Visualization and unsupervised predictive clustering of high-dimensional multimodal neuroimaging data.

Area of Science:

• Neuroscience
• Machine Learning
• Data Science

Background:

• Supervised machine learning in neuroimaging requires labeled data, which is not always available or optimal.
• Unsupervised methods are needed for exploring unknown patterns in high-dimensional neuroimaging datasets.

Purpose of the Study:

• Investigate the utility of t-distributed stochastic neighbour embedding (t-SNE) for unsupervised pattern discovery in neuroimaging data.
• Identify novel sample population structures within unlabeled multimodal neuroimaging scans.

Main Methods:

• Applied t-distributed stochastic neighbour embedding (t-SNE) to pre-processed multimodal neuroimaging scans from 92 healthy subjects.
• Utilized K-means clustering for further analysis of t-SNE-identified patterns.
• Compared t-SNE performance against principal component analysis.

Main Results:

• t-SNE successfully separated subjects into two distinct clusters based on unlabeled data, corresponding to gender (silhouette index=0.79).
• An unsupervised clustering model derived from t-SNE clusters achieved 93.5% accuracy in identifying subject gender.
• t-SNE demonstrated superior performance in pattern identification compared to principal component analysis.

Conclusions:

• Unsupervised t-SNE is a powerful tool for identifying hidden structures in high-dimensional neuroimaging data without prior labels.
• This approach holds promise for data-driven discovery of disease phenotypes and patient subgroups in neuropsychiatry.
• The method facilitates the exploration of complex neuroimaging datasets for novel biological insights.