Multi-Label Classification of Bilingual Doctor Responses in Online Medical Consultations Using Deep Learning

Online health consultations (OHCs) have become an integral component of modern healthcare delivery. However, significant challenges remain in multilingual and low-resource contexts such as Indonesia, where language barriers and digital disparities hinder effective doctor–patient communication. Ensuring the quality of such interactions requires the identification of six key communicative functions: building relationships, gathering and providing information, decision-making, promoting disease- and treatment-related behaviour, and responding to emotions. While existing research has largely focused on English-language OHCs, studies analysing these communicative functions in Indonesian remain limited due to the lack of annotated datasets and linguistic complexity. To address this gap, we propose a deep learning framework for multi-label classification of communicative functions in bilingual (Indonesian/English) doctor response texts. The dataset used in this study was annotated by medical professionals with six predefined communicative function labels. We conducted a comprehensive comparative evaluation of three deep learning architectures namely Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), and Convolutional Neural Networks (CNN) equipped with cross-language word embedding to improve multilingual generalization. Model performance is evaluated through four complementary perspectives: example-based, label-based, ranking-based, and multifaceted metrics, ensuring a holistic assessment. Result show that the fine-tuned LSTM model achieved the highest precision (0.972) on Indonesian texts, while Bi-LSTM obtained the best results on English texts with 0.890 accuracy and 0.980 precision. The LSTM model also reduced false positives in Indonesian classifications, whereas Bi-LSTM improved diagnostic reliability in English, confirming the models’ cross-lingual adaptability. These findings highlight the potential of deep learning to improve communication effectiveness in bilingual and resource-constrained OHC settings.