• Aucun résultat trouvé

4. DISCUSSION

4.1 The early diagnosis of AD using task-based fMRI

4.1.1 Using an associative memory task in the early diagnosis of AD

A previous study of our group validated the use of an associative memory task for highlighting memory deficits in MCI patients (van der Meulen, Lederrey et al. 2012), particularly in recollection-based processes, but also in familiarity-based recognition, to a lesser extent. We corroborate the behavioral results in a partially overlapping cohort of subjects, and report very high AUCs (between 0.71-0.96) in discriminating between EC and MCI patients for both accuracy and reaction time in all recognition trials. These data are in accordance with studies

98

reporting that MCI patients are impaired in both recollection and familiarity (Wolk, Signoff et al.

2008, Algarabel, Escudero et al. 2009, Ally, Gold et al. 2009, Algarabel, Fuentes et al. 2012, Wolk, Mancuso et al. 2013), although recollection-based processes are usually found to be more affected. In fact, recollection-based recognition is consistently found to be impaired in aging, independently of the methods used (Prull, Dawes et al. 2006). Familiarity, on the other hand, is altered in some tests, but not in others, in healthy elderly individuals (Prull, Dawes et al. 2006, Wolk, Mancuso et al. 2013). These findings led Algarabel and colleagues to suggest that familiarity impairment could represent a more specific marker of AD than recollection, as it could potentially better separate normal aging from prodromal AD (Algarabel, Escudero et al.

2009). Another argument for familiarity being a good early marker for AD is the fact that it relies upon the integrity of the perirhinal cortex, which is the first brain region affected by tau pathology (Braak and Braak 1991). Yet, we found that accuracy in conditions relying upon recollection yielded higher AUCs (AUC between 0.94-0.96) than conditions relying upon familiarity (AUC between 0.71-0.86); however, reaction times in familiarity conditions (AUC between 0.90-0.95) were a little more discriminative than in recollection conditions (AUC 0.87-0.92). Therefore, we conclude that both processes add informative value in the between-group discriminability.

Our associative memory task also allowed us to manipulate the semantic relationship between the pictures within a pair in order to test whether it facilitated the recognition of the target picture. In a previous study of our group, elderly controls were less accurate in trials with semantically related compared to unrelated pairs, suggesting that unrelated pairs required a deeper and more elaborate encoding than related pairs (van der Meulen, Lederrey et al. 2012). We tested the effect of this variable (in the fourth study), but did not find any main effect of semantic relatedness on either accuracy or reaction time.

4.1.1.2 From an imaging perspective 4.1.1.2.1 ROI-based classification

We found that task-related imaging patterns allowed to distinguish MCI from EC, but to a lesser degree than behavioral measures. When using activity within ROIs as features for our classifiers, we noted that the experimental conditions that generated the highest imaging-based predictive accuracies were not the trials that relied upon recollection-based recognition. Indeed, the two

99

conditions relying upon familiarity (RRN and RUN) induced significantly more discriminative activity in several regions, and that with two different classifiers. One possible explanation for this finding is that differences in activation patterns may be easier to detect when using a higher number of trials. As we analyzed only images corresponding to correct trials, conditions that relied upon familiarity generated more correct responses and therefore more images.

Notably, the highest accuracy (86% with SVM, survived to FDR correction) was found using activity of the right PCC during encoding of semantically related pairs. The accuracy was also significant (76% with SVM, survived to FDR correction) when using activity of all encoding trials, both semantically related and unrelated. This finding could represent the impaired deactivation of medial parietal regions that is typically reported in MCI patients during encoding (Lustig, Snyder et al. 2003, Celone, Calhoun et al. 2006, Petrella, Wang et al. 2007).

Furthermore, activity in the left PHG during encoding of related pairs yielded a significant classification accuracy (71% using SVM, did not survive FDR correction). This result could reveal either hypo- or hyper-activation in MCI patients in this MTL region, as both aberrant patterns have been demonstrated in MCI patients (Small, Perera et al. 1999, Machulda, Ward et al. 2003, Dickerson, Salat et al. 2004, Dickerson, Salat et al. 2005, Celone, Calhoun et al. 2006, Johnson, Schmitz et al. 2006, Hamalainen, Pihlajamaki et al. 2007, Petrella, Wang et al. 2007, Hanseeuw, Dricot et al. 2011, Putcha, Brickhouse et al. 2011, van der Meulen, Lederrey et al.

2012), depending on the stage patients are in (early or late MCI). These findings add further evidence to previous reports of pathological processes taking place in these two regions.

Importantly, the dysfunctional patterns of fMRI activity in these regions had never been used for individual classification before.

