GeoAI-LYME Early Care Navigator: Lyme AI-ert

by Jisung Chang | 1 week ago | in Lyme Innovation: TOPx Submission
Approved

GEBA Consulting LLC developed GeoAI-LYME Early Care Navigator: Lyme AI-ert, an operational web and mobile platform for location-aware Lyme alerts and early-care navigation. Our interdisciplinary team includes Dr. Jisung Chang (project lead, GeoAI and data science), Renah Kim (care-support communication), Joseph Chang and Joshua Chang (student interns supporting web and mobile app development), and Issac Hyesung (design). The platform brings together three complementary AI approaches: GeoAI uses CDC surveillance and environmental data to predict annual Lyme cases; TrendAI adds monthly and weekly regional search-interest signals; and VisionAI provides non-diagnostic rash-pattern screening. GeoAI and TrendAI are combined to generate weekly county-level alerts on a 0–5 scale. Users can enter a ZIP code or county name, explore interactive maps and archived alerts, access symptom checklists and trusted Lyme information, and find nearby-care resources. The goal is to turn U.S. open data into understandable, actionable Lyme alerts and early-care information—not to diagnose individuals.

Solution Overview (MVP Demonstration):

User-facing Application

Live demo, prototype, or working product

Live demo URL: https://gebaconsulting.com/lyme

Access instructions (if needed)

No login is required. Explore the interactive Lyme alert map or enter a ZIP code or county name to view local alerts. Additional features include high-risk counties, VisionAI rash-pattern screening, Lyme information, and nearby care. The website also provides Google Play and Apple App Store links to download the smartphone app.

Results for the American People:

Lyme AI-ert helps people across the United States turn open data into practical next steps through visual Lyme alerts and early-care information. It provides weekly lyme alerts for counties across the contiguous United States, showing local alert levels and changes over time. The primary users are people with possible tick exposure, rash concerns, or Lyme-like symptoms, along with families and caregivers asking, “Should I be concerned, and what should I do next?” A typical user begins by searching online for information about a tick bite or symptoms. After opening Lyme AI-ert, the user enters a ZIP code or county name to check alerts for their area or a place they recently visited. The interactive map displays geographic alert patterns, while the archive shows changes over time. Users can then access VisionAI for non-diagnostic rash-pattern screening, review an external symptom checklist, find nearby care, and explore public-health resources in one place. This workflow is designed to help users gather relevant information, prepare questions for a healthcare professional, and consider appropriate next steps. Available through a website and mobile apps on Google Play and the Apple App Store, Lyme AI-ert connects online searching, local awareness, and early-care guidance. County alerts provide community-level context, not an individual’s infection probability, and VisionAI does not provide a diagnosis.

Data, Tech & Evidence Driven Approach

Most Lyme surveillance data describe past cases. Lyme AI-ert combines surveillance records, environmental data, and recent search trends to generate current-year predictions and weekly county-level alerts. GeoAI uses CDC county-level Lyme case counts as the target and environmental and Earth observation data from NASA, USGS, USDA, and NOAA as predictors. CDC/NNDSS surveillance provides population-level context, not evidence of an individual's infection or diagnosis. Inputs are aligned by county and time period. Indicators of tick habitat and human exposure support a machine-learning model predicting annual Lyme cases. Five-fold random cross-validation and state-holdout validation yielded R² values of 0.86 and 0.72, respectively, on log-transformed case counts. Monthly cutoffs assessed early prediction of full-year cases. The 2026 predictions and annual changes provide the baseline. TrendAI analyzes Lyme-related search trends from 2004 to the present. Seasonal alignment and lagged correlations characterize relationships between search patterns and reported cases, without treating search interest as evidence of infection. Monthly anomalies capture departures from seasonal patterns; one- and two-week changes capture recent shifts. The platform fuses the annual GeoAI baseline with monthly and weekly TrendAI signals into a weekly county-level alert on a 0–5 scale. Alert levels indicate community alert priority, not individual infection probability or diagnosis. A manuscript describing the framework was submitted to a Q1-ranked journal in September 2026 and is under peer review. VisionAI uses a lightweight MobileNet model trained on Lyme rash images and the SCIN dataset for non-diagnostic rash-pattern screening, supporting fast responses on web and mobile.

