ARS DWK Innovation Grant 2026 — Unified Project Narrative (v2)

Working title for the grant: A Surgeon-Owned Computer Vision Platform for Objective Efficiency Analytics in Endoscopic Sinus Surgery

Internal codename: Pharyvac Public framing: Whoop for sinus surgeons, built by a surgeon, for surgeons.

PI: Jaymarc Iloreta, MD (Practicing Rhinologist, ARS member) Mechanism: ARS David W. Kennedy, M.D. Innovation in Rhinology Grant, up to $10,000 direct, 12 months Application due: May 29, 2026 · Finalist pitch (if selected): July 24, 2026 (Summer Sinus Symposium)

Changes from v1: this version is grounded in your actual preliminary data (n=16 cases, working YOLO pipeline at 92.4% precision / 85.3% recall), shifts the aims from “build a dataset” to validate, scale, and open-source, and threads the Whoop / surgeon-by-surgeon framing through the entire narrative.

1. North-star positioning (the one-liner)

Medicine’s data infrastructure was built for billing codes and complication tracking, for payers, regulators, and risk managers. The surgeon doing the work is the one person the data was never built to serve. Pharyvac is part of a new wave of quality data built for the surgeon, not extracted from them, a surgeon-owned camera + computer vision pipeline that turns FESS into an objective per-case efficiency report. It works on 16 cases today. We’re asking for $10,000 to validate it across more surgeons, scale the model, and release it open-source with privacy built in for both patient and surgeon.

That paragraph does six jobs at once: it names the bigger paradigm shift (data for surgeons, not from them), frames the project as surgeon-driven (agency, not surveillance), states it’s already working (lower perceived risk), names the three concrete uses of the money (validate / scale / release), differentiates from the commercial closed players, and signals the moral framing (privacy by design). Every other piece of the application should ladder up to it.

The two-sentence pitch version (memorize this):

“Every athlete with a Whoop knows their recovery score. Every surgeon walks out of a four-hour case with no idea where the time went, because medicine’s data was built for billing, not for the person holding the instruments. We built the surgeon’s version.”

2. Where we are today (the prelim-data anchor)

This is the section that converts skepticism into trust. State it plainly and specifically:

16 bilateral full primary FESS cases analyzed end-to-end (Nov 2025, Mar 2026), spanning simple (n=7) and complex (n=9, with septoplasty / turbinate / extended frontal work).
Surgeon-owned GoPro + YOLO instrument-detection pipeline. A simple camera the surgeon owns and controls — framed on the instruments and the surgeon’s hands — capturing the surgeon’s own performance data, not the hospital’s video system.
Model performance: 92.4% precision, 85.3% recall, ≈88.7% F1.
Per-case decomposition is reliable today: forceps 26.8% ± 4.3%, nav suction 13.2% ± 4.2%, microdebrider 12.8% ± 3.5%, non-nav suction 9.5%, nav probe 3.8%, suction bovie 2.1%. These proportional metrics are unaffected by the known detection-flicker artifact.
Directly-measured “non-instrument time” (the dead-time metric): 22% of recorded time, ≈36 min / case, this is the headline efficiency number, measured straight from the video, no model assumptions required.
Cost framing: at academic-center OR rates ( $60-$ 100/min), each case costs $8,100-$ 16,300; the 22% dead time alone represents $2,160-$ 3,600 / case of non-productive cost.
Known limitations and identified fixes:
- Raw instrument-swap counts are inflated 2–4× by detection flickering between visually similar suction-class instruments. Fix: temporal smoothing with a 1-second minimum bout filter (already designed, not yet deployed). Drops swap count from ~434 → 150–200 true physical swaps.
- 25.4% of frames are unlabelled, mostly hand-only / transition moments. Fix: add a “transition” class to YOLO, plus lens-distortion correction and motion-blur augmentation for the wide-angle GoPro view.

That paragraph alone tells a reviewer two things they value highly: the work is real, and the PI knows what’s wrong with it. A PI who can name their artifacts is a PI who will deliver.

