About InvisibleSignal
Methodology, data sources, and scientific approach
Making invisible illnesses visible through AI and open federal data.
InvisibleSignal is an AI-powered surveillance and decision support platform that transforms fragmented federal health data into actionable intelligence for infection-associated chronic conditions (IACCs). These conditions -- including Lyme disease, Long COVID, ME/CFS, Alpha-gal Syndrome, POTS, and fibromyalgia -- affect over 36 million Americans yet remain systematically undercounted in health surveillance, undertreated in clinical settings, and underfunded in research budgets.
36M+
Americans affected
4.4 yr
Avg diagnostic delay
6
Conditions tracked
8
Federal data sources
| Dataset | Agency | What It Provides | Years |
|---|---|---|---|
NNDSS (National Notifiable Diseases Surveillance System) wonder.cdc.gov | CDC | Lyme disease and tickborne illness confirmed/probable case counts by state, year, and age group | 2018-2022 |
Household Pulse Survey (HPS) census.gov/data/experimental-data-products/household-pulse-survey.html | U.S. Census Bureau | Long COVID prevalence by demographics (age, sex, race/ethnicity, state, disability status, gender identity) | 2022-2024 |
Public Use Files (PUF) data.cms.gov | CMS | Medicare/Medicaid claims volume, ICD-10 billing code utilization, provider specialty, geographic distribution | 2018-2023 |
NCEI Climate Data (National Centers for Environmental Information) ncei.noaa.gov | NOAA | Monthly temperature, humidity, and precipitation by county for environmental risk modeling | 2018-2022 |
EnviroAtlas enviroatlas.epa.gov | EPA | Forest cover percentage, land use classification, green infrastructure metrics for tick habitat modeling | 2019-2021 |
RECOVER Initiative recovercovid.org | NIH | Long COVID clinical phenotypes, symptom clusters, biomarker data from multi-site cohort study | 2021-2025 |
HCUP (Healthcare Cost and Utilization Project) hcup-us.ahrq.gov | AHRQ | Emergency department visits, inpatient stays, and cost-to-charge ratios for infection-associated conditions | 2018-2022 |
NLCD (National Land Cover Database) mrlc.gov/data/nlcd-land-cover-conus-all-years | USGS/MRLC | 30m resolution land cover classification (deciduous/mixed forest, shrubland) for tick habitat suitability scoring | 2019-2021 |
Geographic Risk Index
0.882
Spearman ρ, out-of-sample
424
Counties evaluated
4
Published weights
0
Learned parameters
This block previously reported an AUC, an RMSE, ~54,000 training samples and five-fold cross-validation. None of that described anything that shipped — the training script built its target variable from its own features, and no split or calibration was ever performed. What runs is a deterministic index with four published weights, whose geographic component was validated against a held-out year with leave-one-state-out controls: see the model card.
Model Features
| Feature | Description |
|---|---|
incidence_log | Log-transformed CDC case incidence rate |
month | Calendar month (seasonal encoding) |
seasonal_risk | Composite seasonal tick activity score |
temperature_f | Mean monthly temperature (NOAA) |
humidity_pct | Mean relative humidity (NOAA) |
forest_cover | County-level forest cover % (NLCD/EPA) |
tick_habitat | Tick habitat suitability (derived from land cover) |
population_log | Log-transformed county population |
symptom_burden | Weighted symptom overlap score |
exposure_level | Composite environmental exposure index |
SignalBot: Clinical NLP Reasoning
SignalBot is powered by Anthropic Claude for clinical reasoning and natural language symptom analysis. It performs symptom pattern matching across 6 overlapping conditions (Lyme, Long COVID, ME/CFS, AGS, POTS, fibromyalgia), identifying shared and distinguishing symptom clusters. SignalBot provides informational guidance only and does not diagnose or replace clinical judgment.
SIMD-Optimized SoA Memory Layout
Risk scoring arrays use Structure-of-Arrays (SoA) layout instead of Array-of-Structures (AoS), enabling SIMD vectorization across patient cohorts. This allows the CPU to process 4-8 risk scores per instruction cycle.
4-8x throughput on risk batch scoring
Morton Z-Order Curve Spatial Indexing
Geographic coordinates are encoded as Z-order (Morton) curves, preserving spatial locality in a single 64-bit integer. This enables O(1) spatial lookups for county-level risk queries without tree traversal overhead.
O(1) spatial queries, cache-friendly geographic iteration
BLAS-Accelerated Similarity Search
Symptom pattern matching uses BLAS (Basic Linear Algebra Subprograms) matrix operations for computing cosine similarity across the 6-condition symptom space, leveraging hardware-optimized GEMM kernels.
Hardware-accelerated symptom matching at scale
Branch-Free Scoring via FMA Arithmetic
Risk score computation uses fused multiply-add (FMA) instructions with no conditional branches, eliminating branch misprediction penalties in the scoring hot path. Thresholds are encoded as arithmetic masks.
Zero branch mispredictions in scoring pipeline
No Personal Health Data
InvisibleSignal does not collect, store, or process any individual health records, PHI, or PII. All data is aggregate-level federal open data.
Open Federal Data Only
Every dataset is publicly available through federal agency portals (CDC, Census, CMS, NOAA, EPA, NIH, AHRQ, USGS). No proprietary or restricted data is used.
Informational, Not Diagnostic
All AI outputs -- including risk scores, symptom assessments, and SignalBot responses -- are informational tools for awareness and research. They do not constitute medical advice or clinical diagnosis.
WCAG 2.1 AA Accessible
The platform meets WCAG 2.1 Level AA standards including keyboard navigation, screen reader support, sufficient color contrast, and semantic HTML throughout all modules.
InvisibleSignal was built for the HHS TOPx Tech Sprint 2026, a federal innovation competition challenging teams to leverage open data and emerging technologies to address critical gaps in public health surveillance. InvisibleSignal was selected as a Phase 1 winner for its novel approach to making infection-associated chronic conditions visible through AI-driven analysis of federated open datasets.
Built by AI Cowboys