Prediction of 7‐year's conversion from subjective cognitive decline to mild cognitive impairment

ABSTRACT Subjective cognitive decline (SCD) is a high‐risk yet less understood status before developing Alzheimer's disease (AD). This work included 76 SCD individuals with two (baseline and 7 years later) neuropsychological evaluations and a baseline T1‐weighted structural MRI. A machine learning‐based model was trained based on 198 baseline neuroimaging (morphometric) features and a battery of 25 clinical measurements to discriminate 24 progressive SCDs who converted to mild cognitive impairment (MCI) at follow‐up from 52 stable SCDs. The SCD progression was satisfactorily predicted with the combined features. A history of stroke, a low education level, a low baseline MoCA score, a shrunk left amygdala, and enlarged white matter at the banks of the right superior temporal sulcus were found to favor the progression. This is to date the largest retrospective study of SCD‐to‐MCI conversion with the longest follow‐up, suggesting predictable far‐future cognitive decline for the risky populations with baseline measures only. These findings provide valuable knowledge to the future neuropathological studies of AD in its prodromal phase.


| INTRODUCTION
Alzheimer's disease (AD) is a neurodegenerative disorder with a slow and lengthy progression. The neuropathological process of AD is initiated at least 15-20 years before the first symptom of cognitive impairment (Jack et al., 2018;Tondelli et al., 2012). Current consensus has emphasized the need for early detection of AD based on neuroimaging (PET and MRI; Guo, Landau, & Jagust, 2020) or invasively acquired CSF biomarkers (Frisoni et al., 2017). Detection of amnestic mild cognitive impairment (MCI), a prodromal stage of AD, maybe still too late for early intervention as the massive neuron loss and irreversible cognitive impairment may have already incurred at this stage (Petersen, 2009). It is crucial to explore early markers to predict possible AD conversion at an even earlier stage. Subjective cognitive decline (SCD), defined as a subjectively experienced decline in cognitive capacities in the absence of objectively measurable neuropsychological deficits, may serve as a symptomatic indicator of preclinical AD years before MCI (Jessen et al., 2014), besides indicators of other diseases. Recently, the National Institute on Aging and Alzheimer's Association (NIA-AA) updated the research guideline for AD and defined SCD as a probable clinical stage 2 in the Alzheimer's continuum, that is, with normal performance within expected ranges on objective cognitive tests, a distinctive transitional stage between asymptomatic or preclinical (Stage 1) and symptomatic MCI (Stage 3; Jack et al., 2018).
Accumulating epidemiologic studies support the inclusion of SCD as a pre-MCI stage and a high-risk cohort of future progression (Buckley et al., 2016;Jessen et al., 2020;Mitchell, Beaumont, Ferguson, Yadegarfar, & Stubbs, 2014;Reisberg, Shulman, Torossian, Leng, & Zhu, 2010;Snitz et al., 2018). A meta-analysis study showed that the annual conversion rate from SCD to MCI is 6.6% (Mitchell et al., 2014). However, SCD can manifest in healthy elderly or people with other diseases as well. Some studies regarded subjective cognitive complaints as a benign symptom that may not lead to severe consequences (Hessen et al., 2017), while others reported that a small part of SCD might develop to cerebrovascular disease, Parkinson's Disease, or non-Alzheimer degenerative dementia (Jessen et al., 2020). The evidence altogether indicates that SCD could be a heterogeneous cohort, including those who will eventually become AD, stay stable, and show related symptoms to other diseases. It is extremely challenging to differentiate SCD with progressive ADrelated symptoms (pSCD) from SCD with stable cognitive performance (sSCD) at the baseline. PET or CSF can detect Aβ or tau-related biomarkers for early AD diagnosis, but both are invasive and expensive, making it unsuitable for large-cohort screening (Jack et al., 2018).
Besides neuroimaging markers, clinical and demographic features are also considered to be strong risk factors for dementia. The Lancet International Commission on Dementia Prevention reported nine risk factors for dementia (Livingston et al., 2017). They are less education, physical inactivity, low social contact, smoking, hearing loss, depression, diabetes, hypertension, and obesity. They may interact with each other as a whole and could jointly increase the risk of AD progression.
However, none of them has been investigated in any SCD progression study, and whether they are strong enough to independently differentiate pSCD from sSCD at the baseline is yet unclear. Different sMRI indicators and clinical/demographic features may likely have complex relationships, and they could be jointly used to predict SCD-to-MCI progression. Hence, in this study, we aimed to use machine learning, an advanced multivariate pattern recognition method, to not only detect the progression-related markers jointly and objectively but also conduct an individualized differentiation between pSCDs and sSCDs.
Of note, the study was based on the China Longitudinal Aging Study (CLAS) of Cognitive Impairment (Xiao et al., 2013;Xiao et al., 2016), which was followed up for 7 years. As there was no incidence of AD case, we only focused on predicting SCD conversion to amnestic MCI (the pSCD is hereby defined accordingly from here on). To validate our model and further test the efficacy of the identified contributive markers, we further compared the detected markers among various normal control (NC)/SCD/MCI groups with independent datasets.

