Saltar al contenido principal

Escribir un comentario

PREreview del A Multi-Track Cognition Framework for Global Integrative Medicine: Breaking Paradigm Incommensurability Through System-Level Mapping Across Medical Systems

Publicado
DOI
10.5281/zenodo.24314752
Licencia
CC BY 4.0

Reviewer Report: Multi-Track Cognition Framework for Global Integrative Medicine

Manuscript: "A Multi-Track Cognition Framework for Global Integrative Medicine: Breaking Paradigm Incommensurability Through System-Level Mapping Across Medical Systems" (G. Yang; Preprints.org, doi:10.20944/preprints202603.0470.v1, posted 5 March 2026)

Type of report: Point-by-point reviewer report, prepared to help the author strengthen the work before journal submission.

The manuscript proposes a computational "multi-track cognition" (MTC) framework that lets data from different medical systems (traditional Chinese medicine, biomedicine, Tibetan medicine, with others said to be addable) be compared and fused without forcing one system's vocabulary onto another. The author's stated aim is to move beyond format-level standardisation towards semantic alignment, and beyond single-point linear correspondence towards system-level mapping.

Importance of the research area. Integrative medicine is widely practised and increasingly supported by health policy, including the World Health Organization's strategic work on traditional and complementary medicine. A persistent barrier to generating integrated evidence is that systems describe the patient and the intervention in different conceptual languages. Examples are pattern differentiation, nature and flavour, and meridian tropism on one side, and disease categories, biomarkers and molecular targets on the other. A general and extensible way to connect these descriptions would be useful for evidence synthesis, chronic-disease management and individualised care. The author appropriately targets semantic, and not merely syntactic, interoperability, which is a question that merits careful methodological work.

Approach proposed. The framework has three layers (Figure 1; Section 2):

1. Cognitive tracks (Section 2.1). Each medical system has its own track with a native feature set, normalisation rules and a native "systemic effect" index.

2. Homeostatic representation network, HRN (Section 2.2). A directed acyclic graph of 28 laboratory indicators across eight physiological systems serves as the shared anchor. Its topology comes from literature and three expert physicians. Its parameters are learned with a Bayesian network on MIMIC-III.

3. System-level mapping (Section 2.3). A random forest, guided by three rules (cluster correspondence, network emergence, context dependence), is trained with a mean-squared-error loss to predict post-intervention changes in the HRN indicators. Performance is judged by Semantic Alignment Accuracy (SAA, the share of cases in which two mapped vectors reach cosine similarity ≥ 0.85), R², MAE and RMSE (Section 2.4).

Evidence reported (Section 3). Four public datasets are used (Table 1), with a 70/30 split and 5-fold cross-validation. In the TCM–Western dual-track setting the framework reports SAA 91.27% and R² 0.852, against 59.13–76.58% SAA for three baselines (Table 2). A three-system extension adding Tibetan medicine reports SAA 86.34% and R² 0.821 (Section 3.4). Agreement with three clinical experts is reported as κ = 0.86, and a Lung–Spleen Qi Deficiency case illustrates the workflow (Section 3.5). The author concludes that the framework is lossless, general, and ready to support clinical integration.

Strengths. The following strengths provide a sound basis for revision:

• The problem is timely and clinically meaningful.

• The modular idea, with track-specific components and one shared core, is attractive and in principle testable.

• An outcome-level physiological anchor is a practical way to compare interventions that act through different theories.

• The model, loss, hyperparameters and pseudo-code are stated explicitly, which allows scrutiny.

• The author attempts expert annotation, effect sizes and a multi-system extension.

• The limitations section acknowledges the dependence on paired cross-system data.

Requirement for revision. In its present form the manuscript describes a promising concept, but the empirical results do not yet support the strength of the conclusions ("lossless", "universal", "significantly outperforms", "clinical practical value"). Several important items are missing or unclear: the provenance of the outcome labels, sample sizes, the exact definition and validation of SAA, the role of the HRN in the pipeline, ablation of the three rules, and leakage-free evaluation. I therefore recommend major revision. The comments below (13 major, then minor) each give the location and the reason it matters, with a suggestion where one is likely to help. They are offered in the spirit of helping the work reach a form in which its contribution can be properly recognised.

Recommendation: major revision

Major comments (M1–M13) affect the validity or interpretation of the conclusions. Minor comments (m1–m25) concern clarity, completeness and presentation. Each comment gives its location and the reason it matters. A suggestion is added only where one is likely to help, and the author is free to address a point in another way. Locations refer to section numbers and the Abstract, Introduction, Figure 1, Tables 1–2 and the reference list of the preprint.

