Heterogeneity of cognitive decline in dementia: taking into account variable time-zero severity
NOTEThe biases seen in simulations of Temporal Clustering appear to be even worse in real applications. While this was a novel and interesting approach, ultimately this work has been discontinued. I feel the biases are due to propagated error in estimating individual level offsets based on a single noisy measure, amplified by the fact that high MMSE scores change very slowly and therefore many estimates are from near an asymptote on the left hand of the model. We continue to work in this area, with other approaches showing significantly more promise.\n\nUnderstanding heterogeneity in Alzheimers disease (AD) progression is critically important for the optimal design of trials, allowing participants to be recruited who are correctly diagnosed and who are likely to undergo cognitive decline. Current knowledge about heterogeneity is limited by the paucity of long-term follow-up data and methodological challenges. Of the latter, a key problem is how to choose the most appropriate time zero to use in longitudinal models, a choice which affects results. Rather than a pre-specified time zero we propose a novel methodology - Temporal Clustering - that defines a new time zero using individual offsets inferred from the data. We applied this to longitudinal Mini-Mental State Examination (MMSE), where this approach ensures that individuals have similar estimated MMSE scores at this new time zero. Simulations showed that it could accurately predict cluster membership after the application of a filter. Next we applied it to a cohort of 2412 individuals, with large variability in MMSE score at first visit. Temporal Clustering was used to split individuals into two clusters. The group showing faster decline had higher average levels of AD risk factors: cerebrospinal fluid tau and APOE [isin]4. Cluster membership predicted by Temporal Clustering was less affected by individuals cognitive ability at first visit than was the case for clusters found using Latent Class Mixture Models. Further application and development of this method will help researchers to identify risk factors affecting cognitive decline.