Auditing Clinical Explanations Against Injectable-Gel Physics

When a model tells a clinician that a prediction was driven by feature A rather than feature B, almost nobody can check whether that is true. The stated reason for a sepsis alarm or a metastasis risk score cannot be tested by intervention, so explanation methods are judged by proxies such as agreement with known biomarkers or how fast the prediction collapses when inputs are deleted. We propose a clinically relevant setting where the reasons can be checked. The setting is the injection of a shear-thinning hydrogel (a gel that thins and flows under the shear of a needle, then re-solidifies once it comes to rest) into liver tissue. The question is whether the injected depot stays where it was placed or leaks away.

Every input that plausibly drives that outcome can be set by hand at the bench and changed one at a time. That includes how strong the gel is at rest, how quickly it rebuilds after the needle, how fast it is pushed, how much the surrounding tissue is flowing, and whether there is a crack for it to escape into. That makes it possible to measure the true effect of each input and score explanation methods against it. The note makes two claims. First, a small model built on two physically derived ratios, comparing the gel's strength to the tissue's pushing pressure and its rebuild time to the time it takes to be carried away, will predict leakage about as well as black-box models. Second, its explanations will match the measured effects of intervening on those inputs, while post-hoc explanations of image-based black boxes will credit visual properties that are only correlated with gel mechanics.

Why this setting, and what is missing

Injectable depots are a real clinical problem. Delgado et al. built a poloxamer hydrogel that is liquid at room temperature and gels in under 6 seconds at body temperature. It carries microbubbles for ultrasound and iodixanol for CT, and they tracked depots in swine liver (cross-section 1.7 ± 0.4 cm² on ultrasound versus 2.2 ± 0.4 cm² on CT). They note that intratumoral injections often fail by leaking along low-resistance paths. Their monitoring is still manual contouring, and they do no predictive modeling at all. Yu et al. supply a second material with very different kinetics: a hydrogen-bonded humic acid/PVP gel that forms in under 10 seconds and recovers about 80% of its mechanical properties only after 10 minutes. Together the two materials give us gels whose rebuild times differ by orders of magnitude.

Delgado et al. also report a detail that makes this testbed useful for interpretability. Adding microbubbles raises both ultrasound brightness and the gel's storage modulus and complex viscosity. A model that watches ultrasound video can therefore learn "bright depots stay put" without learning anything about mechanics. Its post-hoc explanation would be wrong, and only an intervention can reveal that. Here an intervention means matching gel stiffness by another route while keeping microbubbles fixed or absent.

On the interpretability side, the file contains three relevant patterns. Kim et al. showed that standard LIME on clinical sequence models credited patient gender for sepsis predictions simply because it sat at the first time step. Their positional explainer PaX corrects this, with up to 0.03 gain in insertion AUC on MC-MED. That paper motivates our concern: post-hoc attributions can be confidently wrong, and the authors could only show it indirectly. Parikh et al. supply a method we reuse in spirit. They built patient states from pharmacokinetic differential equations and then matched patients on them, with an estimated 23.6 percentage-point reduction in poor outcomes. Their mechanism is well-mixed and has no spatial component; ours replaces drug concentrations with rheological ratios. Maronge et al. (RAITR, a model that starts linear and admits nonlinear terms only when they earn their place) and Rueckert and Schnabel (physics layers that keep imaging networks honest) constrain the design. The first warns that flexible nonlinear terms get over-selected. The second shows that physics priors help when the physics is right.

The nearest work joining flow and learning is WSSflow. It recovers near-wall shear stress from particle images and assumes a fluid of constant viscosity. It is an estimation tool, not a test of explanations, and it does not handle shear-thinning materials.

Core Claim

Leakage of an injected shear-thinning depot is governed mainly by two dimensionless ratios. The first is a yield ratio $Y = \tau_y / p_t$. Here $\tau_y$ is the gel's yield stress, the stress it can resist before flowing, taken from a Herschel–Bulkley fit to rheometer flow curves. $p_t$ is the pressure the tissue exerts to push the gel away, set by imposed perfusion flow and measured with a pressure probe. The second is a recovery ratio $R = t_{rec} / t_{esc}$. Here $t_{rec}$ is the time the gel takes to regain 90% of its rest stiffness after high shear, measured in a step-strain test. $t_{esc}$ is the time the local flow would take to carry fluid one needle-tip radius away. We hypothesize

$$\Pr(\text{retained}) = \sigma\big(\beta_0 + \beta_Y \log Y - \beta_R \log R + \beta_F F\big),$$

where $\sigma$ is the logistic function, $F$ indicates a fissure in the target region, and the $\beta$ are fitted coefficients. The logistic-in-log form follows from treating retention as a threshold event: the gel holds if it resists the pushing stress before it is swept away.

The claim is falsifiable in three ways. If adding raw formulation variables to this model improves held-out prediction substantially, the ratios are not enough. If black-box video models clearly outpredict it, the physics is incomplete. If its coefficients disagree in sign or rank with the effects we measure by intervention, its explanations are not faithful either.