Major Comments

M1. The scope of the homeostatic anchor and the claims of "lossless" alignment and preserved native logic

Location: Abstract; Introduction (final paragraph); Section 1 (definitions 3–4); Section 2 (opening paragraph) and 2.2; Section 4.1; Section 5; Figure 1 ("The Only Mapping Anchor and Value Consensus").

Reason:

• The anchor is made of 28 routine laboratory indicators from the biomedical system. Mapping every system onto it is a legitimate and practical choice, but it is not paradigm-neutral. It is in tension with the stated aim of avoiding "interpreting TCM with Western medicine" (Introduction; Section 4.1).

• Projecting a high-dimensional feature set onto a 28-dimensional vector is many-to-one, so information is lost by construction. "Lossless" can only be meaningful relative to a defined task, and no measure of information retention is reported.

• The premise that "the ultimate goal of intervention is to regulate human homeostasis" (Section 2.2) is asserted rather than argued. Surgery, vaccination, psychological therapy and palliative care, and many TCM aims such as symptom relief and constitution regulation, may not be visible in laboratory homeostasis. TCM patterns often present without laboratory abnormality, so they may be invisible to this anchor.

• Definition 3 describes homeostasis as a dynamic equilibrium, "rather than the index stability under disease state". It is then operationalised as laboratory changes in ICU patients. The concept (homeostasis versus allostasis and individual set-points) and its measurement are not the same thing.

• Figure 1 extends the framework to cognitive behavioural therapy, ACT, sports medicine, clinical nutrition and digital-twin medicine. None are tested, and it is unclear how a psychological therapy maps to 28 laboratory indicators.

Suggestion: Describing the anchor as a shared physiological outcome interface, with explicit statements of what it can and cannot represent, would match the evidence better than "lossless" or "only". A short subsection on what is preserved and what is lost, supported by one operational test (for example, comparing track-specific tasks before and after mapping) and a sensitivity analysis with an alternative anchor, would strengthen the argument.

M2. Provenance of the outcome labels and feasibility of the paired data

Location: Section 2.3.2 (input, output and loss function); Section 2.4 (SAA, "for the same health state"); Section 3.1 and Table 1; Section 3.4; Section 3.5 (final paragraph).

Reason:

• The model target Y is the "real change of homeostatic indicators after intervention" for each medical system (Section 2.3.2). According to Table 1, TCMSP is a herb–compound–target database with no patients or outcomes. MIMIC-III is an ICU database of conventionally treated patients with no TCM interventions. The CACMS clinical dataset and the "Public Dataset of Clinical Research on Tibetan Medicine" are not identified further, and the latter is attributed to a journal rather than a repository.

• It is therefore unclear which source supplies post-intervention laboratory changes for TCM and Tibetan interventions, and how pairs of "intervention data from two different medical systems for the same health state" were formed. Without this, none of the performance figures can be interpreted.

• No sample size is reported anywhere (patients, records, interventions, pairs, test cases M in the SAA formula), yet Section 3.5 states that validity was "verified by large-scale dataset experiments".

• Section 3.3 refers to "simulation experiments". If any data were simulated, this must be stated and explained.

Suggestion: A "Data and cohort construction" subsection with a flow diagram would resolve most of this: each dataset's version, access route, population and N, which of the 28 indicators it contains, how Y was computed, and how cross-system pairs were matched. If truly paired data do not exist, a clearly labelled synthetic study with known ground truth, or published trial effects as an external reference, would be a reasonable alternative.

M3. Definition, naming and validation of Semantic Alignment Accuracy (SAA)

Location: Section 1 (definition 6); Section 2.4 item 1; Section 3.3 (Table 2 and the κ paragraph); Abstract.

Reason:

• SAA is the proportion of test cases in which two vectors produced by the same model have cosine similarity ≥ 0.85. It is an agreement rate, not accuracy against a ground truth, so the name may lead to over-interpretation.

• High cosine similarity can arise without any semantic equivalence. Random forests shrink predictions toward the population average, so two outputs can be similar simply because both resemble the mean change. Cosine ignores magnitude and is sensitive to centring. No null distribution (randomly paired or mismatched cases) is reported.

• The metric treats agreement as correct. If two interventions legitimately act differently, a good model should produce different vectors. The rationale for rewarding agreement needs to be stated.