| Study cohorts
The CLAS study is a community-based study initiated in 2011 (Xiao et al., 2013;Xiao et al., 2016). The current data constitute samples from Shanghai where all the subjects received a baseline T1-weighted MRI scan. Some of the data was used in a cross-sectional study on SCD (Yue et al., 2018). In the current study, as the 7-year follow-up has recently finished, we could include all the subjects with both baseline SCD diagnosis and the follow-up visit. Of all the 111 SCD subjects defined at the baseline, the 7-year follow-up was completed in 92 (82.8%) subjects (Table 1) All the SCD participants were assessed by self-report. Based on (Jessen et al., 2014), SCD may refers to any cognitve domain, which is not restricted to memory. However, this study was started in 2011; at that time, only "subjective memory complaint" were recorded and investigated. Meanwhile, the evidence of an association of preclinical AD with problems of memory functioning may be the strongest (Jessen et al., 2014), so we mainly focused on their memory complaints. The SCD was diagnosed according to the following criteria (Jessen et al., 2014;Molinuevo et al., 2017): (a) the onset age of memory decline is >60 years old; (b) presence of gradual memory decline has persisted for ≥6 months; 3) objective memory performance at baseline is within the normal range. More details on the SCD diagnosis procedure are listed elsewhere (Yue et al., 2018). MCI was clinically diagnosed according to the Peterson's criteria (Petersen et al., 1999) with consideration of comorbid conditions. In addition, a battery of neuropsychological tests was carried out for MCI diagnosis , including MMSE, MoCA, a Chinese version of the Rey Auditory-Verbal Learning Test (RAVLT), and Activities of Daily Living (ADL). At the followup visit, we had 24 pSCD subjects who had progressed to MCI and the 52 sSCD subjects who kept cognitive normal, none of which converted to AD at the follow-up visit.

| Subject assessment
We performed a comprehensive battery of sociodemographic and physical health measurements for each subject. As potential risk factors of SCD progression (Livingston et al., 2017), sociodemographic data (gender, age, years of education, years of smoke, years of drink, and Body Mass Index [BMI]) and status of somatic diseases (years of hypertension, hyperlipidemia, diabetes mellitus, sleeping disorder, incident stroke, heart disease, and any type of surgery) at the baseline were collected, as well as the new onsets of the above somatic diseases and years of hearing loss acquired over the follow-up period. Three psychological assessments were carried out at the baseline, including MoCA (measuring overall cognitive performance), geriatric depression scale (GDS, also acquired at the follow-up due to the close association between depressive symptoms and cognitive performance), and social support questionnaire (measuring perceptions of social support and its satisfaction).  Table S1.

