The existence of such a group of individuals at risk of developing AD has encouraged research on diagnostic tools that could discriminate those who will progress to AD. Because of the demands of clinical settings, such discriminative tests ought to be relatively inexpensive, fairly quick to administer and opaque to other factors that can affect cognitive functions such as depression [6,7]. Memory loss being the core symptom of AD, research has focused largely on neuropsychological tests that assess this symptom. However, memory loss can be easily confounded with encoding and attention deficits seen in other conditions such as depression [8-10]. More sophisticated memory tests may need to be administered by neuropsychologists. At the same time, more studies show that episodic memory deficits might not be the only feature of mild AD [11]. For example, studies have found subtle visuoperceptual deficits in AD, particularly changes in word perceptual threshold [7]. Perceptual tests, being for less attention demanding, are likely to be less influenced by depression – a major advantage for an AD tests. One explanation of the deficits observed in word-recognition tests could be the degeneration of the “word-form area” in the occipital-temporal extrastriate cortex [12] also affected, albeit to a lesser degree than memory areas, in AD [13,14]. Another possible explanation could be that patients with AD show poorer performance on more demanding perceptual tasks because they have slower central processing abilities. For instance, a masked word recognition test is more demanding on the central processing abilities than an unmasked letter recognition test, and therefore, the former test would be more sensitive in the earlier disease stages.

Detailed procedures for the WRT assessment are described in Massoud et al. [15]. In brief, four to six letters nouns are presented on a computer screen, matched across blocks for frequency. These nouns are presented in blocks of ten words, each block with an increasing presentation time. Forward and backward masking consists of a series of number signs of the same size and length as the target words. Subjects are asked to read each word out loud. The percentage of the words in a block that were correctly read is calculated. The WRT score is defined as the presentation duration at which 50% of the words in a block are read correctly. Figure ​Figure11 depicts the presentation of the stimuli on the screen.

To calculate the sensitivity and the specificity of the WRT at the suggested cut-off score of 85 ms, we compared NC and AD patients as follow. Sensitivity is TP/(TP + FN) and Specificity is TN/(TN + FP), where TP is the number of true positives; FP, the number of false positives; FN, the number of false negatives; and TN, the number of true negatives. We also carried a Receiver-Operating Characteristic (ROC) analysis to evaluate the performance of WRT in classifying appropriately NC and AD, and MCIp and MCInp.

Finally, to verify whether the WRT is opaque to other cognitive domains, we correlated the WRT scores with tests of memory, language and attention. We tested the subjects on Logical Memory (LM) I and II subscales and Visual Reproduction (VR) I and II subscales of the Weschler Memory Scale-Revised (WMS-R [18]), Rey Auditory-Verbal Learning Test (RAVLT [19]), Boston Naming Test (BNT [20]), Controlled Word Association Test (COWA [21]), the Trail Making Test, subscale A (TRAIL [22]), and Stroop Test [23].

We found significant differences in the groups’ performance on WRT (F_(2,119) = 12.7, p < 0.01). Bonferroni’s post-hoc analyses revealed that significant differences existed between the three groups (p < 0.03). Figure ​Figure22 displays individual WRT scores in each group. Using the cut-off determined in Massoud’s study (85 ms), we found a diagnostic sensitivity of 72% and a specificity of 76%. ROC analyses showed that WRT differentiate between NC and AD about 81% of the time (p < 0.001). However, the test does not perform better than chance alone at differentiating MCIp and MCInp (Area Under the Curve, AUC = 45%, p = 0.51). Indeed, WRT test correctly classified 6 out of the 19 MCI who had progressed to AD after 8 years of follow-up. Logical Memory (LM) delayed and Ray Auditory Verbal Learning Test (RAVLT) delayed, both tests assessing delayed verbal memory, were better at discriminating between MCI subjects who progressed (MCIp) and those who did not (MCInp). The AUC of ROC analyses comparing the neuropsychological tests between MCIp and MCInp are represented in Table ​Table11 and Figure ​Figure33.

