgement_List.pdf/. For additional details and up-to-date information, see http:// www.adni-info.org.