Abstract
Our previous studies have shown that longitudinal reduction in retinal nerve fiber layer (RNFL) thickness is associated with cognitive deterioration. However, whether the combination of longitudinal reduction in RNFL thickness with baseline episodic memory performance can better predict cognitive deterioration remains unknown. Therefore, we set out to re-analyze the data obtained from our previous studies with 78 elderly adults (mean age 74.4 ± 3.83 years, 48.7% male) in the community over a 25-month period. The participants were categorized as either stable participants whose cognitive status did not change (n = 60) or converted participants whose cognitive status deteriorated (n = 18). A logistic regression analysis was applied to determine a conversion score for predicting the cognitive deterioration in the participants. We found that the area under the receiver operating characteristic curve (AUC) for the multivariable model was 0.854 (95% CI 0.762–0.947) using baseline story recall as a predictor, but the AUC increased to 0.915 (95% CI 0.849–0.981) with the addition of the longitudinal reduction of RNFL thickness in the inferior quadrant. The conversion score was significantly higher for the converted participants than the stable participants (0.59 ± 0.30 versus 0.12 ± 0.19, p < 0.001). Finally, the optimal cutoff value of the conversion score (0.134) was determined by the analysis of receiver operating characteristic curve, and this conversion score generated a sensitivity of 0.944 and a specificity of 0.767 in predicting the cognitive deterioration. These findings have established a system to perform a larger scale study to further test whether the longitudinal reduction in RNFL thickness could serve as a biomarker of Alzheimer’s disease.
Keywords
INTRODUCTION
Alzheimer’s disease (AD) is one of the greatest public health problems in the world, and it begins with a long, asymptomatic period (pre-clinical stage of AD), when only its neuropathogenesis progresses [1]. The criteria to define AD have been revised by introducing a new diagnostic guideline that underscores the concept that AD represents a continuum [2–4]. These criteria emphasize the importance of in vivo biological evidence, which could be demonstrated by biomarkers, in the early diagnosis of AD.
A loss of retinal ganglion cells and optic nerve degeneration have both been established as windows for detecting neuropathology, and have also been suggested as potential biomarkers of neurodegenerative diseases including AD [5–10]. Our previous studies have shown that longitudinal changes in retinal nerve fiber layer (RNFL) thickness are associated with the cognitive deterioration in mild cognitive impairment (MCI) and AD patients [11]. Specifically, we found that the reduction in RNFL thickness over time (e.g., 25 months) is associated with the conversion from normal cognitive function to MCI, or from MCI to dementia in the participants [12, 13]. However, utilizing the longitudinal reduction in RNFL thickness and the baseline episodic memory measure together to predict cognitive deterioration has yet to be attempted. Therefore, we set out to re-analyze the data obtained from our previous studies [12, 13] in order to establish multivariable logistic regression models and to calculate a conversion score to predict the cognitive deterioration in the participants of our previous studies. The hypotheses of the current studies were: (1) the logistic regression model combined with the longitudinal RNFL thickness reduction and the baseline episodic memory impairment would better predict the cognitive deterioration; and (2) there would be an optimal conversion score (obtained from the multivariable logistic regression model) for predicting the cognitive deterioration.
MATERIALS AND METHODS
Participants in the studies
This study was approved by the Human Research Ethics Board of an affiliated Tongji Hospital of Tongji University in Shanghai, P. R. China [LL (H)-09-04]. The baseline assessments of cognitive function were conducted from October 2010 to November 2010, and the follow-up assessments of cognitive function were conducted from November 2012 to December 2012 (a 25-month follow-up study). All participants signed the written informed consent before being enrolled in the study. The participants were recruited from three communities in Shanghai, P.R. China via a flyer distribution. More details related to data collection and the eligibility and exclusion criteria have been introduced in our previous studies [12, 13].
Optical coherence tomography (OCT) evaluation
OCT examinations were conducted on the same days as the cognitive assessments, as described in our previous studies [11]. Briefly, a default Optic Disc Cube 200 × 200 protocol (software version 4.6) was used to determine the RNFL thickness. Layer-seeking algorithms found the RNFL inner (anterior) boundary and RNFL outer (posterior) boundary for the entire cube, except for the optic disc. A scan was saved only if the fundus image was sufficiently visible to distinguish the optic disc and the scanning circle, and if there were no obvious movement artifacts with missing data at the acquired scan pattern. Images with eye movements during scans, poor focus, low analysis confidence, or signal strength less than 4/10 were excluded. RNFL thickness was measured three times per quadrant using repeat scan protocols and the average of the 12 values was used for each eye. The RNFL thickness (global, superior, inferior, nasal, and temporal quadrant) in each participant was the average from both eyes, or from one eye, if the data from the other eye were lost.
