Plasma and cerebrospinal fluid amyloid beta for the diagnosis of Alzheimer's disease dementia and other dementias in people with mild cognitive impairment (MCI)
Abstract
Background
According to the latest revised National Institute of Neurological and Communicative Disorders and Stroke and the Alzheimer's Disease and Related Disorders Association (now known as the Alzheimer's Association) (NINCDS‐ADRDA) diagnostic criteria for Alzheimer's disease dementia of the National Institute on Aging and Alzheimer Association, the confidence in diagnosing mild cognitive impairment (MCI) due to Alzheimer's disease dementia is raised with the application of biomarkers based on measures in the cerebrospinal fluid (CSF) or imaging. These tests, added to core clinical criteria, might increase the sensitivity or specificity of a testing strategy. However, the accuracy of biomarkers in the diagnosis of Alzheimer’s disease dementia and other dementias has not yet been systematically evaluated. A formal systematic evaluation of sensitivity, specificity, and other properties of plasma and CSF amyloid beta (Aß) biomarkers was performed.
Objectives
To determine the accuracy of plasma and CSF Aß levels for detecting those patients with MCI who would convert to Alzheimer's disease dementia or other forms of dementia over time.
Search methods
The most recent search for this review was performed on 3 December 2012. We searched MEDLINE (OvidSP), EMBASE (OvidSP), BIOSIS Previews (ISI Web of Knowledge), Web of Science and Conference Proceedings (ISI Web of Knowledge), PsycINFO (OvidSP), and LILACS (BIREME). We also requested a search of the Cochrane Register of Diagnostic Test Accuracy Studies (managed by the Cochrane Renal Group).
No language or date restrictions were applied to the electronic searches and methodological filters were not used so as to maximise sensitivity.
Selection criteria
We selected those studies that had prospectively well defined cohorts with any accepted definition of cognitive decline, but no dementia, with baseline CSF or plasma Aß levels, or both, documented at or around the time the above diagnoses were made. We also included studies which looked at data from those cohorts retrospectively, and which contained sufficient data to construct two by two tables expressing plasma and CSF Aß biomarker results by disease status. Moreover, studies were only selected if they applied a reference standard for Alzheimer's dementia diagnosis, for example the NINCDS‐ADRDA or Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM‐IV) criteria.
Data collection and analysis
We screened all titles generated by the electronic database searches. Two review authors independently assessed the abstracts of all potentially relevant studies. We assessed the identified full papers for eligibility and extracted data to create standard two by two tables. Two independent assessors performed quality assessment using the QUADAS‐2 tool. Where data allowed, we derived estimates of sensitivity at fixed values of specificity from the model we fitted to produce the summary receiver operating characteristic (ROC) curve.
Main results
Alzheimer's disease dementia was evaluated in 14 studies using CSF Aß42. Of the 1349 participants included in the meta‐analysis, 436 developed Alzheimer’s dementia. Individual study estimates of sensitivity were between 36% and 100% while the specificities were between 29% and 91%. Because of the variation in assay thresholds, we did not estimate summary sensitivity and specificity. However, we derived estimates of sensitivity at fixed values of specificity from the model we fitted to produce the summary ROC curve. At the median specificity of 64%, the sensitivity was 81% (95% CI 72 to 87). This equated to a positive likelihood ratio (LR+) of 2.22 (95% CI 2.00 to 2.47) and a negative likelihood ratio (LR–) of 0.31 (95% CI 0.21 to 0.48).
The accuracy of CSF Aß42 for all forms of dementia was evaluated in four studies. Of the 464 participants examined, 188 developed a form of dementia (Alzheimer’s disease and other forms of dementia).The thresholds used were between 209 mg/ml and 512 ng/ml. The sensitivities were between 56% and 75% while the specificities were between 47% and 76%. At the median specificity of 75%, the sensitivity was estimated to be 63% (95% CI 22 to 91) from the meta‐analytic model. This equated to a LR+ of 2.51 (95% CI 1.30 to 4.86) and a LR– of 0.50 (95% CI 0.16 to 1.51).
The accuracy of CSF Aß42 for non‐Alzheimer's disease dementia was evaluated in three studies. Of the 385 participants examined, 61 developed non‐Alzheimer's disease dementia. Since there were very few studies and considerable variation between studies, the results were not meta‐analysed. The sensitivities were between 8% and 63% while the specificities were between 35% and 67%.
Only one study examined the accuracy of plasma Aß42 and the plasma Aß42/Aß40 ratio for Alzheimer's disease dementia. The sensitivity of 86% (95% CI 81 to 90) was the same for both tests while the specificities were 50% (95% CI 44 to 55) and 70% (95% CI 64 to 75) for plasma Aß42 and the plasma Aß42/Aß40 ratio respectively. Of the 565 participants examined, 245 developed Alzheimer’s dementia and 87 non‐Alzheimer's disease dementia.
There was substantial heterogeneity between studies. The accuracy of Aß42 for the diagnosis of Alzheimer's disease dementia did not differ significantly (P = 0.8) between studies that pre‐specified the threshold for determining test positivity (n = 6) and those that only determined the threshold at follow‐up (n = 8). One study excluded a sample of MCI non‐Alzheimer's disease dementia converters from their analysis. In sensitivity analyses, the exclusion of this study had no impact on our findings. The exclusion of eight studies (950 patients) that were considered at high (n = 3) or unclear (n = 5) risk of bias for the patient selection domain also made no difference to our findings.
Authors' conclusions
The proposed diagnostic criteria for prodromal dementia and MCI due to Alzheimer's disease, although still being debated, would be fulfilled where there is both core clinical and cognitive criteria and a single biomarker abnormality. From our review, the measure of abnormally low CSF Aß levels has very little diagnostic benefit with likelihood ratios suggesting only marginal clinical utility. The quality of reports was also poor, and thresholds and length of follow‐up were inconsistent. We conclude that when applied to a population of patients with MCI, CSF Aß levels cannot be recommended as an accurate test for Alzheimer's disease.
Plain language summary
Proteins in blood and cerebrospinal fluids for early prediction of developing Alzheimer’s disease or other dementia in people with cognitive problems
The numbers of people with dementia and other cognitive problems are increasing globally. A diagnosis of the pre‐dementia phase of disease is recommended but there is no agreement on the best approach. A range of tests have been developed which healthcare professionals can use to assess people with poor memory or cognitive impairment. In this review, however, we have found that measuring protein in cerebrospinal fluid (CSF amyloid beta (Aβ40) or CSF Aβ42), as a single test, lacks the accuracy to identify those patients with mild cognitive impairment who would develop Alzheimer's disease dementia or other forms of dementia.
Authors' conclusions
Implications for practice
In 2009, Petersen wrote that the use of CSF biomarkers was of value in clinical research, specifically stratification for clinical trials, but lacked the necessary accuracy for clinical practice. This meta‐analysis endorses that view. The proposed criteria for prodromal dementia would allow a diagnosis of prodromal Alzheimer's dementia to be made in the presence of low CSF Aβ in a person with significant amnestic symptoms. This review does not support this. It is possibly the case that in concert with other biomarker tests, be they imaging or other CSF markers, a reasonable accuracy can be obtained. The proposed criteria should be adapted at this stage in light of our findings that as a single test CSF Aß lacks the accuracy to diagnose Alzheimer's disease in patients with cognitive impairment but no dementia. That said, sequential testing with different tests is limited by the important assumption that each of the tests is conditionally independent of each other; this would not be the case for example with PET‐PiB and CSF Aβ, which are highly correlated measures of the same pathological process. To illustrate the lack of value of this test, if over a three‐year period 20% of people with MCI will go on to develop Alzheimer’s dementia, the pre‐test probability is 20% on the basis of the clinical diagnosis alone (Mitchell 2009). In these patients, a low level of CSF Aβ42 and a LR of 2.22 would yield a post‐test probability of 36% or an increment in risk of developing Alzheimer's dementia of 16%. Conversely, given a negative test result and a negative LR of 0.31, the post‐test probability of that person developing Alzheimer’s dementia is 7.2%; a decrement in risk of developing Alzheimer's dementia of 12.8%. There have been recent initiatives to screen primary care populations for cognitive impairment and refer those who screen positive to specialist care for further evaluation. In effect this may identify large populations with cognitive impairment but no dementia who are not seeking health care themselves. It is known that when applying the Petersen criteria in population and primary care settings the conversion rate to Alzheimer's dementia is lower than in secondary care services (Mitchell 2009). The effect therefore of identification of an at risk population in primary care would be to in effect reduce the pre‐test probability. As likelihood ratios are a function of the test, the effect of reducing the pre‐test probability on the incremental benefit of abnormal CSF Aβ42 is shown in Figure 11. This shows that the incremental benefit of an abnormal CSF Aß level in a patient identified through screening of primary care populations is substantially less than the already low figures achieved in a secondary care population. This is a theoretical conclusion as no studies in our sample were conducted in an explicitly defined primary care population, however in primary care it is likely that the pre‐test probability for Alzheimer's dementia is lower than it would be in a secondary care 'memory clinic' population. The optimal pre‐test probability for incremental benefit of a positive test is 40%, and 64% for a negative test, but even at this optimum level the benefits are modest. The results of this review are that abnormal CSF Aß levels are of little value in patients referred to specialists services and of even less value in those in whom cognitive impairment but no dementia is identified through primary care screening.
