Sample characteristics
Table 2 shows the demographic, pathophysiological and clinical characteristics of the cohort by CSF profile. There were no significant differences between the groups in age, length of education, novel CSF protein levels or gender frequencies. Boxplots comparing distributions in CSF protein levels (NFL, YKL-40, S100B, GFAP) between profile groups are presented in Additional file 1, S1a-d. The CSF AD profile group showed significantly worse performance on the MMSE, RAVLT, Story, ROCF immediate recall and Verbal fluency animal tests compared to the non-AD group (p<0.05).
Table 2 Subject demographics, CSF marker levels and neuropsychological test scores by CSF profile
|
|
|
CSF profile
|
|
|
|
|
Non-AD
T-tau/ Aβ42 ≤ 0.52
(n=24)
|
AD
T-tau/ Aβ42 > 0.52
(n=28)
|
p
value ᵃ
|
Demographics
|
|
|
|
|
|
Gender (M/F)
|
16/8
|
17/11
|
0.66
|
|
|
Age, years
|
67 (46-80)
|
70 (51-84)
|
0.17
|
|
|
Education, years
|
14.0 (9-20)
|
12.5 (6-17)
|
0.11
|
Clinical diagnosis
|
|
|
|
|
|
SCI/MCI/AD/LBD
|
10/13/0/1
|
2/9/16/1
|
N/Ab
|
CSF measures
|
|
|
|
|
|
Aβ42 (pg/ml)
|
703 (374-2332)
|
454 (160-822)
|
N/Ac
|
|
|
T-tau (pg/ml)
|
173 (100-722)
|
416 (132-838)
|
N/Ac
|
|
|
NFL (ng/ml)
|
1.9 (0.9-6.5)
|
2.5 (1.2-4.5)
|
0.15
|
|
|
YKL-40 (ng/ml)
|
165 (83-399)
|
203 (124-367)
|
0.12
|
|
|
S100B (pg/ml)
|
215 (132-335)
|
230 (129-458)
|
0.17
|
|
|
GFAP (ng/ml)
|
1.0 (0.1-7.1)
|
1.3 (0.5-21.3)
|
0.09
|
Cognitive domains
|
|
|
|
|
Global cognition
|
|
|
|
|
|
MMSE, score
|
28 (24-30)
|
27 (24-30)
|
0.01
|
|
Verbal episodic memory
|
|
|
|
|
|
RAVLT immediate recall, score
|
36 (23-66)
|
26.5 (13-51)
|
0.003
|
|
|
RAVLT delayed recall, score
|
6.5 (0-15)
|
1.5 (0-12)
|
<0.001
|
|
|
RAVLT recognition-fp, score
|
9.0 (3-15)
|
5.5 (-3-15)
|
0.003
|
|
|
Story immediate recall, score
|
13.5 (5-17)
|
8 (1-18)
|
0.005
|
|
|
Story delayed recall, score
|
12.0 (1-19)
|
5.5 (0-16)
|
0.002
|
|
Non-verbal episodic memory
|
|
|
|
|
|
ROCF immediate recall, score
|
13.3 (0-27)
|
7.3 (0-26)
|
0.04
|
|
|
ROCF delayed recall, score
|
12.8 (0-25)
|
8.5 (0-26)
|
0.07
|
|
Language
|
|
|
|
|
|
Verbal fluency animal, score
|
20 (8-33)
|
14 (4-27)
|
0.02
|
|
|
Verbal fluency H+S, score
|
24.0 (14-48)
|
25.5 (6-63)
|
1.00
|
|
Processing speed
|
|
|
|
|
|
TMT-A, sec.
|
43.5 (21-133)
|
48.0 (27-116)
|
0.22
|
|
|
Stroop – part I, sec.
|
23.5 (20-42)
|
24.5 (17-34)
|
0.64
|
|
Executive functions
|
|
|
|
|
|
TMT-B, sec.
|
109 (44-340)
|
153 (60-343)
|
0.06
|
|
|
DSST, score
|
8.5 (3-51)
|
7.0 (2-61)
|
0.24
|
|
|
Stroop 4th/3rd part, sec.
|
2.1 (1.4-4.0)
|
2.1 (1.6-5.8)
|
0.25
|
Abbreviations: AD Alzheimer’s disease, CSF Cerebrospinal fluid, DDST Digit symbol substitution test, fp false positives, LBD Lewy body dementia, MCI Mild Cognitive Impairment, MMSE Mini-Mental State – Examination, N/A Not applicable, RAVLT Rey Auditory-Verbal Learning Test, ROCF Rey–Osterrieth complex figure, SCI Subjective Cognitive Impairment, TMT Trail Making Test
Values are shown as median (range) or as numbers per group, ᵃMann-Whitney U non-parametric tests used for continuous variables and Chi-Square tests for categorical variables, p-values not applicable for bclinical diagnosis due to CSF profiles being part of the diagnostic criteria for AD and cAβ42 and T-tau due to their values used for defining CSF profiles
Pearson‘s correlations between CSF markers
Pearson‘s correlations between the CSF markers, age and length of education are presented in Fig. 2, respectively. Inflammatory markers YKL-40 and S100B and neurodegeneration markers NFL and T-tau all correlated positively and significantly with each other. The highest correlation was found between NFL and YKL-40 (NFL: r=0.62, p<0.001). GFAP did only significantly correlate with the CSF marker S100B (r=0.53, p<0.001). No CSF markers correlated significantly with Aβ42. All the CSF markers, except for Aβ42, correlated positively with age. Length of education correlated weakly and negatively with T-tau (r=-0.29, p=0.03).
