Mei Chen Spent Two Years on an FDA Filing for Soraya. Then the Rules Changed.

## The First Year of Building

I had three co-founders in early 2023. None are with Soraya anymore for various good reasons. The team in 2026 is mostly people I hired in late 2023 and 2024: six ML and clinical engineers, four practicing gynecologists on the medical advisory board, two regulatory full-timers plus a former FDA reviewer who consults.

Data acquisition was the hardest part. We struck data-use agreements with four academic medical centers between June 2023 and March 2024. Each took seven months to negotiate. Total imaging volume by mid-2024: 47,000 studies. Small by computer-vision-research standards but appropriate for the FDA's diversity expectations.

Validation on a held-out test set from a fifth institution: sensitivity 91%, specificity 86%, AUC 0.94 across both classifications. Performance held up across racial and ethnic subgroups. By the November 2024 filing, we had $11M total capital raised across seed and small Series A.

## The May 2025 Change

In late April 2025, the FDA's CDRH issued draft guidance on AI-enabled diagnostic devices. The guidance proposed a more rigorous evaluation track for "high-impact AI diagnostic-support systems," including women's health applications, on the rationale that historical underdiagnosis made model errors particularly consequential.

The most consequential new requirement for us: clinical validation must include at least one community-care site, not only academic medical centers. Our four-academic-site study did not meet this. The draft guidance wasn't yet finalized, but our FDA reviewer signaled informally we should address the community-site question before our submission would advance.

I called our regulatory consultant. We were probably going to need to expand the study to include at least one community site. Another year. Another $1.8M.

## The Decision

Two paths. Withdraw, expand, refile in 12-18 months, and raise an emergency bridge round. Or push the original filing through, hope the FDA didn't apply draft guidance retroactively, and risk a rejection that would damage future submissions.

I chose the first path. It was the right choice. It was also the choice that almost ended the company.

## The Bridge Round Nobody Wanted to Fund

Expanding the study meant raising another $4M. We started conversations in June 2025. The women's health AI market had cooled: broader healthcare-AI pullback, visible regulatory delays in the category, increased friction from the FDA draft guidance.

Twenty-three first meetings between June and October. Eleven seconds. Two term sheets, both with aggressive terms: 1.5x participating preferred, anti-dilution rights more aggressive than my Series A, and one with a blocking right on any future regulatory pivot. I declined both.

By October 2025 I was at seven months of runway with no closed lead and a clinical study that needed to start Q1 2026. I had a serious board conversation about an acquisition inbound from a larger imaging-AI company. The exploratory conversation lasted three weeks. The proposed terms would have valued Soraya below Series A post-money and absorbed our team into a roadmap that didn't prioritize endometriosis. I shut it down.

The board supported keeping raising. A smaller, women's-health-specialist fund became our lead in late December at a $52M post-money valuation. Flat with our Series A but with reasonable terms and a partner who'd built and exited a women's-health diagnostics company. The round closed January 14, 2026, three weeks before our cash-out date.

## The Internal Toll

The six months between the FDA guidance change and the bridge close were the hardest of my professional life. I lost twenty-three pounds. I had insomnia that resisted every intervention. I cried in my car between meetings more than once. My partner is a clinical oncologist with her own demanding hours; we had stretches that summer when we'd see each other for an hour at the end of the day and that was the whole interaction.

We instituted a hiring freeze in July. Two excellent engineers left for larger companies between July and November. The signal of the bridge round failing would have been clear to anyone reading cap-table updates.

The hardest internal conversation was with our clinical operations lead, Nikki, who'd been with us since the seed and had architected the original four-site study. The expansion required re-architecting parts of her original work. She handled it with grace I did not deserve. The new plan was better than the original. The work to get there was brutal.

## The Patient Letters

The thing that kept me functional was a slowly-growing email folder called "patient letters." We'd quietly opened a contact form in early 2025. The form clearly said we were pre-clearance. Patients used it anyway. There are 1,247 letters in the folder as of last week. I've read every one. Teenagers watching symptoms appear after a mother's diagnosis. Women diagnosed after twelve years writing to ensure Soraya would reach others. Clinicians tired of telling their patients they had nothing more to offer.

On the worst days of the bridge round, I'd read three or four letters before opening my email. They weren't consolation. They were a reminder that the regulatory work I was failing my way through, the fundraising humiliations, the operational compromises, all of it served a population whose waiting was the actual cost being paid. The letters made it possible to keep going when nothing else did.