Pre‐ and post‐ tests probabilities of Alzheimer's dementia using low CSF amyloid beta 42 as a diagnostic test.
Implications for research
The observations we have made regarding the utility of CSF Aß in identifying Alzheimer's disease prior to the onset of dementia suggest that at least at the MCI phase of illness there is little utility in its use. However, future research may identify that at even earlier stages (younger individuals) an accumulation of Aß in the brain as manifested through low CSF Aß will have greater specificity for indicating pathology rather than being a consequence of normal ageing in the elderly. Moreover, more uniform approaches to thresholds, analysis and study conduct, in particular uniformity of length of follow‐up, may provide a more homogenous estimate than was available here. The exploration for better, more accessible and more accurate biomarkers with, in particular, better specificity are urgently required. This research effort should be stimulated by the observation that CSF Aß does not appear to have the accuracy to draw to a conclusion the search for more accurate biomarkers for Alzheimer's disease. However, where clinical trials using specific anti‐amyloid therapies would benefit from small incremental changes towards improved post‐test probabilities, then CSF Aß may be useful in enriching the population. Our work, summarised in Figure 11, perhaps illustrates best the additional accuracy that can be achieved by using this test and the trial sponsors can apply this empirical evidence in their selection criteria for studies.
Summary of findings
| What is the diagnostic accuracy of plasma and CSF amyloid biomarker levels for detecting Alzheimer's disease pathology in patients with mild cognitive impairment (MCI), and identifying those MCI participants who would convert to Alzheimers disease dementia or other forms of dementia over time | |||||||
| Patient population | Participants diagnosed with MCI at baseline using any of the Petersen criteria or CDR = 0.5 or any 16 definitions included by Matthews (Matthews 2008) | ||||||
| Prior testing | The only testing prior to performing the plasma and CSF biomarkers was the application of diagnostic criteria for identifying participants with MCI | ||||||
| Settings | Participants were recruited from i) secondary care – outpatient clinic (n = 12); ii) secondary care – inpatients (n = 2); iii) community care (n = 2) and mixed setting (n = 1) | ||||||
| Index tests | Plasma or CSF Aß 42, Aß 40, or Aß 42/Aß 40 ratio | ||||||
| Reference standard | NINCDS‐ADRDA or DSM or ICD criteria for Alzheimer's disease dementia; McKeith criteria for Lewy body dementia; Lund criteria for frontotemporal dementia; and NINDS AIREN criteria for vascular dementia | ||||||
| Target condition | Alzheimer’s disease dementia or any other form of dementia | ||||||
| Included studies | 17 studies (2228 participants) of prospectively well defined cohorts with any accepted definition of MCI were included | ||||||
| Quality concerns | Patient selection and conduct of the index and reference standard were poorly reported. Applicability concerns were generally low. Regarding the inclusion criteria set in the review, each included study did match the review question: 'Could Plasma and CSF Aßbiomarkers identify those MCI participant with Alzheimer’s disease pathology at baseline who would convert clinically to dementia at follow up?' However, due to limited number of included studies and levels of heterogeneity it is difficult to determine to what extent the findings from a meta‐analysis can be applied to clinical practice. | ||||||
| Limitations | Limited investigation of heterogeneity due to insufficient number of studies. There was a lack of common thresholds and poor reporting of thresholds. | ||||||
| Test | Studies | Cases/participants | Median specificity from included studies | Sensitivity (95% CI)1 at median specificity | Consequences in a cohort of 100 | ||
| Median percentage converting (range)2 | Missed cases3 | Overdiagnosed 3 | |||||
| Alzheimer's disease dementia | |||||||
| CSF Aß 42 | 14 | 436/1349 | 64 | 81 (72, 87) | 38 (9 to 56) | 7 | 22 |
| All forms of dementia | |||||||
| CSF Aß 42 | 4 | 188/464 | 75 | 63 (22, 91) | 45 (27 to 53) | 17 | 14 |
| Non‐Alzheimer's disease dementia | |||||||
| CSF Aß 42 | 3 | 61/385 | 65 | No meta‐analysis | 16 (15 to 19) | ||
| Investigation of heterogeneity: the planned investigations were not possible due to the limited number of studies available for each analysis. We were unable to investigate the effect of duration of follow up due to substantial variation in length and reporting. | |||||||
| Conclusions: from our review, abnormally low CSF Aß levels has very little diagnostic benefit with likelihood ratios suggesting only marginal clinical utility. The quality of reports was also poor, and thresholds and length of follow up were inconsistent. We conclude that when applied to a population of patients with MCI, CSF Aß levels cannot be recommended as an accurate test for Alzheimer's disease. | |||||||
| 1 Meta‐analytic estimate of sensitivity derived from the HSROC model at a the median value of specificity computed from the included studies. Summary estimates of sensitivity and specificity were not computed because the studies that contributed to the estimation of the summary ROC curve used various thresholds. 2 The median percentage converting and range were computed using all the studies included in the analysis for each target condition. 3 Missed and overdiagnosed numbers were computed using the median percentage converting for each target condition. | |||||||
Background
Alzheimer’s disease is an incurable, progressive, neurodegenerative condition which accounts for over 50% of dementias, afflicting 5% of men and 6% of women over 60 years old worldwide (World Health Organization 2010). As age is the principal risk factor, the general ageing of the global population, despite some suggestions that dementia incidence is falling, means that the total number of people suffering from Alzheimer's dementia is likely to still be increasing. The prevalence increases exponentially with age as Alzheimer's dementia affects < 1% of people aged 60 to 64 years old and 24% to 33% of those aged over 85 years (Ferri 2005).
It is important to distinguish Alzheimer's disease (which is the underlying pathology) from Alzheimer's dementia, which is the final stage of a clinical syndrome that develops as a result of the pathology. The diagnosis of pre‐dementia Alzheimer's disease is currently being debated and the best nosology to define this stage of illness lacks consensus, especially with regards to the use of various terms (for example prodromal dementia) in clinical practice. It is also important to be able to distinguish between the various subtypes of dementia as early as possible in the course of illness to maximise the impact of treatment and risk modification.
The target condition being diagnosed by the testing of cerebrospinal fluid (CSF) and plasma for amyloid beta (Aß) 40 or 42 and the 40/42 ratio is Alzheimer's disease both as a distinct entity from normal ageing as well as from other subtypes of dementia (for example vascular dementia, Lewy body dementia and frontotemporal dementias). In this review we aimed to determine whether CSF or plasma Aß levels were diagnostic of Alzheimer's disease in people with cognitive impairment but no dementia. Subsequent reviews will explore the accuracy of this test to discriminate between subtypes of dementia (Kokkinou 2014).
The first complaints that people make (either to their general practitioner (GP) or to family) are often regarding subjective cognitive complaints such as memory lapses or getting lost. This may lead to a diagnosis of mild cognitive impairment (MCI) being made if formal testing reveals objective evidence of cognitive impairment. In some people, this will progress to Alzheimer's dementia, though currently we are unable to accurately predict those who will progress. This has led to the development of numerous research programmes seeking to define clinical, cognitive and biomarker tests that can identify those with MCI who will develop dementia. The presence of Alzheimer's disease, which may be indexed by low levels of CSF or plasma Aß, in patients with early cognitive complaints is thereby considered to be a very strong risk factor for developing Alzheimer's dementia in the future. Therefore, the presence of Alzheimer's disease within patients with MCI is proposed as being a diagnostic test for the delayed verification of Alzheimer's disease dementia (Davis 2013).
