Hierarchical Diagnostic Reasoning Tree (HDRT)

• HDRT is a diagnostic framework that organizes reasoning into hierarchical, coarse-to-fine levels, enabling progressive narrowing of hypotheses.
• It underpins systems in clinical semantic intelligence and medical test recommendations by using explicit intermediate node reasoning and uncertainty handling.
• Implementations range from six-level diagnostic sequences in oral disease diagnosis to three-layer architectures in outpatient systems, offering flexible diagnostic control.

Hierarchical Diagnostic Reasoning Tree (HDRT) denotes a structured reasoning framework in which diagnosis proceeds from broad screening or routing decisions to progressively narrower hypotheses, leaf-level decisions, and confidence-based confirmation. The term is used explicitly in Clinical Semantic Intelligence (CSI), where HDRT is the core reasoning engine of the system’s Standard Mode for oral disease diagnosis, and closely related HDRT-style architectures appear in outpatient medical test recommendation, explainable question answering, probabilistic model-based diagnosis, and hierarchical auditing of reasoning traces (Mashayekhi et al., 20 Jul 2025, Yang et al., 6 Jan 2025, Zhang et al., 2023). This suggests that HDRT is best understood as a family of hierarchical diagnostic formalisms and implementations rather than as a single canonical algorithm.

1. Terminological status and historical lineage

In the cited literature, the exact term Hierarchical Diagnostic Reasoning Tree appears in "Clinical Semantic Intelligence (CSI): Emulating the Cognitive Framework of the Expert Clinician for Comprehensive Oral Disease Diagnosis" (Mashayekhi et al., 20 Jul 2025). There, HDRT is presented as a computational model of expert clinical reasoning that performs a stepwise, reductive narrowing process from broad screening to specific disease diagnosis, with interactive uncertainty handling.

Earlier and adjacent work establishes the principal ingredients of HDRT-like reasoning without using the same name. "A Probabilistic Approach to Hierarchical Model-based Diagnosis" translates hierarchical functional schematics into Bayesian networks and performs diagnostic inference through a hierarchy-aware join-tree algorithm (Srinivas, 2013). "Knowledge-Based Decision Model Construction for Hierarchical Diagnosis: A Preliminary Report" formulates hierarchical complete diagnosis as a stochastic process using influence diagrams and incremental decision-model construction (Yuan, 2013). "Reasoning over Hierarchical Question Decomposition Tree for Explainable Question Answering" introduces a hierarchical question decomposition tree and recursive probabilistic reasoning over that tree (Zhang et al., 2023). "Tree-based RAG-Agent Recommendation System: A Case Study in Medical Test Data" presents HiRMed as a three-layer hierarchical diagnostic reasoning pipeline for medical test recommendation (Yang et al., 6 Jan 2025).

Taken together, these works indicate that HDRT-like systems share three persistent commitments: hierarchical decomposition, local reasoning at intermediate nodes, and explicit control over diagnostic refinement. What changes across papers is the substrate—Bayesian networks, influence diagrams, retrieval-augmented generation, multimodal LLM pipelines, or typed reasoning graphs.

2. Structural anatomy of an HDRT

The clearest explicit HDRT instantiation is CSI’s six-level diagnostic sequence. In Standard Mode, the hierarchy begins with primary classification into normal or abnormal, continues through lesion characteristics with 8 classes, clinical context with 8 classes, diagnostic categories with 10 broad diagnostic groups, disease identification across 118 possible diseases, and ends with diagnostic confirmation producing 118 final probabilities (Mashayekhi et al., 20 Jul 2025). The hierarchy is therefore not merely organizational; each level reduces the space of plausible diagnoses.

HiRMed provides a more compact three-layer architecture. Its hierarchy consists of a root layer, a department layer, and an item layer. The root receives patient symptoms and relevant clinical data, produces candidate departments and their priorities, the department layer performs specialty-specific reasoning, and the item layer outputs final ranked test recommendations with urgency or importance weights (Yang et al., 6 Jan 2025). The associated knowledge is likewise hierarchical: a department-level knowledge base and a department testing item knowledge base.

RoHT formalizes a related structure as a 3-ary ordered tree $T$, whose root is the original complex question, non-root nodes are sub-questions, and leaves are atomic questions. It distinguishes natural language questions, bridge questions, and symbolic operation questions, with bridge and symbolic operation questions restricted to leaves (Zhang et al., 2023). Although RoHT is not a medical diagnostic system, it supplies a precise template for hierarchical reasoning trees in which different levels correspond to different semantic granularities.

