The main goal of the current study is to develop mathematical models of typical disease trajectories of AD, from its presymptomatic to symptomatic stages - that is, to develop algorithms for use to quantitatively compare patients undergoing experimental treatment for secondary prevention of dementia due to AD with their own anticipated untreated disease course. We used the data for 397 MCI subjects from the ADNI database as available in October 2009. The examples presented here concern a univariate (endpoint-related) approach - that is, an algorithm that predicts the MCI subject group's performance scores on the ADAScog 24 months after their baseline examination - and a multivariate (trajectory-related) approach - that is, an algorithm that forecasts the decline of performance during 36 months, from baseline to the last examination after 3 years - on the composite score of a neuropsychological battery as described previously (NP-Batt) [15
]. Both outcomes, a cognitive performance score after 24 months and the trajectory of cognitive change over 36 months, could be of use in studies with experimental drugs for secondary prevention of dementia due to AD. A total of 11 demographic, neuropsychological and biological measures established at baseline, plus their interactions, were included as potential predictors in the univariate and multivariate analyses.
In the univariate model, the strongest predictors of the ADAScog scores as measured after 24 months were the ADAScog scores established at baseline. Other significant predictors were (in decreasing order of importance) the composite scores on the NP-Batt, the MMSE scores, gender and obesity. There was also a significant interaction of the FAQ with the relationship between the NP-Batt at baseline and the ADAScog scores at 24 months, suggesting that higher scores on the FAQ (indicating some functional restriction at baseline) mainly affected 24 months ADAScog scores of MCI subjects with normal NP-Batt scores at baseline, but not those with lower baseline performance.
In the multivariate analysis, the strongest predictors of the NP-Batt trajectory over 36 months were the NP-Batt scores established at baseline. Other significant, albeit weaker, predictors are age, the FAQ scores at baseline and obesity. Of particular interest is the interaction between the number of apolipoprotein E4 alleles and time (Figure ), indicating that the negative impact of this genetic marker upon cognitive performance was not significant at baseline but developed over time.
Taken one by one, none of these findings is novel. For the baseline cognitive measures (ADAScog, MMSE, NP-Batt) predicting cognitive performance after 24 months as well as the trajectory of cognitive decline over 36 months, numerous studies show that cognitive performance established at some point in time (memory function and other cognitive domains) is a strong predictor of subsequent cognitive decline and of dementia. This was found, at different levels of performance, for older subjects who were cognitively healthy at baseline [20
], for MCI subjects [22
] and for AD patients [6
]. The concept of cognitive reserve [24
] captures these observations in a more general hypothetical construct.
The effect of gender was significant in the univariate approach, but not in the multivariate model. While it is known from epidemiological studies that AD occurs more frequently in aged women than in men [26
], this general finding does not explain the difference in the two models.
Obesity was found to have a protective effect against cognitive decline in both the univariate model and the multivariate model. Cronk and colleagues, who also worked with the ADNI database, have already reported a favorable impact of higher body mass index baseline values on the development of MMSE, ADAScog and NP-Batt scores as early as 1 year after baseline [15
]. Other authors also found a protective effect of obesity upon cognitive performance in persons of older age [27
]. Although overweight in middle age was identified as a risk factor for dementia several decades later [29
], this relationship appears to be reversed in persons beyond 70 years of age (the obesity paradox [28
]). To what extent this paradox and the apparent protective effect of obesity are due to an underlying selection factor - for example, higher mortality in overweight people lacking a hypothetical protective factor that also supports cognitive maintenance - cannot be deducted from the current data.
There was a significant interaction between the FAQ scores and the NP-Batt scores as predictors of the performance on the ADAScog at 24 months. FAQ scores at baseline were a significant determinant of cognitive decline in the multivariate analysis when NP-Batt scores were still normal (that is, z
= 0). As mentioned earlier, the ADNI MCI sample cannot be considered a pure selection of MCI subjects since some subjects had restrictions in activities of daily life and instrumental activities of daily life, as indicated by baseline FAQ scores >0 (Table ), and 60% of the subjects were taking anti-AD drugs (cholinesterase inhibitors and/or memantine) at some point in the study. With regard to impaired activities of daily life/instrumental activities of daily life in some subjects, it is of interest that the baseline FAQ scores turned out to be a significant and, at least in the multivariate model, independent predictor of cognitive decline over the ensuing 36 months. According to Pérès and colleagues, some decline in instrumental activities of daily life is seen in aged subjects as early as 10 years before a clinical diagnosis of dementia is made, and may thus constitute a very early marker of dementia [31
]. Dickerson and colleagues found that, in mildly impaired aged individuals who did not meet strict MCI criteria as implemented in clinical trials, the degree of cognitive impairment in daily life and performance on neuropsychological testing impacted the likelihood of an AD diagnosis within 5 years [32
As for the possible effect of anti-AD medication, separate analyses for MCI subjects with and without use of cholinesterase inhibitors and/or memantine at any time provided very similar predictive models. Although the patients taking any of these drugs at some time showed somewhat inferior cognitive performance at baseline - that is, higher average ADAScog scores and lower scores on the NP-Batt than subjects not taking any of these drugs - the major predictors for the ADAScog scores after 24 months and the trajectory of the NP-Batt over 36 months were the same for both subgroups (data not shown).
