To evaluate the capacity of MRI to improve Alzheimer's disease early diagnosis at the individual level, we designed a fast, automated method for assessing whether the MRI measurement of cortical thickness could predict progression to Alzheimer's disease from its prodromal stage at individual level. Applying our cortical thickness based Normalized Index (NTI) to a large sample of patients from the ADNI database, we found evidence that structural MRI can be used to detect subtle structural changes that help predict the subject's outcome up to 24 months before clinical diagnosis. We also investigated the cognitive reserve hypothesis on our population and showed that subjects with a higher education level had a significantly thinner cortex than less educated subjects with the same level of cognitive performance. For these subjects, structural imaging may thus appear more sensitive to the incipient disease than cognitive testing.
The preliminary validation process we used showed that our methodology was reliable. The cortical values measured in this study are consistent with reported measurements of cortical thickness, in particular regarding the significant effects of age and ApoE genotype, or the non-significant effect of gender, as well as the differences in cortical thickness regional patterns observed between healthy controls, amnestic MCI and Alzheimer's disease patients (Salat et al
; Lerch and Pruessner, 2005; Singh et al
; Shaw et al
; Teipel et al
). Interestingly, our results in terms of regional atrophy between healthy controls, amnestic MCI and Alzheimer's disease patients are very similar to those of other studies based on the ADNI data (Fan et al
; Hua et al
). The NTI may also be considered as an accurate and robust index of Alzheimer's disease, since it yielded a high cross-validated accuracy (85%) for distinguishing Alzheimer's disease patients from healthy controls on a set of subjects that was distinct from the one used for the NTI computation.
For the construction of the NTI, we retained the right posterior cingulate, the right medial temporal zone and the left lateral temporal zone as the optimal combination of zones. However, this choice of zones is not critical regarding the predictive accuracy, since numerous combinations can lead to the same accuracy as illustrated in the supplementary Table
. This can be explained by the fact that widely spread cortical atrophy was observed in the progressive MCI population, which may indicate that these subjects were already at an advanced stage of the disease from an anatomical point of view, despite the fact that they had still not been clinically diagnosed as Alzheimer's disease patients. Had these patients been at an earlier stage of the disease, a more focal atrophy might have been observed, resulting in a smaller subset of discriminative zones and therefore a smaller number of optimal combinations. Further studies on other amnestic MCI populations (or possibly on a preclinical population) are still required to analyse the respective predictive accuracies of these combinations at earlier stages of the disease.
We found that the NTI was able to accurately predict evolution from amnestic MCI to Alzheimer's disease up to 24 months before clinical criteria of Alzheimer's disease are met, with a cross-validated predictive value of 76%. Results from the literature give varied assessments of the ability of structural imaging and of cognitive scores (alone or combined) to predict conversion to Alzheimer's disease. Korf et al
) showed that atrophy in the medial temporal lobe could predict conversion to Alzheimer's disease with a global accuracy of 69%. Sarazin et al
) showed that cognitive scores could predict future conversion to Alzheimer's disease with high accuracy (94%), but the age difference alone between their two groups classified 72% of the individuals. Devanand et al
) found that a combination of cognitive scores and hippocampal and entorhinal cortex volumes could predict conversion to Alzheimer's disease with an accuracy of 87.7%; however, age alone correctly classified 71.9% of the subjects, and their follow-up period (5 years in average) was much longer than ours. Visser et al
) showed that a combination of cognitive scores, hippocampal and parahippocampal volumes and visually assessed atrophy of the medial temporal lobe could predict conversion to Alzheimer's disease for 81% of the subjects, but age on its own classified 78% of the subjects. In our study, age was not a relevant predictor, since we showed that it classified subjects nearly randomly (Area under the ROC = 0.52). More importantly, our present study has the crucial advantage of using a cross-validation procedure that guarantees higher validity of the estimation of predictive values than those reported by other studies which tested predictive values within a single sample of subjects. Cross validation, which usually gives lower predictive values than those obtained on the learning sets, is the recommended way of evaluating the predictive accuracy of a given marker at the individual level. Our results may thus better reflect the true predictive value of structural imaging on a 24-month period. With regards to this predictive capacity, it is important to acknowledge, first, that the conversion rate over 2 years in our amnestic MCI sample was very high, reaching 59% (72/122). Such high annual conversion rates of about 27–30% in amnestic MCI subjects recruited within the framework of clinical settings such as the ADNI study have been reported recently, reaching 28% in the study by Schmidtke and Hermeneit (2008
) and 34% in the study by Rozzini et al
). Such high rates are likely to result from the way subjects are recruited and the diagnosis criteria used for the enrolment of amnestic MCI subjects. It is interesting to compare these rates to those derived from population cohorts or community-based studies: the annual conversion rate observed in the population cohort studied by Lopez et al
) was 18% for probable MCI, and the annual rate reported in the community-based study by Fischer et al
) was 19%. Thus, the conversion rate in the ‘population’ of MCI subjects could be considered to be about 20% per year, compared with 30% in MCI samples ‘enriched’ with patients with early Alzheimer's disease as recruited for clinical studies.
