🤖 machine learning

HantaWatch: Federated Learning for Hantavirus Genomic Surveillance

HantaWatch is a federated learning framework that enables decentralized Hantavirus genomic surveillance by collaboratively training models across laboratories without sharing raw data, thereby addressing data heterogeneity and expert capacity constraints while providing prioritized risk scores for expert review.

Original authors: Shanika Iroshi Nanayakkara, Shiva Raj Pokhrel

Published 2026-07-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shanika Iroshi Nanayakkara, Shiva Raj Pokhrel

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where a tiny, invisible spy is hiding in the woods, jumping from rodent to human, causing serious sickness. This spy is the Hantavirus. Scientists need to keep a close watch on it, but there's a catch: the virus changes its "costume" (its genetic code) constantly, and the people trying to catch it are scattered all over the globe. Some are in big city labs, others in remote clinics, and they can't just hand over their secret files to a central boss because of privacy rules and safety concerns. It's like trying to solve a giant puzzle where every piece is locked in a different safe, and no one can open their safe to show the others. To solve this, scientists use a clever trick called "Federated Learning." Think of it like a group of students taking a test. Instead of everyone bringing their notebooks to the teacher to be copied, the teacher sends a single textbook to each student. The students study their own pages, write down their own answers, and send only the answers back. The teacher then combines the answers to make a better textbook for the next round, without ever seeing the students' private notes. This paper is about building a super-smart version of this system specifically for Hantavirus, so experts can spot dangerous outbreaks faster without breaking any privacy rules.

The researchers behind this project, Shanika Iroshi Nanayakkara and Shiva Raj Pokhrel, created a new tool they call HantaWatch. Their main goal was to solve a tricky problem: when the "students" (the different labs) have very different types of virus data, the teacher (the central computer) often gets confused and makes mistakes. Sometimes, it misses the really dangerous virus strains because the data looks too weird compared to what it's used to. HantaWatch is designed to be a "decision-support layer," which is a fancy way of saying it's a smart assistant that helps human experts decide which virus samples to look at first. It doesn't replace the doctors or scientists; instead, it acts like a traffic cop, pointing a red flag at the most suspicious samples and a green light at the boring ones.

Here is how HantaWatch works, broken down into its coolest parts:

The "Smart Traffic Cop" System
Imagine a massive pile of virus letters (genomic data) arriving from different labs. HantaWatch doesn't just read them; it sorts them into a priority list. It asks three big questions: "Is this dangerous?" "Are we sure about this?" and "Does this look like an outbreak?" If a sample looks risky, uncertain, or weird, HantaWatch flags it for an expert to review immediately. If it looks like a routine virus, it gets sent to the back of the line. This saves experts from drowning in paperwork and helps them catch the bad guys faster.

The "Adaptive Coach" (DU-FedProx)
One of the biggest headaches in this system is that the labs are all different. One lab might mostly see a specific type of virus, while another sees a totally different one. In a standard system, this causes the "teacher" to get confused, leading to unstable learning. HantaWatch introduces a special "Adaptive Coach" called DU-FedProx. Think of this coach as a personal trainer for each student. If a student is struggling with a specific topic, the coach doesn't just yell "try harder!" Instead, the coach adjusts the training plan: maybe slowing down the learning speed, adding extra practice, or changing the focus. In the paper's experiments, this coach helped the system stay steady and accurate, even when the data was messy and uneven.

The "Safety Check" Before Deployment
The authors are very careful not to let a bad model run the show. Before HantaWatch lets any model start sorting virus samples, it runs a strict "Safety Check." It doesn't just look at how many answers were right (accuracy); it looks at how many dangerous viruses were missed (false negatives). If a model is good at guessing but misses the scary stuff, HantaWatch says, "Nope, not ready yet." This ensures that the system only recommends models that are truly reliable for public health.

What They Found
The team tested HantaWatch with two different scenarios. First, they simulated a messy classroom where every student had different books. Second, they used real-world data from different sources, like different countries or research groups.

The Results: In many cases, HantaWatch's "Adaptive Coach" (DU-FedProx) did a great job of keeping the system stable and reducing errors. For example, in spotting high-risk viruses, it achieved a perfect "recall" score of 1.000, meaning it didn't miss a single dangerous case in their tests.
The Catch: The system isn't perfect for every single task. When the virus had too many different "costumes" (many different types to classify at once), the adaptive coach sometimes got a bit too cautious and slowed things down. The paper suggests that while the tool is very promising, it needs more tuning for these super-complex situations.
The Verdict: HantaWatch successfully turned a chaotic pile of virus data into a clear, ranked list for experts. It proved that you can train smart AI models across many different labs without sharing private data, and that this system can effectively tell experts, "Hey, look at this one first!"

In short, HantaWatch is a practical, privacy-friendly tool that helps the world's health experts stay one step ahead of Hantavirus. It doesn't make the final decisions—that's still up to the humans—but it gives them a powerful, organized way to see the danger before it spreads. The authors suggest that with a bit more work to handle the most complex virus types, this could become a standard part of how we watch out for dangerous outbreaks in the future.

Technical Summary: HantaWatch

Problem Statement

Hantavirus genomic surveillance faces critical operational bottlenecks: sequence data is distributed across disparate laboratories and repositories, creating privacy and governance barriers to centralization. Furthermore, surveillance data is inherently non-IID (non-independent and identically distributed) due to variations in host distributions, viral clades, genome segments, and regional transmission patterns. Standard centralized machine learning is often infeasible due to data sovereignty constraints, while existing federated learning (FL) approaches (e.g., FedAvg, FedProx) often struggle with client drift, unstable local updates, and aggregation mismatches in such heterogeneous environments.