Alzheimer's pathology is associated with a central amyloidosis and for many years the amyloid cascade hypothesis (Hardy 1992) has been used to describe the disruption of a probably normal cerebral process which is associated with ageing and how this disruption can lead to Alzheimer's disease. The model describes a pathway by which Aß monomers are produced, aggregate into oligomers and then are sequestered and potentially inactivated into the characteristic extracellular amyloid plaques. This may explain why on neuroimaging some cognitively ‘normal’ individuals have amyloid plaques after the age of 65 years (Aizenstein 2008). In Alzheimer's disease, however, the equilibrium of this system is biased to pathological, away from a physiological function, and an increased conversion to the Aß oligomers from the ‘inert’ monomers results in a toxic pool of soluble amyloid species which are central to the neurotoxicity of Alzheimer's disease. The amount of oligomeric Aß is too great to be sequestered into plaque and thereby remains present in the extracellular space to exert its toxic effects. As a consequence of neuronal disruption from these oligomeric species, intracellular cascades may lead to the hyperphosphorylation of tau, compromising the cytoskeletal protein tubulin with the development of neuro‐fibrillary tangles (NFTs). Toxic oligomers may also compromise synaptic activity leading in part to the clinical symptoms of MCI and eventually Alzheimer's disease dementia. There is also an associated inflammatory response from glial cells causing oxidative stress, synaptotoxicity and (via excess glutamate release) neurotoxicity. At the latter stages of disease progression the senile plaques are believed to also be pro‐inflammatory and thereby stimulate a further neurotoxic inflammatory response.
The amyloid hypothesis has yet to be decisively proven however (Shankar 2008) and one conflicting argument states that the increased production of Aß in Alzheimer's disease is a protective mechanism, due to its potent antioxidant ability, to counter a primary upstream effect, with mitochondrial dysfunction being heavily implicated (Castellani 2004). Moreover, the amyloid hypothesis has been questioned in light of recent studies and drug trials which have indicated no impact on the clinical course of disease after significant reduction in post‐mortem cerebral amyloid following active immunisation (Holmes 2008).
There are two main pathological isoforms of Aß, defined by their amino acid length: 40 and 42. Aß42 has a greater capacity to form oligomers and thereafter fibrils (the main constituent of amyloid plaques) and therefore has a higher neurotoxicity than its shorter counterpart Aß40. Successive ß‐ and ƴ‐secretase cleavage of the ubiquitous amyloid precursor protein (APP) is responsible for Aß42 production. The ɑ‐secretase cleavage of APP does not form either of these amyloid proteins but produces a much more soluble shorter moiety, Aß17, and its action is relatively suppressed in Alzheimer's disease. Despite the increased production of Aß42 in Alzheimer's disease, the trafficking of Aß42 into plaques may in fact lead to an observable lowering of the CSF and plasma levels of Aß42 with maintenance of the smaller, more soluble species in Alzheimer's disease compared to age‐matched controls. Therefore studies have shown that consistently there is a lowering of CSF levels of Aß42 in Alzheimer's disease dementia compared with controls, which may be indicative of extant central Aß pathology. Hypothetically, as the Aß42 levels decrease to a greater degree than the Aß40 levels, it follows that the ratio of Aß42:Aß40 will decrease, which may prove to be a more accurate test for Alzheimer's disease pathology than the absolute values of each protein in plasma or CSF.
It is worth noting that the Aß found in plasma is likely to have emanated from the alpha granules of platelets, where APP is also found in abundance, and therefore may be less closely related to the Alzheimer's disease central pathology. Little concurrence has been found between plasma and CSF Aß levels (Mehta 2001), therefore CSF Aß assays may prove to be a more accurate reflection of central amyloid pathology associated with Alzheimer's disease. However, as plasma is a much more accessible bodily fluid than CSF the accuracy of plasma Aß in diagnosing Alzheimer's disease pathology merits review.
There have been numerous attempts to clinically define the pre‐dementia phase of neurodegenerative disease. MCI is a heterogeneous condition which has been defined, using clinical criteria, by several authors, for example Petersen 1999. There have been over a dozen different definitions used to describe cognitive impairment that exceeds in extent and is somehow qualitatively different from normal ageing. Whilst remaining attentive to the widely differing prognostic implications of the differing terms, in this review the term MCI will be used to collectively describe the 16 conditions included in the research by Matthews 2008 (Matthews 2008). This leads to the proposal that from within these populations additional tests may help support the notion that there is central pathology that will mediate (over time) progression to one of the dementia syndromes. Knowing that a person with MCI has Alzheimer's disease as the cause of their symptoms would allow for the targeting of (future) disease modifying therapies, risk modifying strategies as well as psychosocial management.
There is currently no pre‐mortem gold standard test for Alzheimer's disease pathology. The observation that a patient with MCI, when followed carefully over time, develops Alzheimer's disease dementia can be taken (for the purposes of this review) that they had Alzheimer's disease pathology at baseline (that is when diagnosed with MCI) and it was this Alzheimer's pathology that caused the symptoms observed in MCI. However, since vascular events and Lewy body pathology may not be independent of amyloid pathology, it is also important to examine whether the index test predicts all causes of dementia and whether the accuracy in predicting these outcomes is different from the accuracy of predicting Alzheimer's disease dementia.
Inevitably this means that studies exploring the diagnostic accuracy of a test in MCI for Alzheimer's disease pathology will have a long time interval between the diagnostic test and the reference standard being applied (that is conversion to Alzheimer's disease dementia). However, as the course of this illness cannot be interrupted or affected by any therapeutic intervention, this is not as problematic as it would be in other conditions where interventions have an effect on the course of disease. Accordingly, studies must be long enough to allow 'conversion' from MCI to Alzheimer's disease dementia to occur and must use standardised criteria both to define the baseline population (MCI) and the conversion to Alzheimer's disease dementia.
In essence this review aims to understand the diagnostic test accuracy of Aß40 or Aß42, or their ratio, in CSF or plasma in patients with any of the described forms of MCI to identify them as having Alzheimer's disease pathology.
Target condition being diagnosed
-
Alzheimer's disease dementia
-
Other forms of dementia
Index test(s)
Studies that assessed the accuracy of plasma or CSF measurements of:
Aß42, or
Aß40, or
Aß42 to Aß40 ratio.
Aß is measured in ng.l‐1 or pg.ml‐1, which generate the same values.
The assays most commonly used are the conventional Innogenetics INNOTEST® beta‐amyloid1‐42 kit or the multiplexing INNO‐BIA AlzBio3 for CSF, or INNO‐BIA plasma Aß forms for plasma.
Clinical pathway
Dementia develops over a trajectory of several years. There is a presumed period when people are asymptomatic, and when pathology is accumulating. Individuals or their relatives may notice subtle impairments of recent memory. Gradually more cognitive domains become involved and difficulty planning complex tasks becomes increasingly apparent. In the UK, people often present to their GP when they or a family member or friend note memory deficits. The GP will potentially refer them to a memory clinic after taking a history of the problem and conducting a brief assessment of cognitive function. However, many people with dementia do not present to their GP for several years after the first onset of symptoms and will follow a different pathway to diagnosis, for example being identified during an admission to a general hospital for a physical illness.
Access to diagnostic assessment pathways may vary in other healthcare settings and diagnoses may be made by a variety of specialists including neurologists and geriatricians, who may rely to a greater or lesser degree on tests to assist with the diagnosis. In recent years there have been attempts across the world to identify dementia at the earliest stage, with an emphasis on the GP being vigilant in observing their patients for cognitive decline and then making referrals to specialist services. Accordingly, there is a growing interest in the accuracy of cognitive tests to support a referral and then of biomarkers (such as in imaging and in plasma or CSF) applied in specialist centres.
Alternative test(s)
We did not include alternative tests in this review because there are currently no standard practice tests available for the diagnosis of dementia.
This review is one of a series of diagnostic test accuracy reviews of biomarkers and scales being conducted by the Cochrane Dementia and Cognitive Improvement Group (CDCIG):
-
CSF (cerebrospinal fluid analysis of tau and tau/Aß ratio);
-
sMRI (structural magnetic resonance imaging);
-
18F‐2‐fluoro‐2‐deoxy‐D‐glucose positron emission tomography (18F‐FDG‐PET);
-
positron emission tomography Pittsburg Compound‐C (11C‐PIB‐PET);
-
neuropsychological tests (Mini‐Mental State Examination (MMSE); Mini‐Cog; Montreal Cognitive Assessment (MoCA));
-
informant interviews (IQCODE; AD8);
-
apolipoprotein Ɛ4 (APOEƐ4);
-
regional cerebral blood flow single photon emission computerised tomography (rCBF SPECT).
