DoPI: Proactive TCM Diagnostic LLM

• DoPI is a TCM-oriented LLM designed for proactive, multi-turn interrogation that separates symptom elicitation and final diagnosis using a dual-model architecture.
• It leverages a dynamic knowledge graph to rank candidate symptoms and formulate targeted follow-up questions, achieving 84.68% diagnostic accuracy.
• A synthetic multi-turn dialogue dataset supports evaluation, demonstrating DoPI’s superior performance over traditional medical LLMs in structured TCM diagnosis.

DoPI, short for Doctor-like Proactive Interrogation LLM for Traditional Chinese Medicine, is a Traditional Chinese Medicine (TCM)-oriented medical LLM system designed for proactive, multi-turn diagnostic interrogation rather than single-turn question answering. Its central design separates the tasks of asking clinically useful follow-up questions and making the final diagnosis: a guidance model conducts the dialogue and dynamically selects questions through a knowledge graph, while an expert model built on the Sunsimiao framework produces the final disease diagnosis and treatment recommendation. The system is coupled to a synthetic multi-turn doctor-patient dialogue dataset and an evaluation methodology intended to test diagnostic recovery from partial initial symptom descriptions without relying on manually collected real-world consultation logs. The reported headline result is 84.68% Diagnostic Accuracy for DoPI-7B (Sun et al., 7 Jul 2025).

1. Clinical motivation and problem formulation

DoPI was introduced to address a specific failure mode of current medical LLMs in TCM settings: they may answer knowledge-based questions well, yet remain weak at doctor-like, proactive, multi-turn interrogation. In TCM diagnosis, patients rarely provide complete symptom descriptions in one turn. A practitioner therefore must elicit additional symptoms progressively, narrow candidate diseases or syndromes, and integrate the resulting information with domain knowledge and, in some cases, tongue diagnosis.

The paper identifies several reasons this setting is difficult. Real patients often provide only partial, vague, or symptom-fragmentary descriptions. Proactive questioning is essential, but many LLMs mainly respond to the current query instead of steering the consultation through targeted follow-up questions. TCM diagnosis also depends on structured interrogation logic rather than generic conversational fluency: questions are asked according to clinical importance and pattern differentiation. The paper further emphasizes that high-quality real multi-turn TCM consultation data are difficult to collect because of privacy, annotation cost, and clinical complexity, and that directly fine-tuning a medical model on dialogue data risks degrading medical expertise.

A central claim of the DoPI framework is therefore that medical dialogue competence and professional diagnostic competence should not be collapsed into a single undifferentiated model behavior. The system is organized around the premise that eliciting symptoms and making a final diagnosis are related but separable functions.

2. Two-model collaborative architecture

DoPI uses a two-model collaborative framework composed of a guidance model and an expert model. The guidance model is a smaller LLM responsible for interaction and interrogation. Its tasks are conducting multi-turn dialogue with the patient, parsing patient responses, aligning colloquial language with standardized symptom terms, and using the knowledge graph to decide what to ask next. It acts as the system’s doctor-like questioning component rather than as the final diagnostic authority.

The expert model is a stronger TCM model built on the Sunsimiao framework. It is fine-tuned on high-quality TCM data and is responsible for final disease diagnosis, treatment recommendation, and updating knowledge graph relations after the dialogue. By invoking the expert model only after the interrogation stage is complete, DoPI is presented as avoiding the trade-off between improving dialogue ability through fine-tuning and losing medical expertise.

Operationally, the workflow is serial. The patient first provides an initial symptom description. The guidance model parses that input, and the knowledge graph maps the extracted symptoms to candidate diseases. The guidance model then asks targeted follow-up questions; patient responses update the symptom record; and once similarity or confidence is high enough, the questioning stops. The collected symptoms together with the candidate disease are then passed to the expert model, which outputs diagnosis and treatment advice. The paper characterizes the two models as independently specialized but connected through the knowledge graph.