Fig. 2 Pearson‘s correlation matrix between CSF markers, age and length of education. Colored squares indicate statistical significance (p<0.05). CSF measures were natural log-transformed
Accuracy of CSF markers and cognitive domains in distinguishing between CSF profiles
Accuracies for distinguishing between CSF AD and non-AD profiles were based on univariable ROC analyses (Table 3). AUCs for novel CSF markers ranged from 0.61 - 0.64, with a lower limit of each confidence interval below the value of 0.5. In comparison, neuropsychological tests reflecting verbal episodic memory had the highest accuracy compared to other measurements, with all AUCs over 0.70, which is considered fair [52]. The scores for the Verbal episodic memory composite test (AUC=0.80, CI: 0.69-0.92) and RAVLT delayed recall (AUC=0.80, CI: 0.68-0.93) distinguished the best between the CSF profile groups. A similar trend in results was found when ROC analyses were stratified by gender (Table S1, Additional file 1), although AUC coefficients were overall higher for women (n=19) compared to men (n=33). LASSO logistic regression with stability selection was performed for the selection of variables distinguishing between the CSF profile groups with the highest consistency. Nine possible predictors could be selected, the four novel CSF markers and the five composite tests presenting each cognitive domain. Only the test reflecting verbal episodic memory was selected as a predictor, with selection frequency (96%) above the cut-off value. All other possible predictors had a much lower selection frequency (≤ 20%).
Table 3 Accuracy in distinguishing between CSF AD and non-AD profiles
|
|
|
Univariable
ROC analyses
|
Multivariable LASSO logistic regressionb
|
|
|
|
AUC
|
95% CI (AUC)*
|
Stability selection (%)
|
CSF measuresᵃ
|
|
|
|
|
|
GFAP (ng/ml)
|
0.64
|
0.48-0.79
|
10
|
|
|
YKL-40 (ng/ml)
|
0.63
|
0.47-0.78
|
18
|
|
|
NFL (ng/ml)
|
0.62
|
0.45-0.78
|
2
|
|
|
S100B (pg/ml)
|
0.61
|
0.46-0.77
|
20
|
Cognitive domains
|
|
|
|
|
Verbal episodic memory
|
|
|
|
|
|
Composite z-score
|
0.80
|
0.69-0.92
|
96c
|
|
|
RAVLT delayed recall, score
|
0.80
|
0.68-0.93
|
-
|
|
|
Story delayed recall, score
|
0.75
|
0.62-0.89
|
-
|
|
|
RAVLT immediate recall, score
|
0.74
|
0.61-0.88
|
-
|
|
|
RAVLT recognition-fp, score
|
0.74
|
0.61-0.87
|
-
|
|
|
Story immediate recall, score
|
0.73
|
0.59-0.86
|
-
|
|
Non-verbal episodic memory
|
|
|
|
|
|
Composite z-score
|
0.65
|
0.50-0.81
|
14
|
|
|
ROCF immediate recall, score
|
0.66
|
0.51-0.81
|
-
|
|
|
ROCF delayed recall, score
|
0.65
|
0.49-0.80
|
-
|
|
Executive functions
|
|
|
|
|
|
Composite z-score
|
0.64
|
0.49-0.80
|
16
|
|
|
TMT-B, sec.ᵃ
|
0.66
|
0.50-0.81
|
-
|
|
|
DSST, scoreᵃ
|
0.60
|
0.44-0.75
|
-
|
|
|
Stroop 4th/3rd part, sec. ᵃ
|
0.59
|
0.43-0.75
|
-
|
|
Language
|
|
|
|
|
|
Composite z-score
|
0.60
|
0.44-0.76
|
4
|
|
|
Verbal fluency animals, score
|
0.68
|
0.54-0.83
|
-
|
|
|
Verbal fluency H+S, score
|
0.50
|
0.34-0.66
|
-
|
|
Processing speed
|
|
|
|
|
|
Composite z-score
|
0.56
|
0.39-0.72
|
9
|
|
|
TMT-A, sec. ᵃ
|
0.60
|
0.44-0.76
|
-
|
|
|
Stroop test – part I, sec. ᵃ
|
0.54
|
0.38-0.70
|
-
|
Abbreviations: AD Alzheimer’s disease, AUC Area under curve, CI Confidence Intervals, CSF Cerebrospinal fluid, DDST Digit symbol substitution test, fp false positives, LASSO Least absolute shrinkage and selection operator, RAVLT Rey Auditory-Verbal Learning Test, ROCF Rey–Osterrieth complex figure, TMT Trail Making Test
AUC is the probability that a randomly selected pair of subjects from each CSF profile group is correctly classified, *Confidence intervals calculated with DeLong method, ᵅValues are natural log-transformed, bLASSO logistic regression model was fitted on 100 subsamples, with different predictors (CSF measures and composite test scores) possibly selected into each model. Numbers present the frequency (%) of each possible predictor selected. The per-family error rate (PFER) was set at 1, and the cut-off value at 75% for stability selection. cThe composite test for verbal episodic memory was the only measure to have selection frequency above the cut-off value
Fig. 3 illustrates the ROC curves for the two cognitive domains and the CSF measure with the highest AUC from Table 3. Verbal episodic memory (AUC=0.80) was superior in distinguishing between CSF AD vs. non-AD profiles compared to non-verbal episodic memory (AUC=0.65) and CSF GFAP (0.64).