3. The story arc (the through-line for both proposal and pitch)

Beat 1 — The problem (the hook)

Medicine’s data infrastructure was built for the wrong customer. CPT and ICD-10 codes serve billing. M&M conferences and complication registries serve risk management. NSQIP and quality dashboards serve regulators and payers. The surgeon doing the work is the one entity the data system was never designed to serve.
The result: surgeons end every case with subjective impressions and a single number, total OR time. That’s a noisy proxy that conflates anatomy, pathology, trainee involvement, and surgeon performance.
Trainee assessment uses expert-rater scales (OSATS-style) that are slow, expensive, and don’t scale.
Hospital admins see efficiency through finance dashboards built around case length, not what happens during the case.
Net effect: rhinology, and surgery generally, has no objective dial for efficiency built for the surgeon. We can’t improve what we can’t measure, and we haven’t measured what the surgeon needs to see, because the data has always been built for someone else.

Beat 2 — Why now

Surgical computer vision is no longer experimental. Phase recognition in laparoscopic cholecystectomy reaches >90% F1 (EndoNet, CholecT45/50). Tool detection and gesture analysis are off-the-shelf in general surgery. Surgical foundation models have arrived.
Rhinology is the gap. FESS is the highest-volume endoscopic procedure in otolaryngology, the camera is already in the field, and almost none of the video is being computationally analyzed.
The consumer-camera angle is now feasible. Consumer-grade GoPros now produce surgical-grade footage. The personal performance-data pattern that Whoop normalized for athletes — owned by the individual, focused on their own performance — is finally available for surgeons.
The technical risk has collapsed. The clinical opportunity has not.

Beat 3 — The innovation (what’s actually new)

Five layers, in order of rhetorical priority:

A new category of data: quality data built for the surgeon, not extracted from them. The existing data infrastructure (codes, complication registries, finance dashboards) was built for billing, regulation, and risk management. Pharyvac is part of an emerging category, surgeon-owned, in-the-moment, performance-focused data designed for the practitioner doing the work. This is the framing that elevates the project from “tool” to “movement.”
Whoop for sinus surgeons. A surgeon-owned camera + computer vision pipeline that delivers a per-case efficiency report directly to the surgeon. Surgeon-owned. Voluntary. Performance-focused. Not diagnostic. This is the metaphor the audience will repeat in the hallway.
Built by a surgeon, for surgeons. The PI is a practicing rhinologist who designed and uses this in his own OR. That solves the surgical-AI adoption problem the closed commercial players (Theator, Activ Surgical) have not, surgeons trust tools built by surgeons.
Privacy by design, both sides of the camera. Patient anatomy de-identified before any frame leaves the OR; only metadata (instrument timestamps, derived metrics) ever travels. Surgeons retain ownership of their own data; cohort benchmarking is opt-in and aggregated. This is the differentiation that matters in 2026.
Open-source toolkit. The model weights, post-processing scripts, and analysis pipeline get released. The dataset itself is private by design, but the toolkit is a multiplier for the field, exactly the language the DWK grant uses (“moves the field forward”).

Beat 4 — Specific Aims (the spine)

Aim 1. Validate the Efficiency Index across multiple surgeons. Expand the existing n=16 single-surgeon corpus to a multi-surgeon cohort (target: 3–5 surgeons, 50+ additional cases). Demonstrate that the directly-measured 22% dead-time metric, the smoothed instrument-swap count, and the composite Efficiency Index discriminate as expected across surgeon experience levels. This is the construct-validity study reviewers want to see.

Aim 2. Scale the model with the known fixes. Deploy the temporal-smoothing post-processor (drops swap count inflation from 2–4× to 1.0×). Add the “transition” class to YOLO and retrain with lens-distortion correction and motion-blur augmentation, targeting <15% unlabelled frames (vs. 25.4% today) and >92% recall (vs. 85.3% today). Add per-class precision/recall and confusion-matrix tracking.

Aim 3. Release the toolkit open-source under a privacy-by-design framework. Publish (a) the YOLO weights, (b) the post-processing and analytics pipeline, (c) the privacy framework (patient de-identification spec, surgeon data-ownership model, opt-in cohort benchmarking schema). Companion methods paper to International Forum of Allergy and Rhinology.

The three aims are deliberately staged: Aim 1 is the validity study that produces the IFAR figure, Aim 2 is the engineering work that makes the tool defensible, Aim 3 is the community deliverable that fulfills the DWK “innovation that moves the field” mandate.

