Building AI for Women's Health: How Hertility Combined Bayesian Diagnosis and Scan Automation

This episode of Just Now Possible features three Hertility team members discussing how they build AI diagnostic tools for women's health in the UK and Ireland. Hertility combines online health assessments, at-home hormone testing, and clinician-reviewed reports, built on years of linked symptom, blood, and ultrasound data. The conversation covers two products in development: Gyn.AI, a Bayesian network that outputs probabilities rather than binary diagnoses, and a scan automation pipeline that classifies ultrasound images and drafts clinical letters with self-checking. The discussion matters because it shows how to bring AI into a sensitive, heavily regulated healthcare setting while earning clinician trust, guarding against automation bias, and handling patient data responsibly.

01Key takeaways

• Present model outputs as probabilities with visible reasoning, since this builds more clinician trust than binary labels.
• Build explicit guardrails against automation bias, such as holdout sets and independent review, alongside the model itself.
• Minimize and pseudonymize sensitive data before it reaches any model to reduce risk.
• Run models in infrastructure you govern when handling health data, so controls stay under your own regulatory oversight.
• Treat regulatory requirements as product constraints from day one to make AI products more scalable rather than slower.

02Key sections

Hertility's unique dataset and business

The team explains how linking symptoms, blood results, and pelvic scans over seven years for over a million women creates a rare foundation for diagnostic AI. The data is the core asset that makes the products possible.

Gyn.AI and probabilistic diagnosis

Gyn.AI uses a Bayesian network to express diagnoses as probabilities and shows clinicians the reasoning behind them. This transparency is presented as a way to build trust and speed up triage.

Scan automation pipeline

The pipeline classifies ultrasound images, detects follicles, measures ovarian volume, and drafts clinical letters. An agentic loop checks drafts against patient data before any human review.

Guarding against automation bias

The team describes holdout sets and independent, fresh-eyes review as safeguards so clinicians do not over-rely on model outputs. Guardrails are treated as equally important as the model.

Privacy, infrastructure and regulation

Patient identifiers are pseudonymized and minimized before reaching models, and models run in-house on AWS Bedrock. Regulation is handled as a design requirement from the start rather than a late-stage addition.

03From the post

“Listen to this episode on: Spotify | Apple Podcasts How do you build trustworthy AI diagnostic tools in one of medicine's most historically under-researched areas? In this episode of Just Now Possible, Teresa Torres talks with Tulsi Patel (Director of Product and Technology), Lorna Brightmore (Head of”

04Frameworks mentioned

Summary and takeaways written by PM Atlas; quotes are short excerpts. © the original author.