An increasing body of evidence suggests that Alzheimer’s disease (AD) pathophysiology can be identified using biomarkers [1, 2]. AD is characterized by abnormal patterns in structural and functional imaging as well as by a pathological cerebrospinal fluid (CSF) signature [3]. The pathological CSF signature is defined by decreased CSF concentrations of the peptide amyloid-beta 1–42 (Aβ42) and increased levels of the proteins total Tau (t-Tau) and Tau phosphorylated at threonine 181 (p-Tau). It is of note that biomarkers reflect AD neuropathological changes with relatively high accuracy [4, 5]. In clinical practice, CSF biomarkers aid clinicians with decision-making, embody a key tool in the differential diagnosis especially of atypical dementia syndromes and increase diagnostic confidence [4, 69].

Biomarkers enable the identification of AD pathophysiology in pre-dementia stages of the disease, such as the stage of mild cognitive impairment (MCI) [10]. MCI is a clinical entity characterized by cognitive deficits which are so mild that activities of daily living remain largely unaffected [10]. It is a heterogeneous clinical syndrome with regard to aetiology, clinical appearance and prognosis. MCI can be caused by different diseases (e.g. AD, cerebrovascular disease, depression, frontotemporal lobar degenerations, etc.). As a consequence, only some patients with MCI progress to dementia due to AD or to other dementias within limited time periods, while cognitive functioning remains stable or even reverses to normal in others [1115]. Biomarkers embody a valuable instrument in estimating the likelihood that MCI is engendered by AD. Interestingly, biomarkers are an integral part of the recently proposed National Institute on Aging–Alzheimer’s Association (NIA-AA) guidelines for the diagnosis of MCI in clinical research settings [16]. These guidelines suggest categorizing MCI according to the individual likelihood of underlying AD pathophysiology and the associated risk of developing AD dementia in the future; the highest likelihood category is characterized by biomarker findings pointing to the presence of AD pathophysiology, whereas the lowest likelihood category is characterized by findings not typical for AD. However, the criteria do not consider conflicting biomarker constellations, although they are very common in MCI [1719].

A large number of studies have investigated the prognostic utility of established AD biomarkers (for instance [10, 2028]), but none of them has taken into account the NIA-AA guidelines in combination with exclusive consideration of all three established CSF biomarkers. To shed light on this grey area, we conducted a retrospective study focused on differences in progression to dementia due to AD of patients with MCI and different CSF biomarker constellations.



The study sample encompassed data from all phases of the AD Neuroimaging Initiative (ADNI) (ADNI 1, Go and 2), a collaborative project of academic institutions and private corporations across the USA and Canada which began in October 2004. The study is coordinated by the Alzheimer’s Disease Cooperative Study at the University of California, San Diego. The ADNI data are disseminated by the Laboratory for Neuroimaging at the University of Southern California. Data used in this study were obtained from the ADNI database ( on 27 August 2014. Patients with MCI, fulfilling international diagnostic criteria [16], and with available CSF Aβ42, t-Tau and p-Tau values at baseline and clinical follow-up data were included. In the ADNI, patients with MCI had Mini-Mental-State Examination (MMSE) scores between 24 and 30, a Clinical Dementia Rating (CDR) score of 0.5, memory complaints and objective memory deficits on the Wechsler Memory-Scale-Logical Memory II test. They were not significantly impaired in their activities of daily living. Patients with diagnoses other than MCI at baseline, controls and patients with MCI but not all CSF biomarker findings available at baseline were excluded from the study. Patients diagnosed with AD dementia at follow-up met the NIA-AA diagnostic guidelines for dementia due to probable AD [1]. Regarding MCI patients who had not progressed to dementia but discontinued participation in follow-up visits or died, the data of their last follow-up visit were considered in the analysis.

CSF collection and analysis

CSF collection, shipping, aliquoting, storage and analysis took place according to ADNI standard operating procedures (SOPs) [29]. It is noteworthy that some early CSF samples were mistakenly collected into inappropriate CSF collection tubes at the ADNI sites. However, this was corrected rapidly and the exposure time to any inappropriate CSF collection tube was of limited significance due to the short time the CSF was in contact with the transfer tubes (approximately 25.7 min) [29]. ADNI baseline CSF samples were analysed at the ADNI biomarker core laboratory at University of Pennsylvania according to published methods [5, 30]. CSF samples were put into the freezer at –80 °C. The CSF concentrations of Aβ42, t-Tau and p-Tau were measured using the multiplex xMAP Luminex platform with Innogenetics immunoassay kit-based reagents (INNO-BIA AlzBio 3; Ghent, Belgium) [30].

