Adaptive respiratory signal prediction using dual multi-layer perceptron neural networks

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Accurate prediction of respiratory signals is crucial for monitoring patient health.
• Previous multi-layer perceptron neural network (MLP-NN) models showed limitations in prediction accuracy for irregular breathing patterns due to fixed, one-time training data.
• Adapting models to changing respiratory dynamics is essential for improved performance.

Purpose of the Study:

• To enhance the prediction accuracy of respiratory signals by adapting the MLP-NN model to dynamic changes.
• To investigate a continuous learning technique using updated training data to overcome limitations of fixed-data training.
• To compare the performance of a dual-MLP-NN adaptive approach against a single-MLP-NN model.

Main Methods:

• Developed a continuous learning technique employing dual MLP-NNs for respiratory signal prediction.
• Implemented an adaptive approach where one MLP-NN predicts while the other trains on updated data.
• Evaluated prediction performance using root-mean-square-error (RMSE) on 202 patients' 1-minute respiratory recordings.
• Investigated four network configurations: single MLP-NN, high-computation dual MLP-NNs (U1), and two mixed-computation dual MLP-NNs (U2, U3).

Main Results:

• The dual-MLP-NN continuous learning approach significantly improved prediction accuracy compared to one-time training.
• The U1 method (high-computation dual MLP-NNs) achieved the best performance, reducing RMSE by 34% compared to the single MLP-NN.
• RMSE reductions of 19% and 10% were observed for U1 compared to U2 and U3 methods, respectively.
• Continuous training with updated data demonstrated superior performance in handling irregular breathing patterns.

Conclusions:

• Continuous learning with an adaptive dual-MLP-NN configuration substantially improves respiratory signal prediction accuracy.
• The dual-network approach effectively adapts to changing respiratory patterns, outperforming static, single-network models.
• This adaptive strategy offers a promising advancement for real-time respiratory monitoring and prediction.