WildAlert: A Real-Time, AI-Driven Early Warning System for Wildlife Health and Ecological Threat Detection

WildAlert is a scalable, AI-driven early warning system that integrates natural language processing and anomaly detection on real-time wildlife rehabilitation data to proactively identify diverse ecological and health threats, such as zoonotic diseases and environmental hazards, often before they are confirmed by traditional surveillance methods.

The Gist

Imagine the natural world as a giant, complex orchestra. Usually, the music is harmonious: birds sing, turtles swim, and mammals roam. But sometimes, a single instrument starts playing out of tune, or a section of the orchestra goes silent. In the past, we only noticed these problems after the music had completely stopped and the damage was done.

WildAlert is like a super-smart, 24/7 sound engineer sitting in the control room, listening to the orchestra in real-time. Its job is to hear that first "off-key" note and sound the alarm before the whole symphony collapses.

Here is how this paper explains that system in simple terms:

1. The Problem: We Were Listening to Static

Wildlife rehabilitation centers (places that nurse injured animals back to health) are like the front-line nurses of the animal kingdom. They see sick animals every day. However, for a long time, the data they collected was like a messy pile of handwritten notes.

• The Old Way: Scientists had to read these notes one by one to find patterns. It was slow, like trying to find a needle in a haystack by looking at one straw at a time.
• The Gap: By the time a scientist realized, "Oh no, there's a disease outbreak," it was often too late. The animals were already dead, or the disease had spread.

2. The Solution: The "Super-Reader" and the "Pattern Detective"

The researchers built WildAlert 2.0, a system with two main "brains" working together:

Brain A: The Super-Reader (The NLP Model)

Think of this as a robot librarian who has read every book in the universe.

• What it does: It takes the messy, unstructured notes from wildlife centers (e.g., "Bird found on road, shaking, can't fly") and instantly translates them into clean, organized categories (e.g., "Neurologic Disease," "Vehicle Collision").
• The Magic: It uses a type of AI called BERT (which is like a very advanced version of a spell-checker that understands context). It doesn't just look for keywords; it understands the story behind the words. It can tell the difference between a bird that is "clinically healthy" and one that is "sick but acting normal."

Brain B: The Pattern Detective (The Anomaly Detector)

Once the notes are organized, this brain looks at the timeline.

• The Analogy: Imagine a river. Usually, the water flows at a steady pace. Sometimes it rains, and the water rises a little (seasonal). Sometimes a rock falls in, causing a small ripple (random noise).
• The Job: This detective knows exactly what the river should look like. If the water suddenly turns into a tsunami, or if the river suddenly runs backward, the detective screams, "Something is wrong!"
• The Tools: It uses two different methods to be sure:
  1. Isolation Forest: Like a bouncer at a club who kicks out anyone who looks different from the crowd. It's great at spotting sudden, loud spikes.
  2. Autoencoder: Like a musician who knows the song so well they can hear when a single note is slightly off, even if the song is changing tempo. It's great at spotting slow, creeping changes.

3. The Results: Catching the "Silent Killers"

The paper shows that WildAlert didn't just work in theory; it caught real disasters in the wild.

• The "Ghost" Flu: It spotted signs of Highly Pathogenic Avian Influenza (HPAI) in birds. In some cases, the system flagged an anomaly before the government labs officially confirmed the virus was there. It was like hearing a cough before the fever broke.
• The Red Tide: In Florida, the system noticed a strange spike in seabirds acting weird (neurologic disease). It turned out to be Brevetoxicosis (poison from toxic algae). The system helped scientists realize the algae bloom was happening earlier than expected.
• The Frozen Turtles: It detected a massive event where over 1,000 sea turtles were "cold-stunned" (frozen by cold water) in Florida, triggering a massive rescue effort.
• The Distemper Outbreak: It spotted a cluster of raccoons and skunks acting sick, leading to the discovery of a Canine Distemper outbreak.

4. Why This Matters: The "Canary in the Coal Mine"

For decades, we've used birds as "canaries in the coal mine" to detect danger. But usually, the canary dies before we realize the mine is unsafe.