| SCD progression prediction
Due to the potential redundancy in the 223 (25 clinical + 198 MRI) features, it is necessary to select the most predictive features to train a concise model and avoid overfitting. We used a three-stage feature selection scheme. First, a Relief algorithm was employed to rank the features according to the extent of associations with the labels (i.e., pSCD and sSCD), which has been widely applied in machine learning tasks to improve classification performance, for example, genetic analysis for diseases, Parkinson's disease diagnosis and contributive feature identification (Urbanowicz, Meeker, La Cava, Olson, & Moore, 2018). Considering contextual information during ranking, Relief can properly handle strong dependency among features. Specifically, it generates a robust feature ranking by identifying the feature value differences between nearest-neighbor instance pairs. The number of neighbors was selected from a range between 1 and 20 at a step of 1. Therefore, we generated 20 ranking lists of features. Second, an overall ranking for all the features was obtained by integrating the 20 ranking lists and calculating the occurrence frequency of each feature in the top-50 of each ranking list. Finally, the classification performance based on cost-sensitive support vector machines (CSVM), which better handles the imbalanced samples than the traditional SVM was derived to evaluate the ranked feature set and determine the predictive features (Park, Luo, Parhi, & Netoff, 2011). F1 score (defined as the harmonic average of the precision and sensitivity, that is, F1 score = 2 × (precision × sensitivity)/(precision + sensitivity)) was taken as the classification performance evaluation metric for further feature selection and parameter optimization of CSVM. A sequential forward selection strategy was adopted in this step, which sequentially added the features according to their ranks until adding a new feature did not improve the CSVMs' performance. The CSVMs with the same parameters were used for feature selection and SCD progression prediction. We set a higher misclassification penalty (with a freely estimable cost ratio, pSCD:sSCD) for the pSCD samples compared to those for the sSCD samples in the CSVMs to alleviate the problem caused by unbalanced samples, as pSCDs were much fewer than sSCDs.
The prediction model was validated by nested leave-one-out cross-validation (LOOCV). The feature set selected by the inner LOOCV according to the above method was used to train the classifier and the final prediction performance (including accuracy, sensitivity, specificity, and F1 score) was obtained by the outer LOOCV. In each iteration of the outer LOOCV, one subject was left out as a testing sample and the remaining 75 subjects were for training (these 75 subjects were fed into the inner LOOCV where the predictive features were selected). The trained model was applied to the testing sample for generating a predicted label to compare with the ground truth. The outer LOOCV went through all the subjects. The frequency of each feature being selected across the 76 outer LOOCV iterations was used to assess feature importance. A feature was regarded as a contributive feature if this frequency is higher than 95% (i.e., selected from 72 out of 76 LOOCV iterations). Figure 2 shows the flowchart of the entire analysis.
To facilitate clinical application, after the contributive features were identified, we re-trained a refined SCD progression prediction model with only these contributive features as a representative prediction model. Based on this model, we drew a receiver operating characteristic (ROC) curve and used the area under the curve (AUC) to measure the discriminant ability in the prediction of SCD progression. We further derived an intuitive "decision score" for each subject according to the scoring function learned by such a representative prediction model.

| Statistical analysis
Group comparisons of the demographic and clinical data were conducted using two-tailed independent samples t-tests for continuous

| The most contributive features in SCD progression prediction
Five features were identified and consistently selected as contributive features in pSCD vs. sSCD classification (see details of the selection frequency in Table S1). They are baseline stroke history, years of education, baseline MoCA score, baseline volume of the left amygdala, and baseline WM volume at the wmSTSbanks. Figure 3a and Table S4 show the group comparison results for the five features between pSCDs and sSCDs. In addition to education level, both MRI features (see Figure S1 for locations) show significant group differences

| Relationship between MRI features and cognitive score
The partial correlation analysis between the two important MRI fea-

| Structural connections of the right wmSTSbanks
For possible WM connections to the right wmSTSbanks (i.e., the only WM feature from the MRI detected by the machine learning-based pSCD vs. sSCD differentiation). dMRI tractography with an independent dMRI data from a healthy elderly cohort (Experiment S1) showed that there are many WM fibers could pass through the right wmSTSbanks and linking to the right amygdala and the right hippocampus, two regions targeted by the AD pathology ( Figure S2).
Besides, there are many other cortical regions, especially the right middle, superior, and inferior temporal gyri, as well as the right interior parietal and supramarginal areas and right putamen, which could be reached by the WM fibers that passing through the right wmSTSbanks ( Figure S3).

