Transformer-Based Automatic Speech Recognition Development for Transcription and Clinical Information Extraction in Indonesian on Electronic Health Records

Integrating Automatic Speech Recognition (ASR) into Electronic Health Records (EHR) has the potential to boost the efficiency of clinical documentation by automating the transcription of doctor-patient conversations. However, high-quality transcription doesn't necessarily mean the clinical information will be classified accurately. This study aims to analyze the quality of ASR transcriptions and compare the performance of Machine Learning algorithms in classifying clinical units based on Indonesian clinical texts. The dataset used consists of 1,000 medical records distributed across five clinical units: Cardiology, ENT, General, Internal Medicine, and Pediatrics. Transcription quality was evaluated using Word Error Rate (WER), while classification was done through Term Frequency-Inverse Document Frequency (TF-IDF) feature extraction and four algorithms: Random Forest, Logistic Regression, Decision Tree, and Naive Bayes. Model evaluation was carried out using accuracy, precision, recall, and F1-score on 200 test data. The research results show that the average WER is 0.40%, indicating very good ASR transcription quality. Random Forest achieved the highest accuracy at 44.50%, Logistic Regression 42.00%, Decision Tree 38.00%, and Naive Bayes 32.00%. Random Forest performed best in the Internal Medicine Clinic with a precision of 0.58 and an F1-score of 0.53. These findings suggest that the main challenge lies in the representation of textual features, mainly due to the similarity of medical vocabulary across clinical units. This study highlights the need to develop approaches based on context, clinical entities, ICD-10 codes, and Transformers to improve classification accuracy and support the automation of clinical information management in EHRs. This approach is expected to strengthen semantic understanding, reduce classification errors, and improve system reliability as a supporter of digital clinical documentation in healthcare settings