Text2MDT: Extracting Medical Decision Trees

• Text2MDT is a task that transforms narrative clinical texts into structured medical decision trees, enabling practical clinical decision support.
• It employs both an end-to-end LLM approach and a decomposed pipeline (triplet extraction, node grouping, tree assembling) to automate MDT construction.
• Empirical results show end-to-end models, enhanced by chain-of-thought prompting, often outperform pipelines while offering a trade-off with parameter efficiency.

Text2MDT is a task for automatically extracting medical decision trees (MDTs) from medical texts such as medical guidelines and textbooks. It was introduced to address a central bottleneck in clinical knowledge engineering: knowledge of the medical decision process, which can be modeled as MDTs, is critical to build clinical decision support systems, yet current MDT construction methods rely heavily on time-consuming and laborious manual annotation (Zhu et al., 2024). The task converts narrative clinical guidance into a normalized tree-structured representation. In a later formalization following Zhu et al. (2024), the mapping is written as $X \rightarrow T$, where $X$ is a clinical text and $T$ is a pre-order traversal sequence of structured decision nodes (Li et al., 6 Oct 2025).

1. Clinical motivation and scope

Text2MDT emerged from the need to transform unstructured medical knowledge into a form usable by downstream decision-support infrastructure. The core premise is that clinical guidelines and textbooks often encode diagnostic branching, treatment selection, and conditional care pathways in prose, whereas deployment in a clinical decision support system requires an explicit tree structure (Zhu et al., 2024).

The task is defined over medical texts rather than over already segmented rules or pre-annotated logic fragments. This makes it broader than conventional relation extraction. A later task decomposition makes this distinction explicit by separating triplet extraction, node grouping, and tree assembling, thereby treating MDT construction as a hierarchical structured-prediction problem rather than a flat information-extraction problem (Li et al., 6 Oct 2025).

A common misconception is that MDT extraction is only about diagnostic branching. In the later formalization, MDT nodes include both conditional assessment and therapeutic decision nodes, indicating that the target structure spans diagnosis and intervention rather than diagnosis alone. This broadens the operational meaning of “decision tree” in Text2MDT beyond symptom-to-diagnosis routing.

2. Formal representation of the target structure

The original work states that it normalizes the form of the MDT (Zhu et al., 2024). A later formulation following Zhu et al. specifies the target as a pre-order traversal sequence

$T = [N_1, N_2, \dots, N_{n_{\text{node}}}],$

where each $N_i$ is a structured decision node (Li et al., 6 Oct 2025).

In that formulation, each node is defined as

$$\text{Node} = \{\text{Role}, \text{Triplets}, \text{Logical\_Rel}\}.$$

The field $\text{Role}$ distinguishes node type, with $\Diamond$ denoting a conditional assessment and $\Box$ denoting a therapeutic decision node. $\text{Triplets}$ is a set of subject-relation-object triples, and $X$0 belongs to $X$1, with $X$2 used when the node contains at most one triplet (Li et al., 6 Oct 2025).

The tree structure is not represented by an explicit adjacency matrix in that later account. Instead, the hierarchy is captured implicitly by the pre-order traversal together with node content, and parents, children, and branches are reconstructed from the sequence. This suggests that Text2MDT is simultaneously a content-generation task and a structural serialization task. A plausible implication is that normalization reduces ambiguity in how equivalent clinical logic is encoded across different source texts.

3. Dataset construction and annotation

The original Text2MDT work created an annotated Text-to-MDT dataset in Chinese with the participation of medical experts (Zhu et al., 2024). The texts come from medical guidelines and textbooks, which places the dataset closer to formal clinical knowledge sources than to conversational clinical notes. The dataset is open-sourced at https://tianchi.aliyun.com/dataset/95414, and the source codes are open-sourced at https://github.com/michael-wzhu/text2dt (Zhu et al., 2024).

Later work identifies the same corpus as the Text2MDT dataset from Zhu et al. (2024), constructed from Chinese medical guidelines and textbooks, and states that gold MDTs are constructed via annotation and quality-checked (Li et al., 6 Oct 2025). That later paper does not re-list the specific domains, sizes, or full annotation protocol, instead deferring those details to the original work. Even so, two features of the corpus are clear: it is schema-driven, because the MDT form is normalized, and it is expert-involved, because medical experts participated in annotation.

The use of expert annotation is consequential. MDTs require more than lexical extraction; they require adjudication over node boundaries, logical relations, and branch organization. This implies a higher annotation burden than many standard biomedical IE benchmarks and helps explain why automation was positioned as the central contribution of the task.

4. Modeling paradigms

The original paper investigates two methods for Text2MDT. The first is an end-to-end framework that only relies on a GPT style LLM instruction tuning to generate all the node information and tree structures. The second is a pipeline framework that decomposes the task into three subtasks (Zhu et al., 2024).

A later description of the original formulation specifies those three subtasks as triplet extraction, node grouping, and tree assembling (Li et al., 6 Oct 2025). In this view, the pipeline first extracts atomic clinical facts, then clusters them into decision nodes, and finally connects the nodes into a full decision tree. The end-to-end approach instead asks a single instruction-tuned LLM to emit the full structured MDT directly.