| Validation based on independent data sets
From the CLAS database, we further identified sNC, pNC, and sMCI subjects and extracted the same features from them. Together with the sSCD, all four groups were separately compared with the pSCD group using two-sample t-tests with Bonferroni correction. We found a similar trend that sNC, sSCD, and sMCI, who were not potentially affected by AD pathology, tend to have lower baseline wmSTSbanks volume compared with pSCD, with the pNC specifically sitting inbetween the three stable groups and the sSCD group. Specifically, we found a significant difference (p corrected < .05) between sNC and pSCD in addition to the significant difference between sSCD and pSCD, as well as a trend-to-significant difference (p corrected < .1) between sMCI and pSCD (Figure 4b, see details in Experiment S2). Furthermore, the combined progressive group (pSCD + pNC) also showed a greater baseline right wmSTSbanks volume than the combined stable group (sSCD + sNC; Figure S4). From the ADNI data, although no significant difference was found from the comparison of the baseline volume of the right wmSTSbanks between sNCs and pNCs (p = .24) due to lower statistical power caused by small sample size (due to short follow-up), we still spotted a similar trend to the main results, that is, pNCs had a higher right wmSTSbanks volume than the sNCs (0.62 ± 0.13 vs. 0.59 ± 0.09, Figure 4c, see details in Experiment S3).