• The threshold θ = 0.85 was tuned by ROC on 100 expert-annotated pairs (Section 2.4). The κ = 0.86 in Section 3.3 is then computed against "double-blind annotation results of 3 clinical experts", which appear to be the same annotations. Agreement measured on the data used to set the threshold is optimistically biased. The same paragraph also mentions an "expert-annotated clinical gold standard from the public clinical dataset", which is a different source and is not described elsewhere.

• The annotation protocol is incomplete. It does not say what the experts were shown, inter-rater reliability among the three experts, or how disagreements were resolved. With 100 pairs, sensitivity (92.3%) and specificity (88.7%) have wide confidence intervals.

• It is not explained how SAA is computed for Pearson, LASSO and knowledge-graph baselines, which do not output homeostatic vectors (see M8).

Suggestion: Renaming the metric (for example, cross-system concordance rate), adding mismatched-pair controls and a threshold-sensitivity curve, and fixing θ on a development set before computing κ on an independent set of expert-annotated pairs (with confidence intervals and inter-rater agreement) would address most of these points.

M4. Role of the HRN in the pipeline, the DAG structure and the circular use of MIMIC-III

Location: Sections 2.2.1 and 2.2.2 (including the factorisation formula); Section 2.3.2; Algorithm 1; Section 3.1 (MIMIC-III paragraph); Figure 1.

Reason:

• The HRN is presented as the key to overcoming incommensurability, but after construction its edges E, weights W and the "normal / effective / safety-risk" intervals (Section 2.2.1) do not appear in the mapping model. The random forest takes X and outputs Y without reference to G, and SAA is computed on output vectors. A flat 28-output regressor would produce identical results, so the experiments do not test the HRN.

• A directed acyclic graph cannot represent the feedback loops that define homeostasis (for example the HPA axis, glucose–insulin, renin–angiotensin). A structure learned from cross-sectional data on stable ICU patients also reflects ICU physiology and treatment effects.

• Section 2.2.2 says expert consultation (three physicians) sets the topology, and also says the construction is designed "to avoid subjective bias to the greatest extent" and "fully conforms to the objective laws of human physiology". These statements are difficult to reconcile. No inter-expert agreement is reported.

• The displayed formula is the factorisation of a joint distribution, but is described as "the conditional probability distribution formula". The text speaks of optimising the "weight" and its "direction", while a Bayesian network with fixed structure learns conditional probability tables. Missing details include discretisation or Gaussian assumptions, software, missing-data handling, sample size and validation of the learned parameters.

• The same MIMIC-III subset appears to provide the anchor parameters, the Western-track features and the "real" homeostatic values (Table 1), which is circular.

• The 28 indicators are never listed. Several indicators used in Section 3.5 (SOD, IgG, cortisol) are not routine ICU laboratory tests and are likely sparse or absent in MIMIC-III. Reproductive-system indicators are likely to be sparse as well.

Suggestion: Please show explicitly how G enters the mapping (for example as a graph-structured output layer or a consistency regulariser) and report an ablation with and without G. Listing the 28 indicators in a table, and learning the network on data separate from the evaluation data, would also help.

M5. Operationalisation of the three mapping rules and absence of an ablation

Location: Section 2.3.1; Section 2.3.2; Algorithm 1 (step 5); Section 3.3; Section 3.5 (item 2); Abstract.

Reason:

• The rules are described as constraints with "clear operability and theoretical support". In the implementation, however, cluster correspondence means passing the features of one systemic effect to the model as a group, network emergence means interaction terms that are "automatically learned through the model", and context dependence means stratified sub-models by syndrome or constitution label. A random forest already sees all features jointly and learns interactions, so the first two rules do not change what an ordinary random forest does. How feature groups are formed (expert, clustering, ontology) is not stated.

• Without an ablation, the benefit of each rule cannot be attributed.

• Context dependence needs syndrome or constitution labels. It is unclear which dataset contains them (they are unlikely to be in MIMIC-III). Stratifying reduces the training size per stratum, and the handling of unlabelled new samples is not described.

• "Emergence" is a strong term. Statistical interactions in a random forest are not evidence of emergent properties, and the claim to "completely restore" multi-factor synergy cannot be tested without ground truth.

Suggestion: Defining each rule formally, testing interactions with explicit measures (for example SHAP interaction values or Friedman's H-statistic), and adding an ablation table (full framework, minus each rule, and a plain random forest) would show which rules contribute.

M6. Model specification: unit of analysis, inputs, outputs and causal interpretation

Location: Section 2.1; Section 2.3.2 (items 1–2); Section 2.4 item 2; Section 3.1 (preprocessing and MIMIC-III paragraphs); Section 3.5.

