Syn-Diag: Multi-Source Diagnostic Synthesis

• Syn-Diag is a diagnostic paradigm that integrates heterogeneous evidence sources, models, and decision-makers into a closed-loop, human-centered process.
• It leverages the coordinated interplay of LLMs, medical knowledge graphs, and clinical experts to enhance workflow efficiency, safety, and explanatory synthesis.
• Empirical evaluations demonstrate improvements in diagnostic accuracy, reduced workload, and timely detection across clinical, industrial, and cybersecurity applications.

Syn-Diag denotes a family of diagnostic formulations centered on synergy or synthesis among heterogeneous evidence sources, models, and decision-makers. Its most explicit definition appears in DiagLink, where it is presented as a “synergistic diagnosis paradigm” coordinating LLMs, medical knowledge graphs (KGs), and clinical experts across patient and physician workflows, with diagnosis treated not as a single-shot output but as a closed-loop, human-centered process (Zhou et al., 28 Jan 2026). Across adjacent literatures, the same label is also attached to synthetic rare-disease evaluation pipelines, multimodal clinical evidence benchmarks, cloud–edge industrial fault diagnosis, modular simulation diagnostics, TCP SYN attack detection, and Traditional Chinese Medicine syndrome differentiation. This suggests that Syn-Diag is less a single canonical architecture than a recurring design pattern for diagnostic systems that combine complementary representations under explicit decision logic (Nishibayashi et al., 24 Jun 2025, Chen et al., 25 Feb 2026, Jia et al., 7 Oct 2025).

1. Scope and terminological range

The cited literature uses Syn-Diag in several technically distinct but structurally related ways. In clinical AI, it denotes end-to-end diagnostic assistance that links history taking, differential generation, evidence presentation, and expert correction. In benchmarking work, it denotes synthetic or simulated pipelines for estimating diagnostic performance before deployment. In engineering and security, it denotes system-level diagnosis through multi-signal fusion, time-embedded experimentation, or cloud–edge collaboration. This suggests that the common denominator is not domain content, but the deliberate coupling of multiple evidence channels, reasoning layers, or actors (Zhou et al., 28 Jan 2026, Jiang et al., 13 Mar 2026, Jia et al., 7 Oct 2025).

Setting	Representative work	Core diagnostic object
Clinical assistance	DiagLink	Patient–physician diagnosis workflow
Rare-disease evaluation	Synthetic vignette simulation	Pre-deployment symptom-checker performance
Multimodal benchmarking	MEDSYN	DDx generation and FDx selection
Industrial diagnosis	Cloud–edge Syn-Diag	Few-shot fault classification
System simulation	ADIOSS	Faulty modules in modular workflows
Network security	SYN flood / scan detection	Malicious traffic diagnosis

A common misconception is to treat Syn-Diag as synonymous with a symptom checker. The broader record does not support that equivalence. Some works use it for patient-facing clinical support, but others use it for synthetic evaluation, multimodal evidence synthesis, automotive simulation workflows, cybersecurity telemetry, or syndrome differentiation from clinical text (Zhou et al., 28 Jan 2026, Nishibayashi et al., 24 Jun 2025, Ren et al., 2022).

2. DiagLink as the canonical clinical formulation

DiagLink operationalizes Syn-Diag around three complementary actors: LLMs for natural-language understanding, guided history taking, hypothesis generation, and explanatory synthesis; KGs for structured reasoning, evidence tracing, and transparent knowledge presentation; and medical experts for validation, correction, and continual knowledge evolution (Zhou et al., 28 Jan 2026). The system is dual-user by design. Patients receive an emotionally supportive guided dialogue that progressively builds a semi-structured history without exposing provisional diagnoses, whereas physicians receive a role-adaptive interface that aggregates patient history, KG evidence, and LLM rationales with explicit provenance.

The history-taking workflow follows a CARE-grounded template that includes chief complaint, history of present illness, past/personal/family history, findings, perspectives, and interventions/outcomes. Dialogue proceeds in stages labeled Main, Other, and Ddx, with one to two focused questions per turn, progressive disclosure from general to specific questions, and contextual framing before sensitive topics. A dynamic “History Panel” visualizes the semi-structured record as it is built, allowing patients to verify what the system has captured (Zhou et al., 28 Jan 2026).

For physicians, DiagLink provides a three-column dashboard: patient information, diagnostic KG, and reasoning and recommendations. The KG view lays out the top-3 differentials as polygon vertices, aggregates common symptoms centrally, summarizes shortest-path counts, colors entities by type, and supports focus-plus-context interactions. A navigation bar encodes relative likelihood and the potential severity of missing a disease; changing the active diagnosis synchronously updates reasoning and recommendations. Accountability is explicit: system hypotheses are handed off to physicians, who own final diagnostic decisions (Zhou et al., 28 Jan 2026).

