Deep Neural Network Driven Speech Classification for Relevance Detection in Automatic Medical Documentation

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Speech Processing

Background:

• Automating medical documentation can reduce healthcare costs and time.
• Automatic Speech Recognition (ASR) and deep learning show promise for this automation.
• The efficiency of ASR systems is limited by the volume of processed speech, with over half being irrelevant in follow-up examinations for Intra-Vitreal Injections.

Purpose of the Study:

• To evaluate Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks for a speech classification module.
• To identify speech relevant for medical report generation.
• To analyze the impact of various topology parameters and speaker attributes on model performance.

Main Methods:

• Development of a speech classification module using CNNs and LSTMs.
• Testing various network topology parameters.
• Analysis of model performance across different speaker attributes (gender, accent, unknown speakers).

Main Results:

• CNNs outperformed LSTMs in speech classification accuracy.
• Achieved a validation accuracy of 92.41% using CNNs.
• The CNN model demonstrated satisfactory generalization across diverse speaker attributes.

Conclusions:

• CNNs are effective for classifying relevant medical speech, significantly improving documentation efficiency.
• The developed model shows robustness to variations in speakers, making it suitable for real-world healthcare applications.
• This approach can enhance the automation of medical documentation, leading to substantial time and cost savings.