Multiple additive regression trees with hybrid loss for classification tasks across heterogeneous clinical data in...

Area of Science:

• Machine Learning
• Bioinformatics
• Data Science

Background:

• Multiple Additive Regression Trees (MART) are common for classification but struggle with overfitting in big data.
• Overfitting in distributed, heterogeneous, and imbalanced big data remains an underexplored challenge for MART.

Purpose of the Study:

• To investigate and resolve overfitting effects in distributed MART for disease classification.
• To evaluate a hybrid loss function approach for enhanced model training on complex clinical datasets.

Main Methods:

• Utilized distributed MART with a hybrid loss function for disease classification.
• Employed lexical and semantic analysis to harmonize 10 heterogeneous clinical datasets.
• Applied data augmentation techniques to mitigate class imbalance issues.

Main Results:

• Achieved an 80% overlap in terminology matching across heterogeneous datasets.
• Data augmentation yielded virtual data with a goodness of fit of 0.01.
• Demonstrated improved performance with an average increase of 7.3% in accuracy, 6.8% in sensitivity, and 10.4% in specificity.

Conclusions:

• The proposed distributed MART approach effectively resolves overfitting in big data disease classification.
• Hybrid loss and data augmentation are crucial for handling heterogeneous and imbalanced clinical data.
• This method offers a robust solution for improving diagnostic model performance in real-world healthcare settings.