Create Jira Ticket
This idea cannot be sent to Jira.
InvisibleSignal — SENTINEL: Lyme Intelligence for the Bite-to-Clinic Journey
InvisibleSignal is built by The AI Cowboys. We work in the open: our code, our data provenance, and our limitations are public. Several of us have watched family members go years without a Lyme diagnosis; that experience shapes every product decision we make.
Lyme disease is the most common vector-borne illness in the United States, and the least visible. An estimated 31 million Americans are bitten by ticks each year. CDC estimates roughly 476,000 Americans are diagnosed and treated for Lyme annually, yet only about 89,000 cases were reported through national surveillance in 2023. Most of the gap is never counted anywhere.
For patients, the invisibility is personal. Two-tier serology has under 50% sensitivity in early disease, and up to 30% of patients never develop the classic bull's-eye rash. In patient surveys, 51% report it took more than three years to be diagnosed, and Lyme-treating clinicians cite prior misdiagnosis as a leading cause of delay. Patients cycle through labels — anxiety, fibromyalgia, chronic fatigue — while the underlying exposure history goes unasked.
The data problem compounds the human one. Federal surveillance is published annually, county coding is incomplete (Texas reports cases with zero counties coded; Connecticut changed county definitions mid-series), and no real-time federal tick-borne feed exists. A patient who wants to bring evidence to a ten-minute appointment faces fragmented datasets, jargon, and tools that were never built for them.
The result is a population that is sick, dismissed, and undocumented — and a public-health signal that systematically understates where Lyme risk actually is.
InvisibleSignal is SENTINEL: a Lyme bite-to-clinic toolkit, live in production today at invisiblesignal.theaicowboys.com. It walks a patient from "I found a tick" to "my clinician has what they need" in one connected path, no account required for its public tools.
The path has five working stages. First, an epidemiological map renders CDC county-level surveillance as an interactive choropleth: 152,018 reported cases, reconciled exactly to published state totals. Counties CDC could not code are labeled as undercounted rather than shown as zero — absence of measurement is never presented as absence of risk. Second, an exposure-first screener asks what a short appointment rarely has time for: tick exposure, season, geography, symptom duration, and the diagnoses given instead. It reports symptom overlap as overlap, never as a probability of disease, and red-flag symptoms route to emergency care before any model runs. Third, a cited risk report assembles county surveillance, seasonal tick activity, and a transparent four-weight index — and includes an explicit section titled "What this brief does not establish." Fourth, a lab-records module lets patients upload or paste a report; every extracted value is shown with the exact text it was read from, values that cannot be traced to the document default to rejected, and an on-device de-identification step removes names and dates in the browser before anything is sent. Fifth, the confirmed record exports as an HL7 FHIR R4 document bundle, coded with LOINC, SNOMED, and ICD-10, that a clinician's EHR can ingest. The offline-first Field Log extends the same path beyond coverage — a bite pinned to GPS and an on-device photo at the trailhead, synced when service returns.
Three architectural commitments define the MVP. The first is provenance: every figure on the site carries its federal dataset and the caveat that travels with it, and a 33-finding evidence corpus with DOI citations backs the design. The second is verified AI: language-model features (lab extraction, symptom normalization, evidence retrieval, narrative drafting) operate behind a deterministic verification boundary. Extracted values must be re-located in the source text; generated answers must cite known corpus entries or refuse; narrative prose may only use numbers injected by the deterministic layer. Golden-file evaluation suites gate every change, with extraction precision held at 1.0. The third is operation with no internet: exposure happens where coverage doesn't, so the installable Field Mode carries the full 3,231-county CDC risk dataset on-device — bite logging, GPS, photos, symptoms, and county risk all work offline and sync when coverage returns.
The MVP is hardened for real users: authenticated access to health data with row-level security, rate-limited APIs, strict security headers, a complete PHI access audit log, consent captured before any upload, and one-command deletion of all lab data. More than one thousand automated tests pass on every commit.