Reason:

• The inputs are intervention descriptors (herb nature and flavour, meridian tropism, active ingredients, targets), while the outputs are patient-level laboratory changes. The unit of observation is never defined. It could be a herb, a formula, a prescription or a patient–treatment episode. If herb-level class attributes are the inputs, it is unclear how multi-herb formulas and doses are aggregated.

• Baseline patient state (baseline laboratory values, comorbidity, concurrent therapy) is not listed as an input, although change scores depend strongly on baseline values and on regression to the mean.

• Laboratory changes after an intervention in observational or ICU data reflect the disease course and concurrent treatments as well as the intervention. A model trained to minimise MSE is predictive and associational, yet Section 2.4 speaks of "the real impact of interventions" and Section 3.5 of treatment recommendation.

• The levels are mixed in Section 3.5. "Sweet and warm nature and flavor" is a herb attribute but is listed as a patient feature, and the Western-track features there are patient laboratory results, whereas Section 2.3.2 describes drug ingredients and targets.

• Encoding is described inconsistently. Section 2.3.2 gives ordinal values from −3 to +3, whereas Sections 2.1 and 3.1 say all features are mapped to [0,1]. Random forests are invariant to monotone rescaling, so the emphasis on normalisation needs a reason.

• The outcome window ("after intervention") is not defined. For MIMIC-III, the intervention that follows "stable vital signs within 24 hours of admission" is not identified.

M7. Evaluation protocol and the risk of data leakage

Location: Section 3.1 (preprocessing items 1–3); Section 2.3.2 (model training process); Algorithm 1; Section 3.3.

Reason:

• Section 3.1 lists imputation and min-max standardisation before the 70/30 split, which implies preprocessing fitted on all data. Algorithm 1 (step 5) says preprocessing is performed on the training set. The two statements conflict, and multiple imputation fitted on the whole dataset leaks information.

• The split is random. MIMIC-III contains repeated ICU stays for the same patient, and herbs and formulas recur across records. A random record-level split can place the same patient, herb or formula in both training and test sets, which inflates performance. An R² of 0.852 for predicting laboratory changes is unusually high for this task and warrants checking for leakage.

• Both a 70/30 split and 5-fold cross-validation are described, but it is unclear whether Table 2 reports the cross-validated mean or the held-out test result. The grid-search ranges are not given, nested cross-validation is not used, and the "optimal" values (100 trees, depth 10, minimum split 5) look close to conventional settings.

• No external or temporal validation is reported.

Suggestion: Fitting imputers and scalers within training folds, splitting at patient (and formula) level, using nested cross-validation or a held-out test set used once, and adding external or temporal validation would address the leakage concern.

M8. Choice and construction of baselines

Location: Section 3.2; Section 3.3; Table 2.

Reason:

• Baselines 1 and 2 are linear, while the proposed model is a nonlinear random forest. The advantage may therefore reflect the model class and not the framework.

• Baseline 3 (knowledge-graph alignment) addresses term alignment, not prediction of post-intervention laboratory change. It is unclear how it yields predictions for R², MAE and RMSE, or homeostatic vectors for SAA. Reference [12] is a general survey of knowledge graphs, not a specific method.

• The statements that these are the "mainstream" methods in the field are not supported by citations.

• Several informative comparators are missing:

o the same random forest without the three rules;

o a mean-change predictor, which anchors R² and SAA;

o modern learners such as gradient boosting or a multi-task neural network;

o an ontology-based alignment through standard terminologies.

Suggestion: Specifying each baseline's inputs, outputs and tuning, adding same-learner ablations and a mean predictor, and including at least one modern learner with an equal tuning budget would make the comparison interpretable. If the knowledge-graph method cannot produce homeostatic vectors, it may be better evaluated separately on term-alignment accuracy.

M9. Statistical reporting and internal numerical inconsistencies

Location: Abstract; Section 3.3 (Table 2, bullet points, paired t-test and κ paragraphs); Section 3.4.

Reason:

• The Abstract says accuracy is "improved by 32.14%", whereas Section 3.3 says 32.14 percentage points (91.27 − 59.13). The relative improvement is about 54%. The Abstract should not leave this ambiguous.

• The Abstract states R² ≥ 0.85, but the three-system result in Section 3.4 is 0.821.

• The unit behind Cohen's d (2.87, 2.31, 1.94) is not stated. If SAA is a per-case binary indicator, the reported proportions imply much smaller standardised differences. My rough calculation with pooled binomial variances gives about 0.8 against Baseline 1 and about 0.4 against Baseline 3. Values near 2 would arise if d were computed on variability across folds or repeated runs. Fold-level values are not independent, which can inflate d. Please specify how d was obtained.

