The goal is to detect counties with risk of transition to a higher risk status for Lyme. It uses two types of signals to make that risk assessment. One detects potential for newly emergent Blacklegged ticks and Western Blacklegged ticks in counties that were previously not at risk. The second detects potential for temporal shifts in human Lyme cases over time. Both forecasting systems are intended for use by health officials involved in monitoring, surveillance, and preventive efforts. As the distribution of Lyme carrying ticks emerge across new geographies, detecting these signals ahead of time can help with timely action, including issuing clinician and public awareness.
We conducted a stakeholder input call with a clinical epidemiologist who collaborates with clinicians and leads patient cohort and case-series research, and a public health data scientist who supports a Lyme dashboard. Both stakeholders use Lyme surveillance data to understand geographic trends, changing endemic areas, and disease burden.
They identified a central challenge motivating LymeSentinel –Lyme surveillance data are valuable, but are often released with a 1-to-2-year delay, limiting the ability of public health agencies, clinicians, researchers, and the public to understand current or emerging risk. They indicated that more timely data, ideally monthly during the seasonal Lyme period, could help identify when local risk is beginning, support clinical preparedness, and inform public awareness and prevention. This need is especially important in counties where tick populations and Lyme risk are newly emerging.
Stakeholders also emphasized that human surveillance data alone are incomplete. Lyme disease can be difficult to diagnose, early diagnostic tests may be unreliable, and reported cases depend on individuals seeking care and clinicians recognizing symptoms. Similarly, a county’s “No Records” tick classification may reflect limited surveillance rather than the absence of ticks. They highlighted the value of integrating complementary data sources, including tick surveillance, climate and environmental data, insurance claims, and electronic health records data. LymeSentinel directly addresses these needs by combining multiple data streams to forecast changes in tick establishment and Lyme risk ahead of delayed annual human case surveillance data.
A GIS map that shows current tick-status at the county level, alongside risk of transition to a higher risk-status in the year ahead. It uses CDC defined classification of tick-status, ‘No Reports’, ‘Reported’, ‘Established’. It also shows reported human case rates in each county, over time, up to latest year, and forecasts for the year ahead. Consensus risk score uses predictions from each model and signal type (ticks and human cases in current version) to rank counties by risk of transition.
The underlying methodologies informing the dashboard predictions will be an ensemble of models, each model built independently by different researchers using different methods, providing more confidence and stability in forecasts. We currently have 3 models for tick transition, and two models for human case shift. Our goal is for this to serve as a hub for scientific collaboration, where researchers can submit their models and forecasts. As Lyme cases are driven by interacting systems, including ecology, epidemiology, climate, and human behavior, forecasting Lyme involves a pipeline of models. We envision a future version where researchers can share stand-alone system modules (e.g., tick ecology, epidemiology of Lyme), for reuse by other researchers in their own modeling pipelines. A documentation page would document methodology, and data sources. The hub would also provide processed data files in a common format, for use by other models. For example, our compiled data was standardized with columns GEOID, Statename, County name, YEAR, Month, followed by columns specific to weather (temperature, precipitation,), land cover (cultivated, developed, etc), socioeconomic (income, poverty, etc). Over time, we envision that this compiled data, including new data potentially generated from other teams in this challenge, will facilitate innovative research directions for forecasting of Lyme and also other vector-borne diseases. The hub is designed as a living dashboard and allows for updates as new data and models become available. The data-driven models are expected to improve in accuracy and precision as more data becomes available. As a future direction, we expect to ensemble the results from the different models contributed to the hub, providing more confidence and stability in forecasts.
We use both tick status and human cases to asses risk, and an ensemble methods in each analysis. We use scientific machine learning that integrate robust math models with AI models, allowing to capture complex dynamics between climate and ecology, while accounting for data bias. We describe below the models driving the two complementary analysis. The first predicts tick-status and identifies counties with newly emergent risk. The second forecasts future cases of Lyme. The goal is for an ensemble analysis using multiple methods. We provide an overview of our current-status below, and hope to expand it in the future
Methods: Neuro Symbolic regression to predict tick status and forecast change in tick-status
Background: CDC’s tick surveillance system tracks and classifies tick status into three classes: No Records (No ticks or not documented), Reported (<6 ticks in 12-month period), and Established (6+ ticks in 12-month period). Publicly available data are at the county level and limited to years 1996, 2016, 2018, 2019, 2020, 2023. The goal is to train a model to predict tick-status for each year each county, to track changes over time, and forecast tick status for the year ahead. We summarized the data sources used in these models at the end of this document.
