Domain Adaptation via Low Rank and Class Discriminative Representation for Autism Spectrum Disorder Identification: A Multi-Site fMRI Study

Abstract

To construct a more effective model with good generalization performance for inter-site autism spectrum disorder (ASD) diagnosis, domain adaptation based ASD diagnostic models are proposed to alleviate the inter-site heterogeneity. However, most existing methods only reduce the marginal distribution difference without considering class discriminative information, and are difficult to achieve satisfactory results. In this paper, we propose a low rank and class discriminative representation (LRCDR) based multi-source unsupervised domain adaptation method to reduce the marginal and conditional distribution differences synchronously for improving ASD identification. Specifically, LRCDR adopts low rank representation to alleviate the marginal distribution difference between domains by aligning the global structure of the projected multi-site data. To reduce the conditional distribution difference of data from all sites, LRCDR learns the class discriminative representation of data from multiple source domains and the target domain to enhance the intra-class compactness and inter-class separability of the projected data. For inter-site prediction on all ABIDE I data (1102 subjects from 17 sites), LRCDR obtains the mean accuracy of 73.1%, superior to the results of the compared state-of-the-art domain adaptation methods and multi-site ASD identification methods. In addition, we locate some meaningful biomarkers: Most of the top important biomarkers are inter-network resting-state functional connectivities (RSFCs). The proposed LRCDR method can effectively improve the identification of ASD, and has great potential as a clinical diagnostic tool.

Research on multi‐center assisted diagnosis of ASD based on multimodal feature fusion

“…The discriminant function for multi‐site source data can be formulated as follows: $$$$ \varPsi (P)={\varPsi}_{same}(P)-\tau {\varPsi}_{diff}(P)= tr\left({P}^T{X}_s\left({D}_{same}-\tau {D}_{diff}\right){X}_s^TP\right) $$$$ where $$$ {\varPsi}_{same}(P) $$$ and $$$ {\varPsi}_{diff}(P) $$$ are the intra‐class and inter‐class distance loss terms, and λ is used to balance these two loss terms. $$$ {D}_{same}\in {R}^{n_s\times {n}_s} $$$ and $$$ {D}_{diff}\in {R}^{n_s\times {n}_s} $$$ are the matrices used to compute intra‐class and inter‐class distances for projected multi‐site source data 26 …”

“…D same R n s Ân s and D diff R n s Ân s are the matrices used to compute intra-class and inter-class distances for projected multi-site source data. 26 Supposing x si and x sj are any two samples from multiple source domains, with labels y si and y sj Ψ same P ð Þ and Ψ diff P ð Þ are expressed as follows:…”

Joint multi-site domain adaptation and multi-modality feature selection for the diagnosis of psychiatric disorders

“…To verify the effectiveness of the proposed HMSDA method, we compared it with alternative DA methods, including: (1) LRR [32], aiming to find a low-dimensional representation for multiple domains, (2) multi-site adaption framework via low-rank representation decomposition (maLRR) [7], learning the common component from specific projection matrices of multiple source domains as the projection matrix of the target domain, (3) transfer subspace learning via low-rank and sparse representation (TSL_LRSR) [33], projecting the data of the two domains into the common subspace to learn the domain invariant feature representation, (4) geodesic flow kernel (GFK) [34], leveraging low-dimensional data structures to address the dissimilarity in data distributions between the source and target domains, (5) low-rank and class discriminative representation (LRCDR) [35],…”

“…The regularization parameters for LRR, maLRR, TSL_LRSR, LRCDR and MSLRDA were selected from the set [1e -5 , 1e -4 , 1e -3 , 1e -2 , 1e -1 ], and the parameter (i.e., dimensionality of the subspace) in GFK was chosen from [5, 10, …, 50] via 5-fold cross-validation. It is worth noting that for maLRR, TSL_LRSR, GFK and LRCDR, we followed an identical experimental setup as Liu et al [35] to perform dimensionality reduction for obtaining the optimal results. We conducted experiments separately on the FBIRN and ABIDE I datasets, each domain was selected as the target domain in turn, and the remaining domains were treated as the source domains.…”