Beat 5 — Feasibility (lean on the prelim data)

This is a one-paragraph trust-builder. Verbatim or close to it:

We have already analyzed 16 bilateral primary FESS cases over five months using a tower-mounted GoPro and a custom YOLO instrument-detection model (precision 92.4%, recall 85.3%). The pipeline produces per-case instrument time decomposition and a directly-measured 22% non-instrument dead-time metric, with corresponding OR-cost estimates of $2,160-$ 3,600 per case in non-productive time at academic rates. Known artifacts (detection-flicker inflation of swap counts; 25% unlabelled frames driven by hand-only / transition moments) have been characterized in detail and have implementable fixes (temporal smoothing, transition-class addition, lens-distortion correction). Hardware and storage are in place; IRB status is [exempt / approved on __ / in submission]. The proposed 12-month plan extends the existing pipeline rather than building from scratch.

Beat 6 — Impact and future trajectory

A 3-year arc, even though the grant is 12 months:

Year 1 (this grant): Multi-surgeon validation + scaled model + open-source release. One IFAR submission (methods + validity).
Year 2: Multicenter expansion (R03 or industry partnership), trainee-assessment substudy, real-world QI integration at a partner program.
Year 3: Real-time intraop feedback prototype; OR scheduling integration; surgical-coaching layer.

Reviewers fund a seed, not an endpoint. Show them the tree this seed grows.

Beat 7 — The ask

$10,000 direct, 12 months. PI takes no salary (per grant rules, state this explicitly).

4. The privacy-by-design framework (a dedicated section — this is one of your strongest differentiators)

Build this into both the proposal text and the pitch. Three commitments:

Layer	Patient	Surgeon
Data ownership	Hospital retains custody of raw video; patient consent governs use.	Surgeon owns their own efficiency data. Hospital does not.
What leaves the OR	Only de-identified metadata (instrument timestamps, derived metrics). Raw frames never travel.	Aggregated, opt-in cohort metrics. Individual surgeon data stays with the surgeon.
Open-sourced artifact	Toolkit (model weights, scripts). not raw video or PHI.	Privacy framework spec, so any group adopting the tool inherits the same protections.

This is the slide reviewers and audience members will quote back to you. It’s also the reason a hospital legal team will say yes to adoption, and the reason commercial closed-platform competitors can’t easily match it.

5. Budget skeleton ($10,000 direct, 0% indirect)

Line	Approx.	Justification
Annotation labor (RA / med student, ~150 hrs for transition-class labeling + new-surgeon cases)	$3,000–3,500	Largest cost; needed for retraining and multi-surgeon validation
Cloud GPU compute / storage for retraining + dataset hosting	$1,500–2,000	Includes temporal-smoothing reprocessing of all cases + new-class retraining
Annotation software (CVAT pro / Encord)	$500–800	Versioned annotation, reviewer workflow, audit trail
IRB amendment + multi-surgeon onboarding (data-use agreements, opt-in consents)	$500–1,000	Critical for scaling to additional surgeons; underpins the privacy framework
Open-source release infrastructure (GitHub Pro, model hosting, documentation)	$500	Required deliverable of Aim 3
Travel to ARS Summer Sinus Symposium for presentation (per Section H)	$1,000	Required by grant terms
Secure storage, HIPAA-compliant pipeline tooling, miscellaneous	$1,000	Privacy framework requires real infrastructure, not promises
Buffer	$500
Total	≈ $10,000

Two reviewer-pleasing moves to keep: (1) explicitly state PI takes no salary, (2) show the largest line item is annotation labor, money goes into the substrate, not overhead.

6. How the narrative satisfies each review criterion

Criterion	What it gets in the proposal/pitch
Innovation	§3 Beat 3. Whoop-for-surgeons framing + privacy-by-design + open-source release. Non-traditional and explicitly non-hypothesis-driven; matches DWK mandate language directly.
Approach	§3 Beat 4 + §2, three specific aims grounded in named technical fixes (temporal smoothing, transition class, lens correction); per-class metrics and confusion-matrix tracking.
Feasibility	§2 + §3 Beat 5, n=16 cases analyzed, working YOLO at 92.4% precision; this is extension, not creation.
Future potential	§3 Beat 6. 3-year arc, IFAR pipeline, R-series setup, trainee assessment, real-time feedback.
Budget appropriateness	§5, labor-heavy, infrastructure-light, no PI salary, every line traceable to an aim.
Likelihood of completion	Defined deliverables: validated multi-surgeon cohort + retrained model + open-source release + IFAR submission. The pipeline already runs; the grant funds extension, not invention.