3. Knowledge-graph-driven interrogation mechanism

The knowledge graph is the core mechanism governing dynamic question generation. It contains disease nodes, symptom nodes, and weighted edges between diseases and symptoms as well as between related symptoms. The initial edge weights are based on symptom-disease co-occurrence frequency in the source data. The system uses the currently known symptoms to estimate which diseases are most probable, then asks about the most informative unknown symptoms associated with those candidate diseases.

The patient symptom state is represented as

$P=(p_1,p_2,\ldots,p_n), \quad p_k= \begin{cases} 1, & \text{If $k$ is a known symptom} \ 0, & \text{otherwise.} \end{cases}$

For each disease $i$, the system uses a disease symptom vector

$$D_i=(d_{i1},d_{i2},\ldots,d_{in}), \quad d_{ik}=w_{ki},$$

where $w_{ki}\in[0,1]$ is the graph edge weight between symptom $k$ and disease $i$. Disease importance is measured by the cosine similarity between $D_i$ and $P$; this score serves as the disease’s influence factor. For an unknown symptom $j$, the importance score is

$Score(j)=\sum_{i=1}^{N} w_{ji}\cdot S_i,$

and a Gaussian perturbation is added for robustness:

$i$0

The paper states that $i$1 decreases as dialogue rounds increase, so early rounds are more exploratory and later rounds more focused.

The interrogation loop follows an interactive diagnostic algorithm. The patient query is first encoded into symptoms by Graph Encoder(Q). In each questioning epoch, the system finds probable diseases from the graph, sorts them by importance, gathers associated symptoms, ranks candidate symptoms by importance, selects a few high-priority symptoms, asks the patient about them, updates symptom memory, and computes cosine similarity between the current symptoms and the candidate disease. The loop terminates when the cosine similarity of the current most probable disease exceeds a predefined threshold $i$2. After the final diagnosis, the expert model proposes graph updates by strengthening correct symptom-disease and symptom-symptom links and weakening misleading ones. Because this update occurs after the dialogue, it is not intended to slow the consultation.

4. Synthetic dialogue dataset and evaluation protocol

Because real consultation data are scarce, DoPI is trained and evaluated with a synthetic multi-turn doctor-patient dialogue dataset constructed from TCM knowledge and LLM role-play. The construction begins from binary tuples of disease + symptom list. For each tuple, the setup assumes that the patient has that disease and exhibits all listed symptoms. Qwen-plus is then used to simulate both doctor and patient and generate full dialogues.

Each synthetic dialogue contains the patient’s initial symptom description, several rounds of doctor questions and patient answers, and a final diagnosis. The generation process is constrained by three requirements. Common Sense requires initial patient descriptions to be realistic and focus on obvious symptoms. Colloquialism requires doctor questions to be understandable and patient replies to remain symptom-focused. Honesty requires patient responses to accurately reflect whether a symptom is present. The dataset contains more than two thousand high-quality multi-round doctor-patient conversations. The original data are not publicly available due to patient privacy.

The evaluation protocol avoids manually collected real-world consultation logs by giving models only the initial partial symptom description and testing whether they can recover the correct final diagnosis through interaction. Three quantitative metrics are defined. Diagnostic Accuracy is

$i$3

where $i$4 is the number of all patients and $i$5 is the number of correctly diagnosed patients.

The metric printed in the paper as Q{paper_content}A Ratio is

$i$6

where $i$7 is the number of rounds where the model asks questions, $i$8 is the number of rounds where the model answers or diagnoses, and $i$9 is the number of test patients. It is intended to measure how questioning-heavy and proactive the model is.

Interrogation Distance is

$$D_i=(d_{i1},d_{i2},\ldots,d_{in}), \quad d_{ik}=w_{ki},$$0

where $$D_i=(d_{i1},d_{i2},\ldots,d_{in}), \quad d_{ik}=w_{ki},$$1 and $$D_i=(d_{i1},d_{i2},\ldots,d_{in}), \quad d_{ik}=w_{ki},$$2 denote the doctor’s questioning and answering rounds in the dataset. This metric measures how closely a model’s dialogue behavior matches the expert doctor’s interrogation pattern.

