Optimizing neural networks for medical data sets: A case study on neonatal apnea prediction

Area of Science:

• Neonatal care
• Computational neuroscience
• Medical data analysis

Background:

• The neonatal period is critical for infant development and health.
• India faces high rates of preterm births, with low birth weight infants susceptible to conditions like apnea.
• Real-time neonatal medical data is complex, noisy, and nonlinear, challenging disease prediction.

Purpose of the Study:

• To optimize neural network architectures for predicting apneic episodes in neonates.
• To develop a generic framework for selecting optimization algorithms and architectures for complex medical datasets.
• To enhance predictive performance for neonatal intensive care unit (NICU) data.

Main Methods:

• Kernel Principal Component Analysis (PCA) for dimensionality reduction.
• Hyper-parametric and parametric optimization of neural networks (learning rate, regularization, activation functions, gradient descent, depth).
• Comparison of optimized Multilayer Perceptron (MLP), Deep Belief Networks, Stacked Auto-encoders against Support Vector Machine (SVM), K Nearest Neighbor, Decision Tree (DT), and Random Forest (RF).

Main Results:

• An optimized eight-layer MLP model achieved an AUC of 0.82 for predicting neonatal apnea.
• The optimized MLP performed comparably to a Deep Auto-encoder (AUC 0.83) and outperformed conventional machine learning models.
• The optimization process significantly improved the predictive performance of the MLP model.

Conclusions:

• Step-wise optimization significantly enhances MLP model predictive performance for noisy, nonlinear medical data.
• The optimized MLP model demonstrates accuracy comparable to deep neural networks and superior to traditional algorithms.
• The proposed optimization framework offers a valuable diagnostic tool for neonatologists.