Fig. 3 Comparison between ROC curves of the two cognitive domains and the CSF marker with the highest area under the curve (AUC) coefficients
Selection of predictors for scores on each cognitive domain
LASSO linear regression with a stability selection was applied for identifying a set of variables (CSF markers and demographic variables) predicting cognitive scores with the highest consistency (Fig. 4). Two analyses were performed for each of the five domains, one including all subjects (n=52) and the other only among those with a CSF AD profile (n=28). Variables with stability selection above 75%, were considered reliable predictors. GFAP (78%) was selected as a predictor for executive functions (Fig. 4a) and age (95%) as a predictor for non-verbal memory (Fig. 4b) within the whole cohort. Among subjects with a CSF AD profile, GFAP (87%) and age (81%) were selected as predictors for processing speed (Fig. 4c) and NFL (80%) for verbal episodic memory (Fig. 4d). No variables reached the stability selection criteria as predictors of scores reflecting language (Fig. 4e).
Fig. 4. LASSO linear regression - stability selection analyses for prediction of composite z-scores reflecting a) verbal episodic memory, b) non-verbal episodic memory, c) language, d) processing speed and e) executive functions. Two analyses were created for each domain, one including all participants (n=52) and the other only the CSF AD profile group (n=28). The cut-off selection value was set at 75% and the per-family error rate (PFER) at 1 for all analyses.
Pearson‘s correlations between selected CSF markers and cognitive domains
Relationships between CSF measures and cognitive domains, as selected with LASSO regression – stability selection analyses (Fig. 4), were visualized using scatter plots. It is well established that normal aging, level and quality of education can influence cognitive test performance [53]. Composite z-scores were therefore adjusted for age and education prior to Pearson‘s correlations calculations.
CSF NFL levels did not significantly correlate with verbal episodic memory among all subjects (r=-0.26, p=0.06, Fig. 5a). Analysis by CSF profile (Fig. 5b) revealed moderate, significant correlation among subjects with a CSF AD profile (r=-0.43, p=0.02) compared to none among those without (r=-0.05, p=0.82). Correlations between the NFL levels and individual neuropsychological tests reflecting verbal episodic memory are presented in Additional file 1, S2a-e. T-tau did not reach the selection criteria for any cognitive domain. It is, none the less, of interest to compare the results of T-tau to NFL as both proteins are markers of neurodegeneration. The association between T-tau and verbal episodic memory was similar to NFL within the whole cohort (r=-0.28, p<0.04, Fig. 5c) but did not reach significance within the CSF AD group (r=-0.15, p=0.45) when analyzed by CSF profile (Fig. 5d).
Correlation between CSF GFAP levels and processing speed did not reach significance within the whole cohort (r=-0.27, p=0.06, Fig. 5e) or among those with a CSF non-AD profile (r=0.02, p=0.94, Fig. 5f). A moderately strong correlation was, on the other hand, detected among those with a CSF AD profile (r=-0.68, p<0.001, Fig. 5f). A weak, negative correlation was found between CSF GFAP levels and executive functions, both within the whole cohort (r=-0.37, p=0.01, Fig. 5g) and among subjects with a CSF AD profile (r=-0.39, p=0.04, Fig. 5h). The corresponding correlations between CSF GFAP levels with individual neuropsychological tests reflecting processing speed and executive functions are presented in Additional file 1, Fig. S3a-e. Additional file 1 also includes scatter plots identical to those shown in figure 5 without adjustment for age and education (Fig. S4a-h) and Pearson‘s correlations between CSF markers, age and education and composite scores of each cognitive domain, both unadjusted and adjusted for age and education (Table S2).
Fig. 5. Scatter plots presenting Pearson‘s correlations between CSF levels of NFL and verbal episodic memory (a,b), T-tau and verbal episodic memory (c,d), GFAP and processing speed (e,f) and GFAP and executive functions (g,h) within the whole cohort and by CSF profile. *Cognitive domains were adjusted for covariates (age and education). Without the bottom corner GFAP outlier in the CSF AD profile group, Pearson‘s correlations were a slightly lower for f) processing speed (r=-0.58, p=0.001) and h) executive functions (r=-0.28, p=0.15)