SELENE: Why Women’s Healthcare AI Must Be Open, Local, and Privacy-First

#ai #womenshealth #wecoded

When it comes to women’s midlife health, particularly the neuroendocrine transition of menopause, there is a massive historical data gap. For decades, the standard of care has been reactive and fragmented. AI has the immense potential to close that gap by processing longitudinal symptom tracking, but the tech industry’s default response has created a new problem: the privacy deadlock.

We are building SELENE (Scientific Evidence-based Longitudinal Endocrine Neuropsychological Engine) to prove that high-utility medical AI does not require a data center. Currently in its alpha prototype phase, SELENE is a local-first medical intelligence engine. More than just a tool, it is an argument for why healthcare AI must be open source, open code, and fiercely protective of user privacy.

Here is why we built it, how the architecture works, and why we are calling on the open-source community to help us take it to the next level.

The Clinical Gap: The "Invisible" Transition

Menopause is not a singular medical event but a complex transition affecting the brain, cardiovascular health, and metabolic function over seven to ten years. Yet, clinical guidelines are often based on cross-sectional studies that fail to capture this fluctuating reality.

A woman might experience a constellation of symptoms, for example, disrupted sleep architecture, transient cognitive impairment ("brain fog"), and anxiety that don't align neatly with a 15-minute primary care visit. Because clinicians lack high-fidelity, structured data mapping symptom velocity over time, care often becomes guesswork. Women’s brain health deserves rigorous, longitudinal tracking that validates their lived experiences.

The Failure of Cloud-First AI

To solve this data deficit, the default tech solution is a cloud-centralized tracking app. But in today's digital landscape, the aggregation of intimate reproductive and neurocognitive health data on central servers poses an unacceptable risk. Asking women to trade their privacy to third-party data brokers for health insights is a non-starter.

When you decouple health intelligence from cloud surveillance, an interesting technical benefit emerges: you completely eliminate the attack vectors associated with transmitting sensitive data, and you don't even have to worry about securing cloud API keys or guarding against middle-man breaches because there are no external API calls. The data never leaves the hardware.

SELENE: A Neuro-Symbolic, Privacy-First Architecture

SELENE runs entirely on the user's device. We utilize the MedGemma 1.5 (4b) variant, optimized for edge deployment via Ollama, giving it a nuanced understanding of clinical terminology that general models lack.

The user experience centers on capturing structured data and qualitative notes, which are processed through a strictly local Retrieval Augmented Generation (RAG) pipeline:

Vector Database: We utilize ChromaDB running locally to manage semantic search for both a medical knowledge base and the user’s long-term history.
Retrieval & Synthesis: SELENE consults a locally stored, vector-embedded library of peer-reviewed research and endocrine protocols, synthesizing the user's logs with retrieved medical literature.
Knowledge Patches: To keep the AI current without exposing user data, we use a one-way update mechanism. The system pulls vector chunks of new research Server → Device. No user queries or chat history ever flow Device → Server.

Why Open Code is Non-Negotiable

Women's health has suffered for decades under opaque medical paradigms. AI cannot afford to replicate that with "black box" proprietary algorithms. We need open source and open code for three critical reasons:

Audibility of Safety Guardrails: LLMs are probabilistic, but patient safety must be deterministic. By making the code open, the community can verify our safety layers. For example, SELENE uses a deterministic math layer to analyze symptom velocity. If it detects a red flag (e.g., rapid deterioration in sleep combined with severe headaches), it forces the LLM to output a hard-coded referral warning.
Mitigating Bias: Proprietary models hide their training biases. Open weights like MedGemma allow researchers and clinicians to look under the hood and ensure the model isn't perpetuating gender-based medical biases.
Verifiable Trust: Open code is the ultimate proof that an application isn't quietly exfiltrating sensitive data in the background.

Call for Collaboration: Join the Alpha

SELENE is currently an alpha-first prototype. It proves the concept that we can build robust, locally-run, neuro-symbolic AI for midlife women's health. But to push this forward, it takes a village.

We are looking for open-source contributors:

ML Engineers: To help optimize quantized local models and improve our local RAG pipelines.
Frontend Developers: To refine our UI and ensure accessibility.
Clinical Researchers: Passionate about women's brain health, neurocognition, and endocrinology to help curate our Knowledge Patches.

We don't have to accept a future where healthcare AI is a black box that harvests our data. We can build it in the open, keep it local, and ensure it actually serves the people who need it most.

Check out the repository and live demo and let’s build AI that understands women’s health safely, openly, and securely.