• The paired t-test ("p<0.01") does not say what was paired, how many pairs there were, or how multiplicity was handled (three baselines × four metrics). No standard deviations or confidence intervals appear in Table 2, and statements such as "extremely significant" and "significantly lower MAE and RMSE" are not backed by reported tests.

• A Cohen's d above 0.8 is a convention for standardised difference. It does not by itself establish clinical value.

• For κ, a p-value adds little, and a confidence interval is needed. "Extremely strong" is not a standard descriptor ("almost perfect" is the conventional label for κ > 0.80).

Suggestion: Stating the unit of analysis for each test, using patient-level bootstrap confidence intervals, applying McNemar's test or a paired bootstrap to SAA with correction for multiple comparisons, and judging clinical relevance against reference-range width would resolve these points. The Abstract should also say "percentage points".

M10. Evidence for generality and extensibility to "any" medical system

Location: Abstract; Introduction (final paragraph); Section 3.4; Section 4.1 (third "breakthrough"); Section 5; Figure 1.

Reason:

• Universality is inferred from three systems (TCM, biomedicine, Tibetan medicine). Two of them are encoded with the same ordinal and one-hot scheme, and the Tibetan dataset is not identified. Statements such as "any number of independent medical systems" and "extremely strong universality" go well beyond this. Ayurveda, Mongolian medicine and functional medicine are mentioned but not tested.

• It is not explained how SAA is defined for three systems (pairwise average, or a three-way criterion). The fall in SAA from 91.27% to 86.34% is not analysed.

• The effort needed to build a new track (feature dictionary, normalisation rules, expert time) is not reported, although the claim is that only the track needs to be constructed.

Suggestion: Moderating the wording to "extensible in principle, demonstrated for three systems", reporting each pairwise result, and adding a leave-one-track-out experiment would align the claim with the evidence.

M11. The case demonstration and the clinical claims

Location: Section 3.5; Section 4.3; Figure 1 ("Planning and Decision-Making Phase", "zero-compromise principle"); Section 5; Abstract (final sentence).

Reason:

• The case study contains no numerical output (no mapped vector, similarity value or before-and-after HRN profile), so the reader cannot follow how the abnormal dimensions were derived. Abnormal IgG, SOD and cortisol are said to "correspond to" the immune, endocrine and digestive systems, but the digestive link is not explained.

• The levels are mixed (see M6), as herb attributes are presented as patient features.

• The case states that the framework outputs an "integrated intervention scheme recommendation" naming a formula (modified Sijunzi Tang), nutritional and lifestyle advice, and a "safety assessment". None of these outputs is defined in the Methods or evaluated. Treatment recommendation requires causal evidence, interaction and safety checks (including herb–drug interactions) and appropriate oversight.

• The closing sentence says the model's universal validity and statistical efficiency "have been verified by the aforementioned large-scale dataset experiments". This is not supported by the information provided (see M2).

• Figure 1 refers to a "zero-compromise principle" and "paradigm conflict resolution". These are not defined in the text.

Suggestion: Presenting either an actual de-identified case (or a clearly labelled synthetic case) with numerical outputs at each step, or recasting the section as a conceptual schematic, would be sufficient. The treatment recommendation could be removed or labelled as outside the scope of this study.

M12. Reproducibility, transparency and required statements

Location: Entire manuscript; Section 1 (opening sentence: definitions are "consistent with the core connotation of the corresponding invention patent"); Section 3 (introductory paragraph); Section 3.1 (final paragraph); back matter.

Reason:

• No code, data-availability statement, indicator list, hyperparameter grid, software versions or random seeds are provided. Without them the results cannot be reproduced.

• The statement that the datasets comply with "medical ethics norms and data use requirements" is asserted without detail. MIMIC-III requires credentialed access under a data use agreement. The conditions for the CACMS and Tibetan datasets are not stated.

• The reference to an invention patent indicates a potential competing interest and bears on novelty. It is not disclosed in a competing-interests statement.

• The preprint shows no funding, author-contribution or AI-tool-use statement, which most journals require.

M13. Positioning in the literature and support for the stated research gaps

Location: Introduction (paragraphs 1–3); Section 4.1; reference list.

Reason:

• The central argument rests on claims about existing work ("most existing studies", "most schemes ... single-point linear correspondence", "reductionism transformation trap"). These carry no supporting citations or examples.