We found significant correlations, after Bonferroni’s corrections for multiple comparisons, with memory as measured with Visual Reproduction (VR) II (r = −0.32, p = 0.003) and RAVLT delayed (r = −0.35, p = 0.001). Scores on WRT also correlated with language tests scores, as measured by Boston Naming Test (BNT; r = −0.32, p = 0.002), Controlled Word Association (COWA) lexical subtest (r = −0.35, p = 0.001) and semantic subtest (r = −0.34, p = 0.001). However, these correlations might have been spurious, as they did not remain statistically significant when the correlations were performed in the diagnostic groups independently, with the exception of COWA lexical subtest in the NC (p = 0.002). WRT was found to correlate with tests of attention, such as the Trail Making subscale A (TRAIL; r = 0.48, p < 0.001), Stroop dots (r = 0.39, p < 0.001). The correlation between TRAIL and WRT remained significant in NC (r = 0.623, p = 0.0002), whereas, in the MCInp, the correlation was marginally significant (p = 0.06). Also, one individual with AD had a TRAIL score of 800 ms; that is, 5 standard deviations below the average score of AD on that test. This poor performance was not explained by visual impairment and was therefore treated as an outlier. Taking that subject out, revealed a significant positive correlation (p = 0.0002). The correlation between STROOP dots and WRT remained significant in the NC and MCInp. Table ​Table22 summarises the correlation coefficients for each tests with WRT in each groups.

It is worth noting that we found slightly bigger standard deviations in the NC and MCI groups, and smaller ones in the AD group than what was previously reported, indicating that our groups might have been less homogenous. Examining the individuals scores from Massoud’s study [15] and our own replication study illustrates this difference clearly (Figure ​(Figure2).2). In the pilot study, only 1 out of 12 NC had a WRT score greater than the 85 ms, compared to 8 out of 33 in our follow-up study. Similarly, in Massoud’s study, only 3 out of 13 AD patients scored below the cut-off, compared with 8 out of 25 in the current study. We simply failed to maintain this striking separation between NC and AD patients in a larger cohort. Even though the age and dementia severity, measured by MMSE, of the groups were similar in the two studies, we must conclude that the pilot study, being smaller, simply showed less heterogeneity than was found in our replication study.

The ROC analyses showed that WRT does not perform better than chance level in classifying MCI individuals who will progress to AD or not (AUC = 0.45, p > 0.1). Indeed, in the current study, WRT only correctly classified 32% of the MCI individuals who progressed to dementia, which is less impressive than what Massoud and his colleagues had found, where all of the MCI subjects who had scored ≥ 85 ms on WRT had progressed to AD after 2 years of follow-up. It is important to note that the follow-up period in our study was of 8 years. Massoud’s study included 4 individuals with MCI who progressed to AD, which is within the expected published rates of progression to AD (15% a year). The current study, in contrast, yielded a larger sample of converters. However, the average rate of progression from MCI to AD in our study (about 4% a year) is lower than in most studies [17]. In addition, examination of the AUC for other neuropsychological tests reveals that the delayed recall of Logical Memory and Ray Auditory Verbal Learning Tests were better at discriminating MCIp from MCInp (Table ​(Table11).

Massoud and colleagues found that WRT scores correlated with scores on a global cognitive assessment test, the MMSE. In our study, we replicated this finding, but we wanted to evaluate whether WRT scores were related to those of other cognitive tests. We found evidence for an association between WRT scores and measures of memory and language. These associations might be spurious, however, as they were found only when the three groups were collapsed, with the exception of the lexical subtest of the Controlled Word Association test in the NC group. WRT scores were also correlated with measures of attention, more precisely Trail Making Test A and Stroop dots subtest. This suggests that performance on WRT is related to overall mental speed and attention processing. However, we found that the association with Trail Making Test A was maintained in the NC, MCInp, and AD groups, but not in the MCIp group. Similarly, the association with Stroop dots remained significant only in the NC and MCInp groups. Although explaining the absence of correlation in the MCIp or AD groups is beyond the scope of this study, it might suggest that at the MCI stage of AD pathology different cognitive domains are affected to different degrees. As for the absence of correlation between MMSE and WRT scores in the NC and MCI groups, it can easily be explained by ceiling effect. In the NC and MCI groups, MMSE scores do not vary enough, whereas in the AD group, MMSE scores range from 11 to 29.