Automatized FACEmemory® scoring is related to Alzheimer’s disease phenotype and biomarkers in early-onset mild cognitive impairment: the BIOFACE cohort

A total of 97 participants with EOMCI from the BIOFACE study [23] were recruited at the Memory Unit of Ace Alzheimer Center Barcelona. Three were excluded from analysis because they failed to complete the FACEmemory test (one) or whole MRI data was not available for quantitative analysis (two).

As detailed previously [23], a comprehensive neuropsychological, neurological, and functional assessment at the Memory Unit of Ace Alzheimer Center Barcelona, as well as routine analyses of blood and structural brain neuroimaging, were administered to all BIOFACE participants within a time window of 3 months.

Within the diagnosis procedure of Ace Alzheimer Center Barcelona, all participants had a first neurological and neuropsychological visit (with the neuropsychological battery of Fundació ACE (NBACE) [24, 25] administered by one of the neuropsychologists of the team), including an accurate diagnosis to ensure that they fulfilled the BIOFACE inclusion criteria. The participants returned to the Memory Unit for an additional neurological assessment, followed by the FACEmemory visit, which was performed in another office with a psychologist who invited each patient to sit at a table with a tablet and complete the FACEmemory test with minimal supervision (only technological issues were addressed). Finally, in the same week, each participant came back for an additional neuropsychological assessment administered by a neuropsychologist (A.P.) in the same order. Thus, the neuropsychological assessment included three visits: (1) the NBACE, (2) additional neuropsychological tests to detect prodromal AD, and (3) the FACEmemory test.

Inclusion criteria were diagnosis of MCI according to the Petersen [3] criteria (including the amnestic and non-amnestic types), a Clinical Dementia Rating of 0.5, age of onset under 65, Mini-Mental State Examination (MMSE) score ≥ 24, educational level of at least elementary school, capacity to provide written informed consent, and fluent Spanish language skills. Exclusion criteria were contraindication for brain MRI, presence of an underlying medical or neurological illness that could account for cognitive impairment based on lab tests or brain imaging, major psychiatric disorder, active drug abuse, and severe auditory/visual abnormalities.

In addition to FACEmemory, the following classical memory tests were administered: the Spanish version of the Word FCSRT [11, 26], the Word List from the Wechsler Memory Scale, third edition (WMS-III) [27], and the Spanish version of the Rey-Osterrieth Complex Figure Test (ROCF) [26].

The FACEmemory test was administered via a tablet computer with voice recognition and a touchscreen, ensuring its standardized administration and immediate automatized scoring and registration of the results in an anonymous database.

As detailed elsewhere [17], the temporal sequence of FACEmemory was the following: two learning trials, a short-term memory task, and a long-term memory task that included face, name, and occupation memory recognition. The total test duration was approximately 30 min. Briefly, the first learning trial consisted of a total of 12 faces, each one associated with a name and an occupation that appeared beneath it for 8 s. The second learning trial followed the same procedure as the first one, but the 12 faces appeared in a different order. Participants were instructed to read the name and occupation appearing beneath each face aloud and to try to remember it. Then, the application asked the participants to press the red microphone button and say the name and occupation they remembered as being associated with each face. Two minutes after the second learning trial began, the application again asked the participants to say the name and occupation they remembered as being associated with each face. Finally, 20 min after the second learning trial, long-term memory assessment was initiated, which involved free recall and recognition tasks. First, the participants were instructed to recognize, from three faces, the face that had appeared in the first learning trial and to touch it. Then, the correct face appeared and the participants were asked to say the name and occupation they remembered for each face. After each answer, a screen appeared showing the correct face, and two rows below the face were three name options and three occupation options. The participants were instructed to touch the name and occupation they recognized as being associated with that face.

With regard to the Spanish version of the Word FCSRT [11], the task consists of learning 16 words that are presented on four pages. Participants were asked to read each word aloud and then say which of the words corresponded to a given semantic category. After this initial learning and encoding procedure, there were three recall trials, each preceded by a non-semantic interference task (countdown of 20 s). For each trial, participants were asked to remember as many words as possible, and then, the semantic category was provided for those items that were not recalled. The same recall procedure was repeated 20 min later.

The Word List from the WMS-III [27] was administered using a traditional procedure in which the list was presented orally in four consecutive learning trials, without using any interference list. After a 20-min delay, a free recall trial, followed by a yes/no recognition task, was administered [24].

With regard to the ROCF [26] test, a complex figure was placed in front of the participant, who was asked to copy it as accurately as possible. Then, the figure was removed from view. After a 20-min delay, the subject was asked to reproduce the figure from memory.

Scans were acquired in a Siemens MAGNETOM VIDA 3T scanner (Erlangen, Germany) using a 32-channel head coil from Clínica Corachan (Barcelona). T1-weighted images, for the morphological and the volumetric studies, were acquired using a gradient-echo 3D MPRAGE sequence with the following parameters: TR 2200 ms, TE 2.23 ms, TI 968 ms, 1.2-mm slice thickness, FOV 270 mm, and voxel measurement 1.1 × 1.1 × 1.2 mm. To complete the acquisition, a 3D isotropic FLAIR, an axial sequence T2-weighted and an axial sequence T2*-weighted gradient recall echo were performed to detect vascular brain damage and microbleeds. All images were stored in a PACS and submitted to an automated process of deidentification.