• The reference list has 12 items, six of which are Chinese-language journal articles. Several areas directly relevant to the proposal are not engaged:

o terminology and ontology alignment (for example the ICD-11 Traditional Medicine chapter, WHO international standard terminologies on traditional medicine, SNOMED CT, UMLS);

o multi-view and multimodal learning, domain adaptation and representation alignment;

o knowledge-graph and network-pharmacology work that already links TCM and biomedical data;

o the physiology literature on homeostasis and allostasis;

o later discussions of incommensurability (Kuhn's later "local incommensurability" and translatability) and the epistemology of integrative medicine.

• Section 4.1 attributes ideas to named individuals ("the holistic principles championed by Academician Fan Daiming", "on the basis of Professor Li Shao's ... analysis"). Scholarly attribution would cite specific works and state what is built upon, and the claim of a "computational instantiation" needs to be shown.

• "Breaking paradigm incommensurability" in the title is a strong claim. The paper does not define what test would show that incommensurability has been overcome.

Suggestion: A short related-work section organised by theme (terminology and ontology alignment; knowledge graphs and network pharmacology; representation learning for heterogeneous data; epistemology of integrative medicine), together with a more measured title such as "A multi-track framework for outcome-level mapping across medical systems", would position the contribution more clearly.

Minor Comments

m1. Title.

Location: Title page. "Breaking Paradigm Incommensurability" and "Global" promise more than the study demonstrates (see M1, M10, M13).

m2. Abstract.

Location: Abstract. It gives no description of datasets, sample sizes or study design, and "strong consistency with clinical expert judgments" is stated without the number of experts, pairs or κ. The phrase "homeostatic representation network of multiple dimensions of human eight physiological systems" is hard to parse. The percentage-point and R² inconsistencies are noted in M9. A structured abstract (Background, Methods, Results, Conclusions) with N and confidence intervals would help.

m3. Introduction, framing of the problem.

Location: Introduction, paragraphs 1–2. The opening is framed by the "Healthy China" strategy, whereas the title and aims are global, and the "consensus in the medical community" rests on references [1,2], which are commentary articles. The WHO strategy is cited as 2023–2032, but the current document is the Global Traditional Medicine Strategy 2025–2034 adopted by the 78th World Health Assembly (WHO).

m4. Introduction, aims and contributions.

Location: Introduction, final paragraph. The paragraph states that the framework "can be seamlessly adapted to any medical system with a complete theoretical system" but gives no explicit research questions or list of contributions, so the reader cannot tell what is a hypothesis, a design choice or a result. Two or three stated aims would help.

m5. Definitions and glossary.

Location: Section 1 (definitions 1–6). "Cognition" suggests cognitive science, whereas the construct is a data-representation channel, and "Multi-track Cognition" is called an "original core theory" although what is proposed is a computational framework. Definitions 2 and 4 each repeat the same statement as a "formal definition", definition 4 calls the HRN "the only" benchmark, definition 3 ends with an ambiguous clause, and definition 5 is not used again. Terms such as "Qi deficiency" and "nature and flavor" are not explained for international readers, and terminology varies ("Western medicine" and "biomedicine"; "node", "dimension" and "indicator"). A short glossary would resolve this.

m6. Figure 1.

Location: Figure 1 and its caption (which continues on the next page). The figure shows a three-phase clinical workflow (including a "zero-compromise principle") and several systems (mind–body, sports and rehabilitation, clinical nutrition, Bantu medicine, digital twin) that are not described in the Methods or tested. It does not show the actual pipeline (X → random forest → Y → SAA). Some labels are very small (the ACT node) or broken across lines ("Rehabilitatio n", "Biomedical/R eductionist"), and grouping clinical nutrition and sports medicine under "reductionist" is debatable. Redrawing it around the tested pipeline, with future extensions clearly marked, would help.

m7. Feature dimensions and track scope.

Location: Section 2.1, items 1–2. The example dimensions (nature and flavour, meridian tropism, five-source attribute, six tastes) are listed without sources, and it is unclear whether tracks describe herbs, patterns, formulae, acupuncture, diet or all of these. The "grading standard of herbal medicine property intensity" and the "quantitative scoring standard of TCM syndromes" are not cited. Citing the standard for each dimension set, and tabulating dimensions with their encodings, would help.

m8. Native systemic-effect index versus the training objective.

Location: Section 2.1, item 3; Section 2.3.2. Item 3 says the native effect index (for example harmony of viscera function, yin–yang balance) "is used as the optimization objective of mapping model training", but Section 2.3.2 trains only on the MSE of laboratory changes, so these indices do not appear in the model. A quantitative "harmony degree" is also not standardised. The statement could be removed, or the index added as an auxiliary target with a defined measurement.

m9. The 28 indicators and reference ranges.