The expert layer closes the loop. Evolution events are triggered when a candidate disease is missing from the KG, when its subgraph is unused or obsolete, or when expert review is due by time threshold. The LLM drafts disease-centric text under strict templates; experts correct and enrich the draft, enforce category constraints, and commit it. Structured triples are then extracted, redundancy and near-duplicate checks are applied using semantic similarity thresholds, and the resulting disease subgraph immediately affects future KG reasoning and RAG ranking (Zhou et al., 28 Jan 2026).

3. Algorithms, interfaces, and decision logic

DiagLink’s history-taking logic is formulated as a state machine with four states: Main, Other, Ddx, and Done. In Main and Other, the LLM asks guided questions based on dialogue history and template gaps. In Ddx, it proposes preliminary differentials and asks discriminative questions to refine likelihoods and reorder the differential until convergence or a question limit is reached (Zhou et al., 28 Jan 2026).

Symptom extraction and linking combine LLM recognition with dense retrieval over KG entities. The paper defines symptom extraction as

$$\mathcal{E}_s = \{s_1, s_2, \dots\} = LLM\bigl(Prompt_{recog},\; History\bigr)$$

and aligns each extracted symptom to the KG by embedding similarity and FAISS-based retrieval:

$$\hat{e}_i = \arg\max_{e_j \in \mathcal{E}_g} \mathrm{sim}\bigl(\mathrm{emb}(s_i),\; \mathrm{emb}(e_j)\bigr).$$

Candidate diseases are obtained from one-hop disease neighbors of the linked symptom set, then scored by path-informed relevance:

$$\mathrm{score}(d_i) = \sum_{\hat{e}_j \in \hat{\mathcal{E}}_s} \frac{1}{\mathrm{dist}(d_i,\; \hat{e}_j)}.$$

The KG candidates are merged with the LLM’s preliminary differential set into $\mathcal{D}_{comb}$, and RAG-based ranking returns the final top-3:

$$\mathcal{D}_{final} = \mathrm{TopK}\Bigl(LLM\bigl(Prompt_{rank},\; History,\; \mathcal{K},\; \mathcal{D}_{comb}\bigr),\; 3\Bigr).$$

This dual-track LLM–KG design is central to the Syn-Diag formulation because breadth from internal LLM knowledge and structure from graph traversal are combined rather than substituted for one another (Zhou et al., 28 Jan 2026).

Prompting is tightly constrained. Prompt_recog is instructed to return only symptoms and signs with qualifiers such as onset, duration, and severity, while excluding diagnoses. Prompt_rank supplies disease definitions, matched symptom neighbors, shortest-path summaries, drug relations, and the full patient history, and asks the LLM to justify a 0–10 likelihood score per disease, cite KG-linked evidence, and avoid inventing facts. Safeguards include template slot constraints, coverage checks, withholding provisional differentials from patients, expert-validated KGs as the authoritative source, and timestamped traceable edits during knowledge evolution (Zhou et al., 28 Jan 2026).

Implementation details reinforce the same design logic. The reported stack uses GPT-4.1, PrimeKG with approximately 130k nodes and approximately 4M edges, Flask, Neo4j, FAISS, Graphology, Sigma.js, and a Nuxt.js/Vue 3 frontend. The framework is described as KG-agnostic, and proprietary KGs can be substituted (Zhou et al., 28 Jan 2026).

4. Empirical performance, usability, and safety

DiagLink was evaluated in a within-subject controlled study with 12 scenario packs, 12 simulated patients with mean age 24.2, and 12 practicing physicians with mean experience 5.3 years. Three conditions were compared: DiagLink; Baseline A, which used a matched UI with an independent KG viewer and LLM chat but no integrated synergy; and Baseline B, which was like DiagLink but revealed system predictions only after explicit physician confirmation (Zhou et al., 28 Jan 2026).

The study reports patient-side gains on Likert 1–5 items via paired t-tests. Significant improvements were observed across Relationship & Respect, Participation & Understanding, Emotional Support & Welfare, and Revisit Intention. Reported examples include Q4 “relaxed and at ease” with Diff = 2.25 and $p < .001$, Q6 “understand condition” with Diff = 0.75 and $p = .0001$, and Q10 “revisit intention” with Diff = 1.42 and $p = .0001$ (Zhou et al., 28 Jan 2026).