Across these systems, the recurring architectural pattern is a root-level problem representation, one or more intermediate refinement layers, and leaf-level outputs. This suggests that HDRT is defined less by any fixed branching factor or depth than by explicit coarse-to-fine diagnostic progression.

3. Reasoning mechanisms and control policies

HDRT implementations differ most sharply in how node-level reasoning is executed. In HiRMed, each node is not a passive retrieval point but a RAG-based reasoning module. The pipeline encodes the patient query with OpenAI embeddings, retrieves evidence from a FAISS-based vector database, uses GPT-O1 to reason over retrieved evidence, and applies a fine-tuned LLaMA3.2-3B weight model to rank departments and tests. At the item layer, a memory component consolidates earlier decisions and flagged symptoms, allowing the final recommendation to reflect urgency, clinical utility, and likely diagnostic value (Yang et al., 6 Jan 2025).

RoHT makes its control mechanism explicit through a recursive reasoning function

$$f(q^i, p^i, G, C) \rightarrow R^i : \{(ans_j^i, p_j^i)\},$$

where candidate answers at each node are scored probabilistically. Node certainty is defined as

$p^i = l^a \cdot l^i,$

and a scheduler decides whether a node should be answered from a knowledge base, from text, or by recursively solving child questions (Zhang et al., 2023). This architecture makes decomposition uncertainty and source-selection uncertainty first-class objects of inference.

The probabilistic model-based diagnosis literature implements analogous control with different mathematics. In the Bayesian-network formulation, each component has inputs, an output, and a discrete mode variable $M$; hierarchical abstraction is encoded by an abstraction function $Ab$ relating lower-level subcomponent modes to a higher-level mode:

$$P(m_h \mid m_{l1}, m_{l2}, \dots, m_{ln}) = \begin{cases} 1 & \text{iff } m_h = Ab(m_{l1},m_{l2},\dots,m_{ln}),\ 0 & \text{otherwise.} \end{cases}$$

Diagnostic inference then consists of posterior updating after observations are entered into the network (Srinivas, 2013). In the influence-diagram approach, hierarchical diagnosis is controlled by expected-cost evaluation, with the functional component using

$$V_{fun} = C(test) + C(treatment) + C(status\ following\ treatment),$$

and meta-level control deciding which causal pathway to pursue next (Yuan, 2013).

CSI adds an explicitly interactive uncertainty gate. If the difference between the logarithmic probabilities of the top two diagnoses is below 0.3, the system flags uncertainty and requests clarification:

$$\left| \log p_1 - \log p_2 \right| < 0.3 \Rightarrow \text{request additional user input}.$$

This makes uncertainty handling operational rather than merely descriptive (Mashayekhi et al., 20 Jul 2025).

4. Representative implementations across domains

The HDRT label is most directly associated with clinical diagnosis, but the underlying design pattern has been instantiated across several neighboring research areas.

Work	Hierarchical object	Domain
CSI	HDRT	Oral disease diagnosis
HiRMed	Root–department–item hierarchy	Medical test recommendation
RoHT	Hierarchical Question Decomposition Tree	Explainable QA
ReasoningLens	Macro graph plus micro subgraphs	Reasoning-trace auditing
HeaRT	Hierarchical Circuit Reasoning Tree	AMS design optimization

CSI provides the most direct performance evidence for an explicit HDRT. On a 431-image internal test set, Fast Mode achieved 73.4% overall accuracy, while Standard Mode, which uses the full HDRT, achieved 89.5%. On an external hold-out set of 176 images, Fast Mode reached 64.8% and Standard Mode 85.2%. The zone-specific internal results were 85.4% → 94.3% for Zone 3, 63.7% → 89.5% for Zone 2, and 35.8% → 58.1% for Zone 1 (Mashayekhi et al., 20 Jul 2025).

HiRMed applies an HDRT-style design to outpatient test recommendation. On 125,000 outpatient visits, with clinical validation by 12 clinicians reviewing 500 random cases, HiRMed reported Coverage Rate 92.3%, Accuracy 88.7%, Miss Rate 2.1%, and Clinical Relevance Score 4.3, outperforming Flat-RAG and Traditional Vector Similarity (TVS). Its ablations also show large degradations for w/o Memory, w/o Department Layer, and Single Knowledge Base (Yang et al., 6 Jan 2025).

RoHT extends the same structural logic to explainable QA over heterogeneous sources. On Musique, the paper reports that RoHTmix achieved 63.6 F1 on the test set, outperforming the best public result of 52.3; removing the scheduler reduced performance by 5.8 EM on KQA Pro and 7.4 EM on Musique (Zhang et al., 2023).

