Validation of Krakencoder’s ability to predict cognition and demographics across lifespan datasets

1505 

Abstract Submission 

Qinxin Wang¹, Keith Jamison², Amy Kuceyeski²

¹Cornell university, ITHACA, NY, ²Weill Cornell Medicine, New York City, NY

Qinxin Wang  
Cornell university
ITHACA, NY

Keith Jamison  
Weill Cornell Medicine
New York City, NY

Amy Kuceyeski  
Weill Cornell Medicine
New York City, NY

Krakencoder is a simultaneous, multi-atlas, multi-modality brain connectivity set of encoding models that addresses the challenge of integrating diverse brain connectivity types that each carry unique information and possible noise (Jamison OHBM 2023). Its architecture includes a set of encoders and decoders designed to transform one connectivity type into another through the shared latent representation (See Figure 1). It was trained on resting-state functional and white-matter structural connectivity data from 700 healthy young adults in the Human Connectome Project (HCP-YA), optimizing for reconstruction accuracy and within-subject similarity in latent-space. Building on recent studies that demonstrated the effectiveness of kernel regression models for predicting demographics from brain connectomes (Li 2019, Ooi 2022), we trained kernel regression or classification models to predict sex, age and cognition. Using held-out subjects from HCP-YA, as well as two out-of-sample datasets, we compared the predictions of subject demographics from latent representations and the raw brain connectivity data, demonstrating Krakencoder's generalizability to data with diverse subjects and acquisition parameters.

Our study utilized three datasets from the Human Connectome Project, drawn from individuals across the lifespan, including young adults (HCP-YA, 22-37), developing individuals (HCP-D, 8-22), and an aging population (HCP-A, 36-86). Beyond age differences, the HCP D/A acquisitions are approximately half the length of HCP-YA. All datasets include three parcellations (86, 268, and 439 regions), three common functional connectivity (FC) metrics (Pearson correlation, Pearson correlation with global signal regression, and regularized partial correlation), and two structural connectivity (SC) approaches (streamline counts from deterministic and probabilistic tractography), yielding 15 total connectivity types. For Krakencoder, our predictions use the mean latent vectors across FC flavors, SC flavors, or both SC and FC for each subject. For the raw data, we use the mean cosine similarity matrices for FC, SC, or SC and FC. We employed cross-validation in fitting and testing the models, ensuring that family groups are not split.

The analysis of held-out HCP-YA subjects reveals that predictions from Krakencoder's latent representations generally outperform raw connectome data, though raw FC predicts age better than latent FC. FC latent space predicts cognition better than SC latent space. SC latent space better predicts age, and both predict sex with accuracy > 93%. The combination of FC and SC consistently yields high predictive performance for all three variables, outperforming or at least matching the better-performing individual modality.
For out-of-sample HCP-D/A datasets, latent representations perform comparably to raw data. Latent was better in sex prediction but worse in age prediction, with minor differences in cognition prediction. Within latent representations, SC outperforms FC in predicting age, while FC is more predictive of sex and cognition. Combined connectivities also maintain relatively high performance.

Krakencoder effectively reduces the high dimensionality of raw data (677,870 dimensions) into low-dimensional vectors (128 dimensions) within the latent space, facilitating an integrated and aligned representation across diverse modalities. The resulting latent representations largely preserve the predictive properties inherent in the raw data, as evidenced by the marginal disparities in their predictive performances observed in our results. Although Krakencoder was originally trained on the HCP-YA dataset, it exhibits strong generalizability when applied to the HCP-A&D datasets, despite differences in data acquisition methods and substantial variations in the age distribution of subjects. This robust performance across distinct datasets underscores Krakencoder's generalizability and versatile applicability.

Classification and Predictive Modeling ²

Connectivity (eg. functional, effective, structural) ¹

FUNCTIONAL MRI

Machine Learning

NORMAL HUMAN

WHITE MATTER IMAGING - DTI, HARDI, DSI, ETC

Other - functional connectivity; structural connectivity; multimodal; brain-behavior mapping; structure-function coupling