Dynamic Models for Epidemiological Longitudinal Studies of Chronic Diseases – DyMES

The Prediction task aims to exploit the joint modeling theory, landmark prediction models, and statistical learning to develop dynamic individual predictions of clinical events that use all the information previously collected overtime on an individual. For example, to predict the risk of dementia based on repeated measures of cognition, functional dependency, and neuro-imaging markers.
The Causality task consists in proposing approaches to estimate causal associations in a longitudinal context. Among others, an innovative statistical approach is proposed to quantify temporal influences between health processes, and thus unravel physiopathological mechanisms. It should make it possible for instance to understand the relationships between autonomic dysfunction and functional progression in AMS or the dynamic relationships between brain structures and clinical manifestations in AD.
The Long-term exposures task aims to develop statistical methods to simultaneously analyze time-dependent exposures in relation to health markers. These methods will allow, for example, a better understanding of which lifestyle exposures (such as cardio-metabolic factors, promising avenues for prevention in AD) and which temporal windows (e.g., at adulthood) are the most critical for preserving brain health in the elderly.
The Dissemination task aims to provide efficient and user-friendly software solutions (in the form of R-packages) for the wide use of cutting-edge statistical models.
To achieve its objectives, DyMES exploits data from 4 large cohorts in AD: Paquid, 3C, MEMENTO, and the Nurses' Health Study. In MSA, the cohort from the national reference center (based in Bordeaux and Toulouse), one of the largest cohorts in the world, is analyzed.

Thanks to the access to high-quality data and the multidisciplinary project team comprising statisticians, epidemiologists, and neurologists working in close collaboration, the project has the ambition to:

- provide innovative statistical models with efficient associated programs to the statistical and epidemiological communities that are essential for the analysis of multivariate longitudinal data available in modern cohort studies;

- answer burning questions such as unraveling the influences between processes involved in a disease, predicting its progression, or identifying lifestyle behaviors and time windows to optimize prevention.

Although motivated by MSA and AD, the project applies far beyond. It will enhance the statistical analysis of longitudinal epidemiological studies by providing valid reproducible methods, and improve the understanding of chronic diseases with repercussions for public health in the long-term.

Software:
R package marqLevAlg: A Parallelized General-Purpose Optimization Based on Marquardt-Levenberg Algorithm, disponible sur le CRAN – version 2 : cran.r-project.org/web/packages/marqLevAlg/index.html
R package lcmm : Estimation of various extensions of the mixed models including latent class mixed models, joint latent latent class mixed models and mixed models for curvilinear univariate or multivariate longitudinal outcomes using a maximum likelihood estimation method - version 1.9.2 : cran.r-project.org/web/packages/lcmm/index.html

Scientific papers:
Wagner, Grodstein, Leffondré, Samieri*, Proust-Lima* (2020). Joint modeling of time-varying exposure history and subsequent health outcomes: identification of critical windows. Under review
Philipps, Hejblum, Prague, Commenges, Proust-Lima (2020). Robust and efficient optimization using a Marquardt-Levenberg algorithm with R Package marqLevAlg. Under review

With their large amount of multivariate information collected repeatedly over visits, epidemiological cohorts allows for the in-depth study of the dynamic and multidimensional aspects of health phenomena. Yet, valid statistical methods that jointly analyze multiple dynamic processes are lacking.

DyMES aims to develop innovative statistical models and apply them to large cohort data to understand and predict multivariate health processes. We will focus on two neurodegenerative diseases, Alzheimer’s Disease (AD) and Multiple System Atrophy (MSA), which share major methodological challenges such as the coexistence of multiple inter-related progressive impairments.
We will tackle four issues:

Task 1: Prediction. We will develop joint models to analyze multiple longitudinal dimensions and clinical events. From them, we will be able to compute individual dynamic predictions of clinical events that make use of all the information previously collected on an individual. For instance, repeated measures of cognitive functions, dependency to daily activities and neuro-imaging markers will be exploited to predict the risk of dementia.

Task 2: Causality. We will develop an innovative approach to quantify temporal influences between dynamic processes. By using mechanistic models in discrete time, the method can unravel complex causal relationships between processes, and as such help understand underlying physio-pathological mechanisms. In MSA we will for instance identify the relations between autonomic system and functional progression; in AD we will disentangle the dynamic relationships between brain structures and clinical manifestations.

Task 3: Long-term and time-varying exposures. Modifying lifestyle factors is a promising perspective for prevention in AD but little is known on exposure trajectories in relationship with late-life outcomes (e.g., most critical exposures and/or time windows). We will develop statistical methods that will enable the simultaneous analysis of several time-varying exposures (e.g., cardio-metabolic exposures in midlife) linked with health outcomes (e.g., cognitive decline after 70 years old). For exposures and disease markers concomitantly observed in late-life, we will also exploit task 2 developments and separate causal associations between exposures and markers from changes in exposures due to the underlying disease.

Task 4: Dissemination. We will largely improve the diffusion of up-to-date statistical models with validated user-friendly R packages. By implementing a parallelized estimation algorithm, we will speed up all the estimation programs developed by the team in the past and within this project.

DyMES will leverage data from four complementary cohorts in AD: the population-based cohorts Paquid and 3C study, the MEMENTO study, and the large American Nurses’ Health Study with lifestyle exposures collected from midlife. In MSA, we will analyze one of the largest cohorts worldwide from the French national reference center.

Thanks to the access to high quality data and the multidisciplinary project team comprising statisticians, epidemiologists and neurologists working in close vicinity, the project has the ambition to:

- provide innovative statistical models with efficient associated programs to the statistical and epidemiological communities that are essential for the analysis of multivariate longitudinal data;

- answer burning questions such as unravel the influences between processes involved in a disease, predict its progression or identify lifestyle behaviors and time windows to optimize prevention.

Although motivated by MSA and AD, the project will apply far beyond. It will enhance the statistical analysis of longitudinal epidemiological studies by providing valid reproducible methods, and the understanding of chronic diseases with repercussions for public health in the long-term.