WildAlert changes the game. It acts as a digital early warning system.

• It's Fast: It processes data in real-time, not months later.
• It's Broad: It watches everything from rare eagles to common raccoons.
• It's Connected: It links animal health to human health and the environment. If the animals are sick, it often means the water is poisoned, the air is bad, or a new virus is jumping species.

The Bottom Line

WildAlert is like a smartwatch for the entire ecosystem. Just as your smartwatch tells you if your heart rate is too high before you feel sick, WildAlert listens to the "heartbeats" of wildlife rehabilitation centers. When it hears a rhythm that doesn't fit, it alerts scientists immediately, giving them a chance to stop a disaster before it spreads to humans, livestock, or the rest of the natural world.

It turns a pile of messy paper notes into a crystal ball for saving nature.

Technical Summary

1. Problem Statement

Emerging infectious diseases and environmental disruptions pose escalating threats to biodiversity and public health. Free-ranging wildlife often serve as early indicators of ecological instability, yet existing surveillance systems suffer from critical limitations:

• Lack of Real-Time Capability: Most animal disease surveillance relies on retrospective analysis or manual review, hindering rapid response.
• Data Heterogeneity: Participatory surveillance systems (e.g., wildlife rehabilitation centers) generate vast amounts of unstructured clinical data that are difficult to standardize and share due to varying data parameters and collection methodologies.
• Analytical Gaps: Existing platforms lack robust, scalable algorithms capable of processing unstructured text and distinguishing true anomalies from background noise in non-stationary, seasonally influenced time-series data.
• Limitations of Previous Versions: The initial WildAlert system relied on simple "bag-of-words" NLP (losing semantic context) and basic moving averages for anomaly detection, which failed to capture complex patterns or gradual shifts in disease baselines.

2. Methodology

The authors developed WildAlert 2.0, an AI-driven platform that integrates Natural Language Processing (NLP) with unsupervised anomaly detection to process real-time data from the Wildlife Rehabilitation Medical Database (WRMD). The system is currently deployed across five U.S. states (CA, FL, AZ, PA, WA) involving 182 rehabilitation organizations.

A. Natural Language Processing (NLP) Module

• Architecture: Utilizes a fine-tuned DistilBERT model (a transformer-based Large Language Model) to capture contextual relationships between terms, replacing the previous bag-of-words approach.
• Tasks:
  1. Clinical Classification: Multi-label classification of unstructured admission notes into 11 clinical categories (e.g., Physical Injury, Neurologic Disease, Nutritional Disease).
  2. Circumstances of Admission: Classification of 74 distinct admission scenarios (e.g., gunshot, vehicle collision, plane collision).
• Training: The model was trained on 15,636 labeled observations for clinical signs and 35,844 for circumstances, using 10-fold cross-validation and class weighting to handle imbalanced data.

B. Anomaly Detection Framework

The system employs a hybrid computational framework to detect statistical anomalies in time-series data (species + syndrome + location).

1. Preprocessing: Weekly admission time series are decomposed into Trend, Seasonal, and Residual components using additive seasonal decomposition. Anomalies are detected in the residual component to isolate non-systematic fluctuations.
2. Dual-Model Approach:
   • Autoencoder (AE): A feedforward neural network trained to reconstruct residual time-series windows. Anomalies are flagged when the reconstruction error exceeds the 95th percentile threshold. This model is robust to gradual baseline drifts and seasonality.
   • Isolation Forest (IOF): A tree-based ensemble algorithm that isolates outliers in a one-dimensional feature space. This model is highly sensitive to abrupt, high-impact spikes.
3. Ensemble Strategy: The system leverages both models to balance sensitivity to abrupt events (IOF) with robustness to evolving baselines (AE).