The MRIs were processed at the neuroimaging laboratory at Ace Alzheimer Center Barcelona. After a visual inspection for artifacts, cortical and subcortical segmentation of the structural images was performed using FreeSurfer 7.2 (https://​surfer.​nmr.​mgh.​harvard.​edu/​). This procedure allows the segmentation of gray matter, white matter, and other substructures. A surface-based morphometry (SBM) analysis of cortical volume was performed for the total FACEmemory score variable, with age, schooling years, sex, and total intracranial volume as covariables. Surfaces were smoothed with a 10-mm FWHM kernel. Cluster-wise correction for multiple comparisons was done using a z-based Monte Carlo simulation with 10,000 iterations. Surface clusters were looked at with a cluster-forming threshold of p< 0.001 and corrected for multiple comparisons for p< 0.05.

This protocol followed the consensus recommendations established by the Alzheimer’s Biomarkers Standardization Initiative [28]. Briefly, a lumbar puncture was performed by an experienced neurologist with the patients in a seated position and under fasting conditions. After applying local anesthesia (1% mepivacaine) subcutaneously, the neurologist obtained CSF by lumbar puncture in the intervertebral space of L3–L4. The fluid was collected passively in two 10-ml polypropylene tubes (Sarstedt Ref. 62610018). The first tube was analyzed externally for basic biochemistry (glucose, total proteins, proteinogram, and cell type and number). The second tube was centrifuged (2000×g 10 min at 4 °C), and the fluid was aliquoted into polypropylene tubes (Sarstedt Ref. 72694007) and stored at −80 °C until analysis. Time delay between CSF collection and storage was less than 2 h.

On the day of the analysis, the aliquots were thawed at room temperature and vortexed for 5–10 s to determine AD biomarkers in CSF. One aliquot/patient was used to determine the concentrations of Aβ1-42, Aβ1-40, t-tau, and p181-tau using chemiluminescence enzyme immunoassay (CLEIA) with the commercially available Lumipulse G™ reagents in the Lumipulse™ platform (Fujirebio, Europe) at the research laboratory of Ace Alzheimer Center Barcelona.

Using CSF and MRI biomarkers, participants were classified into three categories according to the ATN scheme [29]. Those categories were normal AD biomarkers (A-T-N-), Alzheimer’s continuum (including A+T-N-, A+T+N-, A+T+N+, and A+T-N+), and non-AD pathologic changes (-SNAP-, including A-T+N-, A-T-N+ and A-T+N+), where A refers to aggregated Aβ, T to aggregated tau, and N to MRI neurodegeneration or neuronal injury. Cut-offs from the Fundació ACE biomarker research program (FACEBREP) cohort were used to dichotomize each CSF biomarker into +/− as follows: Aβ1-42/Aβ1-40 ratio< 0.063 for A, p181-tau> 54 pg/ml for T, and presence of neurodegeneration in the MRI for N.

The N classification was obtained using a machine learning method (Naive Bayes) [30]. As a training database, the subjects’ data were selected from the ADNI database (adni.loni.usc.edu) [31] as follows. Only baseline MRI data from individuals diagnosed with dementia and a cognitively healthy status at baseline were selected, for a total of 1222 data points. The volume from the hippocampus, entorhinal cortex, middle temporal cortex, and lateral ventricles, as well as the subject’s age and estimated intracranial volume, was chosen as independent variables. Subjects diagnosed with dementia were considered to be neurodegeneration positive (N+), while cognitively healthy subjects were considered as neurodegeneration negative (N−). To test the algorithm, ADNI data was randomly spliced in two datasets with 70% of the data as the training dataset and 30% as the test dataset. The algorithm so fed was tested and showed an accuracy of 82%, sensitivity of 82%, and specificity of 80% for N classification. So, the full ADNI dataset was used to build the prior probability function. Then, the values of independent variables for the BIOFACE cohort were calculated and the N values were assessed as the dichotomization of the posterior probability function.

Statistical analysis was performed using Statistical Package for the Social Sciences (SPSS) version 26 for Windows (version 26.0; SPSS Inc., Chicago, IL). All data were examined for normality, skew, and restriction of range.

Univariate and multiple linear regression analyses (the stepwise procedure) were carried out to search for traditional memory variables (delayed free recall on the Word List from the WMS-III and Word FCSRT, total free recall/learning on the Word List from the WMS-III and on the Word FCSRT, and long-term visual recall on the ROCF) correlated with the total FACEmemory score, after adjusting for age, schooling years, and sex.

In the subsample with lumbar puncture, univariate linear regression analyses with FACEmemory and each CSF AD biomarker value, adjusting for age, schooling years, and sex, were carried out. Moreover, to search for differences between ATN groups (AD, SNAP, and normal), an analysis of covariance (ANCOVA) was performed.

