A semi-mechanism approach based on MRI and proteomics for prediction of conversion from mild cognitive impairment to Alzheimer’s disease

Mild cognitive impairment (MCI) is a precursor phase of Alzheimer’s disease (AD). As current treatments may be effective only at the early stages of AD, it is important to track MCI patients who will convert to AD. The aim of this study is to develop a high performance semi-mechanism based approach to predict the conversion from MCI to AD and improve our understanding of MCI-to-AD conversion mechanism. First, analysis of variance (ANOVA) test and lasso regression are employed to identify the markers related to the conversion. Then the Bayesian network based on selected markers is established to predict MCI-to-AD conversion. The structure of Bayesian network suggests that the conversion may start with fibrin clot formation, verbal memory impairment, eating pattern changing and hyperinsulinemia. The Bayesian network achieves a high 10-fold cross-validated prediction performance with 96% accuracy, 95% sensitivity, 65% specificity, area under the receiver operating characteristic curve of 0.82 on data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database. The semi-mechanism based approach provides not only high prediction performance but also clues of mechanism for MCI-to-AD conversion.


Results
Biomarkers selection. In this section, the process of biomarkers selection is described. The dataset used in this study contains 518 biomarkers (328 MRI markers and 190 plasma markers). 45 biomarkers (1 MRI marker and 44 plasma markers) are deleted during data checking due to too many missing entries. 75 biomarkers (57 MRI markers and 18 plasma markers) with significant difference between converters and non-converters are identified by ANOVA test. 34 biomarkers (25 MRI markers and 9 plasma markers) related to Alzheimer's disease assessment scale (ADAS-cog) are selected by lasso regression. 7 biomarkers (5 MRI markers and 2 plasma markers) are eliminated during Bayesian network structure learning because they fail to connect to the Bayesian network. In addition, as 2 MRI markers are labeled as "unknown", they are also eliminated. Finally, 25 biomarkers (18 MRI markers and 7 plasma markers) are selected for conversion prediction. The process of biomarkers identification is summarized in Fig. 1. The list of selected biomarkers is shown in Table 1.
Structure and performance of Bayesian network. In this section, we present the results of Bayesian structure learning and the performance of conversion prediction. The Bayesian network structure obtained by max-min hill-climbing (MMHC) is given in Fig. 2. It contains 26 nodes and 43 arcs.
In order to evaluate the performance of Bayesian network, a 10-fold cross-validation is performed to estimate its accuracy, sensitivity and specificity. Furthermore, the performance of Bayesian network is compared to the performances of linear discriminant analysis (LDA) and SOM. The performances of all these methods are evaluated by 10-fold cross-validation. The results are given in Fig. 3. The Fig. 3A shows that the accuracy and sensitivity of Bayesian network are higher than those of LDA and SOM with markers selection. In Fig. 3B, the area under receiver operating characteristic curve (AUC-ROC) of Bayesian network is much higher than that of LDA and SOM with marker selection. Moreover, to evaluate the performance of markers selection, we apply SOM and Bayesian network with or without markers selection and compare their performances. With markers selection, the classification performances of both SOM and Bayesian network are improved.

Network disruption profile.
According to the result of Bayesian network, a group of highest conversion probability patients (high-risk group, n = 11) and a group of lowest conversion probability patients (low-risk group, n = 48) are drawn from the dataset. 11 biomarkers have significant difference (P < 0.05, ANOVA test) between high-risk group and low risk group. The mini network balance map Fig. 4A) shows that the high-risk group may suffer from more severe network disruption than the low risk group. The network disruption parameters coincide with the mini network balance map. Parameters U and ϕ increase significantly in high risk group (P < 0.01, ANOVA test, shown in Fig. 4B) which may suggest that patients with greater U and ϕ may have higher conversion risk.