| DISCUSSION
SCD is recognized as a high-risk status of AD with a normal cognitive level, which is earlier than the stage of MCI (Buckley et al., 2016;Mitchell et al., 2014;Wolfsgruber et al., 2015). In the current study, instead of predicting MCI progression, we focus on the detection of possible AD at an even earlier stage, that is, predicting whether SCD converts to MCI. We were able to do so because of the CLAS dataset that focused on the risky population at an even earlier stage compared to MCI and had an extremely long follow-up time (7 years). MoCA, montreal cognitive assessment; pNC, progressive normal control; pSCD, progressive subjective cognitive decline; sMCI, stable mild cognitive impairment; sNC, stable normal control; sSCD, stable subjective cognitive decline baseline gray-matter and white-matter volumetric data based on sMRI. Our results indicated that incident stroke, fewer years of education, lower baseline MoCA score, smaller left amygdala, and larger white matter at the banks of right superior temporal sulcus jointly favored amnestic MCI progression. We also showed that with a stateof-the-art, data-driven feature selection and classification, the sSCD could be individually separated from the pSCD automatically and objectively with satisfactory accuracy. We then proposed a clinically feasible, much-simplified prediction model with only five (three clinical and two sMRI-derived) baseline features with which an AUC of 0.8 was reached. The strengths of this study include an unprecedently long-term follow-up, a targeted risk population (SCD) from wellcharacterized community samples, an advanced multivariate pattern recognition algorithm that jointly identified the associations among the clinical and brain MRI features toward an effective feature set, and a demonstrated feasibility of early computer-aided diagnosis of pSCD individuals.
While it is important to identify which individuals with SCD eventually develop clinically significant cognitive impairment years later, long term follow-up is crucial because it might take almost a decade for a subject with "compensatory normal cognition" to change from SCD to MCI (Molinuevo et al., 2017), especially for community-based cohorts (Snitz et al., 2018). However, to our best knowledge, only four SCD follow-up studies had investigated the link between baseline brain biomarkers and incident clinical progression, but all of them only followed up for 2-5 years (Dubois et al., 2018;Lim et al., 2019;Tijms et al., 2018;Verfaillie et al., 2016). This might not be sufficient to fully investigate the conversion from SCD to MCI. One study only identified four out of 318 SCD individuals who progressed to prodromal AD within 30 months (Dubois et al., 2018), while another study reported that community-based SCD subjects confer a much lower risk of progression to MCI over 3 years compared to clinical cohorts (Snitz et al., 2018). Not only cutting off sample sizes, insufficient follow-up time could also misidentify many SCDs who could potentially convert to MCI, which could result in a highly mixed and heterogenous sSCD group, further reducing statistical power. Furthermore, compared to the subject recruitment from clinical sites, those recruited from communities can empower the AD early detection by enabling even earlier detection, even for those who have not developed cognitive impairment yet (on contrast, clinical cohorts could have more or less developed symptoms already at the baseline). Our study is an SCD conversion study with a long (>5 years) follow-up, with advanced feature searching strategies from comprehensively engineered features to identify pivotal factors that could provide early predictive value.
Our most important finding is the early enlargement of the right wmSTSbanks in the pSCD vs. sSCD. This feature characterizes the relative WM volume beneath the cortical part of the superior temporal sulcus (STS). STS is a "chameleon" in the human brain and a high-order association cortex that receives connections from different sensory modalities (Hein & Knight, 2008). It is believed to be involved in diverse cognitive functions, such as audiovisual integration, as well as motion, speech, and face processing (Hein & Knight, 2008). The WM under the STS may play a crucial role in efficient information transmission to support the multifunctionality of the STS. This is further proved by our fiber tracking result, which shows dense fiber connections with lateral and medial temporal cortices, an area for multimodal information integration and the language-and memory-related cognitive functions (Ishibashi, Lambon Ralph, Saito, & Pobric, 2011). Meanwhile, STS was consistently reported to be affected by AD neuropathology (e.g., neurofibrillary tangles and neuronal loss) in a very early stage. The neural loss in the STS was found to be correlated with duration of illness, ranging from no measurable loss in the early stage (with a duration of less than 1 year) to more than 75% neuron reductions in the severe AD stage (Gomez-Isla et al., 1997). Another study revealed increased regional cerebral blood flow (rCBF) along the default mode network regions in SCD compared to NCs . These areas are either closed to or include the wmSTSbanks, further supporting our compensation hypothesis.
Another potential reason for the greater WM underneath the STS could be gray-matter atrophy in the adjacent areas. Our study also found decreased superior temporal thickness in pSCD, which may be related to the adjacent increased WM and the early stage of AD (Table S5). A previous study reported decreased gray matter at the banks of STS in the MCI subjects who converted to AD within 3 years (Killiany et al., 2000). A very recent PET study with ADNI data also found that cognitively normal individuals with a high Aβ burden in the cortical region of the banks of STS were at an increased risk of cognitive decline (Guo et al., 2020). Of note, we used a ratio of wmSTSbanks volume over the total white matter volume as it is a conventional way to exclude confounding effects (e.g., global WM atrophy) (Nordenskjold et al., 2013), there could be a possibility that it is the shrunk total white matter volume that led to such a result. However, we found that it is less likely because we found that the pSCD still showed a larger volume than sSCD even without conducting such a global normalization (Table S5).
