Clinical studies increasingly report that disease risk does not rise steadily with a biomarker. Instead it stays flat and then jumps at a "tipping point", or it climbs and then levels off at a "saturation point". Cardiovascular risk in metabolic syndrome is flat below a neutrophil-percentage-to-albumin ratio of 14.21 and rises above it. Diabetes risk climbs with the liver enzymes ALT and GGT up to about 50 IU/L and then plateaus. These cutoffs are found one biomarker at a time, using smooth curve fits such as restricted cubic splines. Real patients, though, move through a space of many correlated markers, and that space is mostly empty: many combinations of values never occur in any person. When researchers fit multi-marker risk surfaces to catch interacting thresholds, the fitted surface is extrapolated into exactly those empty regions, so the thresholds read off it can be artifacts.
This proposal borrows one idea from generative modeling. There, "geodesic interpolation" replaces straight-line paths between two latent points with shortest paths that stay where data actually live. We propose to estimate tipping points and saturation points along such population-supported paths through multi-marker biomarker space, rather than along coordinate axes or across unsupported regions. The same path machinery then yields a second, clinically useful object: for a high-risk patient, a route through realistic patient states that minimizes the total risk accumulated on the way to a low-risk region. Because the route is found by graph search over observed patients instead of by following risk gradients, it still works for patients sitting on a saturation plateau, where the risk gradient is zero and gradient-based methods stall.
Core Claim
The claim is that multivariate risk thresholds should be defined only along paths the population actually supports. A density-respecting shortest-path graph over real patients gives three things. First, it gives threshold estimates that hold up when transferred to another cohort better than estimates from multivariable spline surfaces. Second, it gives an honest picture of how one marker's tipping point shifts with another marker; for example, whether the NPAR threshold moves with the triglyceride-glucose index. Third, it gives recourse paths that avoid both empty regions and transient high-risk zones.
The method is deliberately simple. We build a nearest-neighbor graph over standardized biomarker vectors $x_i$ of real participants, and assign a path $\gamma = (x_{k_0}, \dots, x_{k_m})$ the cost
$$C(\gamma) = \sum_{j} \|x_{k_{j+1}} - x_{k_j}\|^{p}\,\big(1 + \lambda\, \hat r(x_{k_j}, x_{k_{j+1}})\big),$$
where $p > 1$ is the Fermat exponent that makes long jumps through sparse regions expensive, $\hat r$ is an estimated outcome risk averaged over the edge, and $\lambda \ge 0$ trades off path length against accumulated risk.
When $\lambda = 0$, this reduces to the density-weighted Fermat distance of Weighted Geodesic Distance Following Fermat's Principle. That paper penalizes large steps so that paths follow dense "highways", and it found $p$ between 1.7 and 3.2 to work well. When $\lambda = 0$, the graph is used purely to define where thresholds are read. When $\lambda > 0$, it is used for recourse.
The cost depends on position, not on direction. This handles biphasic markers such as AST, which is protective below 35 IU/L and harmful above it according to Association between liver enzymes and type 2 diabetes: a real-world study. Our view is that this sidesteps the asymmetric-metric problem rather than solving it. A path toward the risky side of a U-shape pays for the risk it enters, whichever direction it travels.
Relation to Prior Work
The motivating evidence comes from the clinical threshold literature. The NHANES analysis in Association between neutrophil percentage to albumin ratio and cardiovascular disease in the metabolic syndrome population reports the 14.21 inflection, with an odds ratio of 1.12 (95% CI 1.10–1.15) per unit above it. Association between the triglyceride-glucose index and 1-year major adverse cardiovascular events in patients with coronary heart disease and hypertension reports a triglyceride-glucose (TyG) tipping point near 8.85. The liver-enzyme study above replicates the ALT/GGT saturation across a Chinese hospital cohort, UK Biobank and NHANES. All of these analyses are univariate with covariate adjustment, and all report a single cutoff. We keep their outcomes and cohorts, but ask whether the cutoff is a property of one marker or of a region of the joint space.
