Decentralized distribution-sampled classification models with application to brain imaging.

Area of Science:

• Computer Science
• Machine Learning
• Data Science

Background:

• Big data models necessitate large, often distributed, datasets.
• Transferring data between local sites poses privacy and network load challenges, particularly for sensitive medical imaging data governed by HIPAA.
• Existing decentralized methods often incur significant network traffic.

Purpose of the Study:

• To develop a novel singleshot decentralized classification method.
• To reduce network load in decentralized machine learning while preserving accuracy.
• To address challenges in handling large, distributed datasets, especially in medical imaging.

Main Methods:

• Implemented a singleshot approach for neural networks and support vector machines.
• Each local site estimates its data's statistical distribution.
• Sites resample data from individual distributions and train models on local and resampled data, averaging accuracy for a global result.

Main Results:

• Demonstrated effectiveness in handwritten digit classification.
• Achieved comparable classification accuracy to centralized models in multi-subject brain imaging for schizophrenia.
• Showcased significantly lower network load compared to multishot decentralized methods.

Conclusions:

• The proposed singleshot decentralized method performs comparably to centralized approaches.
• This approach effectively minimizes network traffic compared to existing multishot methods.
• Offers a viable solution for accurate decentralized classification with reduced network burden.