Research Program

We first build the testbed. One part is agarose tissue phantoms with controllable perfusion through embedded channels and optional cut fissures. The other is ex vivo bovine and swine liver, following Delgado et al.'s protocol. Formulations span poloxamer 407 at several concentrations, microbubble fractions from 0 to 0.5% v/v (staying under the roughly 1% level where Delgado et al. saw comet-tail artifacts), and humic acid/PVP gels at several ratios. Each formulation is characterized on a rheometer for flow curve, yield stress and recovery time.

Injections vary rate, perfusion and fissure presence in a crossed design. Each injection is recorded by ultrasound video and a syringe-pressure trace. The outcome is the fraction of injected volume inside the target region on CT at 10 minutes, thresholded at 80% for the binary label. We target roughly 400 injections.

The design deliberately includes "decoupling pairs" so we can measure the true effect of variables that are normally tangled together. One kind of pair matches yield stress with and without microbubbles, using a higher poloxamer concentration to make up the stiffness. Another matches yield stress across poloxamer and humic acid/PVP while recovery time differs. The measured retention changes across these pairs, and across the manipulated injection and tissue settings, serve as ground-truth effects.

We then score every model's explanations against these effects. The scores are rank correlation of attribution magnitudes with measured effect sizes, sign agreement, and a "spurious credit" score: the attribution given to microbubble fraction or image brightness in the matched pairs, where its true effect is zero by construction. For the pressure-trace models we also run PaX to test whether the late recovery phase gets real credit or whether credit piles up at the moment of needle entry, the analogue of primacy bias in this setting.

Baselines

The naive reference is the base-rate classifier, together with the rule of thumb clinicians effectively use: predict retention whenever the depot looks compact on ultrasound at the end of injection, scored from the same videos by a blinded reader. No reported result found on Lacuna for depot leakage prediction from any method, so these references are set empirically within the study.

The strongest black-box competitor is a video classifier trained on the ultrasound clip plus the pressure trace, explained with SHAP and with LIME. Image models dominate depot monitoring in practice. LiTS-style CT segmentation networks are the closest reported image systems (LiTS reports tumor Dice of 0.739 at MICCAI 2018), though they measure where tissue is rather than whether a gel will stay. A gradient-boosted tree on all raw tabular inputs (formulation concentrations, rheometer outputs, injection rate, perfusion, fissure flag) with SHAP is the strongest tabular competitor. RAITR fitted on the same raw inputs represents the best existing inherently interpretable method. Parikh-style matching on the two ratios serves as a case-based variant of our model.

All models use the same grouped cross-validation, with folds split by formulation batch and by tissue specimen so that no specimen appears in both training and test. Each gets the same hyperparameter search budget.

Two ablations isolate the key ingredient. In the first, the logistic model receives the same raw quantities that make up the ratios ($\tau_y$, $t_{rec}$, $p_t$ and flow speed, all logged) and may combine them freely. This tests whether the physical grouping matters or merely the inputs. In the second, the ratios are computed with shuffled pairings of numerator and denominator, keeping their distributions but breaking the physics.

"Beating" the black boxes means held-out AUROC within 0.03 of the best one, using PaX's reported gain as a scale for practically meaningful differences. It also means higher rank agreement with interventional effects and lower spurious credit on the matched pairs. The proposal wins on prediction only if it clearly beats the raw-input logistic ablation, especially when tested on held-out materials.

Why not just…?

Why not just fit a power-law model to rheometer data and use it as the predictor? Fitting flow curves describes the material but not the injection, because leakage depends on the material relative to the tissue. Delgado et al. attribute failures to leakage paths in the tissue, not to the gel alone. The material-only logistic model is part of our ablations, and if it matches the full model the tissue-ratio claim fails. This is a genuine threat, and the held-out-material test settles it.

Why not just train a strong video model and explain it with SHAP? It may well predict best, which is why it is the main competitor. But the microbubble coupling reported by Delgado et al. means its explanations have an easy wrong answer available. Kim et al. show that standard attributions do fall into such traps. The matched-pair spurious-credit score tests directly whether that happens here. If the video model's attributions pass that test, our second claim is wrong and we will report it.

Why not just use synthetic benchmarks with planted ground-truth features? Planted features are independent by construction and lack the measurement noise and coupled variables that make real clinical explanations fail. Our ground truth comes from physical intervention on variables that are naturally entangled, which is exactly the regime where synthetic checks say nothing.

Why not just use RAITR on raw inputs? RAITR is included. Its linear default may serve poorly here: retention plausibly has a sharp threshold, which a linear-first model may smooth away or capture only through splines that are hard to read. Comparing its held-out performance and its agreement with interventional effects against the two-ratio model answers this directly.

Decisive Checks

The decisive experiment trains every model on poloxamer injections and tests on humic acid/PVP injections. That material's 10-minute recovery puts it far outside the training range of $R$. A model that has learned the physics should transfer through the ratios, while models keyed to formulation or image appearance should not.

A second check uses the microbubble-matched pairs: any method that assigns non-negligible credit to microbubble fraction there is shown to be unfaithful. Third, we refit the logistic model on intervention-only subsets and confirm that the fitted $\beta_Y$ and $\beta_R$ keep their signs. If the ratio model loses on transfer, or its coefficients flip, the physics-grounded explanation claim is refuted. The testbed would still remain useful as a ground-truth benchmark for clinical explanation methods.