The influence of such a difference on the accuracy of the NTI used to predict conversion to Alzheimer's disease in the general population of amnestic MCI can be evaluated by comparing the positive and negative predictive values (PPV and NPV) of the NTI using the Alzheimer's disease prevalence observed at 24 months in our ADNI amnestic MCI group (59%) and in the general amnestic MCI population (~40%) (Fischer et al
; Lopez et al
). The computed positive predictive value on the ADNI sample is 83% compared with 69% in the general amnestic MCI population, whereas the negative predictive value increases from 71% in the ADNI population to 85% in the amnestic MCI population. The values obtained in the general amnestic MCI population supports the notion that the NTI still conveys useful information in clinical practice for the prediction of conversion from the amnestic MCI stage to Alzheimer's disease.
As suggested by Stern (2006
), clinical evaluation alone may prove insufficient for the early diagnosis of Alzheimer's disease. Indeed, it suffers from the confounding effect of the cognitive reserve and cannot directly reveal the underlying pathology. Stern stressed the fact that it would be wise to focus on the development of surrogate markers such as neuroimaging, which may be less affected by the cognitive reserve. Our results strongly support this suggestion: indeed, we showed that cognitive testing, even when combined with education level, could not predict the evolution to Alzheimer's disease as accurately as the NTI. Besides, using education as a proxy for cognitive reserve and the NTI as a proxy for Alzheimer's disease pathological burden, we showed that highly educated subjects had a significantly decreased NTI whereas their MMSE was not affected. This result indicates that the NTI, being less affected by the cognitive reserve than an index of global functioning such as the MMSE, may prove more sensitive for the early diagnosis of Alzheimer's disease, in particular in highly educated subjects. This notion is confirmed by the post hoc
analyses, which revealed no differences between progressive MCI subjects and Alzheimer's disease patients in terms of NTI, but a significant difference in terms of MMSE. We also obtained further consistent results by testing the effect of education level on the timeline of conversion to Alzheimer's disease: progressive MCI subjects who converted to Alzheimer's disease earlier than 12 months after baseline had a significantly lower level of education than those who converted later than 12 months after baseline. This result supports the cognitive reserve hypothesis within the progressive MCI group, since it indicates that individuals with a higher degree of education have less apparent symptoms for a longer time. Moreover, our results show that stable MCI subjects with an NTI negative score had a high education level, even slightly higher than that of the progressive MCI subjects who converted to Alzheimer's disease later than 12 months after baseline. Such a high education level may have prevented them from showing clinical signs of dementia 24 months after baseline, while having an advanced Alzheimer's disease pathology. Study of the clinical status 36 months after baseline will indicate if, as predicted by the NTI, most of these subjects eventually converted to Alzheimer's disease.
