Automated annotation of functional imaging experiments via multi-label classification.

Area of Science:

• Neuroscience
• Computer Science
• Information Science

Background:

• Accurate classification of human neuroimaging experiments is crucial for effective meta-analyses.
• Manual classification of experimental methods is time-consuming and resource-intensive.
• Automating this process can significantly enhance research synthesis.

Purpose of the Study:

• To evaluate the performance of machine learning (text mining) methods for automatically classifying neuroimaging experiment abstracts.
• To replicate human expert annotations of experimental stimuli, cognitive paradigms, and response types.
• To assess the feasibility of using off-the-shelf software for literature labeling.

Main Methods:

• Utilized abstracts of published functional neuroimaging papers as the text corpus.
• Employed machine learning algorithms including Naive Bayes (NB), k-nearest neighbor, and Support Vector Machines (SMO).
• Used human expert performance as training data for classification models, with labels from the Cognitive Paradigm Ontology (CogPO).

Main Results:

• Exact match performance varied from 15% to 78% across different methods.
• Naive Bayes methods, particularly with binary relevance transformations, demonstrated strong performance and robustness against overfitting.
• The study achieved notable results with minimal text pre-processing, highlighting the effectiveness of standard machine learning components.

Conclusions:

• Machine learning, specifically text mining, offers a viable solution for automating the classification of neuroimaging experimental methods.
• Naive Bayes methods show particular promise for robust and accurate literature labeling.
• Automated classification can streamline the process of preparing data for neuroimaging meta-analyses.