Neuropsychological assessment
The Mini-Mental State Examination (MMSE) was a brief measure for global cognitive screening, and the Chinese version of the MMSE has proven to have good validity and reliability within the Chinese population [14, 15]. The Chinese version of Activities of Daily Living Scale (ADL) [16], composed of a Physical Self-Maintenance Scale (PSMS; six items) and an Instrumental ADL (IADL; eight items), was administered to evaluate individuals’ daily living functions. The Chinese version of Repeatable Battery for the Assessment of Neuropsychological Status (RBANS, Form A) was used to evaluate specific cognitive domains of the participants, including Immediate Memory, Visuospatial/Constructional, Language, Attention, and Delayed Memory. The reliability and validity of RBANS in the Chinese elderly community has been validated [17]. All neuropsychological assessments were conducted by trained research assistants according to the protocol, as described in our previous studies [11].
According to Petersen’s definition [18], participants were diagnosed with MCI if they have (1) self-reported memory problems for over 3 months, (2) normal function of daily living: the Chinese version of ADL ≤15 [16], (3) normal general cognitive function but abnormal memory functioning evidenced by lower score of MMSE than education-dependent cutoff points (20/21 for 6 years of education or less, and 24/25 for more than 6 years of education) [15]. Dementia was determined according to the DSM-IV criteria [19]. The diagnosis of MCI or dementia was made only if two psychiatrists confirmed the diagnosis. Participants were defined as “stable” if they did not change from normal cognition to MCI (according to Petersen’s definition [18]) or from MCI to AD dementia (according to DSM-IV criteria [19]) at the follow-up assessment.
Statistical analysis
A One-Sample Kolmogorov-Smirnov method was first used to test the normality of all variables. Non-parametric Mann-Whitney U test was used to compare the difference in the score of MMSE between two groups because the data were not normally distributed (determined by One-Sample Kolmogorov-Smirnov). Student’s t-test was used to compare the other cognitive variables including score of RBANS, continuous demographic characteristics (e.g., age), and retinal parameters, because these data were normally distributed. Chi-square test was used for comparing dichotomous factors (e.g., gender).
A univariate logistic regression was performed firstly to screen possible risk factors, including demographic variables (age, gender), clinical variables (blood glucose, cholesterol, and triglyceride), reduction of RNFL thickness and cognitive variables at baseline assessment. A multivariable regression model was then built to determine relevant predictors of cognitive deterioration [20]. The factors obtained from univariate regression, which showed significant associations with cognitive deterioration (p < 0.05), were chosen to enter into the multivariable model. Because education influenced cognitive performance significantly, it was adjusted as a confounder apart from age and gender in both the univariate and multivariable regression models. Regression models containing different risk factors were then established. The overall model-fit was determined by Hosmer-Lemeshow goodness-of-fit tests. A model would be defined as a good one if the subjects who experienced the main binary outcome (e.g., converted from normal cognition to MCI) mostly fell within the higher risk criteria (p > 0.05) [20].
The predictive probability of each model was evaluated by the area under the curve (AUC) of a receiver operating characteristic (ROC) analysis. The probability of cognitive conversion (P) in each subject was calculated. Estimated probability (P) was calculated by the following equation: ln[P/(1–P)] = β0 + β1X1 + β2X2 + … … + β n X n , in which “X” is the influence factor; “n” is the number of influence factors; “β” is the influence coefficient, indicating how strongly the corresponding independent variables contribute to the outcome, and “β0” is a constant [21]. The conversion score equation was: (P) = 1/[1 + e–(β0 +β1X1β2X2 + … … +βnXn) ] [22]. In the final model, constructed with the significant variables in the previous analysis, and adjusted for age, gender and education, an optimal cut-off point according to the Youden index [23] (maximum of [sensitivity +specificity –1]) was obtained.
All analyses were performed using SPSS version 20.0 (SPSS Inc., Chicago, IL), with p < 0.05 as the significance level.
RESULTS
Demographic characteristics
One hundred and four participants were enrolled initially, 82 of whom had normal cognition and 22 of whom had MCI. Among the 82 cognitively normal participants at baseline, 19 participants dropped out from the follow-up assessment, which occurred 25 months after the initial enrollment. Two of the 22 MCI patients dropped out from the follow-up assessment as well. Thus, 83 participants were assessed 25 months after the enrollment. Three of the 83 participants were excluded because they were not able to provide a clear image of RNFL in both eyes by OCT measurements. Another two of the 83 participants were diagnosed with cancer before the final assessment. Therefore, 78 participants (58 of them with normal cognition and 20 of them with MCI at baseline) were included in the final data analysis. The detailed information related to the participants is described in Table 1 and also in our previous studies [12, 13].