Methods: We trained three machine learning models, neural networks, random forests, and XGBoost, to predict tick status as the dependent variable using climate, land cover, terrain, and rural classification as independent variables. Surveillance data labels are incomplete, specifically, ‘No record’ could mean both ‘no-ticks’ or ‘no data’.
Results and Validation: We used F1 score as the performance metric. The F1 score for predicting tick-status was ~65%. Consensus transition risk score assigned to each county was based on each model’s probability of transition and number of models agreeing with these results. Symbolic regression converted blackbox models into interpretable mathematical expressions, to understand what features influenced the model predictions.
We describe two models
Data: We used climate, land cover, terrain, socio-economic factors, and tick-status as independent variables. Response as reported human case rate
We used NeuralCDE to train the model. We chose this method as the data is irregularly sampled, e.g., tick-status were samples for only 5 years between 1996 and 2023, different time intervals, e.g., temperature and precipitation are available as monthly averages, while data such as land cover and forest cover are at annual intervals, and controlled signals, e.g., potential changes in case definition.
Background: High-fidelity AI-based methods are undoubtedly effective for many tasks, including forecasting. However, training time can be considerable. This introduces difficulty in Lyme disease surveillance, where data are regularly updated, as the incorporation of such data, in principle, can require full model retraining. Reduced-order NeuralCDE accelerates this process. Training costs are reduced to a few minutes on conventional hardware, allowing for seamless integration of new/updated data, as well as enabling techniques like Monte Carlo analyses of the models themselves while remaining tractable computationally.
Concept: While relevant time series data across jurisdictions/variables are, in some sense independent, in practice we expect that processes underlying Lyme disease (for example, diffusion of tick and vector populations, precipitation/temperature, etc) to result in substantial correlation across indicators. Intuitively, this approach seeks to, rather than forecast for each quantity individually, to identify a small number of correlations underlying the outputs. Then, we learn and forecast just a handful of time series corresponding to each distinct correlation. For example, to forecast human case rates in 3,310 counties, instead of training on each county, we were able to extract 5 common correlation factors, and train/learn 5 time series rather than 3,310. This reduced training time to minutes. NeuralCDE allowed for accounting for missing and irregularly sampled data, and enable continuous projections.
Proof-of-concept/results: We used county-level annual data on human case rate, together with important indicators (temperature, participation, tick status, as well as changes in case definitions/reporting protocol). Training the model took approximately 4 minutes. Case rates per county for 2022-23 (not included in the training data) were predicted with an R-squared of .825 for 2022 and .757 for 2023.
| Data set | Contains | Years | File | Source |
|---|---|---|---|---|
| CDC tick status | Blacklegged tick (Ixodes scapularis) status per county: No records, Reported, Established (SCAP). Pathogen status (PAC) | 1996, 2016, 2018, 2019, 2020, 2023 | d40f1ac4-1873-489c-a23e-a7c3ecb54619.csv |
CDC Blacklegged Tick Surveillance |
| Lyme disease cases, population, rate | Reported cases, population and rate per county per year. | 2001 to 2025 | 434b261d-ddf9-434b-8c38-1f80344369f2.csv(ae9e9ddc-… is an identical copy) |
Internal / Provided |
| Data set | Contains | Years | File | Source |
|---|---|---|---|---|
| Temperature | Average temperature by county: 12 monthly values plus winter, spring, summer, fall. | 1988 to 2026 | Temperature_1988_2026.csv(built from dd29efcd-…csv for 1988 to 2020 and Temperature_2020_present.csv for 2020 to 2026) |
NOAA NCEI nClimDiv county data, file climdiv-tmpccy: Data Link / Format notes |
| Precipitation | Total precipitation by county: 12 monthly values plus winter, spring, summer, fall. | 1988 to 2026 | Precipitation_1988_2026.csv(built from 8efecc47-…csv and Precipitation_2020_present.csv) |
NOAA NCEI nClimDiv county data, file climdiv-pcpncy: Data Link |
| Data set | Contains | Years | File | Source |
|---|---|---|---|---|
| Land cover | Percent of each county in eight classes: barren, cultivated, developed, forest, herbaceous, shrubland, water, wetlands. | 1996 to 2025 | LandCover_1996_2025.csv(same data as LandCover_by_year.csv, wide format) |
Annual NLCD Collection 1.2 (USGS / MRLC) Download pattern config |
| Forest edge | Forest edge length, edge density, forest-developed edge and its density per county. | 1996 to 2025 | ForestEdge_by_year.csv(static versions: bffaab2e-…csv, de90bf9c-…csv) |
Calculated from the same Annual NLCD rasters. |
| Data set | Contains | Years | File | Source |
|---|---|---|---|---|