This is not a prototype awaiting a data agreement or a pilot site. It is deployed, it works end to end today, and it is built from data the federal government already publishes.
Solution Overview (MVP Demonstration):
User-facing Application
Live demo, prototype, or working product
InvisibleSignal is a live, deployed product — not a prototype — at https://invisiblesignal.theaicowboys.com. Core functionality: (1) interactive county risk map over reconciled CDC surveillance; (2) exposure-first screener with deterministic safety escalation; (3) cited county risk brief with an explicit limitations section; (4) lab-records module with verbatim-verified extraction and on-device de-identification; (5) one-click HL7 FHIR R4 export a clinician's EHR can ingest. The installable Field Mode logs a tick bite fully offline — GPS pin, on-device photo, county CDC risk with no connection — and syncs when coverage returns. Intended use: a hiker or patient documents exposure at the trailhead, arrives at the appointment with an organized, cited history, and hands the clinician a standards-coded document instead of a story that starts from zero. The full path walks end to end in five minutes; it was demonstrated live at the Beta milestone.
Access instructions (if needed)
Public tools require no account: https://invisiblesignal.theaicowboys.com (map, screener, risk brief, model validation, evidence corpus). Field Mode installs from the browser's install prompt or directly at /sentinel/field. Health-data features (lab records, Field Log sync) require a free account; panelists can request a pre-provisioned demo account via the TOPx organizers. Source code: https://github.com/The-AI-Cowboys-Projects/HHS
Results for the American People:
InvisibleSignal is built with lived experience, not just for it. Patient interviews during the sprint changed the product: the screener asks about exposure before symptoms because that is the order patients actually experience; results arrive with anticipatory guidance because research shows unreviewed results without framing cause harm; a manual, no-AI path exists because nine in ten patients say an opt-out is essential to trust.
The people served are the roughly 476,000 Americans treated for Lyme each year, the unknown number still searching for a name for their illness, and the clinicians who see them. A patient arrives with an organized, cited history instead of a shoebox; a clinician receives a standards-coded document instead of a story that starts from zero.
The platform also serves the public directly. It makes the surveillance undercount visible rather than smoothing it over: counties without data are labeled unmeasured, and the gap between treated and reported cases is stated plainly on the front page. That framing turns a private frustration into a legible public fact — one that patients, advocates, and local health departments can cite.
Everything public is free, requires no account, works on a phone, and reads in plain language. The measure of success is not engagement; it is a shorter, better-documented path from tick bite to appropriate care.
Data, Tech & Evidence Driven Approach
InvisibleSignal is built entirely from U.S. open data and published evidence, connected where they were previously fragmented:
- CDC NNDSS Lyme surveillance, state and county, 2018–2023: 152,018 cases reconciled to published totals, with per-state county-coding completeness carried as a first-class field.
- U.S. Census Bureau county population estimates as incidence denominators.
- CDC MMWR / NSSP emergency-department baseline for tick bites (2017–2019), transcribed with its caveats, giving the seasonal model an observed anchor.
- openFDA 510(k) device database: all 105 cleared Lyme/Borrelia diagnostics, searchable, linked to FDA records, and matched to extracted lab tests.
- NIH RePORTER for Lyme research funding context.
- A curated evidence corpus: 33 findings from 7 peer-reviewed and federal sources, each with DOI and a source-strength rating.
Technology choices serve trust. The stack is deliberately ordinary — Next.js, Supabase Postgres with row-level security, and a small FastAPI ML service — so a health department could operate it. Clinical data is coded to LOINC, SNOMED, and ICD-10 and exported as HL7 FHIR R4. Language-model features run behind a deterministic verification boundary with a cite-or-refuse policy: the model may propose, but only verified, source-anchored output reaches the user. A small on-device model (OpenMed, Apache-2.0) de-identifies pasted lab text in the browser before transmission. Retrieval over the evidence corpus is lexical and reproducible rather than embedding-based, so results are stable and auditable. Every AI output is benchmarked against golden-file evaluation suites in CI.