7. Mapping the narrative to the two deliverables

A. The 2-page written proposal (due May 29)

Suggested layout:

Title + Specific Aims box (~25% of page 1), verbatim aims from §3 Beat 4
Background & Significance (~25% of page 1), §3 Beats 1–2, ending with the gap statement and the “by-a-surgeon” frame
Innovation (~10% of page 1), §3 Beat 3, four numbered points
Preliminary Data & Feasibility (~25% of page 2), §2 + §3 Beat 5; this is the trust-building section, lead with the n=16 numbers
Approach (~25% of page 2), sub-sectioned by aim; the figure goes here
Privacy & Open-Source Plan (~10% of page 2), §4 condensed to a table or three bullets
Timeline & Deliverables (~10% of page 2), small Gantt
Future Directions / Impact (~5% of page 2), §3 Beat 6 in two sentences
References, separate page

The single figure (you only get one) should be: a 3-panel composite, (left) the GoPro + YOLO pipeline schematic, (middle) the per-case Efficiency Report mock-up showing instrument-time decomposition + dead-time + cost, (right) a real plot from your n=16 data (e.g., simple-vs-complex case-time comparison or instrument-time stacked bars). One image, three jobs done: pipeline credibility, output legibility, real-data proof.

Format constraints to respect: 11pt+ font, ½” margins, no headers/footers, one figure or table.

B. The Shark-Tank pitch (10–15 min, July 24, if finalist)

The proposal is graded on rigor; the pitch is graded on memorability + credibility + audience response. Same content, different center of gravity.

Suggested 10-slide spine (≈45–75 sec/slide, leaves room for Q&A):

The hook, the data paradigm slide. A two-column visual: left column lists “Data medicine has built about you” (CPT codes, ICD-10, M&M reports, NSQIP, complication registries, OR finance dashboards). Right column shows what the surgeon actually gets at the end of a case (a single number, total OR time). Caption: “Every system above was built for someone else.” This is the moral hook.
What every other industry already gave its practitioners. Three logos: Whoop (athletes), Strava (cyclists/runners), Garmin (pilots). Caption: “Surgeons are still waiting.”
The pitch in one sentence. “Quality data built for the surgeon, not from them. An objective per-case efficiency report. Built by a surgeon, already working on 16 cases.”
What it actually shows you. Mock-up of the per-case Efficiency Report, instrument-time bar, dead-time fraction, cost estimate, surgeon’s own trend line. This is the slide they remember.
Preliminary data, the “we already started” slide. Headline numbers: 16 cases, 92.4% precision, 22% non-productive time directly measured, $2,160-$ 3,600 / case in dead-time cost. One real plot from your data.
Three asks for $10K. Validate (multi-surgeon). Scale (transition class + smoothing). Release (open-source toolkit). Three boxes.
Privacy by design, the table from §4. Patient on the left, surgeon on the right. The reason a hospital will say yes.
Why open-source, and why this is bigger than a tool. “Theator and Activ raised $30M to keep their tools closed. We're spending$ 10K to give ours away. Because the data revolution surgeons deserve isn’t going to come from billing systems or commercial vendors, it’s going to be built by surgeons, for surgeons, in the open.” This is the moral close-out before the ask.
Where this goes (3-year arc). Trainee assessment. QI dashboards. Real-time coaching. Plant the future.
Close. “Quality data, built for the surgeon. $10,000. 12 months. Help us build it for the field.” Then your contact.

Backup slides: methodology depth, IRB status, references, expanded budget, the “known artifacts and fixes” page (this turns a Q&A landmine into a strength).