C. Validation Strategy

• Retrospective Validation: Compared WildAlert anomalies against independently confirmed events (USDA HPAI data, state wildlife agency reports).
• Sensitivity Analysis: Simulated data perturbations (stochastic outliers, linear trends, varying contamination levels) to test model robustness across endemic and sporadic reporting targets.
• Ecological Correlation: Cross-correlation analysis between bird admission anomalies and Karenia brevis (Harmful Algal Bloom) cell counts to assess temporal lags.

3. Key Results

NLP Performance

• Clinical Classification: Achieved a macro-averaged F1-score of 0.92. Performance was highest for "Clinically Healthy" (0.97) and "Physical Injury" (0.94).
• Circumstances: Precision and recall exceeded 0.95 for most categories, with perfect F1-scores (1.00) for distinct events like "gunshot" and "plane collision."
• Conclusion: The BERT-based model successfully extracts structured, biologically meaningful data from unstructured text, enabling scalable syndromic surveillance.

Anomaly Detection Performance

• Model Agreement: The overlap between AE and IOF detections was moderate (~19%), with a Cohen's Kappa of ~0.25, indicating the models detect different types of anomalies (complementary strengths).
• Sensitivity:
  • IOF: Extremely sensitive to contamination levels (injection rate) but less responsive to gradual trends. Best for detecting abrupt spikes.
  • AE: More balanced, showing robustness to non-stationary signals and gradual baseline shifts.
• Hybrid Utility: Combining both models enhances the detection of diverse anomaly types.

Real-World Event Detection (2024–2025)

WildAlert successfully detected and flagged multiple confirmed events, often coinciding with or preceding official confirmation:

• Highly Pathogenic Avian Influenza (HPAI): Detected anomalies in raptors and shorebirds in CA and FL during USDA-confirmed outbreaks. In FL waterfowl, IOF showed statistically significant signal elevation during HPAI periods.
• Harmful Algal Blooms (HABs): Detected neurologic disease alerts in seabirds and shorebirds in FL, later confirmed as brevetoxicosis. Cross-correlation analysis revealed species-specific lag times (e.g., Ruddy Turnstones lagged HABs by ~1 week; Sooty Terns by ~2–3 weeks).
• Other Events:
  • Cold-stunning: Detected a massive stranding of >1,000 sea turtles in FL (Jan 2025).
  • Canine Distemper: Flagged neurologic/dermatologic cases in mesocarnivores (skunks, raccoons) in CA (Feb 2025).
  • Domoic Acid Toxicosis: Detected multi-species seabird strandings in Southern CA (March 2025).
  • Starvation: Identified nutritional disease spikes in Brown Pelicans in CA (April 2024).

4. Key Contributions

1. First Large-Scale AI Wildlife Surveillance: Established the world's largest standardized, near real-time wildlife health surveillance system, transforming unstructured rehab records into actionable intelligence.
2. Advanced NLP Integration: Demonstrated that transformer-based LLMs (DistilBERT) can effectively structure complex, unstructured veterinary clinical data, overcoming the limitations of traditional keyword-based methods.
3. Hybrid Anomaly Detection: Validated a dual-model approach (Autoencoder + Isolation Forest) that effectively handles both abrupt spikes and gradual baseline shifts in ecological time-series data.
4. One Health Framework: Provided a scalable model linking biodiversity monitoring, zoonotic disease preparedness, and environmental hazard response, proving that wildlife rehab data can serve as an early warning system for public health and ecosystem threats.

5. Significance

WildAlert 2.0 represents a paradigm shift from retrospective, reactive wildlife monitoring to proactive, real-time surveillance. By automating the detection of anomalies across taxa and regions, the system enables:

• Early Intervention: Agencies can investigate outbreaks (e.g., HPAI, HABs) days or weeks before traditional reporting methods confirm them.
• Resource Optimization: Targeted investigations are triggered by data-driven alerts rather than random sampling.
• Scalability: The framework is transferable to other regions and species, offering a blueprint for global biodiversity monitoring and pandemic preparedness at the human-animal-environment interface.

The study confirms that integrating deep learning with participatory data streams can significantly enhance the timeliness and accuracy of ecological threat detection, ultimately supporting conservation and public health efforts.