APOE genotyping

APOE genotypes were determined for all ADNI participants through analysis of blood samples using standard polymerase chain reaction methods [31].

Classification of patients with MCI

In line with the NIA-AA algorithm, each patient’s biomarker values were categorized as either positive for AD, negative for AD or borderline. The definition of the range of borderline values was based on biomarker cut-off values and standard deviations (SDs) selected from previous reports on ADNI MCI patients [5]. The range of borderline values was specified with the aim to reach a reasonable compromise between minimizing the chance of an artificial categorization as positive and at the same time classifying less than 20% of the measured values of each biomarker as borderline. Values within 20% of the SD from the respective cut-off point were classified as borderline [32]. Aβ42 concentrations lower than the defined range of Aβ42 borderline values and t-Tau and p-Tau levels higher than the respective borderline ranges were assumed to be AD positive. All other biomarker values were considered negative. Aβ42, t-Tau and p-Tau concentrations <181 pg/ml, >105.2 pg/ml, and >26.6 pg/ml, respectively, were thus regarded as positive for AD. CSF levels of Aβ42 > 203 pg/ml and t-Tau and p-Tau concentrations <80.8 pg/ml and <19.4 pg/ml, respectively, were considered negative for AD.

Patients were classified according to their fluid biomarker profile into the following subgroups:

  • MCI with no positive biomarkers (MCINon+).

  • MCI with all biomarkers positive (MCIAll+).

  • MCI with positive Aβ42 but negative or borderline p-Tau and t-Tau (MCIAβ+).

  • MCI with positive Aβ42 and positive p-Tau or t-Tau (MCIAβ+T+).

  • MCI with negative or borderline Aβ42 but positive p-Tau and/or t-Tau (MCIT+).

The biomarker constellations of the three latter subgroups are so far not being considered in the NIA-AA diagnostic guidelines, because for MCI patients with such biomarker constellations no likelihood grade for the presence of AD pathology is assigned by the NIA-AA criteria [16].

Statistical analysis

The statistical analyses were performed in SPSS v19.0 for Windows (IBM Corp., Somers, NY, USA). Normal distribution of data was checked using the Kolmogorov–Smirnov test. The raw biomarker data of study participants were graphically presented by means of non-negative matrix factorization (NNMF) [32, 33], a data-learning technique particularly suited for analysing positive valued data so that the available information is condensed in a low-dimensional (2D) space. The overall set of measurements:

X i  = {Aβ42, t-Tau, p-Tau}unlikely i ,  i = 1, 2, …, N , where N is the total number of participants, was approximated as:

X [N×3] ≈ W [N×2] B [2×3]

in order to minimize the reconstruction error induced by the Frobenius norm: ||X-WB||2. In this way, the vector of measurements X i associated with the ith participant took the form of:

X i  = w i1 B 1 + w i2 B 2,

where B 1 and B 2 were the unit length vectors for a parsimonious 2D representation and w i1 and w i2 were the corresponding components. Differences between the MCI subgroups regarding demographic and CSF data, MMSE scores, follow-up duration, presence of the APOE ε4 allele and progression rates to dementia due to AD were tested by analysis of variance (ANOVA), Bonferroni post-hoc analysis, Kruskal–Wallis test, Mann–Whitney test and chi-square test as appropriate. Differences in the hazard of progression between the MCI subgroups were analysed using Cox regression models, adjusting for patient characteristics that significantly differed between the subgroups. Two-sided p < 0.05 was considered statistically significant.


