Decomposing Participatory Surveillance Symptom Time Series to Track Respiratory Infections: A Cross-Country Evaluation Using Non-Negative Matrix Factorization

This study demonstrates that Non-negative Matrix Factorization applied to participatory surveillance symptom data can effectively decompose co-circulating respiratory pathogen trends, identifying specific signals for viruses like SARS-CoV-2, rhinovirus, and influenza in the Netherlands and successfully transferring these insights to enhance surveillance in Italy.

The Gist

Imagine you are a detective trying to figure out who is causing the chaos in a crowded city square. You can't see the criminals directly, but you can see the symptoms of the chaos: people coughing, sneezing, losing their sense of smell, or complaining of stomach aches.

Usually, if you see a cough, you might think, "Oh, that's the flu." But in reality, the flu, a cold, and even a new virus can all cause coughing. It's like hearing a siren and not knowing if it's a fire truck, an ambulance, or a police car.

This paper is about a new way to solve that mystery using a mathematical tool called Non-Negative Matrix Factorization (NMF). Here is the story of how they did it, explained simply.

1. The Problem: The "Symptom Soup"

Every winter in Europe, many different viruses (like the Flu, SARS-CoV-2, RSV, and the common cold) circulate at the same time. Traditional health systems are like a sieve: they catch some cases but miss many, or they can only tell you "someone is sick" without knowing which virus is making them sick.

The researchers wanted to look at the "Symptom Soup" (thousands of weekly reports from people saying, "I have a fever and a runny nose") and separate it back into its original ingredients.

2. The Method: The "Magic Recipe" (NMF)

Think of the data as a giant, messy smoothie made of 22 different fruits (symptoms like fever, cough, loss of smell, etc.).

• The Goal: The researchers wanted to figure out the recipe for the smoothie. They wanted to know: "How much of this smoothie is made of 'Flu Fruit' and how much is 'Cold Fruit'?"
• The Tool (NMF): They used a computer algorithm that acts like a super-smart blender in reverse. It looks at the data and says, "Okay, I see that 'loss of smell' and 'fever' always happen together in a specific pattern. Let's call that Group A. I see that 'runny nose' and 'sneezing' happen together in a different pattern. Let's call that Group B."

Because the math only allows for "positive" numbers (you can't have negative symptoms), the results are easy to understand. It's like sorting a pile of mixed Lego bricks into separate buckets based on how they fit together.

3. The Experiment: The Dutch Lab and the Italian Test

The researchers had two main groups of data:

• The Netherlands (The Lab): They had a special group of people who not only reported symptoms but also sent in nose and throat swabs. This meant they knew the "ground truth"—they knew exactly which virus was in the sample.
• Italy (The Test): They had people who reported symptoms but didn't send in swabs. They wanted to see if they could use the "recipes" learned in the Netherlands to guess what was happening in Italy.

The Results:

• The "SARS-CoV-2" Recipe: The algorithm found a specific pattern of symptoms (loss of smell/taste, fever, cough) that perfectly matched the weeks when SARS-CoV-2 was high in the Dutch lab.
• The "Rhinovirus" Recipe: It found another pattern (runny nose, sneezing, no fever) that matched the common cold virus.
• The "Winter Virus" Recipe: It found a third pattern (cough, shortness of breath) that seemed to track Flu, RSV, and other seasonal viruses all at once. It's like a "Winter Cold" bucket that catches everything that circulates in the freezing months.

4. The Big Breakthrough: Cross-Border Translation

Here is the coolest part. The researchers took the "recipes" they learned in the Netherlands and applied them to the Italian data.

• Did it work? Yes! Even though the Italian people didn't send in swabs, the "recipes" from the Netherlands successfully identified when SARS-CoV-2 and cold viruses were peaking in Italy.
• The Analogy: Imagine you learned how to recognize a specific type of bird by its song in London. You then go to Paris, hear a similar song, and realize, "Ah, that's the same bird!" You didn't need to catch the bird in Paris to know what it was; you just needed to recognize the pattern.

5. Why This Matters

This study shows that we don't always need expensive lab tests for every single person to know what viruses are spreading.

• The Future: If one country has a good lab system (like the Netherlands), they can teach other countries (like Italy) how to "read" the symptoms of their own people.
• The Benefit: This helps health officials spot outbreaks faster, even in places with fewer resources. It turns a messy pile of "I feel sick" reports into a clear map of which viruses are winning the race.

In a Nutshell

The researchers used a mathematical "de-mixer" to separate a noisy crowd of symptoms into distinct groups. They proved that these groups act like fingerprints for specific viruses. By learning these fingerprints in one country, they could successfully track the same viruses in another country just by listening to the symptoms people reported online. It's a smarter, faster way to keep an eye on the flu season.

Technical Summary

1. Problem Statement

Respiratory seasons in Europe are characterized by the co-circulation of multiple pathogens (e.g., Influenza, SARS-CoV-2, RSV, Rhinovirus, Seasonal Coronaviruses). Current surveillance systems face significant limitations:

• Traditional Virological Surveillance: High specificity but resource-intensive, limited coverage, and often delayed.
• Syndromic Surveillance: Broad coverage and timely (e.g., Influenza-like Illness or ILI rates) but lacks pathogen-specificity and standardized case definitions.
• The Gap: There is a need for a method to extract pathogen-specific signals from general symptom data (syndromic surveillance) without requiring laboratory testing for every individual, particularly to enable cross-country surveillance where virological data may be scarce.

