Uncovering Lifespan-Consistent Representations in Cognitive Function through Metric Learning

Poster Session C - Sunday, April 14, 2024, 5:00 – 7:00 pm EDT, Sheraton Hall ABC

Xiaoxiao Sun¹ (xiaoxiao.sun@columbia.edu), Chichi Chang¹, Arunesh Mittal¹, Christian Habeck², Yaakov Stern²; ¹Columbia University, ²Columbia University Irving Medical Center

Changes in cognitive function across the lifespan can be captured by four core reference abilities: episodic memory (MEM), fluid reasoning (FLUID), perceptual speed (SPEED), and vocabulary (VOCAB). Stern et al. collected fMRI data across tasks targeting these reference abilities and found no significant age-related differences in Reference Ability Neural Networks (RANNs) across age groups (20 to 80 years). Efforts to distinguish these networks with linear analytical approaches, including principal components analysis (PCA) and linear-indicator regression, resulting in accuracy of 77% (+- 5%, MEM: 76%; FLUID: 82%; SPEED: 79%; VOCAB: 71%). Here, we consider a non-linear approach based on metric learning, which constructs latent spaces from the same high-dimensional fMRI data and is interpretable via sampling. Our metric learning approach yielded a substantial improvement in classification accuracy (87 +- 5%) across all four reference abilities by acquiring the latent embeddings that reduce the intra-class variability while maximizing the inter-class variability. With these embeddings, we were able to extract more robust and interpretable spatial activation patterns. On pairwise classification we achieved an even higher average accuracy of 96% (+-2%) across the six binary comparisons (MEM vs. FLUID: 96%; MEM vs. SPEED: 99%; MEM vs. VOCAB: 97%, FLUID vs. SPEED: 95%; FLUID vs. VOCAB: 97%; SPEED vs. VOCAB: 94%). Age-related changes in networks underlying cognition can provide insights into possible cognitive decline, and alert physicians and caregivers of a need for treatment or intervention. Our findings demonstrate that RANN biomarkers can be made more robust and interpretable using non-linear methods employing metric learning.

Topic Area: METHODS: Neuroimaging