Execution, Efficiency & Delivery

Lyme AI-ert is already live as a website and mobile apps available on Google Play and the Apple App Store. It provides weekly Lyme alerts for counties across the contiguous United States. Users can check local alerts and access early-care information in one place. By entering a ZIP code or county name, users can view alerts for their area or a place they recently visited. The interactive map shows geographic patterns, while weekly updates and the Alert Archive show changes over time. This provides local context for concerns about tick exposure or Lyme-like symptoms. VisionAI offers non-diagnostic rash-pattern screening, alongside an external symptom checklist, public-health information, and nearby-care resources. These features are designed to help users organize their concerns, prepare questions for a healthcare professional, and consider appropriate next steps. The immediate efficiency benefit is easier access to relevant information. Users can view combined results without collecting and interpreting separate surveillance reports, environmental datasets, and search trends. The operational MVP supports a practical workflow: check local alerts, review symptom-related information, and prepare for a healthcare conversation. Its real-world usability and impact will be evaluated through user feedback and pilot testing.

Scalability, Reuse & Societal Impact:

We plan to maintain Lyme AI-ert for the next three years (2027-2029), providing weekly county-level alerts across the contiguous United States from March through November. The existing website and mobile apps provide a practical channel for continued delivery. The AI framework has the potential to scale to other tick-borne diseases and broader geographic regions. Transfer learning could adapt the GeoAI model, while the data-processing pipeline, TrendAI approach, and mapping interface provide reusable components. Each extension would require disease-specific surveillance data, relevant environmental predictors, sufficient search signals, and validation before public deployment. We are also developing a 1-km mapping service for public-health agencies, subject to additional fine-scale validation. A key assumption is that CDC surveillance data adequately represent geographic and temporal patterns in reported Lyme cases for model training and evaluation. Continued operation also depends on timely CDC surveillance updates and access to environmental and search-trend data. Changes in data coverage or reporting practices must be considered when updating models and interpreting alerts. Our next steps are to collect user feedback, improve the clarity and usefulness of alerts and early-care resources, and expand research and outreach partnerships with public-health and Lyme community organizations. These collaborations can help evaluate real-world use, reach more users, and identify priorities for other tick-borne diseases. Partnerships could adapt the platform for underserved populations and support local public-health planning. The intended societal benefit is sustained access to understandable, location-aware information that supports earlier awareness and informed care-seeking across the United States.

Files
final mvp geoai lyme
Idea image
edited on Oct 24, 2026 by Jisung Chang
Public (1)
You will need to login to post a comment
Uri Kartoun 20 hours ago

Jisung, what stands out in Lyme AI-ert is the validation: holding out entire states, which gave an R² of 0.72 alongside 0.86 for random folds, is a much harder and more honest test than random splitting. I also appreciate that search interest is used as a signal about attention and never as evidence of infection, and that the alerts say they describe community priority, not individual risk.

You state your key assumption plainly: that CDC surveillance adequately represents where and when reported Lyme occurs. One published comparison bears on it. Kugeler et al. (Emerg Infect Dis 2021) estimated from insurance claims that 19% of Lyme diagnoses were in low-incidence states, against about 5% in surveillance (chart attached). The authors say the gap mixes under-reporting with possible overdiagnosis and the data can’t separate the two, so it’s a point to weigh, not a verdict. But it suggests that where surveillance and claims diverge, alert levels in low-incidence states deserve extra care, and that a claims-based layer could serve as a second reference point for your 1-km mapping work.

My entry turns published papers into simulators you can change and check against federal data, so a reader can see how each number moves: https://invisibleillness.crowdicity.com/post/896844

Lisa’s CAIRN also validates with whole-state holdouts and reports a held-out Spearman of 0.62, and Chaitra’s LymeSentinel forecasts county-level cases with held-out R-squared of 0.825 and 0.757. Lisa, Chaitra, I’d be glad to hear how your validation approaches compare with Jisung’s.

Wishing you a successful three seasons of alerts.

Users tagged:

Avatar for Jisung Chang Jisung Chang
Avatar for Jisung Chang Jisung Chang
Avatar for Lisa Yang Lisa Yang
Avatar for Chaitra Gopalappa Chaitra Gopalappa

Reply 0

Jisung Chang 20 hours ago

Uri, thank you for the thoughtful feedback and for sharing the Kugeler et al. paper.