A total of 469 MCI patients out of 1729 ADNI participants with available baseline data fulfilled the inclusionary criteria of the study. APOE ε4 and sex distribution, as well as age and MMSE scores, significantly differed between the subgroups (Table 1). In particular, the MMSE scores of the MCINon+ subgroup were significantly higher compared with the scores of the MCIAll+ (p < 0.001) and MCIAβ+T+ (p < 0.01) subgroups. MMSE scores in the MCIAll+ subgroup were significantly lower in comparison with those of the MCIAβ+ (p = 0.02), MCIAβ+T+ (p = 0.03) and MCIT+ (p < 0.01) subgroups. Across the five studied MCI subgroups, approximately 45% of patients had conflicting CSF biomarker constellations. Figure 1, a graphical presentation of participants’ Aβ42, t-Tau and p-Tau CSF levels using NNMF, points to the high variability of the CSF biomarker findings in patients with MCI. Data were available from clinical follow-up visits conducted every 6 months up to 8 years after baseline. In total, 159 patients with MCI progressed to dementia due to AD. No patient progressed to any other form of dementia. The difference between the MCI subgroups in the proportions of patients who developed dementia due to AD within the follow-up period attained statistical significance (p < 0.001), whilst the duration of the follow-up period did not differ.

Table 1 Characteristics of the study sample
Fig. 1
figure 1

Condensed representation, as a 2D scatter plot, of the CSF biomarker values of the study participants. The ensemble of trivariate measurements of CSF Aβ42, p-Tau and t-Tau for all participants was analysed via NNMF and approximated by means of a bivariate data swarm that conveniently represents the total variation in the original data. Labels indicate the different groups and lend semantics to the plot. MCI Non+ MCI without positive CSF biomarkers, MCI Aβ+ MCI with positive Aβ42 and negative or borderline p-Tau and t-Tau, MCI Aβ+T+ MCI with positive Aβ42 and positive t-Tau or p-Tau, MCI All+ MCI with Aβ42 and both t-Tau and p-Tau positive, MCI T+ MCI with negative or borderline Aβ42 and at least p-Tau or t-Tau positive

Cox regression analyses unveiled a significant association between group membership and risk of progression to AD dementia (P < 0.001), whilst sex, age and the presence of the APOE ε4 allele did not exert such an influence. As expected, lower MMSE scores were related to a higher hazard of developing dementia. The differences in progression risk between patients in the MCINon+ subgroup and those with positive Aβ42 values (MCIAβ+, MCIAβ+T+, MCIAll+ subgroups) reached statistical significance. In addition, the MCIAll+ subgroup was at significantly higher risk for progression compared with both the MCIAβ+T+ and MCIT+ subgroups. No further significant differences with regard to progression risk were detected between any other of the compared MCI subgroups (Table 2, Fig. 2). The risk pattern in relation to different biomarker constellations in MCI is presented in Fig. 3.

Table 2 Estimates of variables in Cox regression
Fig. 2
figure 2

Cox regression survival curves for patients with MCI and different constellations of CSF Aβ42 and neuronal injury markers (p-Tau and t-Tau). MCI Non+ MCI without positive CSF biomarkers, MCI Aβ+ MCI with positive Aβ42 and negative or borderline p-Tau and t-Tau, MCI Aβ+T+ MCI with positive Aβ42 and positive t-Tau or p-Tau, MCI All+ MCI with Aβ42 and both t-Tau and p-Tau positive, MCI T+ MCI with negative or borderline Aβ42 and at least p-Tau or t-Tau positive

Fig. 3
figure 3

Risk for progression to dementia due to AD of patients with MCI and different constellations of CSF Aβ42 and neuronal injury markers (p-Tau and t-Tau). AD Alzheimer’s disease, MCI Non+ MCI without positive CSF biomarkers, MCI Aβ+ MCI with positive Aβ42 and negative or borderline p-Tau and t-Tau, MCI Aβ+T+ MCI with positive Aβ42 and positive t-Tau or p-Tau, MCI All+ MCI with Aβ42 and both t-Tau and p-Tau positive, MCI T+ MCI with negative or borderline Aβ42 and at least p-Tau or t-Tau positive


In line with a number of previous reports [1719], but in contrast to others [8, 9, 34, 35], approximately half of the MCI cases in our study had conflicting CSF biomarker constellations. This discrepancy in the frequency of patients with conflicting CSF biomarker results could be possibly attributed to differences in study design. For instance, not all studies considered all three CSF AD biomarkers. In addition, past studies implemented a dichotomization strategy in interpreting biomarker findings, whilst in the present study biomarker values were categorized as positive, negative or borderline in line with the NIA-AA guidelines. Moreover, it is possible that academic, research centres recruit more patients with complex constellations of biomarker findings, whilst more patients with AD-typical CSF profiles and consequently more advanced neuropathology are recruited in non-academic, clinical settings. Interestingly, it has been shown that patients of a non-academic memory clinic suffered from more severe clinical symptoms in comparison with the patients of an academic memory clinic [36].

