Latent Brain Age Predicts Mortality in Amyotrophic Lateral Sclerosis.

2245 

Abstract Submission 

Ayodeji Ijishakin¹, Florence Townend², Edoardo Spinelli³, Silvia Basaia⁴, Parido Schito⁵, Yuri Falzone⁵, Massimo Fillipi⁵, Federica Agosta⁵, James Cole⁶, Andrea Malaspina⁷

¹Centre for Medical Image Computing, London, UK, ²University College London, London, Greater London, ³San Raffaelle Scientific Institute, Italy, Italy, ⁴San Raffaelle Scientific Institute, London, Italy, ⁵San Raffaelle Scientific Institute, Milan, Italy, ⁶University College London, London, London, ⁷Department of Neuromuscular Diseases, UCL Queen Square Institute of Neurology, University College Lo, London, UK

Ayodeji Ijishakin  
Centre for Medical Image Computing
London, UK

Florence Townend  
University College London
London, Greater London

Edoardo Spinelli  
San Raffaelle Scientific Institute
Italy, Italy

Silvia Basaia  
San Raffaelle Scientific Institute
London, Italy

Parido Schito  
San Raffaelle Scientific Institute
Milan, Italy

Yuri Falzone  
San Raffaelle Scientific Institute
Milan, Italy

Massimo Fillipi  
San Raffaelle Scientific Institute
Milan, Italy

Federica Agosta  
San Raffaelle Scientific Institute
Milan, Italy

James Cole, PhD  
University College London
London, London

Andrea Malaspina  
Department of Neuromuscular Diseases, UCL Queen Square Institute of Neurology, University College Lo
London, UK

The neuroanatomical changes associated with neurodegenerative disease progression appear to overlap with natural ageing. This suggests that changes akin to the ageing process might serve as indicators for predicting mortality in such diseases. In fast progressing neurodegenerative diseases such as amyotrophic lateral sclerosis (ALS), methods producing age related brain biomarkers must be particularly sensitive to the neuroanatomical effects of ageing. Generative models offer a solution to this problem as they have achieved state-of-the art representation learning capabilities. In this work, we leveraged a state-of-the art generative model called a diffusion autoencoder, in-order to produce age conditioned representations which are more effective than classic approaches for ALS prognostication.

Our approach combines a diffusion autoencoder model with a multilayer perceptron (MLP) which predicted survival time in days of ALS patients, based on their latent similarity to healthy individuals of the same age. The algorithm is visualised in Figure 1. First, the diffusion autoencoder is trained on MR images of healthy individuals, conditioned on their age, for N epochs. We then fix the weights of the diffusion autoencoder and begin to train the survival prediction MLP. To do so, we sample an individual with ALS of age = k. We then sample all MR images from healthy individuals of age = k, and place the healthy individuals and ALS patients through the encoder. Then the absolute difference between the mean latent representation of the healthy individuals and the latent representation of the ALS patient is computed. This quantity is the latent brain age. The latent age difference vector, then goes through the MLP to predict survival time.

We trained our model on a dataset of, 4621 2D T1-weighted MR images (mean age = 56 years, std = 20.9 years) from 8 publicly available datasets. Our patient group consisted of 72 2D T1w MR images (mean age = 61 years, std = 10.9 years) from the San Raffaelle Hospital in Milan. The diffusion based pre-training period lasted for 400 epochs, and our MLP was trained for another 400 epochs. The MLP was trained on 57 ALS patients, leaving 17 patients in the validation set. All training was performed on an Nvidia GeForce RTX 4090 graphics card, with the Adam optimizer in PyTorch lightning.

We achieved a validation accuracy of 0.77(p<0.01) as measured by the Pearson's r correlation between predicted survival and actual survival in days, with a mean absolute error (MAE) of 8.19 months. Table 1 displays how are method compares to other neural network based approaches to predicting survival with the same training data. The final column is a linear regression approach using, age, sex and the ALS functional rating score. The results demonstrate that we outperform them on our validation set. As the other approaches were not generative models, the pre-training task used was standard brain age prediction on the MR images derived from healthy controls. Table 2 displays results of a survival analysis via a cox proportional hazard regression using age, sex and brain predicted age differences (brain-PADs) of the ALS cohort. None of the covariates were significant predictor's of survival, in contrast to our approach.

The present study was a preliminary exposition of the concept of latent brain age, in which its utility for the prognostication of ALS disease was demonstrated. Our results show that predictions using latent brain age outperform traditional cox proportional hazard analyses using standard brain age measures. Our approach also outperformed non-generative neural network approaches. However, given the small sample size of patient data (train size = 57, test size = 17), an analysis with more subjects is required to ensure the validity of the present findings. Future work should also apply the present approach in other diseases and on 3D MRI volumes.

Classification and Predictive Modeling ²

Informatics Other ¹

Aging

MRI