Emory Hacks 2026 — Athlete recovery

Stop guessing whether the leg is ready.

ATHLYx turns a phone camera into a movement screen: three biomechanical markers, a weighted symmetry score, and a skeleton replay your physiotherapist can check — without one frame of video leaving the browser.

How the score worksPrivacy by design

Camera only · 3 valid reps per leg · 90% LSI target

Session 06 · right knee

Tracked

71.4%

Readiness

Below target

3 reps · both legs

Knee valgus

56.3%

Knee flexion

86.7%

Trunk lean

78.6%

Median of 3 valid reps per leg. Video discarded on-device.

The gap

Cheap tools can’t see the pattern. Clinical tools cost a clinic.

In months four to nine after surgery the risk is not the knee bending — it is the knee collapsing inward under load, and the stiff landing nobody measures. A static angle app sees a pose. Force plates and 3D capture see the pattern, at a price almost no outpatient clinic pays.

01

The injury hides in the pattern

Valgus collapse and reduced knee flexion appear under load. Held stretches and careful walking do not show them.

02

Lab hardware is out of reach

Force plates and 3D motion capture start in the tens of thousands. Most clinics never get near one, and neither do the athletes paying for them.

03

So the decision falls back on the eye

Between appointments, progress is a feeling. Small asymmetries — the ones that precede reinjury — stay invisible on both sides of the relationship.

ATHLYx compares the operated leg with the athlete’s own healthy leg — no normative database, no lab visit, no extra hardware.

The measurement

Three markers. One score. No video.

A single-leg squat, filmed front-on for valgus and trunk lean, side-on for flexion. MediaPipe tracks the landmarks in the browser; the angles are computed on the device.

01 / 45%

Knee valgus

Inward collapse of the knee in the frontal plane. The pattern clinicians watch for after ACL reconstruction.

Lower is better

02 / 35%

Peak knee flexion

How deeply the knee bends — a proxy for shock absorption. Stiff landings show up as low flexion.

Higher is better

03 / 20%

Lateral trunk lean

The tilt of the shoulder-to-hip line from vertical — a quiet compensation at the hip.

Lower is better

Peak value per repetition, then the median across three to five accepted reps per leg. A rep is accepted only when tracking confidence holds — rejected reps are never scored.

The method

A symmetry score you can trace back to the angles.

The limb symmetry index compares the operated leg with the healthy one, where 100% means equal. ATHLYx adapts it to camera-derived angles and shows every step of the arithmetic.

01 — Measure

Peak valgus, flexion and trunk lean for every repetition, on both legs.

02 — Sub-scores

Higher-is-better divides operated by healthy; lower-is-better divides healthy by operated, with a 5° buffer against dividing by zero.

03 — Weight

0.45 valgus + 0.35 flexion + 0.20 trunk. Valgus carries the most weight because it is the pattern clinicians watch.

04 — Flag

Below target under 90%, or when any single marker breaks its floor — a strong area cannot hide a weak one.

Worked example

Six months post-ACL, single-leg squat

Nothing here is a black box. Each row is one marker, measured on both legs, converted to a symmetry sub-score and weighted into the total.

Criteria status is shown to the clinician first. The athlete sees the sub-scores and a neutral summary.

MarkerHealthyOperatedSub-score
Knee valgus4°11°56.3%
Knee flexion75°65°86.7%
Trunk lean6°9°78.6%

71.4%

Readiness score

Below target

Depth is nearly symmetrical. The score is pulled down by inward knee collapse on the operated side — the marker a clinician would follow.

Privacy

The video never leaves the device.

Pose estimation runs in the browser. Frames are analysed and discarded; what reaches the server is a set of angles, medians and a compact landmark track for the skeleton replay.

On-device inference

MediaPipe Pose Landmarker runs locally at 30 fps or better. No frame is uploaded, stored, or sent to a model.

Only measurements are saved

Angles, sub-scores, criteria status and landmark sequences — enough to replay a repetition, nothing that shows a body.

A link the athlete controls

A therapist sees a session only while the athlete keeps the link active, and can be cut off at any time.

Deletable by design

One request removes sessions, notes and plans. The demo runs on volunteer or synthetic data only.

No video. No images. No cloud inference.

The privacy model is the product, not a settings page.

Scope and limits

What this version does — and what it refuses to do.

A tool that measures one thing well, and says so out loud. The boundary is part of the design.

In this version

  • Live camera tracking with a skeleton overlay and a repetition counter.
  • One standardised test: the single-leg squat, both legs, three valid reps each.
  • Three markers, a weighted score, per-marker flags and a criteria status.
  • Check-in schedules, therapist review, rep notes and skeleton replay.
  • Exercise plans written by the therapist, and a completion record.

Out of scope, on purpose

  • Diagnosis, medical clearance or return-to-sport decisions.
  • Force, strength or jump-height measurement.
  • 3D reconstruction or multi-camera capture.
  • Native apps, or storing video of any kind.
  • Real patient data — the prototype is not HIPAA-compliant.

The honest version

ATHLYx is a screening and tracking aid for a care team. It does not diagnose, and one test does not clear anyone to return to sport. Webcam pose estimation is less accurate than lab equipment, so small differences may sit inside measurement noise. The decision belongs to the athlete’s surgeon and physiotherapist.

The build

Small stack, boring on purpose.

Everything that decides a criteria status is deterministic and runs on the device or on one small server. Gemini drafts wording a clinician edits; it never sets a status and never picks an exercise.

Emory Hacks 2026

A camera, a marked spot on the floor, and three reps.

That is the whole setup. The athlete gets a neutral summary; the clinician gets the numbers, the replay and a place to write the note.

See how the score worksRead the limits

Prototype · volunteer and synthetic data only · not HIPAA-compliant · not a medical device.