At the time of the re-assessment (25 months after the initial baseline assessment), 50 of the 58 participants who had been cognitively normal at the baseline remained cognitively normal, and 8 of the 58 participants converted to MCI. Among the 20 participants who had MCI at the baseline, 10 of them still had MCI and 10 of them converted to mild (9) or moderate dementia at the re-assessment. Therefore, the 18 participants who converted from normal cognition to MCI or from MCI to dementia were defined as “converted participants”. The other 60 participants, who neither converted from normal cognition to MCI, nor from MCI to dementia, were defined as “stable participants”. After 25 months, the MMSE score of converted participants declined to 24.00 ± 3.00 [median ± interquartile range (IQR)], which was significantly lower than that of stable participants (28.00 ± 2.00, p < 0.001, Mann-Whitney U test). Moreover, the reduction of MMSE score of converted participants was greater than that of stable participants (–3.50 ± 5.00 versus 0.00 ± 2.00, p = 0.001).
The multivariable logistic regression models to predict cognitive deterioration
The univariate regression showed that only the reduction of RNFL thickness in the inferior quadrant (p = 0.004), but not in other quadrants, was associated with cognitive deterioration. This result is consistent with our previous findings [12, 13]. Among all cognitive variables, score of story recall (r = –0.343), word list learning (r = –0.079), and immediate memory (r = –0.040) at baseline assessment were associated with cognitive deterioration, among which story recall showed most significantly independent-effect on the risk of cognitive deterioration than other cognitive variables. Therefore, we chose RNFL_inferior reduction and baseline score of story recall as potential variables for the establishment of the regression model. We employed three multivariable logistic regression models to determine the best one for the prediction of cognitive deterioration. Model 1 contained the variable of “story recall (score)” at baseline as a predictor, model 2 contained the longitudinal “RNFL_inferior reduction” at 25 months after the baseline as a predictor, and model 3 combined both the baseline “story recall” and the longitudinal “RNFL_inferior reduction” as predictors (Table 2). We also used age, gender, and education (adjusted) as the covariates in all of these models. The model-fit was tested by Hosmer-Lemeshow goodness-of-fit tests. The fitnesses of models 1 and 2 were calculated as χ 2 (8) = 7.174 (p = 0.518) and χ2(8) = 3.557 (p = 0.895), respectively. The fitness of model 3 was calculated as χ2(8) = 2.567 (p = 0.959). These results suggest that all three of the regression models could be used to predict the cognitive conversion. However, model 3, the RNFL_episodic memory model, which combined baseline “story recall” and longitudinal “RNFL_inferior reduction” at 25 months after the baseline, was determined to be the best model of the three for predicting cognitive deterioration in the participants, as evidenced by the fact that the p value in model 3 (p = 0.959) was larger than those in model 1 (p = 0.518) and model 2 (p = 0.895). We also used a ROC analysis to determine the AUC for the three models. We found that the AUC of the regression model 1 and model 2 were 0.854 [95% confidence interval (CI) 0.762–0.947, p < 0.001] and 0.889 (95% CI 0.809–0.969, p < 0.001), respectively. The RNFL_episodic memory model (model 3), which combined baseline “story recall” and longitudinal “RNFL_inferior reduction”, had an AUC up to 0.915 (95% CI 0.849–0.981, p < 0.001) (Fig. 1). These data further suggest that model 3, the RNFL_episodic memory model, is the best one for predicting cognitive deterioration.
The conversion score to predict cognitive deterioration
Next, we generated a formula, based on model 3, to estimate the possibility of cognitive deterioration and generated a conversion score: ln[P/(1 – P)] = 0.241 *Age (year) + 0.811 *Gender – 0.293 *Education (year) – 0.324 *story recall_baseline (score) + 0.137 *RNFL_inferior reduction (μm) – 16.547. Based on this formula, the average conversion score of the converted participants was higher than that of the stable participants: 0.59 ± 0.30 versus 0.12 ± 0.19 (p < 0.001, Mann-Whitney U test) (Fig. 2).
An optimal cut-off point was determined by Youden index (maximum of [sensitivity + specificity –1]). It reflects the best balance of sensitivity and specificity. Thus, we employed a ROC analysis to determine an optimal cutoff of the conversion score for predicting the cognitive deterioration in the participants. Using the optimal cutoff (0.134) obtained from ROC analysis, a sensitivity of 0.944 and specificity of 0.767 were found in the prediction of cognitive deterioration. We then compared the incidence of higher conversion score (over 0.134) of stable and converted participants, which showed a significant difference between stable (14/60, 23.3%) and converted (17/18, 94.4%) participants (χ2 = 26.343, p < 0.001). These results verified the cutoff value of the conversion score and testified to its sensitivity and specificity in identifying the individuals at a higher risk of developing cognitive deterioration.