Discussion
In this study, we propose a semi-mechanism based Bayesian network to predict the conversion from MCI to AD. The proposed method has two contributions. Firstly, the proposed approach achieves relative high prediction performance. Secondly, as the Bayesian network can learn the causal relationships among biomarkers from the database, these causal relationships can provide some more insight into the mechanism of MCI-to-AD conversion.
Scientific RepoRts | 6:26712 | DOI: 10.1038/srep26712 The proposed model is compared to previous researches based on data-driven methods ( Table 2). Comparing with LDA and SOM, Bayesian network has higher accuracy and sensitivity with markers selection. The high sensitivity of Bayesian network may lie in two points. On one hand, the semi-mechanism nature of Bayesian network may provide higher performance because it can learn causal relationships from data and combine these knowledge and data to conduct the prediction task 22 . On the other hand, plasma markers may be highly sensitive in conversion prediction 12 . Though the data-driven methods also achieved high performance, the Bayesian network still has its unique advantage. The structure of Bayesian network may contain the causal relationships of markers which makes it a semi-mechanism method and provide more information beyond the performance of classification.
In addition, with markers selection, the classification performance of Bayesian network is improved. It suggests that Bayesian network should work with an appropriate marker selection strategy. In another words, without markers selection, Bayesian network may produce false positive causal relationships which may not only decrease the performance but also mislead the MCI-to-AD conversion mechanism investigation. Therefore, combining Bayesian network and lasso marker selection strategy is very helpful in improving understanding the conversion mechanism and classification performance.
The semi-mechanism nature of Bayesian network is beneficial on investigating the mechanism of conversion. Structure of Bayesian network shows that 6 markers including volume of left middle temporal, cortical thickness average of right entorhinal, volume of right inferior temporal, AGRP, c-peptide, and fibrinogen may be related to the conversion directly. Our result, that destruction of entorhinal is associated with MCI-to-AD conversion, is consistent with previous research 23 . Previous researches had also reported the variations of temporal, c-peptide level, and fibrinogen level in AD patients 19,24,25 . However, our results suggest that these changes may have happened at MCI stage. It indicates that the conversion from MCI to AD may start with destruction of temporal, entorhnal, increased level of AGRP, c-peptide, and fibrin 25 . Bayesian network identifies the variations of above six markers caused by MCI-to-AD conversion directly. But some of these changings may not be the key factors in the conversion. Therefore a reanalysis based on the    results of Bayesian network is performed to identify the major factors. The subgroups network disruption profile suggests that the progress MCI patient may suffer from more severe network disruption than stable MCI patients. Network disruption may be related to the marker panel including 11 markers. The six markers identified by Bayesian network and marker panel related network disruption share three markers: Cortical Thickness of Entorhinal, Volume of Temporal and AGRP. These three markers may be the key factors in the conversion. In other word, they might be attributed to the warming signals of conversion. Previous researches showed that the destruction of entorhinal and temporal is associated with verbal memory impairment and verbal memory impairment might be the warming indicator of MCI-to-AD conversion 26 . In addition, clinical researches have reported that AD patients have greater preference for high-fat and sweet food than normal groups. However, our results suggested that such change in eating pattern may have happened at MCI stage 27,28 , as the elevated level of AGRP, an orexigenic peptide, in high-risk patients may increase the preference for a high fat diet 29 . The crosstalk between cerebral destruction and plasma markers alteration revealed by Bayesian network can provide more clues for the mechanism of conversion. The crosstalk between C-peptide and cerebral destruction may play a vital role in the MCI-to-AD conversion. C-peptide is a measure of insulin secretion. Elevated C-peptide level represents high peripheral insulin secretion. It is reported that high peripheral insulin secretion can increase the risk of AD. Because high level peripheral insulin secretion impairs amyloid clearance by inhibiting brain insulin production which is a beneficial effect on amyloid clearance 30 . Bayesian network suggests that C-peptide may be related to the destruction of middle temporal, entorhinal, and inferior temporal. It suggests that amyloid may mainly aggregate in the above three regions at the MCI stage which may aggravate their damage. As all above three regions are involved in verbal memory, high level of C-peptide may impair to verbal memory which was confirmed by previous works 26,[31][32][33] .
In summary, the analysis of Bayesian network shows that the conversion from MCI to AD may start with multiple pathological changes such as verbal memory impairment, vascular abnormalities, hyperinsulinemia and eating pattern change. In this study, a high performance semi-mechanism based approach is developed to predict the conversion from MCI to AD by combining MRI and plasma markers. The semi-mechanism based approach provides not only high performance prediction but also more insight into the mechanism of conversion from MCI to AD.