| RUCC (rural-urban continuum codes) | Code 1 to 9 per county, from metro to most rural. | 1996 to 2024 | RUCC_by_year.csv, RUCC_Population_through_2023.csv(source files: dfd2642a-…csv, rucc_2023/) |
USDA ERS Rural-Urban Continuum Codes |
| Population (1996) | County population estimate for 1996. | 1996 | census_pop_1990s/99c8_*.txt |
Census Bureau 1990s county estimates |
| Population (2001 to 2023) | County population and density per year. | 2001 to 2023 | Population_by_year.csv |
Population column of the Lyme disease case file above. |
| Data set | Years | File | Source |
|---|---|---|---|
| Elevation | N/A | 53dc02f9-4097-4c9a-9d75-b7c07dfb79b8.csv |
Johns Hopkins Lyme Data Explorer |
| Slope | N/A | 92e5e960-fcfd-46b1-bd87-be3c6bceae9d.csv |
Johns Hopkins Lyme Data Explorer |
| Dominant soil order | N/A | 4517c207-bd64-4327-b9f9-b3d6dba2a8cb.csv |
Johns Hopkins Lyme Data Explorer |
| Data set | Years | File | Source |
|---|---|---|---|
| Unemployment, median household income, education | 2014-2018 | 52c00715-...csv, 82521e17-...csv, dc053e66-...csv |
Johns Hopkins Lyme Data Explorer |
| Percent change in housing units | 2010-2018 | fdb88e00-...csv |
Johns Hopkins Lyme Data Explorer |
| Insurance claims | 2016 to 2023 | insuranceClaim.csv |
Johns Hopkins Lyme Data Explorer |
| County boundaries | N/A | county_boundaries/cb_2020_us_county_500k.shp |
US Census Bureau cartographic boundary file |
Federal and local public health officials can use LymeSentinel to monitor and review forecasts alongside existing surveillance systems, to communicate future risk with healthcare systems and general public. Public awareness will help prevention. Clinical awareness, especially in previously low risk counties, will help early diagnosis and surveillance reporting, which then feed into improved surveillance data and forecasts, creating a data feedback cycle for improved monitoring and prevention.
For example, currently the Centers for Disease Control and Prevention (CDC) monitors Lyme cases using national surveillance systems and four other data sources: insurance claims, SubLyme that integrates electronic health records, Tick Bite Data Tracker that shares emergency department data on visits for tick bites, Tick Surveillance Data Sets. These are all surveillance systems, tracking current and previous cases. LymeSentinel uses these datasets, and others (noted above) to forecast future risk. LymeSentinel improves prevention and early diagnosis by creating a data sharing loop: Forecasted risk can be communicated back to healthcare systems for clinician and public awareness. Increased awareness can improve surveillance data, to consequently improve forecasting. This data sharing and feedback loop to improve surveillance data, and forecasting, in the long-term, can improve prevention and early diagnosis.
Our work used many publicly available data sources. The list of data sources are included in a table above.
Our goal is to partner with federal partners and other modelers and datasources to 1) integrate surveillance data into our forecasting modeling workflow.2) integrate outcomes from our work into CDC systems workflow for both monitoring and disseminating risk assessment with healthcare systems and public awareness systems. Our group can also assist in updating model training with new data or modeling assistance to address data gaps.
While Lyme disease surveillance systems serve to document actual cases, LymeSentinel projects future cases. As the distribution of Lyme carrying ticks emerge across new geographies, detecting these signals ahead of time can help with timely action, including issuing clinician and public awareness.
Forecasted risk can be communicated back to healthcare systems for clinician and public awareness. Increased awareness can improve surveillance data, to consequently improve forecasting. This data sharing and feedback loop to improve surveillance data, and forecasting, in the long-run, will improve prevention and early diagnosis.
It currently forecasts two signals, tick status transitions using three ML models, and temporal shifts in human cases using two scientific machine learning models. Our goal is for this to serve as a hub for scientific collaboration, where researchers can submit their models and forecasts to improve ensemble analysis.
More intricate models of the interacting systems, such as ticks ecology, Lyme epidemiology, climate, and human behavior can help improve Lyme forecasting. We envision a future version where researchers contribute these independent system models, that can then be used as modules in the forecasting model pipeline to forecast Lyme and other vector-borne diseases.
Our goal is to serve a hub for forecasting models. These models need regular data stream .Our goal is to reach out to partner with other modelers and data sources to use and link models on LymeSentinel to their compiled and processed data. As some of the data are not in a monthly resolution, our goal is to provide a curated monthly dataset, by using modeling where raw data are unavailable. Over time, we envision that this compiled data, including new data potentially generated from other teams in this challenge, will facilitate innovative directions to improve Lyme risk forecasting, prevention, and early diagnosis.