## The Refiled Submission and the Path Forward

We restarted clinical recruitment in February 2026 with two new community sites added to the four academic centers. The six-site study is on track to complete enrollment by Q3 2026. Refiled De Novo submission targeted for Q4 2026. Expected clearance, under the most-likely-case FDA review timeline, mid-2027.

That's still four years from when I started Soraya. By the standards of FDA-cleared medical-device AI, it's on the faster end of the distribution. The product side has continued to mature: clinician-facing UI built, integrations with three major imaging-PACS systems, validation on an additional 14,000 imaging studies. Performance has held up. The clearance is the gating step.

## The Quieter Inflection

The bigger breakthrough was internal. The bridge close, the study pivot, and the regulatory re-strategy together changed my relationship to the company. Before mid-2025 I was running Soraya as a product company that had to clear a regulatory hurdle. After mid-2025 I started running it as a regulatory-and-clinical organization that happened to have a great technical core. The shift changed how I hire, allocate time, and communicate with the board.

The version of Soraya I run today is much closer to one that can actually ship a real product to real patients. The chaos of 2025-2026 corrected an imbalance: too fast on the technical axis, too slow on the regulatory.

Rebecca, the patient from the October 2024 phone call, emailed me in March 2026 to ask how things were going. I wrote her a long, honest reply. She wrote back: "Thank you for telling me the truth. Most of the time when I email companies about this, I get a marketing reply. The truth is much better." I keep her reply pinned in my email.

## What I'd Tell a Founder Considering FDA-Path Healthcare AI

Eighteen months into the regulatory journey, here are the things I wish I'd internalized earlier.

§1. Regulatory cost dominates everything; budget accordingly

Clearing a novel AI diagnostic device through De Novo is $4-8M and 24-36 months minimum. That's the floor. Expected value is higher. Variance is enormous. If your business model assumes a 12-month clearance and a $2M regulatory budget, you are mis-modeling. You cannot bootstrap meaningful regulatory AI in healthcare. Choose investors specifically for understanding of this.

§2. The patient is your North Star; the regulator is your governor

There is constant tension between moving fast for patients and correctly for regulators. Move too fast on the patient axis and you ship a product that destroys the regulatory pathway for the entire category. Move too slow on the regulator axis and your patients wait years longer for a product that could have helped them. Be transparent about the trade-off with both audiences.

§3. Build the team you need in 18 months, not the team you can afford today

I made the mistake in 2023 of hiring strong technical engineers first and adding regulatory and clinical-operations expertise later. The right order is regulatory and clinical-ops as early as you can afford. The technical work can wait. The regulatory work has fixed timelines you can't compress. I would have brought our regulatory lead in six months earlier.

§4. Communicate the regulatory chaos transparently

When the FDA guidance changed in May 2025, I chose transparency over polish. The investors who stayed through the bridge round were the ones I'd kept fully informed. Healthcare-AI regulatory environments are inherently turbulent. Tell the truth about what you don't know. Tell the truth about what you're doing anyway.

§5. The patient letters are not optional

The patient-letters folder turned out to be load-bearing infrastructure for my own ability to keep going. Founders in mission-driven categories often build mission framing for external use and treat it as marketing. The mission framing is also operating infrastructure for the founder. Build the channel for patient voices early. Read what comes in.

§6. The board you have is the board you live with

Our Series A board included one investor who, in 2025, behaved in ways that made the situation worse. Not maliciously: his expertise didn't translate to women's-health AI specifically. The cost of an ill-matched board member during a crisis is much higher than an empty seat. Prioritize board fit over board prestige.

§7. Healthcare AI founders are running marathons that look like sprints

Startup culture in 2026 is built for B2B SaaS timelines. Healthcare-AI timelines work like biotech. Multi-year regulatory pathways. Capital-intensive trials. Outcomes measured in patient impact, not MRR. Founders who operate healthcare-AI companies on SaaS rhythms will burn out, raise wrong rounds, hire wrong teams. Choose investors and employees who understand the actual cadence.

I am still in the middle of this work. Soraya is not cleared. The patients I owe this work to are still waiting. I am tired, I am committed, and the work is real. I'll write a different version of these lessons in 2028, when I hope I can write about the patients we've reached rather than the ones still waiting.