While the number of apolipoprotein E4 alleles was not a significant predictor of the ADAScog scores at 24 months in the univariate analysis, this genetic risk factor of AD showed a significant interaction with time and, consequently, a strong impact upon the trajectory of the NP-Batt scores in the longitudinal analysis (Figure ). Interestingly, the number of apolipoprotein E4 alleles did no longer significantly affect cognitive performance when the Aβ42/T-tau quotient was introduced as a potential predictor; note, however, that the analysis including the cerebrospinal fluid markers was performed with a sample only one-half as large as that for the other calculations. The Aβ42/T-tau quotient is a known early marker of AD [19
] and was recently reported to be a predictor of functional decline in the ADNI MCI sample [34
]. Aβ42 and subsequently tau in the cerebrospinal fluid are considered early markers in the AD pathological cascade [35
In summary, although none of the individually significant predictors identified in our models was unexpected, the specific and weighted combination of predictors is novel in the models presented: the univariate approach predicted cognitive performance (ADAScog scores) 2 years after baseline, and the multivariate approach forecasted the decline of cognitive performance - as measured by means of the NP-Batt - over 36 months. The univariate model for the square root of ADAScog explains 63% of the variance, and the prediction error of a single outcome in cross-validation is 0.67. This is obviously not a precise estimate for single values, as the 95% interval for an estimate of ADAScog score of 20 would range roughly from 10 to 34 for an individual patient. In the application of this model to clinical studies, however, the relevant measure will be the mean of, say, 200 outcomes. The standard error of this mean would be 0.047 on the square-root scale, and the 95% confidence interval for a mean of 20 on the original scale would be as narrow as 19.2 to 20.8. In the multivariate model for NP-Batt, which has a standard normal distribution in a healthy population, the standard deviation of the conditional residuals is 0.2175. Taking the between-patient variability into account, the standard deviation of residuals in the population is between 0.32 and 0.51, depending on the time of observation (0 to 36 months). For a sample of 200 subjects these standard deviations are reduced to 0.023 and 0.035, respectively. Estimated means from such samples appear to be sufficiently precise for group comparisons in clinical trials.
An important point to be addressed in future analyses concerns the possibility of generalizing our models to new, independent datasets and, eventually, their application in clinical trials of experimental drugs intended for secondary prevention of dementia due to AD. As a first step in this direction, the simulation model for the NP-Batt as derived from the ADNI MCI subjects was challenged by applying it to the AD patient sample of the ADNI dataset. NP-Batt data for visits at 6, 12 and 24 months were simulated (there are no data from AD patients at 18 months). A comparison of the scores from the observed and the simulated data is shown in Table and indicates that the observed and the model-based simulated values of the ADNI AD patient sample were indeed very similar. This is a preliminary indication that the mathematical model established from MCI data is also usable for datasets from patients with dementia; that is, over a wide range of presymptomatic and symptomatic AD patients. We are currently testing our models with MCI datasets from other, ADNI-independent projects, which contained partly different assessment criteria. The results of these tests will be reported in due time.
Descriptive statistics of observed and simulated NP-Batt scores for Alzheimer's disease patients at months 6, 12 and 24
Assuming that our models are supported by analyses of further independent datasets, what would principally argue against their use in trials with experimental compounds aimed at secondary prevention of dementia due to AD? Evidently, the 50-year tradition of placebo-controlled study designs in clinical neuropsychopharmacology argues against a new approach like the one suggested here - although several limitations of the conventional designs have repeatedly been pointed out [4
To illustrate one particularly important limitation of RPCTs and its consequences, let us for a moment assume the perspective of an older person who has just learned that he or she shows the characteristics of presymptomatic AD or MCI, implying that he or she is likely to become demented within a few years, and who is offered participation in a long-term phase 3 RPCT with a promising experimental disease-course altering drug. Would one not expect that this individual would be uncertain as to how he or she should decide: agree to participate in a placebo-controlled - that is, a Russian roulette type of trial - or reject participation and hope for a better alternative?
In recent years, clinical investigators - notably in the United States - have reported increasing difficulties recruiting patients into AD clinical trials [36
]. One cannot rule out that some of these difficulties are due to patients' unwillingness to enter trials that entail a high risk for participants of being treated with placebo for months or even years. Thus, apart from the ethical issue of exposing high-risk individuals to admittedly ineffective treatment (placebo), one should also consider that only a self-selected fraction of all trial candidates will eventually enter RPCTs, a fact that seriously jeopardizes the external validity of such trials. In spite of these concerns, current regulatory guidelines [37
] and specialized task forces [38
] keep recommending or even demand RPCTs as proof of efficacy for drugs intended for use in AD, including compounds aimed at secondary prevention of dementia due to AD that require very long studies to prove efficacy. This insistence is surprising, given that placebo-controlled designs were originally developed for clinical studies of analgesics, antidepressants and anxiolytics - that is, for trials in mostly self-limiting, unstable and partly subjective central nervous system indications that differ in important aspects from slowly developing, irreversible, degenerative disorders such as AD.
If supported by further evidence, where in the clinical development process of a new compound aimed at secondary prevention of dementia due to AD could be the place for the proposed PGSA? No doubt some of the earlier (phase 1 and phase 2) trials, which are often performed on healthy subjects and subsequently on AD patients at different levels of deterioration, do require placebo control, notably in order to detect and characterize any relevant safety issue of the new compound. As these earlier studies last only a few months for each patient, and since little is known early in development about a new drug's potentially useful effect in man, there is no ethical concern to use placebo at this stage. Once the proof-of-principle and placebo-controlled safety studies are completed, however, and presumably effective and safe doses of the novel drug need to be tested for long-term efficacy in the target population - that is, in subjects in presymptomatic stages of AD - then a placebo-free approach such as the PGSA should be seriously considered. In addition to its ethical and scientific merits, it also has the potential to save patients, time and money. The next years will show whether the AD research community [3
] and drug regulatory bodies are ready and willing to de-emphasize a traditional study paradigm that has serious shortcomings, and are willing to consider a design that has the potential both to benefit the patients and facilitate anti-AD drug development.