When engineers simulate an implant fixed in bone, they have to decide what happens at the boundary between the two materials. The usual choice is "perfect bonding": the bone and the metal move together exactly, and force passes across the boundary without loss. Everyone knows this is wrong for an implant that is loosening. In practice, though, the choice is made by convention. Nobody checks it against data. Meanwhile, hospitals hold records of thousands of implanted patients, and those records show who later needed revision surgery. This note proposes using those historical outcomes to decide which interface condition a mechanical model should use. Each candidate condition (fully bonded, partially compliant, or sliding with friction) becomes a competing hypothesis. Each hypothesis is scored by how well its simulated stresses and micromotions explain which patients in a retrospective cohort went on to revision. The idea in two sentences: build patient-specific mechanical models from routine postoperative computed tomography (CT) scans under several interface assumptions, and let the cohort's revision outcomes pick the assumption, and its stiffness, that best explains reality. The by-product is a continuous mechanical instability score. It may separate failing from stable implants where radiologists' yes/no loosening reads have too little statistical power.
The opportunity is clearest in reverse shoulder arthroplasty, a shoulder replacement in which the ball and socket are swapped. CT imaging findings in symptomatic patients with and without revision surgery after reverse shoulder arthroplasty compared 52 revised and 47 non-revised symptomatic patients and graded loosening as present or absent. Peg loosening (30% vs 16%, P=0.49), baseplate loosening (15% vs 6%, P=0.16) and superior screw loosening (18% vs 7%, P=0.10) all trended toward revision but missed significance. Only dislocation reached significance (19.2% vs 2.1%, P=0.007). The images held more information than the yes/no grades captured, and no mechanical model was used to extract it.
Core Claim
The choice of interface continuity condition is an empirical question, and a retrospective cohort with outcomes can answer it. We model the bone–baseplate boundary as a spring layer. The traction across it is $\mathbf{t} = K\,[\![\mathbf{u}]\!]$, where $\mathbf{t}$ is the force per unit area transmitted across the boundary, $[\![\mathbf{u}]\!]$ is the displacement jump between bone and implant, and $K$ is the interface stiffness. Letting $K\to\infty$ recovers perfect bonding, $K\to 0$ means full debonding, and a frictional-contact model is a third, separate family. For patient $i$ with clinical covariates $\mathbf{x}_i$, the model computes a mechanical index $D_i(K)$ under a standard load case (peak baseplate micromotion is the primary choice). We then fit
$$\Pr(\text{revision}_i) = \sigma\big(\beta_0 + \beta_1 D_i(K_i) + \boldsymbol{\gamma}^\top \mathbf{x}_i\big), \qquad \log K_i \sim \mathcal{N}(\mu_K, \tau^2),$$
where $\sigma$ is the logistic function and the patient-level stiffnesses are drawn from a population distribution with mean $\mu_K$ and spread $\tau$. The scientific claim is that the posterior over $\mu_K$ and $\tau$ sits far from the perfect-bonding limit. A second claim is that, scored by out-of-sample predictive fit, the compliant-interface family beats both the bonded and the frictional families. If perfect bonding wins, that is also a publishable answer. It would mean the convention behind most implant finite element studies is defensible at the resolution clinical outcomes can see.
How This Relates to Prior Work
The closest existing work is Exploration of Pedestrian Head Injuries—Collision Parameter Relationships through a Combination of Retrospective Analysis and Finite Element Method. It paired 61 historical collisions, analyzed by Fisher's exact test over binned impact speeds, with finite element simulations of a human body model. The two analyses run in parallel. The cohort never informs the model's interface assumptions, and the intracranial layers stay perfectly bonded. Our proposal differs because the cohort outcome is the likelihood that calibrates the interface parameter. The mechanics and the statistics form one model rather than two side-by-side reports.
Two papers supply the method. The Statistical Finite Element Method treats a finite element model as a probabilistic prior with an explicit term for mismatch between model and reality, then updates it with sensor data. We borrow that stance but update against clinical outcomes rather than measured displacements, which is a much weaker and noisier signal. That is why we pool across patients hierarchically. Meta Learning of Interface Conditions for Multi-Domain Physics-Informed Neural Networks already treats interface conditions as something to select rather than assume. It uses a bandit to choose which continuity constraints a neural PDE solver enforces, cutting solution error by up to 88.5% on Burgers' equation. Its selection criterion is numerical accuracy against a known PDE. Ours is agreement with real-world outcomes when the true interface physics is unknown. That paper also constrains our design. It shows that enforcing every continuity condition over-constrains the solver, which is one reason we keep the interface family small and the stiffness low-dimensional.
Several papers motivate the gap. The link-slab bridge analysis, A Method for Analyzing Transverse Stress in Link Slabs of Simply Supported Steel–Concrete Composite Bridges, explicitly lists neglected slip between concrete and girders as a limitation. The molar access-cavity study Effect of access cavities on the biomechanics of mandibular molars: a finite element analysis models restored teeth as perfectly bonded, homogeneous continua. Both show that perfect bonding is assumed even where authors suspect it fails. On the clinical side, Osteoporosis exacerbates perioperative complications in total knee arthroplasty: a 10-Year nationwide analysis found adjusted odds ratios of 1.26 for prosthetic loosening and 2.13 for periprosthetic fracture among 1.33 million knee replacements. It names "biomechanical failure of the bone-implant interface" as the mechanism but has no geometry, no bone density values, and only 4.28% coded osteoporosis. This tells us administrative databases cannot carry this question, and image-bearing cohorts are required.