Framework	Core operation	Reported emphasis
End-to-end	Generate all node information and tree structures directly	GPT style LLM instruction tuning
Pipeline	Decompose into triplet extraction, node grouping, tree assembling	Lightweight encoder-based pretrained models

The relation between the two paradigms is not purely competitive. The original results state that end-to-end LLMs outperform the pipeline methods, but also that a lightweight pipelined method based on encoder-based pretrained models can perform comparably with LLMs with model complexity two magnitudes smaller (Zhu et al., 2024). This establishes an important engineering trade-off: direct generation yields stronger overall performance, while decomposition can remain attractive under tight compute constraints.

Prompting is also part of the modeling landscape. The original paper reports that chain-of-thought prompting improves the performance of the fine-tuned LLMs on the Text2MDT test set (Zhu et al., 2024). Later work further notes that prompt templates for each subtask, as well as an end-to-end chain-of-thought style prompt called COT-Gen-3, are adopted from Zhu et al. (2024) (Li et al., 6 Oct 2025). This places prompt design alongside architecture choice as a material factor in MDT extraction quality.

5. Evaluation and empirical findings

The original Text2MDT experiments report three headline findings. First, the end-to-end method based on LLMs, for models with 7B parameters or larger, shows promising results and successfully outperforms the pipeline methods. Second, chain-of-thought prompting improves performance on the Text2MDT test set. Third, encoder-based pipeline models can remain competitive despite much smaller model complexity (Zhu et al., 2024).

A later evaluation framework makes the metric structure explicit. It uses precision, recall, and F1 for triplet extraction; NG_LR, a Levenshtein ratio, for node grouping; and Tree_Acc, DP_F1, and Tree_LR or Tree_ER for full-tree evaluation (Li et al., 6 Oct 2025). Tree_Acc measures exact tree recovery, while DP_F1 evaluates decision paths from root to leaves.

Representative results from the later FT-MDT benchmark illustrate how Text2MDT performance is now compared under parameter-efficient fine-tuning:

Setting	Baseline	PI-LoRA
Pipeline, Qwen 2.5–7B	LoRA: Tree_Acc 0.764, DP_F1 0.858, Tree_LR 0.952	Tree_Acc 0.772, DP_F1 0.884, Tree_LR 0.967
End-to-end, Qwen 2.5–7B	LoRA: Tree_Acc 0.510, DP_F1 0.646, Tree_ER 0.911	Tree_Acc 0.550, DP_F1 0.679, Tree_ER 0.936

In pipeline mode, the same later study reports that PI-LoRA attains the highest scores on all seven reported metrics among the compared PEFT methods, including triplet precision, triplet recall, triplet F1, NG_LR, Tree_Acc, DP_F1, and Tree_LR. It also reports that PI-LoRA outperforms a GPT-4 API baseline on overall tree metrics, with GPT-4 at Tree_Acc 0.672, DP_F1 0.786, and Tree_LR 0.893, versus PI-LoRA at 0.772, 0.884, and 0.967 respectively (Li et al., 6 Oct 2025).

These later numbers do not replace the original contribution. Rather, they show that Text2MDT matured from a newly proposed task into a benchmark for comparing LLM adaptation strategies under both end-to-end and decomposed formulations.

6. Subsequent developments, limitations, and significance

A major subsequent development is FT-MDT, which adapts pre-trained open-source LLMs to Text2MDT using PI-LoRA, a path-integrated low-rank adaptation method (Li et al., 6 Oct 2025). That work treats Text2MDT as supervised generation, uses Qwen 2.5–7B as the main backbone, and also reports improvements over LoRA and AdaLoRA on Baichuan2–7B and GLM-4-9B-Chat. In its related-work framing, prior work introduced Text2MDT and compared pipeline extraction with end-to-end LLM approaches using standard LoRA, whereas FT-MDT extends the comparison to modern PEFT methods.

The limitations identified in later work clarify the open problems of the task. These include dependence on input quality, difficulty with non-standard formats and ambiguous phrasing, incomplete coverage of rare conditions, sensitivity of $X$3, pruning ratios, and LoRA placements to domain characteristics, and a gap between structural correctness and clinical optimality (Li et al., 6 Oct 2025). Experiments are also limited to Chinese guidelines, leaving cross-lingual generalization unresolved.

These limitations matter because Text2MDT evaluation is heavily structure-centric. A tree may score well on serialized similarity while still being clinically incomplete, outdated, or pragmatically suboptimal. This suggests that structural metrics are necessary but not sufficient for deployment. A plausible implication is that future Text2MDT systems will need tighter coupling between formal tree recovery, clinical validation, and maintenance workflows for evolving guidelines.

Text2MDT is therefore best understood as a foundational task at the intersection of medical information extraction, structured generation, and clinical knowledge formalization. Its original contribution was to define the task, normalize the MDT target form, release a Chinese expert-annotated dataset, and establish both end-to-end and pipeline baselines (Zhu et al., 2024). Its subsequent significance lies in providing a benchmark on which prompt design, LLM scale, encoder efficiency, and parameter-efficient adaptation can be studied in direct relation to the extraction of executable medical decision logic (Li et al., 6 Oct 2025).