Introduction

Obesity and age are well-known risk factors for Alzheimer's disease. Previous investigations showed a reduction in the quantitative Magnetisation transfer (qMT) forward exchange index (k_f) in both grey and white matter areas involved in Alzheimer disease progression. The reduced capacity of macromolecules to exchange magnetisation with surrounding water molecules is believed to be a signal of axonal membranes, myelin damage and neuroinflammation^1-4. We recently proposed a model assuming obesity may produce early limbic grey matter regions changes through the mediation of white matter limbic connections⁵. We focus here specifically on k_f index in Fornix and Hippocampus. We investigated the model through a causal analysis approach. Causal inference⁶ is today used intensively in epidemiology, social sciences and econometrics. The distinct power of causal analysis approaches is to actively use both prior knowledge and experimenter hypothesis to simulate a true-experiment (i.e. an "interventional" study like a random trial) only on the basis of observational data (i.e. by performing an "observational" study).

Aim

Our aim is to answer to the following counterfactual questions: “which percentage of the study population would have presented reduced macromolecular exchange in the hippocampus, given the whole study population would have been forced to be not overweight?”.

Methods

We collected age, waist-to-hip ratio (whr) and Magnetic resonance imaging (MRI) data on a cohort of 151 healthy volunteers (85 females and 56 males, aged 56±8y). MRI acquisitions were performed on a 3T MAGNETOM Prisma clinical scanner equipped with a 32-channel receiver head coil (Siemens Healthcare, Erlangen, Germany). We acquired: 1) a qMT dataset, as reported in⁷; 2) a High Angular Resolution Diffusion Imaging (HARDI) dataset; 3) a high resolution 3D-T1 magnetization-prepared rapid gradient-echo (MPRAGE). From them we extracted respectively: 1) qMT maps, including k_f, as previously reported by Cercignani et al.⁷; 2) fornix tracts, using an home made artificial intelligence based automated tract recognition script and ExploreDTI⁸; 3) hippocampus volumes of interest, using Freesurfer⁹. For each subject, after inter-modality mutual information affine registration by ANTs¹⁰, we calculated the mean k_f on fornix (fornix_kf) and bilateral hippocampus (hc_kf). We performed analysis in both a continuous, by Structured Equation model (SEM), and discretised way, by Probabilistic SEM (PSEM). Continuous approach was realised on the original data (age, whr, hc_kf, fornix_kf) using TETRAD software (Carnegie Mellon University, Pittsburgh, PA, UShttp://www.phil.cmu.edu/tetrad/). Discretised approach was realised on discrete data (gender plus binned age, whr, hc_kf, fornix_kf) using BayesiaLab software (Bayesia S.A.S.,Changé, France https://www.bayesia.com). Firstly, we instantiated a SEM in TETRAD, and evaluated it against our data (see Figure 1). As we know gender may affect whr, hc_kf and fornix_kf, but being a discrete variable cannot be used in our continuous variable SEM, we stratified data by gender (i.e. performed separate analysis for male and female groups). SEM coefficients were computed by multilinear regression, and p-value was estimated, being null-hypothesis a zero coefficient. Model was estimated by comparing the covariance matrix implied by estimated coefficients with the true data covariance matrix. A chi-square statistics was produced and a p-value was determined, being null-hypothesis a correctly specified model. Secondly, we crated a PSEM in BayesiaLab (see Figure 2), and estimated the model Joint Probability Distribution. We examined its fitness by Contingency Table Fit (CFT) score, based on log-likelihood¹¹, and ranging between 0% (completely disconnected network) and 100% (fully connected). We set whr as intervention variable and simulated the effect of eliminating overweight occurrences in males and female subpopulations for both age ranges.

Results

In males, all effects in the SEM were found to be highly significant, and the analysis of induced covariance matrix showed that the null hypothesis cannot be rejected, confirming our causal model is consistent with the data. Differently than for male group, in female group only age was found to cause k_f variations. The CTF score for the model was find to be 85%. For pre-intervention probability distribution and for intervention results see respectively Figure 2 and Figure 3. As expected from TETRAD analysis, intervention on females’ whr did not produce significant effects. On the contrary, on males’ group, the intervention did decrease the occurrence of lower k_f values in the population.

Discussion

Our results show that the age is an important factor in determining k_f decrease, and that the body weight may affect in a different ways males and females, as well as the different ages. In particular, on older than 56 years males, intervention on weight may reduce dramatically the occurrence of low macromolecular exchange rates, possible indicator of inflammation and myelin/axonal damages.

Acknowledgements

We acknowledge Lionel Jouffe and Bayesial S.A.S. for providing a trial licence of ByesiaLab software, as well as Greg D Parker, for sharing with CUBRIC community his automated script for white matter tracts extraction.