Outside diagnosis narrowly defined, "ReasoningLens: Hierarchical Visualization and Diagnostic Auditing for Large Reasoning Models" organizes long Chain-of-Thought traces into a macro-level exploration graph and micro-level execution subgraphs, while "HeaRT: A Hierarchical Circuit Reasoning Tree-Based Agentic Framework for AMS Design Optimization" reports >97% reasoning accuracy and >98% Pass@1 across a 40-circuit benchmark repository (Zhang et al., 22 Jun 2026, Poddar et al., 24 Nov 2025). These systems are not explicit HDRT implementations in the CSI sense, but they show the portability of hierarchical diagnostic reasoning into adjacent domains.

5. Interpretability, evaluation, and hierarchical error analysis

A central rationale for HDRT is interpretability through explicit intermediate structure. CSI’s Standard Mode can expose a concrete diagnostic path such as Abnormal → White Lesion → Adult → Inflammatory/Reactive → Oral Lichen Planus, and its interaction protocol allows follow-up questions when uncertainty remains (Mashayekhi et al., 20 Jul 2025). HiRMed likewise accumulates evidence across layers, so that root-level department priorities, department-level narrowing, and item-level weights jointly determine the final test list (Yang et al., 6 Jan 2025).

RoHT makes interpretability procedural as well as structural. The HQDT specifies how the original problem was decomposed, while node-level candidate answers and probabilities expose how the final answer path was selected (Zhang et al., 2023). ReasoningLens adds a diagnostic auditing perspective by attaching error types to segments of a hierarchical reasoning graph and evaluating reconstruction quality with Node Type Accuracy (NTA) and Graph Edit Similarity (GES) (Zhang et al., 22 Jun 2026).

Hierarchical evaluation has also emerged as a distinct research problem. "H-DDx: A Hierarchical Evaluation Framework for Differential Diagnosis" argues that flat Top-k accuracy is too coarse for differential diagnosis because it treats clinically relevant near-misses and diagnostically distant errors as equally wrong. Its Hierarchical DDx F1 (HDF1) augments each diagnosis with all ancestral nodes in the ICD-10 hierarchy and computes hierarchical precision and recall over those augmented sets (Lim et al., 4 Oct 2025). This does not define an HDRT, but it aligns evaluation with the same coarse-to-fine semantics that HDRT systems operationalize. A plausible implication is that hierarchical reasoning and hierarchical evaluation are complementary: one structures the decision process, while the other measures whether model outputs remain in the correct clinical neighborhood.

6. Limitations, scope conditions, and common misconceptions

A common misconception is that any multi-step diagnostic pipeline is automatically an HDRT. The surveyed systems suggest a stricter criterion: the hierarchy must carry semantically distinct levels of abstraction, and node-level decisions must depend on local reasoning, evidence, or probabilistic control rather than on a single flat prediction reused at multiple stages. This is why HiRMed emphasizes that it is not performing simple vector similarity or flat RAG, but semantic retrieval + clinical inference + hierarchical narrowing (Yang et al., 6 Jan 2025).

At the same time, not every HDRT-like system is a classical tree-search procedure. HiRMed explicitly lacks a formal branching criterion, threshold, or backtracking algorithm; it is described more accurately as a hierarchical reasoning pipeline than as a mathematically specified search tree (Yang et al., 6 Jan 2025). RoHT is more formal, but it is currently restricted to KBs and text, and failures may arise from scheduler mistakes, incorrect text extraction, bridge-resolution errors, or KB coverage limitations (Zhang et al., 2023).

The probabilistic diagnosis literature is even more restrictive in its assumptions. The Bayesian-network approach assumes no feedback loops, non-correlated faults, full fault models, fully specified input distributions, single-output components, a restricted form of hierarchy, and systems without dynamics (Srinivas, 2013). The influence-diagram formulation assumes single fault, complete domain knowledge, provided probabilistic information, and static probe costs, and it performs top-down hierarchical refinement without backtracking (Yuan, 2013). These constraints clarify that hierarchical reasoning is not a substitute for domain modeling; it depends on the availability of an appropriate structural prior.

The literature also shows that hierarchy can appear in different roles: as a reasoning mechanism, as a diagnostic control structure, as an interpretability substrate, or as an evaluation framework. This suggests that HDRT is not a single mature standard but a convergent research motif linking clinical cognition, structured inference, and inspectable decision processes.