Computer Science > Machine Learning
Title:Deconstructing Intraocular Pressure: A Non-Invasive Multi-Stage Probabilistic Inverse Framework
Abstract:We propose the mechanistically hierarchical inverse problem (MHIP) paradigm for biological transport systems, where a macroscopic observation emerges from the multiplicative entanglement of intrinsic micro-scale physics and latent macro-scale geometry. Without in vivo ground truth, solving MHIPs from clinical data is severely underdetermined. Standard statistical sampling degenerates along non-identifiable manifolds, while monolithic machine learning baselines collapse to an uninterpretable conditional mean. We propose a two-stage physics-guided decomposition that turns a globally non-identifiable inverse mapping into conditionally identifiable sub-problems. The first stage is a constrained inverse surrogate that isolates the intrinsic parameter under a fixed reference geometry. The second stage performs empirical structural calibration against population-level clinical data. We also introduce Physics-Calibrated Data Scaling (PCDS), a multi-fidelity strategy that builds large training sets from few high-fidelity simulations without patient-specific 3D modeling. In a glaucoma case study, the framework deconstructs a single intraocular pressure (IOP) measurement to estimate trabecular permeability ($\hat{K}_{TM}$) and conventional outflow facility ($\hat{C}_{trab}$). In retrospective validation against published in vivo data, the non-invasive $\hat{C}_{trab}$ estimates showed low bias against measured outflow facility. In an ablation over 27 clinical cohorts, naive monolithic baselines reach 37.0% diagnostic accuracy. Adding the isolated $\hat{K}_{TM}$ estimate raises this to 85.2%. The full two-stage framework reaches 88.9%. The framework also returns a probabilistic profile of the latent hydrodynamic state, offering a general computational blueprint for MHIPs.
| Comments: | v2: revised and extended version. 48 pages (23 main + 25 supplement), 4 figures |
|---|---|
| Subjects: | Machine Learning (cs.LG) ; Quantitative Methods (q-bio.QM); Applications (stat.AP); Methodology (stat.ME) |
| Cite as: | arXiv:2509.14167 [cs.LG] |
| (or arXiv:2509.14167v2 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2509.14167 arXiv-issued DOI via DataCite |
Submission history
Access Paper:
- HTML (experimental)
Additional Features
- Audio Summary
Current browse context:
References & Citations
- Google Scholar
- Semantic Scholar
BibTeX formatted citation
Bookmark
Bibliographic and Citation Tools
Code, Data and Media Associated with this Article
Demos
Recommenders and Search Tools
arXivLabs: experimental projects with community collaborators
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.