Adaptation of Autoencoder for Sparsity Reduction From Clinical Notes Representation Learning.

Area of Science:

• Machine Learning in Healthcare
• Natural Language Processing for Clinical Text
• Data Mining and Representation Learning

Background:

• Clinical text classification on small datasets often struggles with feature space sparsity.
• Traditional feature selection methods may not adequately address the non-linear dependencies and sparsity in clinical data.
• Multilayer perceptrons show promise but can be further optimized through effective feature representation.

Purpose of the Study:

• To develop an alternative approach for tackling sparsity in clinical representation feature spaces.
• To effectively compress high-dimensional, sparse clinical data, enabling analysis of limited datasets, including French clinical notes.
• To improve the performance of neural network classifiers by evaluating them in a compressed feature space.

Main Methods:

• An autoencoder learning algorithm was proposed to reduce dimensionality and exploit sparsity in clinical note representations.
• The study focused on compressing the feature space of clinical notes to enhance learning representations.
• Classification performance was evaluated using classifiers trained and tested on the compressed feature space.

Main Results:

• The autoencoder approach yielded overall performance gains of up to 3% across test set evaluations.
• The classifier achieved high performance metrics: 92% accuracy, 91% recall, 91% precision, and 91% F1-score.
• The theoretical information bottleneck framework was used to demonstrate the autoencoder's compression and prediction mechanisms.

Conclusions:

• Autoencoder learning effectively addresses sparsity in small clinical narrative datasets, improving downstream classification.
• The algorithm's lossless compression capacity allows for learning optimal data representations, outperforming deep learning models in this context.
• This method significantly enhances the classification ability for sparse, high-dimensional clinical data.