A deep learning framework with hybrid stacked sparse autoencoder for type 2 diabetes prediction.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Sparse numerical datasets are prevalent in applied mathematics, astronomy, finance, and healthcare.
• High dimensionality and zero-value predominance in these datasets complicate feature selection and data analysis.
• Existing methods struggle with optimal feature selection for sparse, high-dimensional data.

Purpose of the Study:

• To introduce a novel deep learning algorithm, Hybrid Stacked Sparse Autoencoder (HSSAE), for improved feature selection in sparse datasets.
• To enhance the efficiency and robustness of sparse data analysis through advanced deep learning techniques.
• To evaluate the performance of HSSAE against traditional and deep learning models for sparse data classification.

Main Methods:

• Developed the Hybrid Stacked Sparse Autoencoder (HSSAE) integrating L1 and L2 regularization with binary cross-entropy loss.
• Incorporated dropout and batch normalization techniques to improve model generalization and training stability.
• Evaluated HSSAE against Decision Tree, Random Forest, KNN, Naïve Bayes, CNN, LSTM, and Stacked Sparse Autoencoder using metrics like Accuracy, F1-score, and AUC.

Main Results:

• HSSAE demonstrated superior performance compared to traditional and deep learning classifiers on sparse datasets.
• Achieved the highest accuracy of 89% on a health indicator dataset and 93% on an EHRs diabetes prediction dataset.
• The algorithm effectively extracts features and enhances robustness for sparse data applications.

Conclusions:

• The proposed HSSAE algorithm is highly effective for feature selection and analysis of sparse datasets.
• HSSAE offers significant advantages in healthcare applications requiring high prediction accuracy.
• The deep learning approach provides a robust solution for challenges posed by high-dimensional, sparse data.