Selective Prediction With Long Short-term Memory Using Unit-Wise Batch Standardization for Time Series Health Data...

Area of Science:

• Health Informatics
• Machine Learning
• Time Series Analysis

Background:

• Accurate classification and confidence assessment are crucial in healthcare systems.
• Selective prediction models are needed for time series health data with confidence levels.

Purpose of the Study:

• To develop a method using long short-term memory (LSTM) models with a reject option for classifying time series health data.
• To enhance the performance of LSTM models in selective prediction by addressing limitations in conventional approaches.

Main Methods:

• Adopted an existing selective prediction method for LSTM models, incorporating a reject option.
• Proposed unit-wise batch standardization to normalize hidden units in LSTM, improving the selection function's performance.
• Evaluated the method on human activity and arrhythmia time series health data.

Main Results:

• The proposed method achieved lower average coverage violations (0.98% and 1.79%) compared to conventional approaches.
• Demonstrated superior performance in selective risk, false-positive rates, and false-negative rates across both datasets.
• Outperformed other normalization methods when using the reject option for classification.

Conclusions:

• The unit-wise batch standardization approach enhances selective prediction for time series health data.
• This technique is expected to increase user confidence in AI classification systems.
• The method can improve human-AI collaboration in the medical field by providing confidence-aware classifications.