DISCUSSION
The well-recognized biomarkers, which may serve for the prediction of the cognitive conversion to MCI or AD, include amyloid-β protein and tau protein in human cerebrospinal fluid [24], region-specific 18F-fludeoxyglucose uptake (using positron emission tomography) [25], and hippocampus volume (using magnetic resonance imaging) [26]. However, the utilization of these biomarkers is limited, owing to the facts that they are either invasive or they are expensive. In the present study, we found that the combination of the longitudinal reduction of RNFL thickness with the baseline episodic memory impairment test could better predict cognitive deterioration. The conversion score deriving from these two markers could quantify the probability of suffering cognitive conversion.
When the baseline episodic memory (story recall) test and longitudinal reduction of RNFL thickness measure (model 3, Fig. 1) were combined, the AUC of the multivariable model increased to 0.915, which was better than employing the episodic memory test alone (AUC of 0.854, model 1, Fig. 1) or the reduction of RNFL thickness alone (AUC of 0.889, model 2, Fig. 1).
Then, we obtained a conversion score, using the method described by Arbizu et al. [22], to predict the cognitive deterioration for individual participants. We found that the average conversion score of the converted participants was higher than that of the stable participants: 0.59 ± 0.30 versus 0.12 ± 0.19 (p < 0.001, Mann-Whitney U test) (Fig. 2). Finally, we found that 0.134 was the optimal cutoff cognitive score to predict cognitive deterioration in the participants. This cognitive score could generate a sensitivity of 0.944 and a specificity of 0.767 for the prediction of cognitive deterioration in the participants. A majority of the converted participants (17/18, 94.4%) had converted scores greater than 0.134, whereas only 23.3% (14/60) of the stable participants had a conversion score greater than 0.134.
These results verified the ability of the conversion score for identifying the individuals who would have a higher risk of developing cognitive deterioration, e.g., from a normal cognition to MCI or from MCI to AD dementia. Taken together, these findings suggest that we can use the conversion score, calculated from the longitudinal reduction of RNFL thickness and the baseline impairment of episodic memory, as a quantitative measurement to facilitate the diagnosis of MCI/AD and potential interventions in future clinical trials.
Different biomarkers, or combinations of biomarkers, have recently been studied with the aim of predicting cognitive conversion, either from normal cognitions to MCI or from MCI to AD [24–26]. Predictors associated with conversion have been defined as those biomarkers that probably reflect disease severity or the probability of an individual undergoing clinical transition [27]. Consistently, our data show that the combination of baseline episodic memory impairment scores with the longitudinal reduction in RNFL thickness measure could provide an optimal way of predicting cognitive deterioration in patients.
Unlike the cerebrospinal fluid biomarker, or brain imaging technologies, the RNFL thickness measurement by OCT is a relatively inexpensive, noninvasive, reproducible, rapid, and well-tolerated method to assess the clinically relevant processes of cognitive deterioration. Longitudinal reduction of RNFL thickness is more reliable than cross-sectional values in reflecting the progress of neurodegeneration in MCI or AD patients [12, 13]. Furthermore, our current studies show that the inclusion of baseline episodic memory scores would enhance the power of longitudinal reduction of RNFL thickness in predicting cognitive deterioration. Finally, a conversion score generated from longitudinal reduction of RNFL and episodic memory impairment scores may prove to be a useful tool to quantitatively predict the risk of developing cognitive deterioration in patients. Pending further investigations, the conversion score would be significant for more accurately characterizing MCI and AD patients with different degrees of risk for further cognitive deterioration.
Our studies have several limitations. The first and foremost limitation is, as a re-analysis of data, the post-hoc nature might limit the value of findings from this study. Nevertheless, we have established a method using conversion score to quantify the risk of cognitive deterioration, which was not reported before. In the future studies, a validation cohort will be included to verify the findings from the current studies. Second, there might be other potential confounders to influence the quantitative score derived from the regression model apart from factors being adjusted (e.g., age or gender), which may lead to an alpha error. Confirmatory investigation with larger sample size and longer term of observation should be conducted to improve the current model. Finally, the participants of the current study were only old adults over 70 years. For old adults less than 70 years of age, the predictive value for cognitive conversion using this model remains unknown. Further studies including participants with wider age spectrum are warranted to explore the feasibility using cutoff score to predict cognitive conversion in different age groups.
In conclusion, we found that employing a combination of the longitudinal reduction of RNFL thickness measure and the baseline episodic memory performance could provide a better method for predicting cognitive deterioration than the employment of the longitudinal reduction of RNFL thickness or baseline episodic memory performance alone. Finally, we calculated and obtained a conversion score, based on the combination of the longitudinal reduction of RNFL thickness measure and baseline episodic memory performance, to predict the cognitive deterioration in the participants. Pending further investigations, the findings and methods established in the current studies could serve as useful tools for future research into the cognitive deterioration of MCI and AD patients, and also for the development of MCI and AD dementia interventions in clinical trials.