Crucially, operational surveillance requires more than just classification accuracy. Limited expert-review capacity necessitates a system that can prioritize which newly submitted sequences require immediate human inspection. Existing tools lack an integrated framework that combines federated training under non-IID conditions with uncertainty-aware decision support to triage records for expert review.

Methodology

HantaWatch is proposed as a federated learning framework designed for decentralized Hantavirus sequence screening and expert-review prioritization. It operates as a decision-support layer rather than an autonomous diagnostic system. The methodology comprises seven primary stages:

Data Harmonization & Feature Extraction: Raw nucleotide sequences are converted into fixed-length numerical representations using $k$ -mer feature extraction (specifically $k=4$ ), creating alignment-independent vectors suitable for machine learning.
Source-Aware Client Construction: The framework models realistic heterogeneity by constructing federated clients based on data provenance (e.g., specific laboratories or repositories) rather than random partitioning. This captures variations in host types, clade compositions, and metadata completeness.
Adaptive Federated Optimization (DU-FedProx): To address client drift in non-IID settings, HantaWatch introduces a Deep-Unfolded FedProx (DU-FedProx) controller. Unlike standard FedProx which uses fixed hyperparameters, this controller dynamically maps client states (previous loss, gradient norm, update norm, class imbalance, and round index) to adaptive local training controls:
- Local learning rate ( $\eta$ )
- Proximal coefficient ( $\mu$ )
- Number of local epochs ( $E$ )
  Two controller variants are evaluated: Single-level (state-based) and Bilevel (optimizing for post-aggregation surveillance performance).
Surveillance-Specific Model Selection: Models are not selected based on accuracy alone. Instead, a suitability layer evaluates candidates using metrics critical to public health: recall, false-negative rate (FNR), F1-score, AUC, and update stability. A model is only deployed if it meets specific suitability thresholds.
Prediction-Only Triage: Once a suitable model is selected, it processes new, unlabelled sequences in "prediction-only" mode. The system generates:
- Predicted labels and class probabilities.
- Risk scores and confidence estimates.
- Uncertainty flags and review rationales.
- A ranked priority list for expert review (e.g., High-Priority, Uncertainty, Routine).

Key Contributions

The paper outlines the following specific contributions:

Formulation: Defining Hantavirus genomic surveillance as a decentralized FL problem focused on expert-review prioritization without centralizing raw data.
Framework: Introducing HantaWatch, which produces risk scores, uncertainty flags, and ranked priorities alongside predictions.
Realistic Heterogeneity: Designing source-aware non-IID client construction that reflects real-world variations in hosts, clades, and metadata quality.
Adaptive Optimization: Proposing the DU-FedProx controller to dynamically adjust learning rates and regularization to mitigate client drift.
Surveillance Metrics: Developing a model selection layer that prioritizes recall, FNR, and stability over raw accuracy to prevent missed high-risk cases.

Experimental Results

The framework was evaluated across binary (high-pathogenicity screening, outbreak association) and multiclass (clade classification, clinical syndrome categorization) tasks under two settings: controlled label-skew partitions and source-representative splits.

Binary Tasks: In high-pathogenicity screening, DU-FedProx-Single achieved the best balance, attaining perfect recall (1.000) and zero FNR while maintaining the lowest update norm (0.160), indicating superior stability. In outbreak-associated screening, FedAvg showed the highest predictive metrics, though DU-FedProx variants maintained lower update norms.
Multiclass Tasks: For clinical syndrome classification, DU-FedProx-Bilevel achieved the highest overall performance (Accuracy: 0.990, Macro-F1: 0.988) with excellent stability. However, for the more complex clade classification task, standard baselines (FedAvg, FedProx) outperformed the adaptive controllers, suggesting that current adaptive controllers may be overly conservative for high-cardinality, highly imbalanced multiclass problems.
Triage Performance: In simulated deployment, HantaWatch successfully translated model outputs into expert-review queues. In one test, it identified all 72 true high-priority records with a triage recall of 1.000 and a review precision of 1.000, effectively reducing the expert workload by filtering routine cases.
Optimizer Sensitivity: The study confirmed that optimizer choice (SGD, Adam, etc.) significantly impacts convergence and update stability, reinforcing the need for the framework's method-agnostic evaluation approach.

Significance and Claims

The paper positions HantaWatch as a practical decision-support layer for decentralized Hantavirus surveillance. Its primary significance lies in bridging the gap between distributed data ownership and the operational need for rapid expert prioritization.

The authors explicitly claim that HantaWatch:

Enables collaborative model training without sharing raw sequence data, addressing privacy and governance constraints.
Converts model outputs into actionable risk scores and ranked priorities, supporting the efficient allocation of limited expert resources.
Provides a safety mechanism by refusing to deploy models that do not meet surveillance-specific suitability criteria (e.g., high FNR).
Is not an autonomous diagnostic tool; it is designed to support, not replace, laboratory, molecular, phylogenetic, and epidemiological interpretation.

The work concludes that while adaptive optimization (DU-FedProx) enhances stability and performance in specific binary and clinical tasks, further calibration is required for complex multiclass scenarios. Future work is identified as focusing on stronger multiclass controller calibration, concept-drift monitoring, and broader validation across multi-pathogen settings.