Although we are conducting reviews on individual tests compared to a reference standard, we plan to compare the results from the reviews in an overview.
Rationale
The two recently proposed diagnostic criteria for Alzheimer’s disease referred to MCI due to Alzheimer’s disease pathology and the prodromal (pre‐dementia phase) of Alzheimer's disease pathology (Albert 2011; Dubois 2010) and incorporate biomarkers based on imaging or CSF measures within the diagnostic rubric. These tests are core to the criteria, assuming they will improve the specificity of the traditional solely clinical criteria. It is crucial that each of these biomarkers is assessed for their diagnostic accuracy before they are adopted as routine tests in clinical practice. It is worth noting that in each of these criteria a single abnormality in any of the proposed biomarker or imaging tests is considered sufficient to make a diagnosis of prodromal Alzheimer’s disease dementia.
Underpinning the new criteria is the assumption that if Alzheimer’s disease pathology can be diagnosed at an earlier, pre‐dementia stage, this could open critical windows for interventions that will have a greater likelihood of success in affecting disease pathways and thereby improving clinical symptoms. Earlier, accurate diagnosis will also help people with pre‐dementia cognitive impairment and their families and potential carers make timely plans for the future. Coupled with appropriate contingency planning, proper recognition of the disease may also help to prevent inappropriate and potentially harmful admissions to hospital or institutional care (Bourne 2007). In addition, the accurate early identification of a dementia syndrome may improve opportunities for the use of newly evolving interventions designed to delay or prevent progression to more debilitating stages of dementia.
Objectives
-
To determine the accuracy of plasma and CSF Aß levels in identifying those participants with MCI at baseline who will convert to Alzheimer’s disease dementia or other forms of dementia over time.
Secondary objectives
-
To determine the nature, extent and impact of heterogeneity on the diagnostic accuracy of plasma and CSF Aß levels.
Methods
Criteria for considering studies for this review
Types of studies
We included all studies that had prospectively well defined cohorts with any accepted definition of cognitive decline but no dementia with baseline CSF or plasma Aß levels, or both, documented at or around the time that the above diagnoses were made. We also included studies which looked at data from those cohorts retrospectively, and contained sufficient data to construct 2 × 2 tables expressing plasma and CSF Aß biomarker results by disease status. The results in those that progressed to clinical Alzheimer's dementia were compared to those who did not progress, improved or developed another dementia.
We excluded any studies which potentially overlapped patient data and recorded them as multiple publications. We also excluded review papers.
Participants
We included participants who were diagnosed with a cognitive decline but with no dementia condition at baseline.
The cognitive decline but no dementia group is defined as patients who have been diagnosed using any of the Petersen criteria, or Clinical Dementia Rating (CDR) = 0.5, or any of the 16 definitions included by Matthews (Matthews 2008).
We excluded those patients or populations with cognitive decline no dementia possibly caused by:
-
current or a history of alcohol or drug abuse;
-
central nervous system (CNS) trauma (e.g. subdural haematoma), tumour or infection;
-
other neurological conditions e.g. Parkinson’s or Huntingdon’s diseases;
-
any other dementia co‐morbidity e.g. frontotemporal or vascular dementias.
We excluded studies that included patients with psychiatric, neurological, metabolic, immunological, hormonal or cerebrovascular disorders, or patients likely to have a genetic cause for their dementia (for example familial autosomal dominant Alzheimer’s disease or frontotemporal dementia, cerebral autosomal dominant arteriopathy with subcortical infarcts and leukoencephalopathy, or Huntington’s disease).
If participants were involved in disease modifying clinical trials we excluded them from the analysis.
Patients with a family history of Alzheimer's disease dementia may be more readily diagnosed with MCI and therefore there can be spectrum bias with these participants. On this basis we examined separately those studies which included those with a known genetic predisposition.
Similarly, early onset Alzheimer's disease dementia, defined as a diagnosis under the age of 50 years, is likely to indicate a different aetiopathogenesis from late onset Alzheimer's disease dementia (including autosomally inherited mutation in the presenilin 1 and 2 genes, or the amyloid precursor protein (Filley 2007)). We excluded studies which included patients below the age of 50 years.
With regard to duration of follow‐up, Bruscoli 2004 indicated that an annual average of 10% of MCI patients become demented and therefore cohort studies would ideally have a minimum follow‐up time of 24 months. This provides a conversion rate to Alzheimer's disease dementia of approximately 19% and gives sufficient time for clinicians to diagnose a cause of MCI. Shorter durations of follow‐up may yield low conversion rates.
Index tests
Studies that assessed the accuracy of plasma or CSF measurements of:
Aß42, or Aß40, or the ratio Aß42/Aß40 were included.
There are currently no generally accepted standards for the plasma or CSF Aß test threshold, and therefore it was not possible to pre‐specify what constituted a positive or negative result. The thresholds that we used in this review were those generated and presented within each included study. It should be noted that where Aß42 and Aß40 data were presented separately, rather than as a ratio, then both tests would be analysed separately.
Measure of index test: Aß level in plasma or CSF, or both (ng.l‐1 or pg.ml‐1).
The assays most commonly used are conventional Innogenetics INNOTEST beta‐amyloid1‐42 kit or the multiplexing INNO‐BIA AlzBio3 for CSF, or INNO‐BIA plasma Aß forms for plasma.
We did not include a comparator test because there are currently no standard practice tests available for the diagnosis of dementia.
Target conditions
There were two target conditions in this review:
-
Alzheimer’s disease dementia (conversion from cognitive decline no dementia to Alzheimer’s disease dementia);
-
any other forms of dementia (conversion from cognitive decline no dementia to other (non‐Alzheimer's) forms of dementia);
-
any dementia (conversion from cognitive decline no dementia to any form of dementia).
Reference standards
The gold standard for Alzheimer's disease dementia is a post‐mortem. There is no ‘gold‐standard’ ante‐mortem diagnostic test for Alzheimer's disease dementia. Therefore, the reference standard in this review was conversion from MCI to Alzheimer's disease dementia based on the National Institute for National Institute of Neurological and Communicative Diseases and Stroke/Alzheimer's Disease and Related Disorders Association (NINCDS‐ADRDA) criteria (McKhann 1984) to define ‘probable’ Alzheimer's disease.
The NINCDS‐ADRDA criteria define three ante‐mortem groups: probable, possible and unlikely Alzheimer's disease dementia, based on evaluation of eight cognitive domains including memory and language.
The post‐consensus (on those cases in which there is a disagreement) specificity of the NINCDS‐ADRDA criteria is 0.84 and its sensitivity is 0.83 (Blacker 1994) against post‐mortem diagnosis. Exploring the impact of using different post‐mortem definitions of Alzheimer's disease, Nagy 1998 showed that the sensitivity of ‘possible’ and ‘probable’ dementia of the Alzheimer’s type according to NINCDS‐ADRDA criteria is 91% to 98% compared to the post‐mortem Khachaturian criteria, Tierney A3 criteria and the CERAD protocol at autopsy. Studies using Diagnostic and Statistical Manual of Mental Disorders criteria (DSM) (DSMIII 1987; DSMIV 1994) or International Statistical Classification of Diseases and Related Health Problems (ICD) criteria (World Health Organization 2010) exclusively to diagnose Alzheimer's disease dementia were also considered though the criteria would be entered as a covariate to assess their impact on results.
For Lewy body dementia as an outcome the reference standard is the McKeith criteria (McKeith 1996); for frontotemporal dementia, the Lund Criteria (Lund 1994); and for vascular dementia the NINDS AIREN criteria (Roman 1993).
Search methods for identification of studies
Electronic searches
The most recent search for this review was performed on 3 December 2012. We searched MEDLINE (OvidSP), EMBASE (OvidSP), BIOSIS Previews (ISI Web of Knowledge), Web of Science and Conference Proceedings (ISI Web of Knowledge), PsycINFO (OvidSP), and LILACS (BIREME). See Appendix 1 for details of the sources searched, the search strategies used, and the number of hits that were retrieved.
We also requested a search of the Cochrane Register of Diagnostic Test Accuracy Studies (managed by the Cochrane Renal Group).