The method comes from manifold-constrained interpolation. Latent Space Oddity: on the Curvature of Deep Generative Models and On Linear Interpolation in the Latent Space of Deep Generative Models show that straight paths cross low-density "dead zones". The second paper reports a 10.20% expected worst-case length improvement for geodesics on MNIST. Fast Approximate Geodesics for Deep Generative Models shows that graph search over encoded training points finds near-geodesics in about 0.09 s, against 200–400 s for continuous optimization, and needs no second derivatives. We adopt this graph approach because it makes saturation plateaus harmless. We test whether a learned latent geometry adds anything, using the encoder-variance metric of A Geometric Perspective on Variational Autoencoders as the latent option. With roughly ten biomarkers it may well add nothing, and we treat that as an acceptable outcome.
The closest bridging work constrains our design. On Data Manifolds Entailed by Structural Causal Models finds counterfactuals by solving a geodesic boundary-value problem with a density-scaled metric, on Adult and COMPAS. Our recourse component is a cousin of that work. It differs in three ways: it needs no known causal graph, it minimizes accumulated risk rather than distance to a classifier boundary, and it uses discrete search rather than a smooth boundary-value solve. Manifold-Aligned Counterfactual Explanations for Neural Networks enforces data support with Local Outlier Factor constraints inside a mixed-integer program, reaching 0% outliers on FICO and German Credit. It is restricted to ReLU classifiers and is a natural baseline for support. Learning spatiotemporal piecewise-geodesic trajectories from longitudinal manifold-valued data models breaks in one-dimensional tumor-burden trajectories. Its breakpoints are events in time; ours are locations in multi-marker space. None of these three uses geodesics to estimate where risk changes shape, and that is the new part.
Research Program
The work has three steps.
The first step builds the graph and the risk model. On NHANES metabolic-syndrome participants, the marker panel will be neutrophil percentage, albumin, triglycerides, fasting glucose, ALT, AST, GGT, BMI and waist. We fit $\hat r$ with a flexible outcome model, either gradient-boosted trees or a tensor-product generalized additive model, under cross-fitting so that a person's own outcome never sets the risk on their own edges.
The second step estimates thresholds. For a chosen marker, say NPAR, we take shortest paths from a low-risk reference set to each patient. Along each path we record risk as a function of arc length and NPAR value, and we locate the breakpoint with the same segmented-spline search the clinical papers use. Paths are then grouped by position on a second marker, such as TyG tertile, which gives a threshold that varies across the joint space. We report a breakpoint only where at least a fixed number of real patients lie within the local neighborhood.
The third step computes recourse. We run shortest-path search with $\lambda > 0$ from each high-risk patient to the nearest region whose risk falls below the published univariate cutoff. For each route we report its length, its accumulated and maximum risk, and how much of it passes through low-density regions.
Two quantities are tuned: $p$ and $\lambda$. We choose $p$ by held-out likelihood of neighbor distances, which partly answers the parameter-selection gap the Fermat paper leaves open. We sweep $\lambda$ and report the whole trade-off curve rather than one chosen value.
Baselines
The naive reference point is the published univariate threshold applied directly. For NPAR this is 14.21 with an odds ratio of 1.12 above it, from the NHANES study. For ALT and GGT it is about 50 IU/L, from the three-cohort liver study. For TyG it is 8.85; adding TyG raised the area under the ROC curve (AUC) from 0.767 to 0.916 in the hospital cohort. We reproduce each with restricted cubic splines on the same data split, so any multivariate method is compared against a re-fit number rather than a quoted one.
The strongest competitor for threshold estimation is a multivariable tensor-product generalized additive model with smoothing chosen by restricted maximum likelihood. Its fitted surface is the standard way to read interacting thresholds. It gets the same covariates, the same cross-fitting folds, and a tuning budget matched to ours on the number of smoothing configurations tried. To beat it, our estimates must do two things in the semi-synthetic test below: recover planted thresholds with lower error, and report fewer spurious breakpoints in low-support regions. On real data they must also transfer better from NHANES to UK Biobank. We found no reported result on Lacuna for multivariate threshold surfaces on these cohorts.
For recourse, the strongest existing method is the support-constrained mixed-integer approach of the manifold-aligned counterfactual paper, run on a ReLU network trained on the same data. Its published figures, 0% outliers on FICO and German Credit and about a 2% optimality gap on Adult, set the bar for support. We also run gradient-based counterfactual search in the style of Wachter, as described in the structural-causal-model paper, on the same risk model. We found no reported result on Lacuna for geodesic recourse on clinical biomarker data.