Physician workload and diagnostic efficiency also improved. NASA-TLX weighted workload decreased from 12.46 in Baseline A to 6.75 in DiagLink, with $p < .001$; the largest reductions were in Temporal Demand at $-11.5$ and Effort at $$\hat{e}_i = \arg\max_{e_j \in \mathcal{E}_g} \mathrm{sim}\bigl(\mathrm{emb}(s_i),\; \mathrm{emb}(e_j)\bigr).$$0, both with $$\hat{e}_i = \arg\max_{e_j \in \mathcal{E}_g} \mathrm{sim}\bigl(\mathrm{emb}(s_i),\; \mathrm{emb}(e_j)\bigr).$$1. SUS was 77.5, corresponding to approximately the 80th percentile versus comparable applications. Correct Top-1 diagnoses increased from 7 to 9, Top-3 from 9 to 11, and mean time dropped from 18.6 minutes to 7.8 minutes (Zhou et al., 28 Jan 2026).

Safety results are notable because the system explicitly avoids autonomous diagnosis. Anchoring effect was absent in the DiagLink versus Baseline B comparison, with McNemar test $$\hat{e}_i = \arg\max_{e_j \in \mathcal{E}_g} \mathrm{sim}\bigl(\mathrm{emb}(s_i),\; \mathrm{emb}(e_j)\bigr).$$2. Relative-likelihood scores discriminated correct versus incorrect diagnoses at $$\hat{e}_i = \arg\max_{e_j \in \mathcal{E}_g} \mathrm{sim}\bigl(\mathrm{emb}(s_i),\; \mathrm{emb}(e_j)\bigr).$$3 versus $$\hat{e}_i = \arg\max_{e_j \in \mathcal{E}_g} \mathrm{sim}\bigl(\mathrm{emb}(s_i),\; \mathrm{emb}(e_j)\bigr).$$4 with $$\hat{e}_i = \arg\max_{e_j \in \mathcal{E}_g} \mathrm{sim}\bigl(\mathrm{emb}(s_i),\; \mathrm{emb}(e_j)\bigr).$$5, $$\hat{e}_i = \arg\max_{e_j \in \mathcal{E}_g} \mathrm{sim}\bigl(\mathrm{emb}(s_i),\; \mathrm{emb}(e_j)\bigr).$$6, and Cohen’s $$\hat{e}_i = \arg\max_{e_j \in \mathcal{E}_g} \mathrm{sim}\bigl(\mathrm{emb}(s_i),\; \mathrm{emb}(e_j)\bigr).$$7. Physician trust was similarly calibrated at $$\hat{e}_i = \arg\max_{e_j \in \mathcal{E}_g} \mathrm{sim}\bigl(\mathrm{emb}(s_i),\; \mathrm{emb}(e_j)\bigr).$$8 versus $$\hat{e}_i = \arg\max_{e_j \in \mathcal{E}_g} \mathrm{sim}\bigl(\mathrm{emb}(s_i),\; \mathrm{emb}(e_j)\bigr).$$9 with $$\mathrm{score}(d_i) = \sum_{\hat{e}_j \in \hat{\mathcal{E}}_s} \frac{1}{\mathrm{dist}(d_i,\; \hat{e}_j)}.$$0, $$\mathrm{score}(d_i) = \sum_{\hat{e}_j \in \hat{\mathcal{E}}_s} \frac{1}{\mathrm{dist}(d_i,\; \hat{e}_j)}.$$1, and Cohen’s $$\mathrm{score}(d_i) = \sum_{\hat{e}_j \in \hat{\mathcal{E}}_s} \frac{1}{\mathrm{dist}(d_i,\; \hat{e}_j)}.$$2 (Zhou et al., 28 Jan 2026).

The same paper also makes its limitations explicit. Evaluation was restricted to text-based, non-emergency adult internal medicine scenarios with simulated patients of basic health literacy. Usability for low-literacy users and assistive technologies remains untested; multimodal interaction is future work. Observed failure modes included template drift, interruptions in emotional support, and graph-view crowding for very rich diseases. Long-term KG evolution effects were not established; the current results show feasibility rather than convergence properties (Zhou et al., 28 Jan 2026).

