Design and Implementation of an AI-Powered Medical Dialogue Assistant

1. Introduction The rapid growth of the technology industry and the explosion of data have driven the resurgence of artificial intelligence, especially in human‑machine dialogue systems. The article introduces the Turing Test as a foundational concept and motivates the need for intelligent medical assistants.

2. Problem Background In online medical services, patients ask questions via web or app while doctors answer, leading to two main issues: insufficient doctor resources and delayed responses. An AI‑based medical dialogue assistant is proposed to mitigate these problems by providing timely feedback and disease probability estimation.

3. Solution Overview The solution consists of two parts: (1) industry data construction and (2) intelligent assistant development.

3.1 Industry Data Construction

3.1.1 Data Acquisition Ideal data would be standard electronic medical records (EMRs) containing pairs. Since EMRs are confidential, two public alternatives are used: multi‑turn Q&A data from online medical websites and disease‑symptom relations extracted from medical textbooks.

3.1.2 Data Processing

3.1.2.1 Data Structuring All sources are converted to a unified format, key fields are extracted, and expert annotation adds labels, especially for disease conclusions.

3.1.2.2 Data Normalization Medical concepts are normalized using SNOMED CT, selecting a standard name and mapping synonyms to it.

3.2 Intelligent Assistant Construction

3.2.1 Overall Framework The assistant’s service logic includes modules for department classification, intent recognition, multi‑turn questioning, automatic QA, and escalation to human agents. Sensitive business modules are omitted for confidentiality.

3.2.2 Key Tasks

3.2.2.1 Disease Range Prediction A hierarchical approach reduces the search space: department → intent → static disease prediction (first turn) → dynamic disease prediction (subsequent turns). The prediction model is a Dynamic Convolutional Neural Network (DCNN) treating disease range as a multi‑class problem.

3.2.2.2 Information Extraction in Dialogue Essential user information (e.g., gender, age, symptoms) is extracted using two complementary methods:

Semantic parsing based on syntactic templates to capture regular symptom expressions.

Bi‑LSTM + CRF sequence labeling to handle colloquial utterances, identifying symptom boundaries (B/I/O tags).

3.2.2.3 Question Generation After determining the next inquiry point, the system generates natural‑language questions using template‑based construction, balancing depth (symptom → sub‑attribute) and breadth (parallel symptoms) to drive conversation convergence.

4. Technical Extensions Building on the core system, reinforcement learning was introduced to further improve dialogue policy. Experiments on an independently annotated dataset showed significant gains over traditional methods, and the results were published at ACL 2018.

References

[1] SNOMED CT – https://www.nlm.nih.gov/healthit/snomedct/ [2] A Convolutional Neural Network for Modelling Sentences – http://www.aclweb.org/anthology/P14-1062 [3] Neural Architectures for Named Entity Recognition – https://www.aclweb.org/anthology/N16-1030 [4] Task‑oriented Dialogue System for Automatic Diagnosis – http://www.aclweb.org/anthology/P18-2033