MeridianFM: A Causal and Safe Foundation Model for Personalizing Traditional Chinese Medicine Therapies

Shuhua Yang; Yunhui Hu; Jianbo Li; Xiangying Xu

doi:10.63313/hmt.9012

Authors

Shuhua Yang The First People's Hospital of Yunnan Province, KunMing 650032, China Author
Yunhui Hu The First People's Hospital of Yunnan Province, KunMing 650032, China Author
Jianbo Li Yongren County People's Hospital Department of Traditional Chinese Medicine, YongRen 651400, China Author
Xiangying Xu The First People's Hospital of Yunnan Province, KunMing 650032, China Author

DOI:

https://doi.org/10.63313/hmt.9012

Keywords:

Foundation Model, Causal Inference, Reinforcement learning, Traditional Chinese Medicine, Personalized Medicine, Multimodal Learning

Abstract

Nonpharmacological therapies like acupuncture, gua sha, and cupping are gaining prominence in integrative medicine, yet their application lacks data-driven personalization and rigorous safety assurances. This paper introduces MeridianFM, a causal and safe multimodal foundation model framework designed to generate personalized prescriptions for these therapies. MeridianFM integrates four key innovations: (1) a meridian-aware graph neural network that encodes the topological and semantic properties of acupoints; (2) a self-supervised multimodal architecture that fuses physiological time series, thermal imagery, electronic health records, and patient-reported outcomes; (3) a causal inference layer employing doubly robust estimation and front-door adjustment to mitigate confounding in observational data and estimate individualized treatment effects; and (4) a constrained reinforcement learning policy optimized with risk-sensitive objectives (Conditional Value-at-Risk) and a mixed-integer programming post-processor to enforce clinical safety and feasibility constraints. To ensure reproducibility, we provide a comprehensive evaluation suite based on synthetic and semi-synthetic data, including all source code. Our experiments demonstrate that MeridianFM surpasses baseline models in optimizing treatment policies, accurately estimating causal effects, and adhering to safety constraints. While this study establishes methodological feasibility and superior performance in simulated environments, it also lays the groundwork for future clinical validation, representing a significant step toward AI-driven precision in traditional medicine.

References

[1] A. J. Vickers and K. Linde, “Acupuncture for chronic pain—update of an individual patient data meta-analysis,” The Journal of Pain, vol. 19, no. 5, pp. 455–474, 2018.

[2] H. Cao, X. Li, and J. Liu, “An updated review of the efficacy of cupping therapy,” PLoS One, vol. 7, no. 2, p. e31793, 2012.

[3] A. Nielsen, Gua Sha: A Traditional Technique for Modern Practice, 2nd ed. Elsevier, 2012.

[4] H. M. Langevin and J. A. Yandow, “Relationship of acupuncture points and meridians to connective tissue planes,” The Anatomical Record, vol. 269, no. 6, pp. 257–265, 2002.

[5] H. M. Langevin, D. L. Churchill, and M. J. Cipolla, “Mechanical signaling through connective tissue: a mechanism for the therapeutic effect of acupuncture,” FASEB Journal, vol. 15, no. 12, pp. 2275–2282, 2001.

[6] J.-S. Han, “Acupuncture analgesia: areas of consensus and controversy,” Pain, vol. 152, no. 3 Suppl, pp. S41–S48, 2011.

[7] E. J. Topol, “High-performance medicine: the convergence of human and artificial intelli-gence,” Nature Medicine, vol. 25, pp. 44–56, 2019.

[8] B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communica-tion-efficient learning of deep networks from decentralized data,” in Proc. AISTATS, 2017.

[9] C. Dwork and A. Roth, “The algorithmic foundations of differential privacy,” Foundations and Trends in Theoretical Computer Science, vol. 9, no. 3–4, pp. 211–407, 2014.

[10] S. Levine, A. Kumar, G. Tucker, and J. Fu, “Offline reinforcement learning: tutorial, review, and perspectives on open problems,” arXiv:2005.01643, 2020.

[11] A. Kumar, A. Zhou, G. Tucker, and S. Levine, “Conservative Q-Learning for offline rein-forcement learning,” in Proc. NeurIPS, 2020.

[12] N. Jiang and L. Li, “Doubly robust off-policy evaluation for reinforcement learning,” in Proc. ICML, 2016.

[13] J. Pearl, Causality: Models, Reasoning, and Inference, 2nd ed. Cambridge University Press, 2009.

[14] U. Shalit, F. D. Johansson, and D. Sontag, “Estimating individual treatment effect: generali-zation bounds and algorithms,” in Proc. ICML, 2017.

[15] C. Louizos, U. Shalit, J. M. Mooij, D. Sontag, R. Zemel, and M. Welling, “Causal effect infer-ence with deep latent-variable models,” in Proc. NeurIPS, 2017.

[16] R. F. Barber, E. J. Candès, A. Ramdas, and R. J. Tibshirani, “Predictive inference with the jackknife+,” Annals of Statistics, vol. 49, no. 1, pp. 486–507, 2021.

[17] J. Lei, M. G’Sell, A. Rinaldo, R. J. Tibshirani, and L. Wasserman, “Distribution-free predictive inference for regression,” Journal of the American Statistical Association, vol. 113, no. 523, pp. 1094–1111, 2018.

[18] A. Vaswani et al., “Attention is all you need,” in Proc. NeurIPS, 2017.

[19] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proc. CVPR, 2016.

[20] F. Shaffer and J. P. Ginsberg, “An overview of heart rate variability metrics and norms,” Frontiers in Public Health, vol. 5, p. 258, 2017.