5. Adjacent medical formulations: dialogue systems, synthetic benchmarking, and multimodal synthesis

Several neighboring medical literatures instantiate Syn-Diag as either diagnostic dialogue, synthetic evaluation, or evidence-synthesis benchmarking. DxFormer models diagnosis-oriented dialogue as symptom inquiry followed by disease diagnosis, decoupling a Transformer decoder for inquiry from a Transformer encoder for classification. It defines symptom recall as

$$\mathrm{score}(d_i) = \sum_{\hat{e}_j \in \hat{\mathcal{E}}_s} \frac{1}{\mathrm{dist}(d_i,\; \hat{e}_j)}.$$3

and stops inquiry when the encoder’s maximum disease probability exceeds a threshold $$\mathrm{score}(d_i) = \sum_{\hat{e}_j \in \hat{\mathcal{E}}_s} \frac{1}{\mathrm{dist}(d_i,\; \hat{e}_j)}.$$4 or the maximum turn budget is reached. On Dxy, MZ-4, and MZ-10, it reported state-of-the-art results in both symptom recall and diagnostic accuracy, with particularly large gains on MZ-10 (Chen et al., 2022).

Rare-disease evaluation work shifts Syn-Diag from assistance to pre-deployment performance estimation. One line generates synthetic vignettes from Human Phenotype Ontology disease–phenotype annotations, maps phenotypes into symptom-checker primitives, simulates interviews under old and new algorithms, and estimates performance changes through $$\mathrm{score}(d_i) = \sum_{\hat{e}_j \in \hat{\mathcal{E}}_s} \frac{1}{\mathrm{dist}(d_i,\; \hat{e}_j)}.$$5 and $$\mathrm{score}(d_i) = \sum_{\hat{e}_j \in \hat{\mathcal{E}}_s} \frac{1}{\mathrm{dist}(d_i,\; \hat{e}_j)}.$$6. For diseases with phenotype frequency information, predicted and actual changes were strongly correlated, with $$\mathrm{score}(d_i) = \sum_{\hat{e}_j \in \hat{\mathcal{E}}_s} \frac{1}{\mathrm{dist}(d_i,\; \hat{e}_j)}.$$7, $$\mathrm{score}(d_i) = \sum_{\hat{e}_j \in \hat{\mathcal{E}}_s} \frac{1}{\mathrm{dist}(d_i,\; \hat{e}_j)}.$$8 for Recall@8 change and $$\mathrm{score}(d_i) = \sum_{\hat{e}_j \in \hat{\mathcal{E}}_s} \frac{1}{\mathrm{dist}(d_i,\; \hat{e}_j)}.$$9, $\mathcal{D}_{comb}$0 for Precision@8 change (Nishibayashi et al., 24 Jun 2025). A complementary line, SYNRARE, uses a Python/Java GUI over Synthea to generate synthetic rare-disease EHR cohorts that differ from majority cohorts by a definable degree through variance magnitude, range shift, modification probability, and BMI-driven comorbidity adjustments (Truong et al., 10 Jul 2026).

MEDSYN evaluates a different failure mode: not failure to enumerate differentials, but failure to synthesize heterogeneous clinical evidence into a final diagnosis. The benchmark contains 452 multilingual complex cases, with a mean of 3.97 distinct clinical evidence types and 8.42 images per case. It formalizes the DDx–FDx gap as coverage minus closed-ended FDx accuracy and introduces Evidence Sensitivity,

$\mathcal{D}_{comb}$1

together with an NMSE-based cross-modal utilization gap. The main finding is that top multimodal LLMs often perform well on DDx generation yet exhibit a much larger DDx–FDx gap than senior physicians, largely because of overreliance on less discriminative textual evidence and underuse of visual evidence, especially microscopy (Chen et al., 25 Feb 2026).

Taken together, these works broaden Syn-Diag from physician-facing assistance to a larger program of evaluation, simulation, and synthesis analysis. This suggests that diagnostic quality is increasingly being operationalized not only by final accuracy but also by dialogue efficiency, evidence sensitivity, calibration, and pre-deployment robustness (Chen et al., 2022, Nishibayashi et al., 24 Jun 2025, Chen et al., 25 Feb 2026).

6. Cross-domain transpositions beyond clinical medicine

Outside medicine, Syn-Diag is repurposed for few-shot industrial fault diagnosis, modular simulation debugging, intrusion detection, and syndrome differentiation. In industrial monitoring, the “Syn-Diag” framework for edge deployment uses three tiers: Visual-Semantic Synergy, Content-Aware Reasoning, and Cloud-Edge Synergy. Spectrograms derived from vibration signals are aligned to LLM semantic space through cross-modal pre-training, a content-aware mixed text embedding is built by similarity-based weighting, and a compact student model is distilled into a shared decision space with the cloud teacher. Across CWRU and SEU few-shot settings, the edge model preserved near-cloud accuracy while reducing model size by 83% and latency by 50% (Jia et al., 7 Oct 2025).