Constellations with conflicting CSF biomarker findings are not currently being considered by the NIA-AA criteria for MCI [16], and our study provides initial evidence on the role of conflicting CSF biomarker constellations for dementia risk estimation. Our findings indicate that compared with the constellation without positive biomarkers, the presence of Aβ42 positivity confers a higher risk for future AD dementia irrespective of t-Tau and p-Tau levels. Thus, it seems that Aβ42 is not only the first marker to become positive in the course of AD [37], but also the decisive marker to determine dementia risk. Despite the absence of general consensus, because a number of past reports point to a higher or almost similar prognostic utility of tau peptides compared with Aβ42 [20, 35, 3842], our observations are in line with several previous studies which showed that Aβ42 has a higher prognostic utility in comparison with tau or failed to find an association between tau and cognitive deterioration [5, 22, 43, 44]. It is important to mention, however, that our results cannot necessarily be generalized to all patients with MCI, because the ADNI MCI cohort is deliberately limited to those with prominent memory deficits in order to enrich the sample with pre-dementia AD cases. MCI is heterogeneous by definition, and the studied biomarkers may react differently in early non-AD cases. Hence, the confounding effects of other brain pathologies associated with increased p-Tau and/or t-Tau levels (such as cerebrovascular changes, Lewy bodies, etc.) are minimized in ADNI, so that the role of Aβ42 as an indicator of AD pathophysiology may be exaggerated. Furthermore, it should be underscored that our findings could have been biased by the artificial definition of the range of borderline biomarker values as well as by the fact that, due to sample size reasons, borderline values were not considered separately from negative values in our analyses. Interestingly, an alternative analysis considering borderline and positive values together resulted in an amelioration of the significance of the role of Aβ42 (the results of the alternative analysis are presented in Additional file 1). Hence, further studies with larger samples enabling the separate consideration of borderline values are warranted before definite conclusions can be drawn.

The higher dementia risk in the MCIAll+ subgroup compared with the other subgroups is in line with the proposed model of a temporal evolution of AD biomarkers as well as with a recently published, probabilistic, data-driven model of biomarker changes in sporadic AD, independent of a-priori patient staging and biomarker cut-off points [37, 45]. Our observation supports the assumption that as clinical symptoms advance and the threshold to dementia is reached, the abnormality of biomarkers becomes evident [46, 47]. The lack of significant difference in terms of dementia risk between the MCIAll+ and MCIAβ+ subgroups is probably a spurious finding, related to the relatively small size of the MCIAβ+ subgroup and/or the definition of the cut-off values. This assumption is supported by the highly significant difference between the MCIAll+ and MCIAβ+T+ subgroups, although the latter subgroup has a higher dementia risk compared with the MCIAβ+ subgroup as illustrated in Fig. 2. Moreover, the alternative analysis, in which borderline and positive biomarker values were treated as one group, unveiled a significant difference in progression risk between the MCI subgroup with all markers non-negative (positive or borderline) and the subgroup with only Aβ42 non-negative (Additional file 1). Thus, the observed lack of significant difference in dementia risk between the MCIAll+ and MCIAβ+ subgroups should be treated with caution. In addition, it is noteworthy that our results confirm the approach of the NIA-AA algorithm to assign the highest likelihood of AD to MCI patients with all biomarkers positive.

Our findings suggest that MCI patients with positive Aβ42 values are at the same risk for AD dementia whether or not they have one positive Tau marker (either t-Tau or p-Tau). Hence, patients with positive Aβ42 and non-positive or conflicting p-Tau/t-Tau levels, who cannot be categorized according to the current NIA-AA algorithm, seem to have the same dementia risk, which lies between that of the lowest and highest risk groups (MCINon+ and MCIAll+ respectively). However, this observation is in contrast with the findings of a large number of previous reports which have shown that MCI patients with two positive CSF markers have a higher risk to progress to dementia compared with MCI patients with only one positive biomarker [20, 22, 25, 28, 40, 44, 4851]. As a result, this finding should be treated with caution. It cannot be precluded that the observation of the present study has been biased by the definition of the cut-off points and/or by the relatively small size of the MCIAβ+ subgroup. Nonetheless, the alternative analysis in which borderline and positive biomarker values were treated as one group (non-negative values) did not reveal significant differences in the progression risk between the MCI subgroup with only Aβ42 non-negative and that with both Aβ42 and p-Tau or t-Tau non-negative (Additional file 1). Further studies are thus required in order to shed more light on the progression risk of the MCIAβ+ and MCIAβ+T+ subgroups.

