Using multiparametric data with missing features for learning patterns of pathology.

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Multimodal classification is challenging when datasets have missing data for some subjects.
• Severe pathologies often lead to incomplete data, limiting traditional classification methods.
• Existing methods may discard subjects with missing data, reducing the classifier's scope.

Purpose of the Study:

• To develop a robust multimodal classifier that handles subjects with incomplete data.
• To broaden the applicability of classifiers to more severe pathologies.
• To maximize the use of available information from multiple modalities.

Main Methods:

• An ensemble-based approach using subsets of complete data.
• Training individual classifiers on these subsets.
• Weighted aggregation of classifier outputs for optimal probabilistic scoring.

Main Results:

• Applied to autism spectrum disorder (ASD) data using magnetoencephalography (MEG) and diffusion tensor imaging (DTI).
• Achieved 83.3% average 5-fold accuracy and 88.4% testing accuracy, distinguishing ASD from controls.
• Outperformed single-modality and complete-data multimodal classifiers (78.3%).

Conclusions:

• The proposed method effectively handles missing data in multimodal datasets.
• This framework enhances classification for conditions like ASD.
• The approach is versatile and applicable to various modality combinations.