Location: Section 2.2.1. "Clinically gold-standard detectable" is non-standard wording, and the eight systems omit others (for example haematological and hepatic function, musculoskeletal and skin). Reference ranges vary by age, sex, assay and laboratory. The "guidelines issued by the Chinese Medical Association" are not cited individually, reference [9] is a 2015 operating manual, and the "intervention effective interval" and "safety risk interval" are not defined or used. A table of the 28 indicators with units, ranges, sources and LOINC codes would address most of this.

m10. Writing and expert process in the network construction.

Location: Section 2.2.2. The Bayesian-network justification is repeated in several sentences ("avoiding subjective bias to the greatest extent", "fully conforms to the objective laws"), there is a missing space ("physiology.Using"), and the selection of the "3 chief physicians" and the method of reaching consensus are not described.

m11. Wording of the three rules.

Location: Section 2.3.1. Rule 1 ends with an awkward "supplementary definition: rule governing synergistic effects of feature clusters", which restates the title. Rule 2 invokes "the emergence theory of complex system science" without citation, and rule 3 refers to "physical constitution" labels without naming a classification system or the datasets that contain it.

m12. Rationale for random forest.

Location: Section 2.3.2 (opening paragraph). The statements that random forest is "not easy to overfit", has "good interpretability" and is "fully conforming to the methodological specifications of medical research [8]" are not supported by Breiman (2001). Interpretability is limited without additional tools such as SHAP, and the handling of the 28 outputs (a joint multi-output forest or separate forests) and the hyperparameter ranges are not described.

m13. Algorithm 1 and the training description.

Location: Section 2.3.2 (model training process); Algorithm 1. The pseudo-code has no grid-search loop and no final evaluation on held-out data, and step 5 ("feature grouping and preprocessing according to the three mapping rules") is vague. The text announces the algorithm twice ("The complete algorithm flow is as follows: The algorithm pseudo-code is as follows:") and the numbered list restarts at "1." for "Model training process".

m14. Descriptions of the evaluation metrics.

Location: Section 2.4 (items 2–4). The descriptions contain claims the metrics cannot deliver. MAE does not "ensure ... without systematic deviation" (the mean signed error would show bias), RMSE penalises large errors but does not by itself "test the stability of the model in extreme clinical scenarios", and R² is a coefficient of determination, not a measure of "interpretability". It is also unspecified whether R² is out-of-sample, which mean is used in the denominator, how the 28 outputs are averaged, and on what scale MAE and RMSE are expressed. Sentences such as "..., This indicator ..." contain comma splices.

m15. Table 1 and dataset description.

Location: Section 3 (opening paragraph); Section 3.1; Table 1. "Massachusetts Institute of Technology" as the source of MIMIC-III omits the originating hospital (Beth Israel Deaconess Medical Center) and the PhysioNet access route. TCMSP has no citation or version, "Chinese Journal of Tibetan Medicine" is a journal and not a dataset, and the CACMS dataset has no name, identifier or access route. The table gives no sample sizes, variables or time periods. Adding columns for version, access, N, variables and period, and citing MIMIC-III (Johnson et al., 2016) and TCMSP (Ru et al., 2014), would help.

m16. Preprocessing details.

Location: Section 3.1 (preprocessing items 1–3 and the MIMIC-III paragraph). The criteria for "stable vital signs and test indicators" are not given, the method and number of multiple imputations are not reported, and treating the mean of an ICU-admission stable period as the "homeostatic benchmark" assumes that a stable ICU patient is homeostatically "healthy". Outlier handling and unit harmonisation are not described, and there is a stray fragment ("Full data preprocessing process:.").

m17. Language of certainty throughout.

Location: Section 2.2 ("fully conforms"), Section 3.3 ("proving", "far exceeding", "extremely significant"), Section 3.4 ("extremely strong universality"), Section 4.1 ("breakthroughs", "solving the core problem", "brand-new"), Section 5 ("excellent", "lossless"). Results from a single study support but cannot "prove" a claim, and strong adjectives reduce credibility. More measured language ("suggests", "is consistent with", "in this dataset") would let the results support the conclusions.

m18. Table 2.

Location: Table 2; Section 3.3. The table shows point estimates only: there is no N, SD or confidence interval, the metric scale is not stated, and the Cohen's d row does not say which metric it refers to (SAA, R² or an error measure), what the unit of analysis is, or that each value compares the proposed framework with the baseline in that column.

m19. Structure of the Discussion.