A few regions that belong to the networks described in familiarity (anterior MTL, lateral temporal and PFC) and recollection-based (posterior MTL, thalamus, medial and lateral parietal) recognition could reliably separate normal aging from MCI. Among familiarity-related regions, activity in the left middle and superior temporal gyrus yielded respectively 71% and 76%

accuracy in the RRN condition (with RF) ; and activity in the left superior temporal pole yielded 71% accuracy in the RUN condition (also with RF). Among recollection-related regions, activity in the right angular gyrus yielded 74% accuracy (with RF) during recognition of related pairs with an old distractor (RRO). These findings complement our behavioral results, and suggest that

100

the neural correlates of both familiarity and recollection-based processes are altered in MCI patients. Finally, discriminative activity patterns were also found in the PFC, particularly during encoding (both conditions) and in one recognition condition (RUO). This finding could represent compensatory processes, as has been previously reported in AD (Grady, McIntosh et al. 2003, Pariente, Cole et al. 2005) and MCI patients, as well as in low-performing EC (Miller, Celone et al. 2008). However, this is hypothetical, as we have not inspected the directionality of discriminative activation patterns.

We were surprised to find that only one MTL region (left PHG, which did not survive FDR correction) could reliably classify our groups. One possibility is that because our analyses were limited to a mask comprising grey matter voxels that were common across all subjects, even if one patient presented severe enough atrophy in the MTL, task fMRI activity in this area could not be used for classification. However, by doing this, we made sure that the discriminative regions that we found were not driven by structural alterations.

To our surprise, the majority of discriminative regions in recognition conditions were located in the occipital lobe (calcarine sulcus, inferior and middle occipital gyrus, lingual and fusiform gyrus). One possibility is that patients rely excessively on their visual pathway to recognize the pairs of images, perhaps by relying on visual imagery (e.g., colors of picture pairs), suggesting poorer strategies for memorizing associations between images. Alternatively, patients may turn to the distinctive features of visual information to enhance their recognition of items. However, these hypotheses imply that patients hyperactivate occipital regions during recognition, which is again only speculative as we have not examined the directionality of activity in these discriminative patterns.

4.1.1.2.2 PLS-based classification

Concerning the PLS analyses, we had high expectations regarding the use of PLS in extracting memory-related networks that could be used to discriminate between our groups. To our surprise, we did not find any significant relationship between task-related fMRI activity and task accuracy.

Nonetheless, brain scores corresponding to the LVs were able to separate the two groups in two conditions (EU and RRN) with 70-73% accuracy. In addition to task accuracy measures, we used neuropsychological scores testing both verbal and visual memory as an alternative behavioral

101

measure. For each encoding condition, we found two significant LVs. In recognition, we found one significant LV in the RRN, RRO and RUN conditions. When using brain scores corresponding to the significant LVs only, we did not reach significant accuracy in separating the two groups. Brain scores only yielded discriminative accuracies when using several LVs of a given condition. In encoding of unrelated pairs (EU) for instance, using brain scores from the first LV (which was significant) yielded a nonsignificant predictive accuracy (57%), while when using brain scores for all 8 LVs for this condition resulted in a significant accuracy (74%). Among recognition conditions, only the RRN condition, which involved familiarity-based recognition with semantically related images, yielded a significant accuracy (67% when using brain scores of all LVs). The combination of conditions, either the two encoding conditions, or the four recognition conditions, or all six conditions, also yielded significant accuracies (67-74%).

One possibility for the lack of a significant relationship between task activity and accuracy is the choice of the images (i.e., an averaged beta map for each condition) that we used to characterize task activity. We did try carrying out a spatio-temporal analysis of fMRI activity according to different experimental conditions using task PLS, followed by behavior PLS with either task accuracy or neuropsychological scores testing memory as the behavioral measure (unshown results). The results we obtained were similar to what we found in our analyses: there were no significant LVs with task accuracy ; alternatively, with neuropsychological scores, we found two significant LVs for the ER condition, no significant LV for the EU condition, and one significant LV for each recognition condition. However, we did not pursue these analyses, because implementing a task PLS followed by a behavior PLS within a cross-validation scheme was more arduous (and did not seem worthwhile) compared to using meanbetas in a « simple » behavior PLS analysis.

Therefore, we were dissatisfied in the use of PLS for feature selection, as a way to extract memory-related networks that could discriminate between normal aging and MCI. Admittedly, the PLS algorithm was informed of the presence of two groups, i.e., the model maximized the covariance between the two modalities across all subjects while allowing while allowing a separate relationship between the imaging data and the memory scores in each group.

Nonetheless, the model did not seek between-group discriminability per se.

102

Previous studies that have used PLS as a dimensionality reduction tool before classifying two groups have tuned the algorithm to maximize group discriminability. For instance, Andersen and collaborators have applied a variant of PLS (oriented PLS), which « orients » dimensionality reduction towards patterns that exhibit group separation, while simultaneously orienting away from within-group covariability (Andersen, Rayens et al. 2012). Furthermore, in several other studies, the PLS algorithm was used to maximize covariance between high-dimensional data (e.g., imaging or genetic) and the subject groups (Nguyen and Rocke 2002, Rosipal, Trejo et al.

2003, Ramirez, Gorriz et al. 2010, Segovia, Górriz et al. 2012). Our strategy was more indirect, as we used PLS to maximize covariance between task activity fMRI patterns and memory measures (which, of course, discriminated our groups). This is because our aim was to inject some a priori knowledge into feature extraction, i.e., we wanted to discriminate our groups using memory-related task fMRI patterns.