Anticipated Q&A (prep answers for these):

“Why GoPro instead of the endoscope feed?” → Surgeon-owned, captures the hands-and-instruments workflow the endoscope can’t see, deployable in any OR without hospital integration. We can fuse the endoscope feed in Year 2.
“How does this generalize beyond a single surgeon?” → That’s literally Aim 1.
“Isn’t this just expensive timekeeping?” → No, it’s per-instrument decomposition + behavioral metrics, plus the validity study that ties them to outcomes. Show the swap-count plot.
“Hospital admins will use this against surgeons.” → Privacy framework, surgeon-owned data, opt-in by design. Same answer Whoop gives to the same question about athletes.
“What about the inflated swap counts you mentioned?” → I want this to come up. It demonstrates that I know what’s wrong with my own data and have an implementable fix. Reviewers love this.

Pitch-specific rhetorical moves:

Open with the Whoop / GoPro side-by-side. Best 10 seconds you have.
Use “by a surgeon, for surgeons” verbatim once per quarter of the talk.
Name IFAR right of first refusal in your close, it’s a built-in reason for the ARS audience to want to fund this.
Don’t oversell real-time feedback in Year 1. Plant it in Year 3.
If you have any surgeon other than yourself willing to lend a 1-sentence quote of support (“I’d use this”), use it as Slide 6.5.

8. Open decisions you still need to make (much shorter list now)

Aim 1 cohort design. Multi-surgeon at one institution, or multi-institution from the start? The first is faster and the IRB is simpler; the second is more impactful but riskier within 12 months. Recommendation: start single-institution, pre-write the multi-institution amendment.
Open-source license. Apache-2.0 is the safest default for academic + clinical adoption (permissive, patent grant). State it in the proposal.
Comparator / benchmark surgeons. Are you recruiting locally, or from a national network? Letters of support from 2–3 surgeons who agree to participate would be a killer addition to the application packet.
IRB status, exact wording. Confirm whether this is exempt, expedited, or full-board, and what amendments scaling to additional surgeons will require.
The internal name. Keep “Pharyvac” as your codename; the grant title should be clinical/descriptive (above). Do you want to debut Pharyvac publicly in the pitch? My instinct is yes, names that the audience can repeat in the hallway after stick.
Co-investigators. Anyone else’s name on this? A computer-vision collaborator or biostatistician would strengthen the application even if they’re a 5% effort partner. Letters of support from key personnel are required; better to have them than not.

9. Action timeline (anchored to May 29)

Date	Action
Apr 26 (today)	Lock §8 decisions: license, IRB wording, supporting surgeons, co-investigators.
Apr 27, May 3	Draft v1 of 2-page proposal from this narrative. Assemble the 3-panel figure (pipeline schematic + report mock-up + real n=16 plot). Pull references.
May 4, May 10	Send v1 to 1–2 trusted readers (one rhinologist, one CV/data person if possible). Update CV. Solicit letters of support.
May 11, May 17	Revise to v2 based on feedback. Confirm IRB amendment plan. Receive letters of support.
May 18, May 25	Final polish, format check (11pt, ½” margins, no headers/footers, single figure), assemble the single PDF (proposal + CV + refs + letters + IRB).
May 26, May 28	Buffer for ARS website upload issues. Submit early.
May 29	Deadline (midnight EST).
Early June	Finalist notification (top 3).
June, July 24	If finalist: build pitch deck from §7B. Practice 5+ times. Have answers locked for the five anticipated Q&A vectors. Bring a printed Pharyvac one-pager to hand out after.

10. What I’d suggest we do next

You’re in a much stronger position than v1 of this narrative assumed. The bottleneck is no longer “what’s the prelim data”, it’s converting this narrative into the two artifacts. Pick one to start:

Draft v1 of the 2-page proposal text. I can write it from this narrative + your data files; you edit. Best move if you want to maximize the time available for revision before May 29.
Build the single figure (the 3-panel composite). Highest single-asset ROI for the written application, one image carries 30% of the rhetorical load.
Build the pitch deck spine in slide form with the 10 slides above as actual outline + speaker notes. Best move if you want the unified narrative pressure-tested in pitch form before you write the proposal, pitches expose weak claims faster than prose does.

My recommendation: (1) first, since the proposal deadline is the binding constraint and the figure can be built in parallel. Tell me to start drafting and I will.

Pharyvac Computer Vision

Explorer

ARS DWK Grant - Unified Project Narrative