Research Program
We would assemble a retrospective, ideally multicenter, cohort of reverse shoulder arthroplasty patients who had a postoperative CT before any revision decision. The outcome window and the covariate set follow the reverse shoulder CT study: age, sex, time since surgery, infection history and symptom severity. From each scan we segment the glenoid bone and register the implant geometry. Local bone stiffness is assigned from CT intensity, which avoids the homogeneous-tissue assumption criticized in the dental study. We then solve a standardized arm-abduction load case under each interface family across a grid of $K$ values. To make the hierarchical fit tractable, we fit a cheap surrogate of $D_i(K)$ per patient, a smooth curve over $\log K$. Fitting proceeds by Markov chain Monte Carlo on the surrogate. Families are compared by leave-one-out expected log predictive density, a cross-validated measure of how well each model predicts patients it was not fitted on. Time-to-revision after the CT is analyzed with a Cox model as a secondary endpoint.
One confound must be handled head-on. Surgeons may decide to revise partly because a radiologist reported loosening. Revision then partly encodes the very reads we compare against. We address this in three ways. The covariate model includes the radiologist reads, so the mechanical index must add information beyond them. The secondary endpoint is time to revision, which is less directly driven by a single report. And revisions done for infection serve as a negative control.
Baselines
The naive reference point is the analysis the field already runs, reproduced on our cohort. Radiologists' binary loosening and dislocation grades plus the clinical covariates go into multivariable logistic regression, exactly as in CT imaging findings in symptomatic patients with and without revision surgery after reverse shoulder arthroplasty. That paper reports group percentages and P-values but no discrimination metric. We found no reported result on Lacuna for the area under the ROC curve (AUROC) of revision prediction after reverse shoulder arthroplasty, so we will establish this number ourselves. The only AUROCs in the file are for a different task, readmission after hip or knee replacement, at 0.846 and 0.822 from clinical-text models in Predicting Unplanned Readmissions Following a Hip or Knee Arthroplasty: Retrospective Observational Study. They give a rough sense of attainable discrimination in arthroplasty outcomes but are not a target.
The strongest competitor is a penalized regression on continuous CT features without any mechanics. Its inputs are exactly those our finite element model consumes: radiolucent line width around peg and screws, peri-implant bone intensity, implant positioning angles, and deltoid volume. It gets the same patients, the same cross-validation folds, and a tuning budget equal to the number of mechanical-model configurations we explore. Beating it means a better held-out log predictive density and AUROC, with a bootstrap interval on the difference excluding zero. If it ties, the predictive claim fails and only the interface-inference claim stands.
The third baseline is standard practice in implant mechanics: the same patient-specific finite element model under perfect bonding, with the index entered into the identical regression. Two ablations isolate the key ingredient. One fixes $K$ at a single population value chosen by cross-validation. The other lets $K$ vary by patient. Together they separate "compliance matters" from "patient-specific compliance matters."
Why not just…?
Why not just regress on continuous imaging features and skip the mechanics? This is the real threat, and it may match us on prediction. The finite element model is a deterministic function of those features plus physics, so a flexible regressor could in principle learn the same mapping. With roughly a hundred to a few hundred patients, it is unlikely to learn a nonlinear stress-transfer function from scratch. The decisive comparison above settles this empirically. More importantly, a regression on features cannot answer the core question of which interface condition is physically consistent with outcomes. That inference is the contribution even if predictive accuracy ties.
Why not just get more patients from a national database? The knee replacement study shows the ceiling of that route. It has enormous numbers, but loosening exists there only as a billing code, osteoporosis is badly undercoded, and there are no images. More patients without geometry cannot inform an interface model.
Why not just train a deep network on the CT volumes? At the cohort sizes these studies reach (99 patients in the shoulder study), an image network would overfit and would say nothing about interface physics. We would only add one if a multicenter cohort reached the thousands, and then as an additional baseline.
Why not just keep perfect bonding and calibrate material properties instead? Bone stiffness and interface stiffness both change the computed micromotion, so one could absorb the other. We test this directly by fitting the bonded family with a free bone-stiffness scaling parameter. If it recovers the compliant family's predictive density, the interface claim is not identifiable from outcomes, and we would report that.
Proposed Experiments and Decisive Checks
First, a recovery test on synthetic cohorts. We plant a known population stiffness, simulate revisions from it with realistic noise and the observed roughly 50% revision rate, and check that the hierarchical fit recovers $\mu_K$ at the target cohort size. This tells us in advance whether a 100-patient cohort is enough or whether a multicenter one is required.
Second, the main comparison. The compliant-interface family should beat the bonded and frictional families in held-out predictive density. It should also beat the feature-regression baseline. The continuous index should separate revised from non-revised patients where binary peg loosening reached only P=0.49.
Third, a negative control. The mechanical index should carry no signal for infection-indicated revisions. If it does, it is proxying for general sickness rather than mechanics.
Fourth, face validity. Patient-level inferred $K$ should be lower where radiologists independently read loosening, and it should fall with time since surgery.
Failing the negative control, or losing to the feature regression while also failing to identify $K$ apart from bone stiffness, would mean outcome data cannot discipline interface assumptions at this scale. We would report that as the result.
References
- CT imaging findings in symptomatic patients with and without revision surgery after reverse shoulder arthroplasty
- Exploration of Pedestrian Head Injuries—Collision Parameter Relationships through a Combination of Retrospective Analysis and Finite Element Method
- The Statistical Finite Element Method
- Meta Learning of Interface Conditions for Multi-Domain Physics-Informed Neural Networks
- A Method for Analyzing Transverse Stress in Link Slabs of Simply Supported Steel–Concrete Composite Bridges
- Effect of access cavities on the biomechanics of mandibular molars: a finite element analysis
- Osteoporosis exacerbates perioperative complications in total knee arthroplasty: a 10-Year nationwide analysis
- Predicting Unplanned Readmissions Following a Hip or Knee Arthroplasty: Retrospective Observational Study