The whole sample consisted of 94 participants, including 37 men and 57 women. Their mean age was 60.74 (standard deviation (SD): 3.42) years, with a mean of 12.05 (SD: 5.08) schooling years. Their mean score on the MMSE was 28 (SD: 1.50), and on the FACEmemory, their mean score was 34.15 (SD: 18.92) points.

With regard to the type of MCI, 46 were amnestic and 48 were non-amnestic. Both groups did not significantly differ in age, sex, or years of formal education. However, as expected, aMCI performed significantly worse than non-amnestic mild cognitive impairment (naMCI) on the FACEmemory total score (mean: 26.87, SD: 16.56, and mean: 41.13, SD: 18.56, respectively; t= −3.92, p< 0.001).

The subset of patients who underwent a lumbar puncture (n= 82) and those who did not (n= 12) were homogeneous on age (t= 0.92, p= 0.357), sex (χ²= 1.19, p= 0.276), schooling years (t= 0.22, p= 0.826), MMSE (t= 0.00, p= 1.000), and total FACEmemory score (t= 0.48, p= 0.629). With regard to the ATN groups, 24 subjects were classified as normal, 15 AD, and 43 non-AD/SNAP. Their demographic and clinical characteristics are detailed in Table 1.

The total FACEmemory score was found to be weakly correlated with age (r= −0.22, p= 0.032), but not with schooling years (r= 0.08, p=0.419) or sex (t= 0.16, p= 0.872).

The performance on the memory tests administered is detailed in Table 2. Convergent validity of FACEmemory was determined based on performance on the FCSRT. Univariate linear regression analyses showed FACEmemory to be significantly associated with each individual traditional memory test (as detailed in Table 3). In a multiple linear regression analysis using classical memory tests as independent variables, adjusted by age, schooling, and sex, long-term free recall on the Word List from the WMS-III and on the FCSRT emerged as the only significant factors associated with the total FACEmemory score (F= 12.830, p< 0.001; beta= 0.42, p< 0.001 and beta= 0.30, p= 0.001, respectively).

The SBM analysis showed the presence of four different clusters where there was a significant correlation between cortical volume and the FACEmemory total score. In the left hemisphere, there were two clusters: the first one comprised the fusiform gyrus and associative visual cortex (p< 0.0002), and the second one was centered in the middle temporal cortex (p< 0.01). The other two clusters were located in the right prefrontal cortex (p< 0.0008) and in the parietal angular and supramarginal gyri (p< 0.004).

For all clusters, there was a positive correlation between memory performance and cortical volume; that is, worse performance on the FACEmemory test was associated with lower cortical volume in those areas. All the clusters are described in Table 4 and shown in Fig. 1.

In the subsample of 82 patients who underwent lumbar puncture, a correlation analysis among demographical variables (age, sex, and education) showed significant associations between age and the Aβ1-42/Aβ1-40 ratio (r= −0.35, p= 0.001), p181-tau (r= 0.32, p= 0.003), and the Aβ1-42/p181-tau ratio (r= 0.39, p= 0.009) levels, but not with Aβ1-42 alone (r= 0.04, p= 0.728). Sex and schooling years were not significantly associated with any CSF variable. Univariate linear regression analyses with FACEmemory and each CSF variable, covariated by age, schooling years, and sex, showed that the total FACEmemory score was moderately correlated with the CSF Aβ1-42/Aβ1-40 ratio, p181-tau, and the Aβ1-42/p181-tau ratio, but only weakly with Aβ1-42 levels (see Table 5 for details).

With regard to ATN analysis, when comparing the total FACEmemory score among normal, AD, and non-AD/SNAP groups, a statistically significant effect was found among groups (F= 9.018, p< 0.001). Post hoc analysis showed that the AD group differed significantly from the normal and SNAP groups (with a mean difference of 23.90 (SE: 5.64), p< 0.001, and 15.73 (SE: 5.14), p= 0.009, respectively) (for details see Table 2).

We acknowledge that our study has several limitations. First, the study design is cross-sectional. However, this is the first step of the BIOFACE project, in which all participants are being followed up annually with clinical and biomarker assessments. Secondly, not all recruited participants gave their consent for a CSF sample, but we recognize that some of them were reluctant to undergo a lumbar puncture due to it being an invasive, not completely innocuous process that can cause pain or discomfort to the patient. Thirdly, ADNI data was used as a training dataset for the neurodegeneration classification algorithm. ADNI data contains subjects with a wider age range which are older, on average, than the current dataset. Although the machine learning algorithm chose the proper prior distribution for any involved variable, it automatically discarded a considerable number of subjects from the training database. The use of a fully representative dataset for the current sample could be worthy of study. Finally, although all participants have been accurately characterized with an exhaustive cognitive neuropsychological assessment and multimodal AD biomarkers, the small sample size may prevent us from making robust conclusions. However, it should be noted that this is a single-center study and patients with an early-onset cognitive impairment are less frequently derived to a memory clinic than those with late-onset cognitive impairment.