We did not apply any language or date restrictions to the electronic searches; methodological filters were not used so as to maximise sensitivity.
Searching other resources
We checked the reference lists of all relevant studies for additional studies.
We also conducted searches in the MEDION database (Meta‐analyses van Diagnostisch Onderzoek) at www.mediondatabase.nl, Database of Abstracts of Reviews of Effects (DARE) at www.york.ac.uk/inst/crd/crddatabases.htm#DARE, Health Technology Assessments Database (HTA Database) at www.york.ac.uk/inst/crd/crddatabases.htm#HTA, and Aggressive Research Intelligence Facility (ARIF) database at www.arif.bham.ac.uk for other related systematic diagnostic accuracy reviews; we searched for systematic reviews of diagnostic studies from the International Federation of Clinical Chemistry and Laboratory Medicine Committee for Evidence‐based Laboratory Medicine database (C‐EBLM). We checked reference lists of any relevant systematic reviews for additional studies.
Data collection and analysis
Selection of studies
We selected studies initially from title and abstract screening by the review authors. We excluded articles on animal studies at this stage. We then obtained the full text for each potentially eligible study. We independently assessed these papers against the inclusion criteria for inclusion or exclusion. We resolved disagreements by discussion with a third author.
Data extraction and management
We extracted data to a study specific proforma which included the following.
-
Author, year of publication and journal.
-
The index test and assay type used (thresholds used to define positive and negative tests).
-
The criteria used for clinical definition for the baseline population.
-
Baseline demographics of the study population (age, gender, apolipoprotein E (ApoE) status, MMSE and clinical setting).
-
The duration of follow‐up (mean, minimum, maximum and median).
-
The proportion of patients developing the outcome of interest (Alzheimer's dementia using NINCDS‐ADRDA criteria) as well as other dementias where standard criteria were used.
-
The sensitivity and specificity of the index test in defining Alzheimer's dementia was used to back‐translate into a 2 x 2 table.
-
Other data relevant for creating 2 x 2 tables (TP = true test positive; FP = false test positive; FN = false test negative; TN = true test negative) e.g. a number of 'abnormal' and 'normal' tests and at baseline; a number of disease 'presence' and disease 'absence' at follow‐up, as well as through scrutiny of scatter plots.
We piloted the proforma against two primary diagnostic studies (Bjerke 2009; Hampel 2004) and amended it as necessary. We extracted data independently with disagreements then resolved by a third author.
Assessment of methodological quality
One review author and an independent assessor performed methodological quality assessments of each study using the QUADAS‐2 tool (Whiting 2011) as recommended by the Cochrane Collaboration. The tool is made up of four domains: patient selection, index test, reference standard and patient flow. Each domain is assessed in terms of risk of bias, with the first three domains also considered in terms of applicability concerns (http://www.bris.ac.uk/quadas/quadas‐2 Appendix 2). The components of each of these domains and a rubric which details how judgments concerning risk of bias are made are detailed in Appendix 3. Certain key areas important for this review regarding quality assessment were participant selection, blinding and missing data.
Statistical analysis and data synthesis
We performed separate analyses for each test and for each form of dementia (Alzheimer's disease dementia, non‐Alzheimer's disease dementia, and all forms of dementia). The disease negatives in each of these analyses were formed by all participants who did not develop the disease of interest. For example, where Alzheimer's disease dementia was the outcome, a participant developing another form of dementia would be in the disease negative group. We conducted exploratory analyses by plotting estimates of sensitivity and specificity from each study on forest plots and in receiver operating characteristic (ROC) space. To summarise test accuracy data across studies, we fitted hierarchical summary receiver operating characteristic (HSROC) models using the NLMIXED procedure in the SAS software package (version 9.2; SAS Institute, Cary, NC). The HSROC model accounts for between study variability through the inclusion of random effects that allow for heterogeneity in threshold and accuracy. Studies with different thresholds can be included (one threshold per study) in the HSROC model for the estimation of a summary ROC curve. A summary point can be identified on the summary ROC curve but summary estimates of sensitivity and specificity only have a clinically meaningful interpretation at a specific threshold. Therefore, in analyses where inclusion of studies was unrestricted by threshold, we used HSROC model parameters to derive sensitivities, with 95% confidence intervals, at median, lower and upper quartile values of the specificities from the included studies. When there were few studies and it was not possible to fit the complete HSROC model, we simplified the model by assuming a symmetrical summary ROC curve or fixed‐effect estimates, or both. In additional analyses, we planned to restrict analyses to only those studies that reported data at a common threshold if there were a sufficient number of studies.
Investigations of heterogeneity
The main sources of heterogeneity considered a priori were:
-
differences in test thresholds;
-
which proprietary laboratory tests were used to undertake the CSF and plasma analyses;
-
duration of follow‐up: we planned to perform a subgroup analysis of short (< 2 years) and longer (2 or more years) duration of follow‐up;
-
criteria used for definition of cognitive impairment and dementia;
-
age of participants.
The HSROC model can be extended to include covariates to assess whether threshold, accuracy, or the shape of the summary ROC curve varies with patient or study characteristics. Where possible, we investigated the effect of each potential source of heterogeneity by using covariates to estimate differences in both the accuracy and threshold parameters, but the underlying shape of the summary ROC curve was assumed to be constant. This assumption was necessary due to the limited number of studies.
Sensitivity analyses
We undertook sensitivity analyses to investigate the impact of bias in the selection of participants, and bias in the conduct and interpretation of the index test, on test performance.
Results
Results of the search
Our search resulted in 37,555 citations, of which 527 full‐text papers were assessed for eligibility. Of the 527, we discarded 437 (82.9%). Of the remaining 90 papers, 17 were included, one paper was identified as an ongoing study, 33 were multiple publications of the included papers, and 39 papers were excluded for various reasons outlined in the PRISMA flow diagram (Figure 1). No extra studies were found through reference checking though usable data for four studies (Galluzzi 2010; Hertze 2010; Vos 2013; Zetterberg 2003) were identified through contacting authors of studies.
Flow of studies through the screening process.
Included studies
The Characteristics of included studies table lists the characteristics of the 17 included studies involving a total of 2228 participants with cognitive decline but no dementia at baseline, of whom 2058 had analysable data. All but two of the included studies used one version or another of the Petersen criteria for MCI; the other two used a CDR score of 0.5 to define cognitive decline but no dementia. The Alzheimer's Disease Neuroimaging Initiative (ADNI) study (Shaw 2009) referenced the full ADNI protocol where it was clear the Petersen criteria were used though this was not clear from the actual manuscript. Of those 2057 participants, 703 developed Alzheimer’s dementia and 206 non‐Alzheimer's dementia. In addition, in one study (Herukka 2007) 33 participants were described as converting to dementia and though the subtypes were described it was not possible (despite contact with author) to generate the required 2 x 2 table for Alzheimer's disease dementia.
The median sample size of the included studies was 79 (range 37 to 588). All the studies were recent publications (2003 to 2013). Most of the studies (14/17) were conducted in Europe (six in Sweden, three in Italy, one in Finland, one in the Netherlands, one in Spain, and one in Greece). There was one study conducted in both Sweden and the Netherlands, two took place in the USA, and one in China. The participants were mainly recruited from secondary care (12 studies from outpatient clinics and two studies from inpatient departments), two studies recruited the participants from the community (Brys 2009; Fei 2011), and one from a mixed setting (Shaw 2009).
Excluded studies
Several hundred studies were discarded because they failed to satisfy two or more inclusion criteria. The proportion of studies excluded because they failed to meet one of the four key inclusion criteria for the review was 39 out of 56 (69.6%) (Figure 1). The main reasons for exclusion were incorrect study design (cross‐sectional rather than longitudinal) and target population (not MCI participants at baseline) (see Figure 1). Two studies used a definition which was unclear for 'cognitive decline no dementia' in their baseline population; two studies did not investigate conversion to Alzheimer’s dementia or other dementias in the MCI sample included, rather they investigated different stages of cognitive decline. Ten studies did not use thresholds for the Aß test instead reporting changes in plasma or CSF Aß42 levels over time. Two studies used combined CSF Aß42 and tau index tests, and one study performed the index test at follow‐up. Seventeen papers were excluded as the data were not presented in a manner that allowed extraction to construct the necessary 2 x 2 table to generate sensitivities and specificities of the test, despite attempts made to contact authors.
Methodological quality of included studies
We assessed methodological quality using the QUADAS‐2 tool (Whiting 2011).
Review authors’ judgments about each methodological quality item for each included study are presented in the Characteristics of included studies table and Figure 2. The overall methodological quality of the included study cohorts is summarised in Figure 3.
Risk of bias and applicability concerns summary: review authors' judgements about each domain for each included study.
Risk of bias and applicability concerns graph: review authors' judgements about each domain presented as percentages across included studies.
In the patient selection domain, we considered eight studies (47%) to be at low risk of bias because participants were enrolled consecutively and inappropriate exclusions were avoided. In the review, we only included data on performance of the index test to discriminate between patients with MCI who converted to dementia and those who remained stable; therefore, we stated that a case‐control design was avoided in all included studies. We considered four (24%) studies to be at high risk of bias mainly because the participants were not consecutively or randomly enrolled. Five studies (29%) scored unclear risk of bias due to poor reporting.
In the index test domain, we considered 10 studies (60%) to be at high risk of bias because the threshold used was not pre‐specified and the optimal cut‐off level was determined from ROC analyses; therefore, the accuracy of the plasma and CSF Aß biomarkers reported in these studies appeared to be overestimated. We considered seven remaining studies (40%) to be at low risk of bias because the threshold used was pre‐specified and the index test results were interpreted without knowledge of the results of the reference standard.
In the reference standard domain, we judged six studies (35%) to be at low risk of bias because the reference standard used was likely to correctly classify the target condition and clinicians conducting follow‐up were not aware of the initial CSF analysis results. We considered eleven remaining studies (65%) to be at unclear risk of bias due to poor reporting.
In the flow and timing domain, we considered nine studies (53%) to be of high concern for risk of bias because not all patients were accounted for in the analysis or the time interval between the index test and reference standard was not appropriate (duration of follow‐up was shorter than one year). We considered eight remaining studies (47%) to be at low risk of bias.
According to the QUADAS‐2 assessment of applicability, we found few concerns that the included patients and setting, index test, its conduct or interpretation, and the target condition (as defined by the reference standard) in each of the included studies did not match the review question: could plasma and CSF Aß biomarkers identify those MCI participant with Alzheimer’s disease pathology at baseline who would convert clinically to dementia at follow‐up? However, due to the limited number of included studies and levels of heterogeneity with respect to the three domains mentioned above, it was difficult to determine to what extent the findings from a meta‐analysis could be applied to clinical practice.
Findings
There were three target conditions for this review: [1] Alzheimer's disease dementia, [2] all forms of dementia, and [3] non‐Alzheimer's disease dementia, and our ability to present data for each was determined by what was undertaken within each of the primary studies. There were also six possible index tests: [1] CSF Aß42, [2] CSF Aß40, [3] CSF Aβ42/Aβ40, [4] plasma Aß42, [5] plasma Aß40, and [6] plasma Aβ42/Aβ40. However primary research for review was not available for most of the possible combinations of target conditions and index tests.
All papers using NINCDS‐ADRDA criteria presented data where probable Alzheimer's disease dementia (8) was the outcome, or it was not stated (8).
The control group in each of the analyses below was formed by all participants who did not develop the outcome of interest, therefore where Alzheimer's disease dementia was the outcome participants developing another form of dementia or MCI non‐converters were in the control group.
CSF Aß42 for detecting Alzheimer's disease dementia
The accuracy of Aß42 for Alzheimer's disease dementia was evaluated in 14 studies. The individual study estimates of sensitivity and specificity from these 14 studies (1349 patients) are shown in Figure 4. The sensitivities were between 36% and 100% while the specificities were between 29% and 91%. The summary ROC curve summarising the accuracy of Aß42 across all 14 studies, irrespective of the threshold used, is shown in Figure 5. Because of the variation in thresholds we did not estimate a summary sensitivity and specificity. However, we derived estimates of sensitivity at fixed values of specificity (see Table 1) from the model we fitted to produce the summary ROC curve. At the median specificity of 64%, the estimated sensitivity was 81% (95% CI 72 to 87). This equated to a positive likelihood ratio (LR+) of 2.22 (95% CI 2.00 to 2.47) and a negative likelihood ratio (LR–) of 0.31 (95% CI 0.21 to 0.48).
Forest plot of study results of cerebrospinal amyloid beta 42 for detection of Alzheimer's disease dementia.
Summary ROC plot for cerebrospinal amyloid beta 42 for detection of Alzheimer's disease dementia. Study estimates of sensitivity and specificity are shown with the summary ROC curve.
| Statistic | Fixed value of specificity % | Estimated sensitivity % (95% CI) |
| Conversion to Alzheimer's dementia (n = 14; cases = 436 ; non‐cases = 913) | ||
| Lower quartile | 57 | 84 (76, 90) |
| Median | 64 | 81 (72, 87) |
| Upper quartile | 72 | 75 (64, 83) |
| Conversion to Alzheimer's dementia sensitivity analyses excluding Kester 2011* (n = 13; cases = 394 ; non‐cases = 855) | ||
| Median | 62 | 82 (73, 88) |
| Conversion to Alzheimer's dementia sensitivity analyses excluding studies at high or unclear risk of bias for patient selection domain (n = 6; cases = 100 ; non‐cases = 299) | ||
| Median | 61 | 78 (55, 91) |
| Conversion to Alzheimer's dementia sensitivity analyses excluding studies at high risk of bias for index test domain (n = 6; cases = 255 ; non‐cases = 534) | ||
| Median | 68 | 76 (60, 87) |
| Conversion to Alzheimer's dementia sensitivity analyses excluding 3 outliers (n = 11; cases = 374 ; non‐cases = 696) | ||
| Median | 67 | 82 (69, 90) |
| All forms of dementia (n = 4; cases = 188; non‐cases = 276) | ||
| Median | 75 | 63 (22, 91) |
* The study did not consider a sample of MCI non‐AD converters in their analysis but excluded them.
The middle 50% of specificities from the included studies were between the lower and upper quartile, i.e. the interquartile range.
CSF Aß42 for detecting all forms of dementia
The accuracy of Aß42 for all forms of dementia was evaluated in four studies (464 patients) (Figure 6). The thresholds used were between 209 mg/ml and 512 ng/ml. The sensitivities were between 56% and 75% while the specificities were between 47% and 76%. The summary ROC curve summarising the accuracy of Aß42 across all four studies, irrespective of the threshold used, is shown in Figure 7. At the median specificity of 75%, the sensitivity was estimated to be 63% (22% to 91%) from the meta‐analytic model. This equated to a LR+ of 2.51 (95% CI 1.30 to 4.86) and a LR– of 0.50 (95% CI 0.16 to 1.51).
Forest plot of study results of cerebrospinal amyloid beta 42 for detection of all forms of dementia.
Summary ROC plot for cerebrospinal amyloid beta 42 for detection of all forms of dementia. Study estimates of sensitivity and specificity (with 95% confidence intervals) are shown with the summary ROC curve.
CSF Aß42 for detecting non‐Alzheimer's disease dementia
The accuracy of Aß42 for non‐Alzheimer's disease dementia was evaluated in three studies (385 patients). Since there were very few studies and considerable variation in test performance, the results were not meta‐analysed. Study specific estimates of sensitivity and specificity are summarised in Figure 8 and Figure 9. The sensitivities were between 8% and 63% while the specificities were between 35% and 67%. The study that did report the lowest threshold had a low sensitivity and the lowest specificity.
Forest plot of study results of cerebrospinal amyloid beta 42 for detection of non‐Alzheimer's disease dementia.
Study estimates of sensitivity and specificity with 95% confidence intervals plotted in ROC space for cerebrospinal amyloid beta 42 for the detection of non‐Alzheimer's disease dementia.
CSF ratio Aβ42/Aβ40 for detecting Alzheimer's disease dementia
The accuracy of the CSF Aβ42/Aβ40 for Alzheimer's disease dementia was evaluated in two studies (199 patients). The sensitivities were between 64% and 88% while the specificities were between 70% and 78%.
Plasma Aß42 and plasma Aβ42/Aβ40 ratio for detecting Alzheimer’s disease dementia
The accuracy of these plasma Aβ biomarkers for Alzheimer's disease dementia was evaluated only in one study (565 and 562 patients respectively). The sensitivity was 86% while the specificities were 50% and 70% respectively.
Investigations of heterogeneity
Investigations of heterogeneity were performed to assess the effect of pre‐specifying the threshold on the accuracy of Aß42 for Alzheimer's disease dementia. Test performance did not differ significantly (P = 0.8) between studies that pre‐specified the threshold (n = 6) and those that only determined the threshold at follow‐up (n = 8). The summary ROC curves for the two groups of studies are shown on the summary ROC plot in Figure 10. There is a paucity of knowledge around thresholds for Aß42 for Alzheimer's disease dementia. Despite best efforts a reliable and valid threshold remains elusive for clinical practice and research. Our findings also raise issues around risk of bias and the appropriateness of artificially determining cut‐offs.
Summary ROC plot for cerebrospinal amyloid beta 42 for detection of Alzheimer's disease dementia. Study estimates of sensitivity and specificity (with 95% confidence intervals) and summary ROC curves are shown according to whether or not studies pre‐specified the threshold for determining test positivity.
We were unable to investigate the effect of duration of follow‐up due to substantial variation in study length and reporting. For instance, some studies measured mean and standard deviation, or the range of duration of follow‐up, while others measured only the median or mean and standard deviation for conversion period. Poor reporting of patient demographic information also contributed to unknown sources of heterogeneity. Due to variation and missing variables, and insufficient number of studies we could not adequately compare patient characteristics between studies. Finally, as the Peterson criteria were used in the majority of studies to diagnose MCI there were not enough studies to investigate different MCI criteria for sources of heterogeneity.
Sensitivity analyses
We undertook sensitivity analysis by excluding one study (Kester 2011) from the analysis of CSF Aß42 for predicting conversion to Alzheimer's disease dementia, because the authors excluded a sample of MCI non‐Alzheimer's disease converters from their analysis. Exclusion of this study made no difference to our results (Table 1). A second sensitivity analysis was undertaken to assess the effect of risk of bias in the patient selection domain on our findings. Eight studies (n = 950 patients) were considered at high (n = 3) or unclear (n = 5) risk of bias. The exclusion of these studies made no difference to our findings. A third sensitivity analysis involved the index test domain; 8 studies (n = 560 patients) were considered to have a high risk of bias. Exclusion of these studies had no impact on our findings.
Discussion
Summary of main results
We included 17 studies in the review. Although we did not specify inclusion based on the use of the term 'mild cognitive impairment (MCI)' to describe 'cognitive impairment but no dementia', all of the included studies defined their baseline population using the MCI criteria. We will therefore use the term MCI in the remainder of this discussion.
The meta‐analysis of the accuracy of CSF Aβ42 for conversion from MCI to Alzheimer's dementia is based on the results from 14 studies. Three studies were included in the meta‐analysis of the accuracy of CSF Aβ42 for conversion from MCI to non‐Alzheimer's disease dementias and four studies for conversion from MCI to all forms of dementia. Only one study reported the use of plasma Aβ42 levels or the plasma ratio Aβ42/Aβ40 for the delayed verification of Alzheimer's disease dementia, which yielded the same sensitivity of 86% (95% CI 81 to 90) and specificities of 50% (95% CI 44 to 55) and 70% (95% CI 64 to 75) for plasma Aβ42 and plasma Aβ42/Aβ40 ratio respectively.
The total number of MCI participants at baseline was 2228. The accuracy of plasma and CSF Aβ biomarkers was evaluated in 2058 participants, of whom 703 developed Alzheimer's disease dementia and 186 non‐Alzheimer's disease dementia. Fourteen studies used one of the Petersen diagnostic criteria of MCI (Petersen 1999; Petersen 2004; Petersen 2006), highlighting the dominance of this description of the individual with objective cognitive impairment but no dementia. Two papers used the CDR value of 0.5 as their definition of cognitive impairment but no dementia. Studies were variable in duration of follow‐up, with a range from six months to eight years (Appendix 4). In general, studies with a longer length of follow‐up tended to show higher sensitivities as a consequence of the greater number converting to dementia in their cohorts as a function of time. The accuracy of low levels of CSF Aß42 in the 17 included studies ranged from specificities of 29% (Shaw 2009) to 91% (Parnetti 2006) and sensitivities from 8% (Galluzzi 2010) to 100% (Papaliagkas 2009). The main utility of CSF Aß42 in the proposed new criteria is the specificity of lowered CSF Aβ42 to identify Alzheimer's disease in people with cognitive impairment but no dementia. We evaluated 14 studies and obtained a summary sensitivity of 81% (95% CI 72 to 87) from the summary ROC curve at the median specificity of 64% (the range of specificity from the included studies was 29% to 91%). This test appears to have low specificity and only modest sensitivity rather than having the desired specificity for Alzheimer's disease proposed in the new criteria.
Strengths and weaknesses of the review
This review took place after extensive discussion within The Cochrane Collaboration regarding the optimal methodology to determine the test accuracy of wet lab biomarkers, imaging modalities and neuropsychological tests for neurodegenerative disease present before the development of dementia. There is a major impetus in this area, driven by both clinical need and the limitations of existing diagnostic criteria, to support discovery of more effective treatments for dementia. Newly proposed criteria have emphasised the integration of biomarker criteria with cognitive criteria with the aim of improving the specificity of diagnosis prior to dementia onset. This review used specific criteria to answer the question, which yielded a reasonably large dataset given the challenges of conducting the primary research in what remains a relatively undeveloped research area. The oldest publication in our data set was from 2003 and 13 of the included studies were published within the last five years. In its own right this would raise concerns about the validity of the proposed criteria, which were published in 2010 and 2011 and therefore relied on a small dataset to form conclusions with no published meta‐analysis of those studies available at the time. This review is the first such systematic review of this emerging literature base. Although there were 17 studies in the review that contributed to the conclusions, the review is limited by the large number of studies that have taken place and which probably had data that could have been used in the review but, despite contact with authors, did not present the results in a manner that could be extracted and used. It is hoped that future revisions of this review will have access to these datasets or data yet to be published from ongoing studies that will be presented in the appropriate format. In these papers, data were presented as mean difference between groups of patients with stable MCI and those who converted to Alzheimer’s dementia, in effect exploratory studies determining statistical associations between disease progression and the biomarker rather than the clinical utility of these markers in diagnostic terms. This observation reflects the predominantly exploratory nature of the use of these biomarkers for identifying disease process rather than as diagnostic tests with clinical utility, in effect representing phase 2 proof of concept studies as opposed to later stage phase 3 clinically useful studies.
Despite this, the findings from our review are entirely consistent with the JAMA paper (Mattsson 2009) that had access to full data sets from seven studies (Bjerke 2009; Brys 2009; Chiasserini 2010; Hansson 2007; Herukka 2007; Kester 2011; Parnetti 2006) of the 17 studies included in our review. Their optimal sensitivity and specificity were virtually identical to the ones we generate from our summary ROC curve, giving our findings external validity (Table 1). It is considered unlikely, therefore, that the inclusion of other studies in our review would have had much impact on our summary of the accuracy of lowered CSF Aβ42 for diagnosing Alzheimer’s disease in a population with cognitive decline but no dementia.
Although the quality of the papers reported in our analysis was generally good, overall the methodological and reporting quality of all considered papers was poor. An international consensus initiative (the STARDdem Initiative) was conducted recently, co‐ordinated by the Cochrane Dementia and Cognitive Improvement Group. This initiative aimed to review the current standard of reporting in diagnostic test accuracy studies and cognitive impairment and to generate enhanced guidance to the existing reporting guideline for diagnostic test accuracy studies (Noel‐Storr 2013; the STARDdem Initiative website). This developing research field would be supported by acceptance of consistent methodologies and reporting which would assist future reviews of the diagnostic accuracy of tests and their synthesis in meta‐analysis.
The small number of studies included in our review precluded formal statistical analysis of the effect of potential sources of heterogeneity. We observed that longer duration studies tended to yield greater sensitivity. As age is the key risk factor for Alzheimer’s disease dementia, it stands to reason that as the cohort ages the incidence of dementia will increase, improving the sensitivity of the test in question. In effect the number of false positives diminishes as a function of time. However, formal analysis of the influence of these factors on the accuracy of CSF Aβ was not possible as age, ApoE status, duration of MCI prior to cohort entry, gender and cognitive function at baseline were not consistently reported.
Most of the studies (14) were conducted in western Europe, two in the USA and one in China. Overall, the ethnic distribution of the population being studied was unclear from the reports. It is also noteworthy that individual studies in the most part generated their own optimal cut points for what constituted a positive or negative test. Such inconsistency is being addressed currently, though the genesis of uniform analysis techniques and thresholds will not alter the diagnostic accuracy of the test but will allow for easier integration of results across studies.
Applicability of findings to the review question
Although there were differences between studies, including test threshold, we suggest that low CSF Aβ42 should not be used as a diagnostic test for Alzheimer’s disease in patients with cognitive decline but no dementia. The test is invasive, costly, suffers from a lack of consensus on sampling and analysis, and is non‐specific. It does not have the necessary accuracy to aid the clinician in making decisions as to which of their patients with cognitive decline but no dementia are likely to develop Alzheimer's dementia.
Recently proposed diagnostic criteria for prodromal AD emphasise low CSF Aβ42 in isolation from any other biomarker test as indicating ‘likely’ Alzheimer’s disease in a patient population with objective cognitive impairment. This proposal is not supported by our findings and the evidence points in the opposite direction where a normal test may have some modest utility in suggesting that normal CSF Aβ42 may rule out Alzheimer’s disease; even in this regard the risks and cost of testing may outweigh the potential benefit.
Pre‐ and post‐ tests probabilities of Alzheimer's dementia using low CSF amyloid beta 42 as a diagnostic test.
Flow of studies through the screening process.
Risk of bias and applicability concerns summary: review authors' judgements about each domain for each included study.
Risk of bias and applicability concerns graph: review authors' judgements about each domain presented as percentages across included studies.
Forest plot of study results of cerebrospinal amyloid beta 42 for detection of Alzheimer's disease dementia.
Summary ROC plot for cerebrospinal amyloid beta 42 for detection of Alzheimer's disease dementia. Study estimates of sensitivity and specificity are shown with the summary ROC curve.
Forest plot of study results of cerebrospinal amyloid beta 42 for detection of all forms of dementia.
Summary ROC plot for cerebrospinal amyloid beta 42 for detection of all forms of dementia. Study estimates of sensitivity and specificity (with 95% confidence intervals) are shown with the summary ROC curve.
Forest plot of study results of cerebrospinal amyloid beta 42 for detection of non‐Alzheimer's disease dementia.
Study estimates of sensitivity and specificity with 95% confidence intervals plotted in ROC space for cerebrospinal amyloid beta 42 for the detection of non‐Alzheimer's disease dementia.
Summary ROC plot for cerebrospinal amyloid beta 42 for detection of Alzheimer's disease dementia. Study estimates of sensitivity and specificity (with 95% confidence intervals) and summary ROC curves are shown according to whether or not studies pre‐specified the threshold for determining test positivity.
| What is the diagnostic accuracy of plasma and CSF amyloid biomarker levels for detecting Alzheimer's disease pathology in patients with mild cognitive impairment (MCI), and identifying those MCI participants who would convert to Alzheimers disease dementia or other forms of dementia over time | |||||||
| Patient population | Participants diagnosed with MCI at baseline using any of the Petersen criteria or CDR = 0.5 or any 16 definitions included by Matthews (Matthews 2008) | ||||||
| Prior testing | The only testing prior to performing the plasma and CSF biomarkers was the application of diagnostic criteria for identifying participants with MCI | ||||||
| Settings | Participants were recruited from i) secondary care – outpatient clinic (n = 12); ii) secondary care – inpatients (n = 2); iii) community care (n = 2) and mixed setting (n = 1) | ||||||
| Index tests | Plasma or CSF Aß 42, Aß 40, or Aß 42/Aß 40 ratio | ||||||
| Reference standard | NINCDS‐ADRDA or DSM or ICD criteria for Alzheimer's disease dementia; McKeith criteria for Lewy body dementia; Lund criteria for frontotemporal dementia; and NINDS AIREN criteria for vascular dementia | ||||||
| Target condition | Alzheimer’s disease dementia or any other form of dementia | ||||||
| Included studies | 17 studies (2228 participants) of prospectively well defined cohorts with any accepted definition of MCI were included | ||||||
| Quality concerns | Patient selection and conduct of the index and reference standard were poorly reported. Applicability concerns were generally low. Regarding the inclusion criteria set in the review, each included study did match the review question: 'Could Plasma and CSF Aßbiomarkers identify those MCI participant with Alzheimer’s disease pathology at baseline who would convert clinically to dementia at follow up?' However, due to limited number of included studies and levels of heterogeneity it is difficult to determine to what extent the findings from a meta‐analysis can be applied to clinical practice. | ||||||
| Limitations | Limited investigation of heterogeneity due to insufficient number of studies. There was a lack of common thresholds and poor reporting of thresholds. | ||||||
| Test | Studies | Cases/participants | Median specificity from included studies | Sensitivity (95% CI)1 at median specificity | Consequences in a cohort of 100 | ||
| Median percentage converting (range)2 | Missed cases3 | Overdiagnosed 3 | |||||
| Alzheimer's disease dementia | |||||||
| CSF Aß 42 | 14 | 436/1349 | 64 | 81 (72, 87) | 38 (9 to 56) | 7 | 22 |
| All forms of dementia | |||||||
| CSF Aß 42 | 4 | 188/464 | 75 | 63 (22, 91) | 45 (27 to 53) | 17 | 14 |
| Non‐Alzheimer's disease dementia | |||||||
| CSF Aß 42 | 3 | 61/385 | 65 | No meta‐analysis | 16 (15 to 19) | ||
| Investigation of heterogeneity: the planned investigations were not possible due to the limited number of studies available for each analysis. We were unable to investigate the effect of duration of follow up due to substantial variation in length and reporting. | |||||||
| Conclusions: from our review, abnormally low CSF Aß levels has very little diagnostic benefit with likelihood ratios suggesting only marginal clinical utility. The quality of reports was also poor, and thresholds and length of follow up were inconsistent. We conclude that when applied to a population of patients with MCI, CSF Aß levels cannot be recommended as an accurate test for Alzheimer's disease. | |||||||
| 1 Meta‐analytic estimate of sensitivity derived from the HSROC model at a the median value of specificity computed from the included studies. Summary estimates of sensitivity and specificity were not computed because the studies that contributed to the estimation of the summary ROC curve used various thresholds. 2 The median percentage converting and range were computed using all the studies included in the analysis for each target condition. 3 Missed and overdiagnosed numbers were computed using the median percentage converting for each target condition. | |||||||
| Statistic | Fixed value of specificity % | Estimated sensitivity % (95% CI) |
| Conversion to Alzheimer's dementia (n = 14; cases = 436 ; non‐cases = 913) | ||
| Lower quartile | 57 | 84 (76, 90) |
| Median | 64 | 81 (72, 87) |
| Upper quartile | 72 | 75 (64, 83) |
| Conversion to Alzheimer's dementia sensitivity analyses excluding Kester 2011* (n = 13; cases = 394 ; non‐cases = 855) | ||
| Median | 62 | 82 (73, 88) |
| Conversion to Alzheimer's dementia sensitivity analyses excluding studies at high or unclear risk of bias for patient selection domain (n = 6; cases = 100 ; non‐cases = 299) | ||
| Median | 61 | 78 (55, 91) |
| Conversion to Alzheimer's dementia sensitivity analyses excluding studies at high risk of bias for index test domain (n = 6; cases = 255 ; non‐cases = 534) | ||
| Median | 68 | 76 (60, 87) |
| Conversion to Alzheimer's dementia sensitivity analyses excluding 3 outliers (n = 11; cases = 374 ; non‐cases = 696) | ||
| Median | 67 | 82 (69, 90) |
| All forms of dementia (n = 4; cases = 188; non‐cases = 276) | ||
| Median | 75 | 63 (22, 91) |
| * The study did not consider a sample of MCI non‐AD converters in their analysis but excluded them. The middle 50% of specificities from the included studies were between the lower and upper quartile, i.e. the interquartile range. | ||
| Test | No. of studies | No. of participants |
| 1 CSF Abeta42 AD Show forest plot | 14 | 1349 |
| 2 CSF Abeta 42 All dementia Show forest plot | 4 | 464 |
| 3 CSF Abeta 42 non‐AD Show forest plot | 3 | 385 |
| 4 CSF Abeta42/Abeta40 AD Show forest plot | 2 | 199 |
| 5 Plasma Abeta 42 AD Show forest plot | 1 | 565 |
| 6 Plasma Abeta42/Abeta40 AD Show forest plot | 1 | 562 |