Location: Section 4.1. The Discussion restates claims instead of interpreting the results. It contains no error analysis (which indicators, systems or strata perform poorly), no comparison with published performance, and no discussion of why a random forest on categorical features should reproduce laboratory changes with R² above 0.85. A per-indicator results figure and a discussion of the weakest indicators would be informative.

m20. Limitations.

Location: Section 4.2. The four limitations are mainly about future extensions. Important ones are missing: provenance and pairing of labels, reliance on a laboratory-based anchor, observational and non-causal data, absence of external or prospective validation, the validity of SAA, the DAG simplification, and transfer between an ICU population and Chinese outpatient settings. Adding NLP for TCM records is future work, not a limitation, and "Small-sample and low-quality datasets will significantly affect the model effect" is unquantified.

m21. Compatibility with medical standards.

Location: Section 4.3. ICD-11 classifies diseases and conditions and does not define laboratory indicators, so the statement that creatinine and eGFR "directly correspond" to ICD-11 diagnostic criteria needs clarification. The ICD-11 Traditional Medicine chapter (Chapter 26) is directly relevant and is not mentioned, the "national standard Classification and Codes of TCM Syndromes" is not cited, and the claim of "seamless" HIS and EMR integration is not demonstrated. Mapping the indicators to LOINC and showing one example mapping from TCM patterns to ICD-11 Chapter 26 would make the claim concrete, or the section could describe these as intended directions.

m22. Conclusions.

Location: Section 5. It repeats "verified", "lossless", "excellent" and "clinical consistency" without referring to the limitations or the open questions about labels and validation.

m23. References.

Location: Reference list; in-text citation order.

• Citations do not appear in order of first mention ([1,2], then [6], then [3–5,7], then [11], [9], [8], [12], [10]).

• The WHO item is given as "WHO global strategy on traditional and complementary medicine 2023–2032" (2022). The current document is the Global Traditional Medicine Strategy 2025–2034, adopted at the 78th World Health Assembly in May 2025 (WHO).

• Some DOIs should be checked. The DOI given for the WHO item uses a prefix (10.4135) that I would not expect for a WHO publication, and the DOIs of references 2, 4 and 10 have near-identical suffix patterns (CJIM.2022.1245, CJIM.2020.1245, CJIM.2021.1546). I could not resolve DOIs while preparing this report, so I ask the author to verify every DOI and bibliographic detail.

• Reference 5 has no DOI, and key sources are missing (MIMIC-III, TCMSP, Bayesian network learning, κ and effect-size methods, the ICD-11 Traditional Medicine chapter).

m24. Equations and notation.

Location: Section 2.2.2 (factorisation); Section 2.3.2 (loss); Section 2.4 (SAA, R², MAE, RMSE). Equations are not numbered, "test samples" and "pairs" are mixed for SAA, and the loss is written for a scalar target although the output has 28 dimensions. The factorisation is described as a conditional probability distribution.

m25. Language, typography and terminology.

Location: Throughout; examples at Section 2.2.2 ("physiology.Using"), Section 2.3.2 (duplicate lead-in sentences, list restart), Section 2.4 (comma followed by capitalised "This"), Section 3.3 ("improvement.." double full stop), and Figure 1 labels. Several phrases read as translated or unedited: "verification" is used where "validation" or "evaluation" is meant, "domestic and foreign scholars" is uncommon in international scientific writing, "ethnic medicine" may be better rendered as "ethnomedicine" or "traditional medicine of ethnic groups", and "Western medicine" and "biomedicine" are mixed. Language editing is recommended.

Concluding Remarks

The author addresses an important question, namely how evidence from different medical systems can be brought into a common, computable form without discarding what makes each system distinctive. The modular separation of track-specific components from a shared core is a sound starting point, and an outcome-level physiological anchor is a practical way to begin testing the idea. With clearer data provenance, a better-defined alignment metric, an explicit role for the HRN, ablations against stronger baselines and leakage-free validation, the work could become a useful methodological contribution.

Competing interests

The author declares that they have no competing interests.

Use of Artificial Intelligence (AI)

The author declares that they did not use generative AI to come up with new ideas for their review.

Puedes escribir un comentario en esta PREreview de A Multi-Track Cognition Framework for Global Integrative Medicine: Breaking Paradigm Incommensurability Through System-Level Mapping Across Medical Systems.

Antes de comenzar

We will ask you to log in with your ORCID iD. If you don’t have an iD, you can create one.

What is an ORCID iD?

An ORCID iD is a unique identifier that distinguishes you from everyone with the same or similar name.

Comenzar ahora