In terms of dementia risk, MCI patients with one or two positive Tau markers but negative or borderline Aβ42 values (MCIT+ subgroup) may be placed between patients without positive biomarkers (MCINon+ subgroup) and those with positive Aβ42 and non-positive or conflicting p-Tau/t-Tau levels (MCIAβ+ and MCIAβ+T+ subgroup, respectively). The term “suspected non-AD pathophysiology (SNAP)” has recently been proposed to designate individuals with abnormal markers of neuronal injury without evidence of amyloid accumulation [10, 18, 52]. According to our findings, the progression risk of the MCIT+ subgroup does not significantly differ from that of the MCINon+ subgroup. Simultaneously, it is no different from that of the MCIAβ+ and MCIAβ+T+ subgroups. However, the dementia risk of the two latter subgroups does in fact differ from the MCINon+ subgroup. As a consequence, the MCIT+ risk can be placed between that of MCINon+ and MCIAβ+, MCIAβ+T+ subgroups (Fig. 3). Nonetheless, this finding should also be treated with caution due to the limited size of the ADNI MCIT+ subgroup and because a prior study showed that the highest proportion of subjects who progressed to dementia was observed not only in the MCI subgroup with both amyloid and neuronal injury markers positive for AD but also in the MCI subgroup with only neuronal injury biomarkers positive [10]. This discrepancy could be explained by the different markers of neuronal injury considered in the two studies (neurochemical vs imaging), especially in light of the pathophysiological character of the former and the downstream topographical character of the latter [53].

The present study should be viewed in the light of some limitations. Owing to the lack of histopathological verification, the main outcome measure was based purely on clinical diagnoses, which are not always confirmed at autopsy [54]. Moreover, the ADNI encompasses individuals recruited at specialized research centres and does not mirror constellations in the community. This is clearly illustrated by the fact that within the follow-up period only conversion to dementia due to AD and not to other forms of dementia was observed. Furthermore, in building the MCI subgroups with distinct biomarker constellations we did not consider borderline values separately from negative values, because such an approach would have expanded the number of MCI subgroups and reduce their size. This limitation could explain the relatively high proportion—in comparison with previous reports—of patients in the MCINon+ subgroup as well as in the MCIT+ subgroup who progressed to dementia due to AD [55], because it is possible that borderline biomarker values became positive shortly after baseline. In addition, it can be reckoned that our observations are biased by the artificial definition of the range of borderline biomarker values. In light of the lack of empirical data with regard to definitions of the range of borderline values, our findings should be treated with caution. Nonetheless, the NIA-AA guidelines clearly specify the presence of borderline biomarker values. As a consequence, further studies considering borderline biomarker values are warranted. A further shortcoming of the study is the collection of some early CSF samples into inappropriate tubes at the ADNI sites. Although the error was corrected rapidly and despite the relatively limited exposure time to any inappropriate CSF collection tube, this error could embody a source of bias because the use of different collection tubes increases intra-laboratory variability [34, 5658]. Moreover, we did not take into account imaging biomarker data. However, it should be underscored that while combining imaging with neurochemical biomarker data may be relevant for research settings, it is rarely applicable to clinical settings because of limitations related to scanner equipment and sophisticated image analyses expertise.


The present study provides a further piece of evidence for the prognostic differences between MCI subgroups with distinct neurochemical biomarker constellations. The study reveals significant differences between subgroups with conflicting biomarkers, on the one hand, and patients with all neurochemical biomarkers positive or non-positive (borderline or negative) for AD on the other. Even though our observations exclusively refer to neurochemical biomarkers and do not consider imaging markers, they point to the necessity of modifying/refining the NIA-AA algorithms for categorizing MCI.