Another contributive sMRI feature is the decreased baseline volume of the left amygdala in pSCD compared to sSCD. Amygdala atrophy in SCD was consistently reported (Hu et al., 2019;Striepens et al., 2010;Yue et al., 2018), especially by a recent study that revealed a smaller left amygdala in the SCDs with CSF biomarker (Aβ 42) compared to those without (Hu et al., 2019). Another long-term follow-up study also suggests that atrophy of the amygdala in cognitively intact elderly people could predict dementia occurred 6 years later (den Heijer et al., 2006). Also, from the cognitively intact elderly cohort, Tondelli et al. (2012) reported the reduced amygdala volume was detectable at least 4 years before any cognitive symptoms. Moreover, the decreased amygdala is also be associated with TDP43-related dementia (Makkinejad et al., 2019). Hence, atrophy in this area in SCD may be an early marker for future cognitive decline.
It is noted that we only found a relatively weak statistical difference (p = .011) in the amygdala by statistical analysis, which means that this region could be undetectable from a large number of candidate features if using traditional mass-univariate analysis. Instead, with advanced feature selection and multivariate pattern analysis in a unified framework of machine learning, we were able to detect such a contributive sMRI feature together with others.
While the pSCD subjects had a smaller baseline amygdala compared to sSCDs, we found a negative correlation between the baseline amygdala volume and the MoCA scores at the follow-up across the pSCD subjects while no such correlation was found from sSCD. This indicates that, for pSCD, a larger amygdala could be associated with worse cognitive outcomes. Such results seem contradictory as positive correlation was originally anticipated and because sSCD (with good outcomes) was found to have a larger amygdala (Soldan et al., 2015;Tang, Varma, Miller, & Carlson, 2017) compared to pSCD. However, as we found that a larger amygdala in pSCD predicted worse outcome, we think that the negative correlation could support a well-known hypothesis that higher cognitive reservation (as reflected by a larger amygdala at the baseline) is associated with a more rapid decline later (and therefore has much worse cognitive outcomes afterward) for the risky populations (Amieva et al., 2014). We further came into an updated hypothesis, that is, as long as the baseline amygdala size is below average, those with a larger amygdala may cover up baseline symptoms and, once they become clinically significant, the subjects could have more rapidly declined cognitive abilities. Based on this hypothesis, for the elderly who has objectively normal cognition and subjectively cognitive complaint with good cognitive reservations and enlarged amygdala, it is suggested to take the biomarker diagnosis (such as CSF or PET) for early AD intervention. Further study with such biomarker diagnostic data is required to validate this hypothesis.
In addition to the sMRI features, the study also revealed three contributive clinical features in distinguishing the cognitive outcome of SCDs, including baseline MoCA, years of education, and history of stroke, all of which have been reported as indicators of poor cognitive prognosis in the previous studies (Amieva et al., 2014;Blom et al., 2019;Nasreddine et al., 2005). Again, traditional statistical analysis could not reveal significant differences in MoCA and stroke history (Table S4). However, this does not contradict our findings, as our prediction model considered these features jointly instead of independently. Our method is quite effective, which is also because we utilized an effective feature selection algorithm using Relief, an algorithm that takes relationships among features into consideration. Furthermore, we derived a simplified model with only five features, all of which can be obtained at the baseline. This renders our model's merit of predicting SCD conversion at the baseline 7 years before noticeable cognitive impairment, suitable for large-cohort screening in the future.
However, this study also has limitations. First, our study does not consist of annual follow-ups, which may be helpful for timely identifying SCD's conversion to MCI. Second, although we followed up on our SCD subjects for a long time compared to the previous studies, it is still likely that more SCD will progress to MCI at a later time (Kaup, Nettiksimmons, LeBlanc, & Yaffe, 2015). Third, our data does not include FDG-PET, Aβ marker, or APOE genotype, making it unclear whether the progression was due to AD or not. Last but not least, our sample size is fairly small and the result needs to be validated in future large-sample studies.
Nevertheless, this is the first sMRI study of SCD subjects with unprecedentedly long follow-up time. We, with advanced machine learning from the comprehensively and thoroughly engineered feature sets, identified key baseline early AD markers comprising both clinical and neuroimaging features with an individualized SCD progression prediction task. We further proposed a much simpler prediction model with only five features and a satisfactory predictive accuracy. Our results provide a potentially feasible and objective early diagnostic tool and help a better understanding of the pathology of AD in its prodromal phase.