Premium
Prediction of the MMSE up to 6 years ahead with cross‐cohort replications
Author(s) -
Maheux Etienne,
Koval Igor,
Archetti Damiano,
Redolfi Alberto,
Durrleman Stanley
Publication year - 2020
Publication title -
alzheimer's and dementia
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 6.713
H-Index - 118
eISSN - 1552-5279
pISSN - 1552-5260
DOI - 10.1002/alz.043541
Subject(s) - logistic regression , statistics , dementia , cohort , linear regression , cognitive decline , ordered logit , logit , regression , mathematics , econometrics , psychology , medicine , disease
Background Most of the studies that focus on the Alzheimer’s disease (AD) prediction forecast a dementia label: cognitively normal, mild cognitive impairments or AD. However these labels cannot always be compared across cohorts, are not perfectly accurate and describe the disease thanks to a limited number of stages (only 3) which, in practice, is a continuous progression. For these reasons, we predict a cognitive test (MMSE) that has a finer discretization and is similarly assessed across cohorts. Method The Bayesian mixed‐effects model introduced in (Schiratti et al. 2015, NIPS) allows to estimate the average temporal progression of a subset of scores from ADNI, out of the individual longitudinal data. This profile is personalized to individuals from the ADNI cohort but also AIBL and PharmaCOG. It allows to predict future values (1.5 to 6 years after the last seen visit). To measure the error of prediction, the value to predict is hidden and then compared to the prediction. This error is compared to the (i) constant prediction (i.e. no MMSE change), (ii) linear and (iii) logit regressions and (iv) the test‐retest MMSE noise ‐ whose standard deviation is 2.8 out of 30 (Clark et al, 1999, Archives of Neurology) ‐ which is the best achievable score. Result The constant prediction, while being trivial, is better than the linear and logit regressions [Figure 1]. Moreover, it is of the order of the noise (i.e. best possible prediction) for horizons of less than 3 years. In any case, our prediction is of the noise order or smaller than the constant prediction, allowing to predict accurately the MMSE at any temporal horizon. Moreover, the accuracy on the AIBL and PharmaCOG [Figures 2 and 3] datasets highlights the potential of the model to generalize to patients from other cohorts but also in real‐life applications. Conclusion The constant prediction outperforms standard regression techniques (linear and logit) due to the important noise level of the MMSE and the fact that the progression of the disease occurs over long periods of time. However, our model is able to beat this prediction for far enough temporal horizons.