Mingzheng — Reproducibility Data Package

Zenodo v2 v2 (2026-05-25): Added a public-cohort transfer stress test on TCM-SD (Supplementary Part I). This release adds a subdirectory, tcmsd_external_validation/ (approximately 270 MB), containing reproducibility artifacts for the TCM-SD external-transfer stress test reported in Supplementary Part I: Qwen3-30B-A3B-Instruct-2507 16-slot structured extraction outputs for 5,316 TCM-SD test records, with an extraction success rate of 99.55%. BGE-m3 dense embeddings of the 16-slot sentences, with shape (5316, 16, 1024). Predictions from 15 frozen LOSO ensemble models and proxy labels for {phlegm-dampness, blood stasis, yin deficiency}. Raw logits for each ensemble member, with shape (15, 5316, 3). Five evaluation reports: A, Mingzheng zero-shot paper-locked thresholds; B, A plus half-retained threshold calibration; C, BGE-slot linear probe with 5-fold cross-validation; D, BGE-only zero-shot with a probe trained on the CRSD source cohort; and E, TF-IDF + SGD self-baseline reference. Reports include two-sided paired-bootstrap p values. All 10 pipeline scripts: L2 LLM extraction, BGE encoding, frozen Mingzheng forward propagation, the monitoring program, and five evaluation scripts. A subdirectory README documenting the workflow, results, per-file licenses, and reproduction commands. Main result (Group A vs Group D; n = 5,316; both true zero-shot, with no TCM-SD labels used): Macro-AUC = +0.020 (95% CI +0.002 to +0.038), two-sided paired-bootstrap p = 0.025, driven mainly by blood stasis (Delta AUC = +0.052, p < 0.001). This advantage did not extend to a statistically detectable macro-F1 difference (Delta F1 = +0.007, p = 0.250), and was not observed for yin deficiency. We position this analysis as a public-cohort transfer stress test rather than the primary external validation, because its operating-point performance is limited by label-transfer effects, especially the approximately 14.8-fold lower prevalence of yin deficiency relative to CRSD. The clinical-grade external-validation evidence for Mingzheng remains the prospective multicenter temporal cohort study (n = 47, collected from n = 105). License for the new subdirectory: Scripts are released under the MIT License, consistent with the Mingzheng GitHub repository. TCM-SD-derived outputs (tcmsd_l2_output.json, tcmsd_bge_4zhen.npz, tcmsd_predictions.json, tcmsd_labels.json, tcmsd_ensemble_logits.npz, tcmsd_test_l2_input.jsonl, tcmsd_eval_v*.json, and *.md) are released under CC BY-NC-SA 4.0, inheriting the ShareAlike requirement of the upstream TCM-SD dataset (Ren et al., CCL 2022; ModelScope OmniData/TCM-SD). Related materials: The corresponding code updates have been committed to the private Mingzheng GitHub repository under external_validation_tcmsd/ (commit 0046d59; https://github.com/Jayden-XL/mingzheng). For now, this repository is accessible only to the editor and editor-invited reviewers. If you are accessing this record during peer review, please contact us through the editor. Mingzheng - Reproducibility Data Package This Zenodo record provides the reproducibility package for the Mingzheng study: a taxonomy-based multimodal artificial intelligence system for interpretable Traditional Chinese Medicine syndrome differentiation in cancer patients with comorbid sleep disorders. This package is independent of manuscript version. It supports reproduction of the main model benchmarks, leave-one-site-out (LOSO) ensemble inference, temporal validation, reader-study model outputs, zero-shot LLM comparisons, and supplementary subgroup analyses reported across manuscript versions. Related paper: Zheng Xiao, Xie Yong, Luo Song, et al. A taxonomy-based multimodal artificial intelligence system for interpretable syndrome differential diagnosis in cancer patients with comorbid sleep disorders. Manuscript, 2026. Zenodo DOI: https://doi.org/10.5281/zenodo.20111083 Code repository: https://github.com/Jayden-XL/mingzheng For now, this repository is accessible only to the editor and editor-invited reviewers. If you are accessing this record during peer review, please contact us through the editor. Contents checkpoints/ - Trained LOSO Ensemble Models, approximately 7.1 GB Trained model checkpoints for the fully deployable version of Mingzheng. The released version contains 15 checkpoints: checkpoints/ |-- loso_seed42/fold{1-5}_best_ckpt.pth |-- loso_seed43/fold{1-5}_best_ckpt.pth `-- loso_seed44/fold{1-5}_best_ckpt.pth This corresponds to 3 random seeds x 5 leave-one-site-out cross-validation folds. For site-matched LOSO inference, a patient from site s is scored only by the three checkpoints whose held-out site is s, one checkpoint per seed. The three predicted probability vectors are then averaged. This scheme prevents the patient's own clinical site from contributing to the training data and is used for deployment-style cross-site inference, temporal validation, and reader-study model outputs. The same checkpoint family can also be used to compute pooled internal LOSO ensemble metrics across the five development sites. Alternative ensemble summaries are provided only as sensitivity analyses where indicated in the analysis files. fold_splits/ - Data Partition Metadata, approximately 13 KB LOSO fold assignments and metadata: fold_splits/ `-- loso_seed{42,43,44}/loso{1-5}/split_info.json Each split_info.json contains: holdout_site test_patient_ids n_train n_val n_test Patient identifiers are de-identified study IDs. analysis/ - Aggregated Results and Manuscript Mapping, approximately 70 KB Aggregated analysis outputs used to reproduce model rankings, ablations, subgroup analyses, and manuscript display items. Key files include: ranking.csv - model ranking table with paper display names and experiment identifiers