Two ablations isolate the key ingredients. Setting $\lambda = 0$ removes risk from the path cost and shows whether risk-weighting matters. Replacing the Fermat exponent with $p = 1$ (an unweighted neighbor graph) shows whether the density penalty matters. A third ablation builds the graph in a VAE latent space instead of raw standardized markers.
Why not just…?
Why not just fit a multivariable spline surface with interaction terms and read thresholds off it? This is the real threat. If the surface is smooth enough and well regularized, it may never wander into empty regions where it matters. The semi-synthetic test settles this directly. We plant thresholds on real NHANES covariates, including some placed in low-support regions, and count the false breakpoints each method reports there. If the spline surface matches us on both recovery error and false-breakpoint rate, the threshold claim fails, and the contribution reduces to the recourse component.
Why not just add a density penalty to gradient-based counterfactual search? A density penalty keeps endpoints realistic but does not help where the risk gradient is near zero. For a patient with ALT at 80 IU/L, deep on the plateau reported by the liver-enzyme study, the gradient gives no useful direction. We will measure the fraction of plateau-starting patients for whom each method reaches the target region. If density-penalized gradient search succeeds as often, the saturation argument is wrong.
Why not just use a VAE pullback or encoder-variance metric? With about ten markers, a learned latent space may add distortion without adding information. A prior-based approximate latent Riemannian metric reports that pullback metrics failed to converge beyond five dimensions. We include the latent-graph variant as an ablation and expect the raw-space graph to be competitive. Anisotropy arguments from image models do not obviously carry over to tabular biomarkers.
Why not just use Fermat distance alone? That is our $\lambda = 0$ setting. It is enough for threshold estimation, but it cannot prefer routes that avoid transient risk peaks. The $\lambda$ sweep shows whether the risk term changes maximum on-path risk at a fixed length.
Proposed Experiments
The first experiment is semi-synthetic. We keep the real NHANES biomarker vectors and simulate outcomes from planted risk functions of three kinds: an NPAR tipping point whose location shifts with TyG, an ALT saturation at 50 IU/L, and a U-shape in AST. The metrics are threshold localization error, false-breakpoint rate in regions below a support cutoff, and coverage of bootstrap intervals for the breakpoint.
The second experiment tests transfer. We estimate ALT/GGT saturation and AST sign-flip regions on NHANES and check them on UK Biobank. The two cohorts already agree on the univariate ~50 IU/L and 35 IU/L cutoffs, which makes a clean test: a multivariate method that breaks that agreement is suspect, and one that adds a reproducible interaction is informative. We report transfer log-likelihood and the agreement of the threshold surfaces.
The third experiment evaluates recourse on NHANES high-NPAR and high-ALT patients. The metrics are success rate, success rate among plateau starters, path length, accumulated risk, maximum on-path risk, and Local Outlier Factor outlier rate at waypoints. We also report wall-clock time; the cited graph-search runtime of about 0.09 s gives an expected order of magnitude.
The fourth experiment is a longitudinal sanity check on CLHLS, where repeated BMI and waist-index measurements exist. We ask whether participants whose observed movement followed low-risk graph routes had lower subsequent cognitive impairment than those who crossed predicted high-risk zones. CLHLS statistics are truncated in our file, so no prior figure exists for comparison.
Two results would count against the proposal. If the tensor spline matches us in the semi-synthetic test, the threshold claim is refuted. If gradient search with a density penalty matches us on plateau starters, the recourse claim is refuted. Either outcome would be reported as such.
References
- Weighted Geodesic Distance Following Fermat's Principle
- Association between liver enzymes and type 2 diabetes: a real-world study
- Association between neutrophil percentage to albumin ratio and cardiovascular disease in the metabolic syndrome population
- Association between the triglyceride-glucose index and 1-year major adverse cardiovascular events in patients with coronary heart disease and hypertension
- Latent Space Oddity: on the Curvature of Deep Generative Models
- On Linear Interpolation in the Latent Space of Deep Generative Models
- Fast Approximate Geodesics for Deep Generative Models
- A Geometric Perspective on Variational Autoencoders
- On Data Manifolds Entailed by Structural Causal Models
- Manifold-Aligned Counterfactual Explanations for Neural Networks
- Learning spatiotemporal piecewise-geodesic trajectories from longitudinal manifold-valued data
- A prior-based approximate latent Riemannian metric