Code-Switching Automatic Speech Recognition for Nursing Record Documentation: System Development and Evaluation

Area of Science:

• Artificial Intelligence in Healthcare
• Natural Language Processing
• Clinical Documentation

Background:

• Taiwan faces nursing shortages due to high healthcare provider turnover.
• Reducing nurse workload is critical for retaining staff.
• Speech transcription offers potential for efficient nursing record documentation.

Purpose of the Study:

• To develop a speech recognition system for nursing records.
• Enable healthcare providers to complete records with minimal typing.
• Improve efficiency in clinical documentation processes.

Main Methods:

• Created the China Medical University Hospital Artificial Intelligence Speech (CMaiSpeech) dataset (100 hours, 525 speakers).
• Applied transfer learning (TL) and meta-transfer learning (MTL) methods.
• Focused on Mandarin-English code-switching scenarios common in Taiwanese nursing records.

Main Results:

• The proposed syllable-based MTL model achieved a word error rate (WER) of 22.20% in code-switching.
• The CMaiSpeech dataset comprises a Mandarin-English corpus.
• The system demonstrated effectiveness in a clinical case with a WER of 31.06%.

Conclusions:

• Established the CMaiSpeech dataset, a valuable Mandarin-English corpus.
• Developed an automatic speech recognition system for nursing records.
• The system can reduce healthcare provider workload and improve handover efficiency.