experiment_map.csv - mapping from display name to experiment name, model file, training script, base experiment, key CLI flags, seeds, and notes per_syndrome.csv - per-syndrome F1/AUC summaries ablation_v3_expert.csv - expert-annotation ablation results ablation_v4_llm.csv - LLM-component ablation results ablation_v4_bge_expert.csv - BGE-embedding ablation results fold_diagnostic.csv - per-fold diagnostic metrics *_cv_summary.json - per-seed or aggregated cross-validation summaries with fold-level F1/AUC breakdowns Because figure and table numbering may differ across journal submissions, use experiment_map.csv and the GitHub repository README to map these files to the corresponding manuscript display items. llm_responses/ - Zero-Shot LLM Outputs, Numeric-Only, approximately 61 KB Per-patient JSON-Lines outputs from the zero-shot LLM baselines evaluated in the study. Free-text fields have been removed before public release. Each record contains only the numeric and administrative fields required by the downstream reproduction scripts, such as: patient_id hashed name field where applicable cancer type parsed per-syndrome confidence scores parsed overall difficulty error status number of attempts Removed fields include raw model responses, supporting/opposing evidence, diagnostic reasoning text, and free-text difficulty explanations, because they may quote patient inquiry-text fragments. Those restricted fields remain under IRB-controlled access and may be requested from the corresponding author subject to institutional approval. llm_responses/ |-- external_n47/ # Temporal validation cohort | |-- deepseek_chat_external.jsonl # DeepSeek-V3.2 | |-- deepseek_reasoner_external.jsonl # DeepSeek-R1 | |-- deepseek_v4pro_external.jsonl # DeepSeek-V4-Pro | `-- qwen36_plus_external.jsonl # Qwen3.6-Plus `-- internal_n478/ # Internal five-site cohort |-- h8_lite_deepseek_chat.jsonl # DeepSeek-V3.2 |-- h8_lite_deepseek_reasoner.jsonl # DeepSeek-R1 |-- h8_lite_deepseek_v4pro.jsonl # DeepSeek-V4-Pro |-- h8_lite_qwen3_30b_zeroshot.jsonl # Qwen3-30B-A3B-Instruct |-- h8_lite_qwen_outputs.jsonl # Qwen3.6-Plus |-- h8_lite_huatuo_o1_8b.jsonl # HuatuoGPT-o1-8B |-- h8_lite_wingpt2.jsonl # WiNGPT2-Llama-3-8B |-- h8_lite_llama31_8b.jsonl # Llama-3.1-8B-Instruct `-- h8_lite_openbio_8b.jsonl # OpenBioLLM-8B These files support reproduction of the zero-shot LLM comparisons, temporal-validation LLM baselines, calibration analyses, and related supplementary analyses. Out-of-Fold Predictions Out-of-fold predicted probabilities for all benchmarked models are available from the corresponding author upon reasonable request, subject to institutional data-sharing approval where required. How to Use with the Code Repository After cloning the code repository and downloading the ZIP archives from this Zenodo record, arrange the files as expected by the reproduction scripts. # 1. Checkpoints for site-matched LOSO inference unzip mingzheng_checkpoints.zip mkdir -p mingzheng/results/cv_v2 for s in 42 43 44; do EXP=v4_full_bge_no_pulse_raw_with_tongue_llm_expert_loso_seed${s} for f in 1 2 3 4 5; do mkdir -p mingzheng/results/cv_v2/${EXP}/loso${f} cp checkpoints/loso_seed${s}/fold${f}_best_ckpt.pth \ mingzheng/results/cv_v2/${EXP}/loso${f}/best_ckpt.pth done done # 2. Zero-shot LLM outputs unzip mingzheng_llm_responses.zip mkdir -p mingzheng/data/external_100/llm_outputs cp llm_responses/external_n47/*.jsonl mingzheng/data/external_100/llm_outputs/ cp llm_responses/internal_n478/*.jsonl mingzheng/data/ # 3. Fold splits and aggregated analysis files unzip mingzheng_fold_splits.zip unzip mingzheng_analysis.zip Then follow the Quick Start and reproduction instructions in the GitHub repository. Reader-Study Protocol Note The reader-study protocol and the revision history for group-order assignment are documented in reader_study/PROTOCOL.md v2.1 in the code repository. The protocol file records a post hoc audit of group-order assignment and explicitly states that the implemented reader-study group order was selected or determined by the readers themselves, rather than being randomly assigned by the investigators or fixed as an unaided-to-aided sequence for all readers. Data Not Included The following data are not publicly released because they may contain identifiable clinical information or fall within restricted institutional approval: raw tongue images raw pulse-waveform recordings inquiry-text transcripts patient demographic information hospital identifiers unrestricted clinical records complete expert-annotation tables when linkage risk is present To request access to restricted data, please contact the corresponding author. Access requires institutional approval, ethics clearance, and a data-sharing agreement. Contact: [e-mail] Ethics The study was approved by the relevant ethics committees of the participating clinical institutions: The First Affiliated Hospital of Zhejiang Chinese Medical University / Zhejiang Provincial Hospital of Chinese Medicine: 2024-KLS-652-01 The Second Affiliated Hospital of Zhejiang Chinese Medical University / Zhejiang Xinhua Hospital: 2025-102-01 Zhejiang Cancer Hospital: 2025-1089 (IIT) The study was registered with the International Traditional Medicine Clinical Trial Registry: ITMCTR2025000994. All participants provided written informed consent. Reporting Framework The related manuscript was prepared in accordance with the TRIPOD+AI 2024 reporting principles for prediction models that include AI or machine-learning components. License Code: MIT License. Released aggregated analysis files, metadata, numeric-only LLM outputs, and model checkpoints: CC BY 4.0, unless otherwise specified in individual files. Raw clinical data are not publicly released and are not covered by this public data license.