Subject and Method
Subject. Patients. In this study, the following criteria are used to select subjects for model developing: • Patients with baseline MRI scan records • Patients with baseline plasma-based biomarker data • Patients with baseline ADAS-cog scores • Patients with MCI due to Alzheimer's disease • Patients with diagnosis records which can be used to determine whether they convert from MCI to AD in 18 months Finally, a data set with complete imaging, plasma-based biomarkers, ADAS data is drawn from ADNI including 316 MCI patients (99 converters and 217 non-converters). The demographic information of subjects is given in Table 3.
Imaging biomarkers. Imaging data in this study is obtained from dataset UCSF-Cross-Sectional FreeSurfer (FreeSurfer Version 4.3). The dataset is available at https://ida.loni.usc.edu/pages/access/studyData.jsp. In this dataset, all scans were acquired on 1.5 T MRI scanners. The imaging data were processed and analyzed with FreeSurfer 4.3 by the UCSF team. The dataset includes 328 MRI biomarkers which can be grouped into 5 categories: average cortical thickness, standard deviation in cortical thickness, the volumes of cortical parcellations (based on regions of interest automatically segmented in the cortex), the volumes of specific white matter parcellations, and the total surface area of the cortex. Details of the analysis procedure are available at http://adni.loni. ucla.edu/research/mripost-processing/.
Plasma-based biomarkers. The plasma-based biomarker data is obtained from dataset Biomarkers Consortium Plasma Proteomics Project RBM multiplex data. The data is available at https://ida.loni.usc.edu/pages/access/ studyData.jsp. The data was acquired by analyzing a subset of plasma samples from the ADNI cohort in a 190 analyte multiplex immunoassay panel. The panel, referred to as the human discovery map, was developed on the Luminex xMAP platform by Rules-Based Medicine (RBM) to contain proteins previously reported in the literature to be altered as a result of cancer, cardiovascular disease, metabolic disorders and inflammation. Details of the assay technology and validation has been described elsewhere (http://adni.loni.ucla.edu/wp-content/ uploads/2010/11/BC_Plasma_Proteomics_Data_ Primer.pdf).

Method.
Considering that ADNI contains more than 500 biomarkers, it is essential to select the more predictive biomarkers to obtain a parsimonious model and avoid the classifier suffering overfitting. Then the Bayesian network is established based on the causal relationships among selected markers for predicting the AD-to-MCI conversion. Finally, a reanalysis of Bayesian network results is performed to profile the network disruption of the patients with highest probability of converting to AD and those with lowest probability. The framework is summarized in Fig. 5.
Biomarkers selection. Biomarkers selection includes two stages. At the first stage, ANOVA test is employed to screen biomarkers with significant difference (P < 0.05) between converters and non-converters. At the second stage, lasso regression is used to filter biomarkers related to ADAS-cog from the selected biomarkers at the first stage.
Lasso regression is a popular technique for feature selection which can continuously shrinks coefficients 34 . It drops biomarkers by shrinking some of coefficients to zero. In this study, a Least Angle Regression (LARS) algorithm is used to solve lasso 35 .  Bayesian network. Considering that the causal relationships among the selected markers may remain unknown, a Bayesian network structure learning algorithm termed as the max-min hill-climbing (MMHC) is employed to learn the causal relationships among the selected markers. MMHC algorithm is a hybrid method, using concepts and techniques from both constraint-based approaches and score-based approaches, which can achieve high quality in structure learning 36 . After the Bayesian network is learned from data, the most popular Bayesian network inference algorithm named junction tree is employed to acquire the conversion prediction 37 .

MCI Converters Non-Converters
Model evaluation. In this study, the receiver operating characteristic (ROC) curve is used to evaluate the performance of Bayesian network. The ROC, which has become established as an important tool for classifier evaluation, is a graph of true positive rate (TPR) against false positive rate (FPR) at various operating points as a decision threshold 38 . The area under the ROC curve (AUC) is a measure of predictive ability 39 . Moreover, three parameters termed as accuracy (number of correctly classified samples divided by the total number of samples), sensitivity (the number of correctly classified converters divided by the total number of converters) and specificity (the number of correctly classified non-converters divided by the total number of non-converters) are calculated and evaluated by 10-fold cross-validation for a further measurement for the model performance 40 .
Network disruption analysis. To get more insight into the mechanism of the conversion, a reanalysis of Bayesian network results is performed using a mathematic method for evaluating the disruption of biology network which was proposed in our previous research 41 . In this study, subjects are divided into two subgroups high risk group and low risk group according to the results of Bayesian network and a mini network balance model is developed to evaluate the network disruption for both high-risk group and low-risk group. The network disruption comparison between these two subgroups may provide more insight into the mechanism of AD-to-MCI conversion. The mini network balance model contains three parameters U, K, and ϕ. U is response to both consistency variation and inconsistency variation comprehensively. K responds to multi-marker consistency variation. ϕ is response to the multi-marker inconsistency variation. These three parameters can be calculated as below: