Our analyses in a large, well-characterized cohort of normal controls found heterogeneity in imaging summary measures and fluid biomarker summaries, indicating that the biomarker profile could not be encompassed by a single, compact, well-defined set of boundaries. Instead, the patterns were suggestive of several distinct clusters of individuals, even though these individuals were cognitively intact at baseline and showed very little clinical progression over three years of follow-up. Regression models suggested that baseline levels of some measures of brain size and CSF biomarkers are associated with cross-sectional differences in cognitive performance, and that increased P-tau181 and the P-tau181/Aβ1–42 ratio may foreshadow slightly more rapid declines in ADAS-cog. The effects, however, were modest, and not improved in multivariate regression analysis. Cluster analysis identified three distinct groups of individuals, based solely on their baseline imaging and biomarker measures without reference to cognitive status or change. One group in particular, comprising about 10% of the NC group, was well separated from the bulk of the NC and lay closer to the centers of the MCI and AD groups, even when some individual marker measurements might be closer to the center for typical NC. This subgroup had strikingly lower baseline scores on the RAVLT, significantly worse scores on the ADAS-cog, and a significantly more rapid deterioration on the ADAS-cog than the typical NC. Membership in this cluster was associated with annual cognitive decline approximately 5 times as rapid as that predicted for a person one standard deviation worse than average on the strongest individual marker. A second group, centered between the atypical group and the more typical NC group, showed structural MRI profiles closer to the MCI group and had somewhat worse baseline performance but showed little difference in cognitive trajectory compared to typical NC. The division into distinct clusters was robust across several clustering algorithms, with the same people consistently identified as a distinct group whose center placed them closer to the profile of MCI and AD. As a complementary approach, unsupervised regression trees were used to generate a proximity matrix which was then plotted on multidimensional scaling axes and colored by cluster membership. The plot showed almost no mixing of clusters, supporting the clusters identified with agglomerative clustering despite the fact that methodologies have almost nothing in common. This suggests that the clusters are not simply an artifact of the clustering process and may instead represent meaningful structure in the data.
Our findings are consistent with previous work suggesting P-tau181
, and their ratio as among the measures most closely reflecting early preclinical neurobiological changes [Stomrud et al. 2007
, Shaw et al. 2009
]. We also found evidence for correlation between brain atrophy measures and cross-sectional measures of cognitive performance, even in this very high-functioning cohort. None of these measures, however, had substantial predictive power for future cognitive decline, either taken individually or as linear combinations. Little previous work has explored the multi-dimensional structure of imaging and CSF biomarkers in clinically normal older people. A recent study (Fagan 2009
) assessed correlations between CSF measures and normalized whole brain volume in cognitively normal elderly subjects and found that Abeta, but not Tau or P-tau181
, correlated inversely with whole-brain volume in elderly control subjects, while Tau and P-tau181
correlated in very mild and mild AD. Such findings have suggested a hypothetical model of the sequence of biomarker changes in the neurobiological process leading to AD, with amyloid changes occurring very early, followed by tau pathology, volumetric and metabolic decline, and with a very gradual onset of measurable cognitive decline, reaching the diagnosis of dementia only very late in the cascade (Jack, 2010
). Our analyses offer additional support for this conceptual model. We identified a subgroup of normal controls, based solely on biomarker profiles, who were notable primarily for amyloid-related CSF abnormalities. Not only were people in this group much closer to the AD group in their CSF profile, but they also had a striking pattern of ADAS-cog decline over 2–3 years, despite being clinically well within the normal range to start. suggests strong similarity to the conceptual model proposed in Jack. Clustering based just on MRI measures failed to isolate such a distinctive subgroup. Our findings are consistent with Cluster 3 representing a group of people who have already progressed so far in amyloid deposition that they are starting to experience the earliest signs of clinical decline, even though their scores remain, for the most part, within the wider normal range and do not yet verge on dementia.
Methods looking at best classification strategies to predict conversion among NC are not yet applicable in ADNI, as there are so few conversions. We are not aware of any work with data mining or other dimension-reduction strategies based on rate of cognitive change as an outcome. Our study represents a first attempt to seek structure based strictly on the imaging and fluid biomarker measures, and to map the observed heterogeneity to distinct subgroups.
One limitation of our approach is that the CSF markers were available only for half the participants, thus reducing the sample size available for clustering. This, in turn, led to a fairly small though stable “atypical” cluster of about 10% of the NC with CSF. The remaining participants split with every clustering algorithm we examined, with some uncertainty in locating the border between a compact, “typical” group and a second group leaning toward the “atypical” group. A larger sample size might have allowed better definition of the margins, as well as the variables most important for identifying the “atypical” cluster. A second limitation is that length of follow-up and the generally healthy nature of this cohort led to only a few conversions to MCI, precluding our assessment of the predictive value of the cluster for change in clinical status. A third limitation is that we used only a fraction of the many potential biomarkers, especially from the imaging summaries, and we included 5 participants with some imaging quality control problems whose effect on our findings appeared minimal in the primary analysis. Our selection was based on previous literature, but may have omitted variables with better prognostic power in the ADNI cohort. A fourth limitation is the approach to standardization of the imaging variables. Dividing by ICV may diminish differences between individuals or groups, although the cluster definitions were more dependent on the CSF biomarkers. Finally, we have not used ADNI’s FDG PET or PiB imaging data because the number of participants with complete data on the MRI, CSF and PET imaging data would be too small for analysis.
Our approach has several notable strengths. First, the ADNI data offer one of the largest, uniformly ascertained databases available, with standardized protocols not only for data collection but also for image processing and biomarker specimen assays.
Second, our approach makes use of 36 months of follow-up with standardized cognitive testing. Third, our clustering strategy was “unsupervised”, that is, based on the imaging and biomarker data without reference to cognitive endpoints or diagnostic categories, so that we address directly the question of whether a strikingly unusual profile on these measures may foreshadow clinical decline, without letting the outcome define the profile of interest. Notably, a recent study on ADNI CSF biomarkers used another unsupervised learning approach that identified nearly the same CSF Aβ cut point as Shaw et al 2009
wherein this cut point was established in subjects with autopsy confirmed diagnoses of AD thereby demonstrating the power of unsupervised analytical approaches (de Meyer et al, 2010
). Supervised techniques such as regression trees or discriminant analysis offer a complementary approach, focused on identifying subgroups of clinical interest based on their outcome, but would require cross-validation, potentially problematic in our sample of under 100 NC with little clinical change as yet. In comparison to univariate or multivariate regression models, our approach takes advantage of the correlated relationship among the biomarkers, rather than being limited by this interrelationship. This strength of the clustering approach may account for the improvement in our ability to detect associations with future cognitive change, compared to the more traditional univariate and multivariate regression models.
Our findings indicate that not just individual abnormalities, but a distinctive pattern of imaging and biomarker deviations from typical healthy older adults may be an early warning sign of neurobiological pathology. Additional follow-up would help to establish the prognostic value for longer-term cognitive decline and conversion to MCI and eventually to AD. A larger sample, with CSF on all participants, would help to confirm the cluster pattern for our “atypical” group and better define the boundary between the other two clusters. In addition, the recruitment of early MCI participants, as currently underway in the recently funded ADNI-GO grant, will offer an opportunity to test whether our atypical group have a profile similar to the earliest stages of clinically-identified impairment. Longitudinal analysis of the imaging and biomarker measures can also address whether the distance of the atypical cluster relative to the center for typical NC is increasing over time, and whether the location of the intermediate cluster shifts toward the atypical cluster’s baseline location over time.