What we're working on
Most of this lab's work concerns a condition that takes away people's ability to walk comfortably and gets remarkably little attention for how common it is. Around that we have built the two kinds of evidence needed to study anything properly: what happens to very large numbers of real patients, and what happens mechanically inside a single hip.
Hip abductor disease
The gluteus medius and minimus sit on the side of the hip and hold the pelvis level every time you stand on one leg, which is to say with every step you take. When those tendons degenerate, walking becomes a limp, stairs become difficult, and sleep is usually the first thing to go, because the painful side is the side you lie on. It affects roughly one in six adults over 50, and about a quarter of women, most of whom are told they have bursitis, given a steroid injection, and sent away. The tendon rather than the bursa is usually the source: in one large series tendinosis accounted for about half of cases and bursitis for a fifth. Left long enough, muscle is gradually replaced by fat, which does not reverse, and by then the options narrow. Anti-inflammatory drugs are risky on older kidneys, so pain that never resolves drifts toward stronger medication. Some patients are offered a hip replacement that was never going to help, because the joint was not the thing that hurt. And even a sound repair may return only a fraction of the function the patient was expecting.
The work runs along several fronts. To ask who develops it, we followed a cohort of 385,868 women to test whether hormone therapy after menopause changes the risk and whether the timing of starting it matters. To ask how recovery should be judged, we catalogued the instruments the field actually uses and found 74 of them across 36 studies, with the strongest predictor of outcome turning out to be a grading scale borrowed from shoulder surgery that was never validated for the hip. In our own repair patients, that same grade, the amount of fat that has replaced muscle on preoperative MRI, tracks the gait-specific parts of recovery, walking without a limp and without a cane, while leaving the overall hip score unmoved, which is probably why earlier studies using composite scores contradicted one another. On treatment, we pooled 38 studies of confirmed gluteus medius and minimus tears into a single staged pathway, since previous reviews had each addressed only one piece of the care. A separate review consolidated nine studies on one under-recognised form of the injury, where the tendon pulls off its footprint beneath fibres that still look intact, which several groups had described independently under different names. The end of all this is a consensus process with radiologists, primary care physicians, and physical therapists, because a disease that cannot be named consistently cannot be counted, funded, or paid for.
Evidence at population scale
Most questions in orthopaedics never get a randomized trial. The conditions are too uncommon, the outcomes take a decade to appear, and a trial large enough to answer the question would cost more than anyone will pay, so surgeons fall back on experience and on studies of a few hundred patients from a single hospital. There is a second limitation that gets less attention. A trial reports an average. If a treatment helps one group of patients a great deal and does nothing for everyone else, the average comes out small and the treatment is written off. That is not a statistical curiosity. It means something that would have worked for a definable group of people is abandoned for all of them, and nobody ever finds out who they were.
Health records now cover tens of millions of patients, which is enough to reach questions a trial cannot. The difficulty is that this data was produced by care rather than by research, so it is full of traps: patients who look healthier because they were selected for treatment, follow-up that simply stops when someone changes insurer, and associations that reflect referral patterns rather than biology. A good part of the work is controlling those problems well enough that the answer means something. We are screening fifty medications at once for association with hip and knee osteoarthritis, comparing how long implants last in adults whose hip disease began with cerebral palsy or spina bifida against adults with ordinary arthritis, and linking record systems that have not previously been combined to test whether spinal fusion in ankylosing spondylitis is associated with gluteal tendon damage. Alongside that we have measured what hip surgery genuinely costs by timing each step of care rather than accepting the hospital's accounting. A further strand turns the same scrutiny on published work: journals now carry machine learning models constantly, very few are ever run on data from a hospital other than the one that built them, and we are beginning to test which of them survive that.
Biomechanics and simulation
A surgeon repairing a torn tendon decides how many anchors to use, where to place them, and whether to reinforce the repair with a graft or a synthetic patch. Those decisions are mechanical, and they are mostly made by feel, because there is no way to try them on the patient in front of you. Repairs of the hip abductors fail often enough to matter, and a failed repair is not a neutral event: it means a second operation on tissue that was already damaged, after a year of rehabilitation that returned the patient to roughly where they started. The same gap between judgement and measurement shows up in knee replacement. Hundreds of thousands are done in the United States every year, and around one in five of those patients is still dissatisfied afterwards, a figure that has barely moved in twenty years, usually blamed on a knee that was not balanced quite right.
We build computer models of individual patients from scans they have already had, so the geometry, the tear, and the state of the muscle belong to that person rather than to an average, and the model shows where load concentrates and which part of a construct is carrying it. Because a simulation is only as good as its assumptions, we test the same constructs in cadaveric tissue, which is harder than it sounds: the muscle tears out of the clamp before the repair itself does, so measuring the strength of a repair means first solving the problem of holding onto the muscle. The third approach is to measure mechanics during the operation. A robotic system plans every cut before the first incision and then records what was actually measured once the surgeon is inside the joint, and almost none of that record is ever examined afterwards. Treating the plan as a prediction and checking it against what the surgeon found shows that it underestimates the gap on the inner side of the knee, and that the error grows the more bowed the leg was to begin with, which a surgeon can correct for once they know it is there.
Where the work is headed
Predicting which abductor tendon repairs fail
Gluteus medius repairs fail at meaningful rates, and outcomes vary widely. Using imaging, intraoperative findings, and longitudinal outcomes from our repair cohort, the goal is a model that identifies repairs at high risk of failure and the factors that drive it.
Data: imaging, operative, outcomes · Methods: ML, survival analysisFrom intraoperative data to a structured operative record
Robotic platforms log detailed data for every case, but little of it reaches surgical registries because the translation is manual. The problem is an agentic pipeline that converts intraoperative and EHR data into structured, auditable documentation a surgeon would sign.
Data: robotic logs, EHR · Methods: LLM agents, FHIRSeparating cause from association in drug-osteoarthritis screens
A medication-wide screen surfaces many drugs associated with osteoarthritis. Association is not cause. The work is to take the strongest signals and test them with negative controls, target-trial emulation, and sensitivity analysis to see what holds.
Data: federated EHR · Methods: causal inference, pharmacoepidemiologyA fast surrogate for patient-specific finite element analysis
A full finite element model of a repair construct takes hours to build and solve. A learned surrogate that predicts the stress field from imaging in seconds, accurately enough to inform a decision, would make this usable in clinic.
Data: imaging, FE models · Methods: surrogate modeling, geometry processingCorrecting the robot's gap prediction across deformity
Robotic knee replacement underpredicts the medial gap as native coronal deformity increases. The problem is to characterize that error across the full deformity range and build a correction a surgeon can use intraoperatively.
Data: robotic arthroplasty · Methods: modeling, data scienceA reproducible classification for concealed abductor tears
Undersurface abductor tears sit beneath intact superficial fibers and are described in the literature under competing names. The work is an imaging-based, reproducible classification, ideally with automated segmentation, that the field can use consistently.
Data: MRI · Methods: segmentation, clinical research