Journal: iScience

Article Title: More reliable biomarkers and more accurate prediction for mental disorders using a label-noise filtering-based dimensional prediction method

doi: 10.1016/j.isci.2024.109319

Figure Lengend Snippet: The overall workflow of the proposed label-noise filtering-based dimensional prediction (LAMP) method using fMRI data of HC-SZ The construction of the LAMP method using FNC features of HC-SZ data is displayed in (A–D), and the evaluation of the method is displayed in (E and F). (A) We extracted the FNC features for each subject from four different fMRI datasets based on previous work, i.e., NeuroMark, which is a fully automated independent component analysis pipeline. (B) We sequentially took one of the four datasets as the independent dataset and the remaining three as the source datasets, namely, the leave-one-dataset-out division strategy. (C) We identified typical subjects in each source dataset. Specifically, we evaluated each subject’s label-dependability that indicates the consistency level between the diagnostic label and fMRI measures via the CRF-based label-noise filtering model, and discarded subjects whose label-dependability was lower than a predefined threshold γ . After that, we thresholded the label-dependability of the remaining subjects, resulting in the typical subjects (i.e., typical HCs and typical SZ patients). Notably, the CRF-based model regards the subjects that are surrounded by the homogeneous subjects as typical subjects. (D) We constructed a dimensional prediction model guided by the typical subjects from various source datasets to provide a dimensional score for each independent subject. In detail, we first predicted a separate score revealing the degree of brain dysfunction for each independent subject according to its relationship to different typical groups in each source dataset. Next, we averaged the three separate scores derived from the source datasets to get a comprehensive score for the independent subject. Then, we relabeled the independent subjects according to the comprehensive score with an adaptive parameter τ and obtained the relabeled HC group, relabeled SZ group, and the Boundary group in which these subjects were mild and could not be categorized into HC or SZ group with enough confidence. (E) We verified the performance of the typical groups under different label-dependability thresholds γ in light of the inter-group separability and intra-group compactness and analyzed the significance and consistency of functional abnormality within the typical SZ group across multiple source datasets. (F) We evaluated the stability of the scores, intra-group compactness, and inter-group separability of the relabeled HC and SZ groups and analyzed the significant differences between the two relabeled groups within the independent dataset. IC denotes the independent component. THC, TSZ, ReHC, and ReSZ represent typical HC, typical SZ, relabeled HC, and relabeled SZ groups in a dataset, respectively. CRF is short for the complete random forest-based label-noise filtering model.

Article Snippet: As shown in A, we first extracted functional network connectivity (FNC) features for each subject using fMRI data via the NeuroMark, a fully automated independent component analysis pipeline.

Techniques: Diagnostic Assay, Construct, Derivative Assay, Functional Assay