We therefore believe that NTI could be very useful to predict conversion to Alzheimer's disease, especially for subjects with a high level of education. The cognitive reserve is particularly confounding for these subjects because it prevents cognitive testing from being able to efficiently detect their underlying disease at an early stage. On the other hand, cognitive reserve seems less confounding with regard to structural changes in these subjects. Therefore, the NTI is able to detect the disease up to 24 months before signs of overt dementia are readily observed.
The cognitive reserve hypothesis has already been investigated for the Alzheimer's disease population in recent studies (Sole-Padulles et al
; Garibotto et al
; Hanyu et al
; Kemppainen et al
), which found that, for a given level of cognitive burden, Alzheimer's disease patients with a higher education level had a lower regional cerebral blood flow (Hanyu et al
), a higher Pittsburgh Compound B uptake (Kemppainen et al
), a lower glucose metabolic rate (Garibotto et al
; Kemppainen et al
) and a lower brain volume (Sole-Padulles et al
) than Alzheimer's disease patients with lower levels of education. Solé Padullés et al
. (2007) extended these results to a small amnestic MCI group, while Garibotto et al
) showed that MCI converters with a higher education level had a relatively lower glucose metabolism. To the best of our knowledge, our study is the first to investigate the relation between cognitive reserve, structural changes and the timeline of evolution to Alzheimer's disease.
The relationships between cortical thickness and education level might be more complex, however, than the cognitive reserve hypothesis suggests. We observed the same effects in healthy controls as in the amnestic MCI group, subjects with a higher degree of education having a lower NTI. Similar results were found by Coffey et al
), who concluded that highly educated healthy subjects managed to remain healthy while having more age-related atrophy. Im et al
) studied the links between cortical thickness, level of education and fractal dimension of the cortical surface for young healthy subjects, and found that high education level was associated with a high fractal dimension of the cortical surface, which itself was associated with small cortical thickness. Their results may lead to the conclusion that a high education level is associated with low cortical thickness, even for young healthy subjects. On the other hand, Solé-Padullés et al
. (2007) found a positive correlation between education level and brain volume on a small population of healthy elderly subjects. It remains unclear whether the effect of education that we observed is related to the pathology at a preclinical phase or whether cortical thickness is directly affected by education level. Future work on younger populations may be required to address this issue.
One limitation of our study comes from the high education level of our population, and the effect of education on people with a low education level (education level ≤10 years) could not really be accounted for. This issue may need further study, including younger healthy subjects. However, the main results of our study were only slightly affected by this limitation since we emphasize that studying anatomical changes may be beneficial, particularly for highly educated subjects.
Another limitation comes from the novelty of our methodological approach. A comparative validation may be required, for example, by comparing our results in terms of predictive value with those of other studies on the ADNI population. However, we have given evidence that our method is reliable, since it had a high predictive value on the Alzheimer's disease population, and an encouraging predictive value on the amnestic MCI population. Also, our results on the cognitive reserve are in accordance with those of the current literature.
Using a new diagnostic criterion based on the NTI, we have highlighted the fact that Alzheimer's disease may be characterized by structural changes that can be detected up to 24 months before the current clinical criteria for Alzheimer's disease are fulfilled. This result supports a recent proposal (Dubois et al
) that emphasizes the need to revise the Neurological Disorders and Stroke-Alzheimer Disease and Related Disorders criteria in order to improve the accuracy of Alzheimer's disease diagnosis and to make the diagnosis at the earliest stages of the disease. The authors propose new criteria which include one or more biomarkers, among which is MRI structural neuroimaging. They however point out that there is still no MRI-based methodology that has been completely validated as suitable for integration in clinical routine. Our study shows that, among subjects meeting MCI criteria, the NTI can anticipate the diagnosis of Alzheimer's disease 2 years before the clinical criteria are fulfilled. The NTI has shown promising results at the individual level on a large population and is an operator independent, fully automated and quick method based on a single 1.5 T MRI scan. It can therefore be easily integrated into clinical routine to improve the early diagnosis of Alzheimer's disease.