I agree that the divergence between CDC surveillance and diagnosis patterns, particularly in low-incidence states, is an important limitation to consider. In Lyme AI-ert, I treat CDC surveillance not as ground truth for the full burden of infection, but as a public-health reference for where and when reported Lyme disease occurs.

GeoAI is designed to learn broader spatial relationships among tick habitat, environmental and climate conditions, human exposure-related characteristics, and reported Lyme patterns, and then test whether those relationships generalize to entirely unseen geographic areas. That is why I included whole-state holdouts in addition to random cross-validation. TrendAI addresses a separate question by examining temporal changes and emerging trends.

I therefore think claims data could be especially valuable as an independent external reference, rather than simply being combined with CDC surveillance as another training target. Areas where surveillance, claims, and GeoAI signals diverge may themselves be informative.

CAIRN and LymeSentinel also use interesting validation approaches. Since our prediction targets, spatial/temporal questions, and metric.s differ, I think the most useful comparison may be how each approach tests generalization rather than directly comparing R² or correlation values.

Thank you again for raising this perspective and opening the discussion

Reply 0

Uri Kartoun 19 hours ago

Jisung, thank you for such a thoughtful answer. I like your framing: treating claims data as an independent external reference, not as another training target, keeps each signal honest, and places where surveillance, claims and GeoAI diverge may be exactly where the most useful questions are. I also agree that the fairest comparison across entries is how each one tests generalization to places and times it hasn’t seen, not the metric values. Lisa, Chaitra, I’d be glad to hear how you each think about that, since your entries both hold out data in their own way.

Users tagged:

Avatar for Jisung Chang Jisung Chang
Avatar for Jisung Chang Jisung Chang
Avatar for Lisa Yang Lisa Yang
Avatar for Chaitra Gopalappa Chaitra Gopalappa

Reply 0

Lisa Yang 10 minutes ago (edited)

Uri, Jisung, thank you for the thoughtful framing. 

Validation discipline and generalization are fundamentally intertwined in how CAIRN is designed, so the two questions belong together.
CAIRN and Lyme AI-ert share the same core holdout discipline: whole-state blocks (GroupKFold), ensuring neither model is evaluated on counties adjacent to its training data. Where the systems diverge is the operational question each is optimizing for, which dictates both metric choice and how generalization is managed across the larger system.
GeoAI predicts magnitude: how many reported Lyme cases a county should expect in a year. R-squared is the appropriate metric because the operational difference between estimating 200 cases versus 20 matters directly for an alert system. CAIRN's area model optimizes for rank order, determining whether a geography's Lyme burden is higher or lower than another. Spearman rho is the right metric because the output is an ordinal band (lower, moderate, higher) rather than a continuous count. At rho 0.62 in held-out states, the model reliably distinguishes high-burden from low-burden geographies it never saw during training. Adding real local tick-establishment data lifts that from 0.46, showing the tick signal performs genuine predictive work beyond regional climate and forest features alone.
To handle generalization cleanly, CAIRN structurally decouples the system into two layers with distinct validation requirements.
The clinical read is a direct implementation of published CDC and IDSA guidelines. Every rule that fires is displayed on screen and applied identically for every user regardless of location. This layer does not need to generalize statistically because it is the published standard; its validation is already encoded in decades of peer-reviewed clinical consensus from the agencies that produced it.
The area model serves strictly as supporting environmental context, not a standalone alert or clinical input. As the Kugeler point highlights, any empirical ranking derived from CDC surveillance data inherits systemic artifacts, including under-reporting that varies state by state. CAIRN addresses this structurally: rather than producing a point estimate the data cannot support, the area model reports a calibrated range and an ordinal tier. On held-out states, that range contained the true local Lyme rate about 9 times in 10, and the tier matched the true activity level at the same rate. The methodology, including log-scale and raw-scale R-squared figures, is in the technical section of CAIRN's model card page.

Reply 1

Uri Kartoun 2 minutes ago

Thank you, Lisa. That’s a clear explanation of how CAIRN separates the clinical read from the area model, and I appreciate you taking the time.

Reply 0