MedAidDialog: A Multilingual Multi-Turn Medical Dialogue Dataset for Accessible Healthcare
Conversational artificial intelligence has emerged as a transformative tool in the realm of healthcare, offering potential solutions for preliminary medical consultations. This is particularly vital in environments where access to healthcare professionals is restricted. However, many existing medical dialogue systems are often limited by their single-turn question-answering format or their reliance on template-based datasets. These constraints hinder the realism of conversations and restrict their multilingual capabilities.
In response to these challenges, a new dataset named MedAidDialog has been introduced. This dataset is designed to simulate more realistic physician-patient consultations across multiple languages. MedAidDialog builds upon the MDDial corpus by generating synthetic consultations through the use of advanced large language models. Furthermore, it expands into a parallel multilingual corpus encompassing seven languages: English, Hindi, Telugu, Tamil, Bengali, Marathi, and Arabic.
Key Features of MedAidDialog
- Multilingual Support: The dataset includes consultations in seven distinct languages, thereby catering to a diverse population.
- Multi-Turn Dialogue: By allowing multiple exchanges between the physician and patient, the dataset simulates a more natural conversation flow.
- Synthetic Consultations: Utilizing large language models to generate realistic dialogue scenarios enhances the dataset’s applicability in various settings.
- Personalization: The framework integrates optional patient pre-context information, such as age, gender, and allergies, to tailor the consultation experience.
Development of MedAidLM
Building on the MedAidDialog dataset, the researchers developed MedAidLM, a conversational medical model that employs parameter-efficient fine-tuning on quantized small language models. This allows for deployment on less powerful computational infrastructure, making the technology more accessible. The model’s design aims to facilitate effective symptom elicitation through multi-turn dialogue, generating diagnostic recommendations that are coherent and contextually relevant.
Evaluation and Impact
To assess the effectiveness of the proposed system, a series of experimental results were conducted. These evaluations demonstrated that MedAidLM can successfully conduct multi-turn dialogues that lead to accurate symptom identification and sound diagnostic suggestions. Furthermore, medical experts were engaged to evaluate the plausibility and coherence of the generated consultations, ensuring that the model produces reliable and clinically relevant outputs.
Conclusion
The introduction of MedAidDialog represents a significant advancement in the field of conversational AI for healthcare. By addressing the limitations of existing medical dialogue systems and providing a robust multilingual dataset, MedAidDialog has the potential to enhance access to healthcare consultations for diverse populations worldwide. As the demand for accessible healthcare solutions continues to grow, innovations like MedAidDialog and MedAidLM will play a crucial role in bridging the gap between patients and medical professionals.