ADIOSS applies Syn-Diag to modular automotive simulations. Here the target is not a disease but a faulty module in a workflow represented as a graph of 0D models. The framework combines DMDc, a MixED-DMD convex program, and autoencoder-based latent embedding. Its minimal configuration requires one extra simulation beyond the reference, while NoDyn uses two simulations and suppresses explicit dynamics modeling. In the BEV case study, embedding-based MixED-DMD correctly localized Battery and Motor, whereas the no-embedding version misattributed one module (Jiang et al., 13 Mar 2026).

Cybersecurity uses the same label for SYN-flood and SYN-scan diagnosis. One SD-IoV pipeline applies StandardScaler, label encoding, and StratifiedKFold with a Random Forest classifier using 20 estimators and depth 10, reporting averaged Accuracy, Precision, Recall, and F1-score of approximately 0.999998 and detection time of approximately 0.24 seconds on CIC-DDoS2019-derived SYN traffic (Zamrai et al., 7 Sep 2025). Classical statistical lines instead use non-parametric CUSUM on per-window SYN counts, with average detection ratio 98.82% and average false alarm ratio 2.46% on DARPA 1998 traces (Zhang, 2012). Challenge-based validation through a “challenging number” in the three-way handshake is used to diagnose spoofed SYN floods (Kavisankar et al., 2011), while active defense work combines time-varying arrival statistics, trace-driven collaboration, and delegated SYN/ACK validation (Kumarasamy et al., 2012). Dendritic Cell Algorithm work further treats process identifiers as antigen and fuses PAMP, danger, safe, and inflammation signals to detect outgoing SYN scans (Greensmith et al., 2010).

Traditional Chinese Medicine work uses Syn-Diag for syndrome differentiation from free text. TCM-SD provides 54,152 real-world clinical records covering 148 syndromes and shows that ZY-BERT, pre-trained on 0.4B tokens of TCM text with Chinese Whole Word Masking, reaches test Accuracy 82.19% and Macro-F1 51.01% (Ren et al., 2022). A related diabetes study models each patient record as a one-dimensional image, generates random region proposals over 186 symptoms, and applies a CNN-based multi-instance multi-task architecture. On its multi-label benchmark, MIMT-CNN reaches MAP 0.573, Coverage 0.321, Subset Accuracy 0.259, and Hamming Loss 0.081, outperforming the reported baselines (Wang et al., 2018).

These transpositions do not establish a single transferable algorithm. They do, however, show a recurring structural commitment: multi-signal fusion, explicit intermediate representations, and decision procedures that remain inspectable enough to support attribution, updating, or mitigation.

7. Limitations, misconceptions, and research directions

The literature consistently resists interpreting Syn-Diag as autonomous end-to-end diagnosis. DiagLink explicitly remains physician-in-the-loop, hides provisional outputs from patients, and places final responsibility on clinicians (Zhou et al., 28 Jan 2026). MEDSYN shows that even strong multimodal models can generate plausible differentials yet fail when asked to collapse heterogeneous evidence into a single correct final diagnosis (Chen et al., 25 Feb 2026). Rare-disease synthetic benchmarking is similarly not a substitute for real-world validation: synthetic vignettes simplify negative evidence and phenotype dependencies, and SYNRARE’s trajectories are acknowledged as “idealized” relative to missingness, confounding, and documentation idiosyncrasies in real EHRs (Nishibayashi et al., 24 Jun 2025, Truong et al., 10 Jul 2026).

Another misconception is that better fusion necessarily yields better calibration. The evidence is mixed. DiagLink reports calibrated trust and an observed absence of anchoring in its controlled study (Zhou et al., 28 Jan 2026), but MEDSYN documents substantial cross-modal utilization gaps even for high-performing models (Chen et al., 25 Feb 2026). In engineering settings, strong cross-condition generalization in cloud–edge fault diagnosis still presumes a known class space and semantically well-formed prompts (Jia et al., 7 Oct 2025). In simulation debugging, detection depends on adequate latent linearization and time-embedded design coverage (Jiang et al., 13 Mar 2026).

Current research directions therefore converge on a narrower but technically sharper question: how to make synthesis itself measurable, updateable, and safe. DiagLink answers with expert-triggered KG evolution and provenance-rich role-adaptive interfaces (Zhou et al., 28 Jan 2026). MEDSYN answers with DDx–FDx gap analysis and Evidence Sensitivity (Chen et al., 25 Feb 2026). Rare-disease evaluation answers with HPO-grounded synthetic interviews before deployment (Nishibayashi et al., 24 Jun 2025), and cloud–edge industrial work answers with a shared decision space that permits head-only online updates (Jia et al., 7 Oct 2025). The cumulative implication is that Syn-Diag is becoming a research program about controlled multi-evidence diagnosis rather than merely a label for prediction accuracy.