2. Methodology

The study utilized an unsupervised machine learning approach to decompose symptom time series into latent components representing specific pathogens.

Data Sources

• Netherlands (Infectieradar): Participatory surveillance platform (Nov 2020–May 2025). Key feature: Participants reporting acute respiratory symptoms could submit self-collected nose/throat swabs (NTS) for multiplex PCR testing. This provided a "ground truth" link between symptoms and specific pathogens (SARS-CoV-2, Flu A/B, RSV, Rhinovirus, Seasonal Coronaviruses, etc.).
• Italy (Influweb): Participatory surveillance platform (since 2008, data used Nov 2020–May 2025). Participants reported symptoms but did not submit self-swabs. Virological data was only available via national sentinel surveillance (GP/Hospital samples).

Data Preprocessing

• Input Matrix: Constructed a symptom-by-week incidence matrix ($V$) for both countries.
• Symptoms: 22 reported symptoms (e.g., fever, cough, loss of smell/taste, rhinitis).
• Normalization: Counts were normalized by active participants to account for reporting variability, and further normalized across weeks to ensure common symptoms did not skew results (values between 0 and 1).
• Time Window: Only symptoms reported within 3–14 days of onset were included to capture the acute phase.

Statistical Approach: Non-Negative Matrix Factorization (NMF)

• Objective: Decompose the non-negative data matrix $V$ into two lower-dimensional non-negative matrices:
  • $W$: The syndromic spectrum (symptom loadings for each latent component).
  • $H$: The time series (incidence of each component over time).
  • Equation: $V \approx W \times H$
• Algorithm: Used Kullback-Leibler (KL) divergence as the loss function. Initialized using Non-negative Double Singular Value Decomposition (NNDSVD) to ensure convergence.
• Model Selection: The optimal number of components ($k$) was determined using the corrected Akaike Information Criterion (AICc) to balance model complexity and fit.
• Validation (Netherlands): Correlated the time series ($H$) of extracted components with weekly pathogen incidence derived from the self-swab PCR data.
• Transferability (Italy): Applied the $W$ matrix (symptom patterns) derived from the Dutch data to the Italian symptom matrix to infer the $H$ matrix (time series) for Italy. These inferred trends were then compared against Italian national sentinel surveillance data.

3. Key Contributions

1. Pathogen-Specific Signal Extraction: Demonstrated that unsupervised NMF can disentangle co-circulating respiratory pathogens from general symptom data without predefined case definitions.
2. Cross-Country Transferability: Proved that symptom patterns (syndromic spectra) learned in one country (Netherlands) can be applied to another (Italy) to approximate pathogen-specific trends, even in the absence of local self-sampling data.
3. Integration of Virology and Syndromics: Successfully linked participatory symptom reports with laboratory-confirmed incidence to validate the "latent" components.

4. Key Results

Component Extraction (Netherlands)

The AICc analysis suggested 8 latent components were optimal.

• Component 5 (SARS-CoV-2): Showed a strong correlation ($r=0.76, p<0.001$) with SARS-CoV-2 incidence.
  • Symptom Profile: Dominated by loss of smell/taste, fever, cough, anorexia, and watery eyes.
• Component 2 (Rhinovirus): Showed the strongest correlation ($r=0.88, p<0.001$) with Rhinovirus.
  • Symptom Profile: Dominated by runny nose, sneezing, sore throat, cough, and vomiting. Notably, fever was absent in this component.
• Component 7 (Seasonal Respiratory Mix): Correlated highly with Influenza ($r=0.68$), Seasonal Coronaviruses ($r=0.73$), and RSV ($r=0.70$).
  • Symptom Profile: Dominated by sputum, dyspnea, and cough. The authors suggest this represents a composite "winter respiratory syndrome" signal due to the high overlap of these pathogens.
• Other Components: Five components did not map clearly to a single pathogen, potentially representing non-infectious conditions, reporting behaviors, or co-circulating pathogens not captured by the specific PCR panel.

Cross-Country Transfer (Italy)

• When Dutch-derived components were applied to Italian data, the inferred time series showed consistency with Italian national surveillance, particularly for SARS-CoV-2 and Rhinovirus.
• Challenges: Magnitude differences were observed, likely due to demographic differences between the Influweb participant pool (mostly mild cases) and the national sentinel system (ILI patients at GPs/Hospitals).
• Independent Italian NMF: Running NMF independently on Italian data yielded 6 components. Despite the different number, key patterns (e.g., the SARS-CoV-2 signature with anosmia/ageusia) were consistent across both countries.

5. Significance and Conclusion

• Scalable Surveillance: This framework offers a scalable solution for countries with limited virological resources. By calibrating the model in one country with robust lab data, other countries can use participatory symptom data to estimate pathogen-specific trends.
• Early Warning: The method provides timely insights into the circulation of specific pathogens, potentially offering earlier warnings than traditional sentinel systems which rely on healthcare-seeking behavior.
• Limitations:
  • Component magnitudes reflect relative signal strength, not absolute incidence counts.
  • Pandemic-era data (2020–2022) may have introduced atypical symptom distributions.
  • High symptom overlap between pathogens (e.g., Flu, RSV, Seasonal Corona) makes disentangling them difficult without specific markers (like loss of smell for SARS-CoV-2).
• Future Outlook: The study validates the potential of unsupervised learning in public health, suggesting that digital participatory surveillance can evolve from tracking "flu-like illness" to tracking specific viral threats, enhancing global preparedness for respiratory epidemics.