Execution, Efficiency & Delivery
The MVP is not a plan; it is running. The full user path — map, screener, cited report, lab records, FHIR export — is live in production and was demonstrated end-to-end at the Beta milestone. The platform has passed every sprint gate on schedule: Phase 1 concept, Alpha working software, Beta hardened product.
Delivery discipline is the differentiator. The repository carries more than one thousand automated tests, including guard tests that fail the build if marketing copy overstates a figure. Security was treated as a launch requirement, not a phase: authenticated and rate-limited APIs, row-level security on every health table, strict content-security and transport headers, and a PHI access audit log. An independent security audit's findings were remediated and re-tested.
Efficiency shows in the operating model. The data pipeline is scripts, not infrastructure: the annual CDC surveillance refresh is a re-run, not a rebuild, and every artifact carries machine-readable provenance metadata. Hosting is commodity serverless. There is no proprietary data dependency and no vendor lock-in that would strand users.
Near-term real-world utility already exists: a patient can use the public tools today with no account, and the clinician-facing export works with any EHR that ingests FHIR R4 documents. Nothing about deployment requires five to ten years of further research — it requires a waiting room.
Scalability, Reuse & Societal Impact:
Lyme is the wedge, not the wall. Every component generalizes: the ingestion scripts, the reconciliation discipline, the verification boundary, and the consent-first lab records flow transfer to other tick-borne and invisible-illness conditions — Alpha-gal syndrome, Babesia and other co-infections, and the post-infection syndromes that follow — by extending data constants, not rewriting software.
Alignment with federal interoperability frameworks is concrete, not aspirational. Clinical exports are HL7 FHIR R4 document bundles coded with LOINC, SNOMED, and ICD-10; the architecture page publishes a criterion-by-criterion self-assessment against the CMS Interoperability Framework, including what is not yet built. The data model follows USCDI conventions, and the export path fits TEFCA-era exchange and Blue Button 2.0-style patient-mediated access.
Reuse extends to government. A state or county health department can re-run the same open-data pipeline, mount the same map, and cite the same sources with no license and no procurement. The platform's discipline — floors labeled as floors, gaps labeled as gaps — is itself a reusable pattern for public-data products.
At national scale, making the Lyme undercount measurable is a policy input: surveillance funding, testing guidelines, and clinician education all depend on numbers the current system underreports. InvisibleSignal makes those numbers visible.
Jacqueline, InvisibleSignal is one of the most careful tools I’ve read in this sprint. Labeling counties CDC couldn’t code as unmeasured instead of zero, a limitations section titled “What this brief does not establish,” and an AI that must cite or refuse are exactly the habits this field needs, and the offline Field Mode shows real thought about where ticks are actually met.
Your 2023 figure also checks out against CDC’s file: 89,469 reported Lyme cases that year, set right beside the 476,000 diagnosed-and-treated estimate, which makes the undercount easy to see at a glance (chart attached). A detail your readers may find interesting: 85,938 of those 2023 cases are “probable” and 3,531 are confirmed, and in 2022 it was 60,357 probable and 2,192 confirmed, after the surveillance definition changed that year.
I carry the same two kinds of numbers in my entry, where each value is labeled as observed, assumed or projected. The claims-based estimate of 476,000 against about 35,000 reported through surveillance is simulator 8 (Kugeler et al., Emerg Infect Dis 2021). My entry is here if you’d like to look: https://invisibleillness.crowdicity.com/post/896844
I’d also be glad to hear Victoria’s view, since her One Health Lyme Gap Atlas also keeps missing data distinct from zero in county surveillance.
Thank you for making your code and data provenance public, and best of luck with the waiting-room deployments.
Users tagged:
Reply 0