The paper also adds an LLM-as-judge qualitative evaluation along four criteria: Knowledgeability, Professionalism, Fluency, and Respectfulness. Two dialogue outputs are compared, each comparison is run five times, and the majority result is used.

5. Empirical performance and comparative results

The principal empirical result is that DoPI-7B achieves 84.68% Diagnostic Accuracy (Sun et al., 7 Jul 2025). The paper presents this as substantial improvement over both general-purpose and medical baselines in the proactive dialogue setting.

Before introducing the full system, the paper reports that existing medical-oriented models perform poorly on the same dialogue task when only partial symptoms are initially available: Sunsimiao-7B: 14.73%, HuatuoGPT-7B: 17.72%, and BianQue-6B: 21.27%. These numbers are used to argue that strong medical QA performance does not by itself translate into successful multi-turn diagnosis.

For the main baseline comparison, the reported metrics are as follows:

• Qwen2.5-Max: Diagnose Accuracy 32.31, Q{paper_content}A Ratio 15.23, Interrogation Distance 2.32
• ChatGPT-4o: Diagnose Accuracy 35.12, Q{paper_content}A Ratio 17.09, Interrogation Distance 2.04
• DeepSeek-v3: Diagnose Accuracy 58.74, Q{paper_content}A Ratio 20.98, Interrogation Distance 1.96
• DoPI-7B: Diagnose Accuracy 84.68, Q{paper_content}A Ratio 21.31, Interrogation Distance 1.84

The reported interpretation is threefold. First, DoPI is far ahead in diagnostic accuracy, which the paper attributes to the cooperative guidance-plus-expert architecture. Second, DoPI has the highest Q{paper_content}A Ratio, suggesting more active questioning behavior. Third, DoPI has the lowest Interrogation Distance, indicating dialogue behavior closer to an expert doctor’s interrogation pattern.

The qualitative LLM-as-judge evaluation further suggests that DoPI outperforms baselines on most of knowledgeability, professionalism, fluency, and respectfulness, with the paper especially emphasizing improvements in knowledgeability and professionalism because the interrogation process is grounded in a structured medical knowledge graph rather than generic conversation.

6. Limitations, interpretive scope, and broader significance

The paper assigns DoPI three main contributions. First, it proposes a TCM LLM system for proactive interrogation that combines a guidance model, an expert model, and a knowledge graph to simulate real diagnostic dialogue while maintaining professional expertise. Second, it constructs a synthetic multi-turn TCM consultation dataset from disease-symptom tuples, reducing the need for manual data collection. Third, it introduces an evaluation method for interrogation quality based on diagnostic accuracy, proactivity-related metrics, and doctor-alignment metrics rather than real-world consultation logs.

At the same time, several limitations are explicit or implied. The original dataset is not publicly released because of privacy constraints. Synthetic dialogue generation may not fully capture the complexity of real clinical interactions. Tongue diagnosis is simplified using a ResNet-based constitution classifier rather than a true multimodal medical reasoning system. The discussion implies the importance of the guidance model, expert model, knowledge graph, threshold-based stopping rule, and graph-update mechanism, but no explicit ablation numbers are shown. In addition, the knowledge graph begins from dataset frequencies, so its quality depends on the underlying source data.

The broader implication drawn by the paper is that medical LLMs should not be judged only by answer quality. In clinical domains, and particularly in TCM, systems also need proactive symptom elicitation, structured interrogative reasoning, controlled stopping criteria, and domain-specific knowledge retention. DoPI frames these requirements as an architectural problem: not merely generating fluent medical text, but organizing consultation as a sequence of inference, targeted questioning, symptom-state updating, confidence-based stopping, and only then diagnosis. A plausible implication is that the framework’s most distinctive contribution lies less in isolated language modeling performance than in its attempt to formalize doctor-like consultation behavior as a knowledge-guided interaction process.