Classifying and Differentiating Individuals with Respiratory Syncytial Virus, Influenza, and COVID-19 Cases in OpenSAFELY

Using the OpenSAFELY platform, researchers developed and validated specific and sensitive computable phenotypes to accurately classify respiratory syncytial virus, influenza, and COVID-19 cases in English electronic health records, demonstrating their alignment with surveillance data while highlighting varying misclassification risks based on phenotype specificity and patient age.

The Gist

Imagine the National Health Service (NHS) in England as a massive, digital library containing millions of patient stories written in a special code. These stories are called Electronic Health Records (EHRs). Usually, these records tell us what happened to a patient (like "fever" or "hospitalized"), but they often don't explicitly say which specific virus caused it, especially if the patient wasn't tested.

This paper is about a team of digital detectives who used a super-secure computer system called OpenSAFELY to write a new set of "search rules" (called phenotypes) to find three specific troublemakers in that library:

1. RSV (a common virus for babies)
2. Flu (the seasonal flu)
3. COVID-19

Here is how they did it, explained with some everyday analogies:

1. The Detective's Dilemma: The "Specific" vs. "Sensitive" Net

The researchers had to decide how to cast their fishing nets to catch these viruses. They tested two different strategies:

• The "Specific" Net (The Sniper): This net has very small holes. It only catches fish that look exactly like the target.
  • Result: It rarely catches the wrong fish (low misclassification), but it might let some actual target fish swim away if they don't look perfect.
• The "Sensitive" Net (The Wide-Mesh Net): This net has huge holes. It catches everything that might possibly be the target.
  • Result: It catches almost all the target fish, but it also scoops up a lot of other fish that look similar but aren't the target (high misclassification).

The Finding: When looking at mild cases (people who just stayed home with a sniffle), the "Wide-Mesh Net" was too messy. It confused mild cases of one virus with another. The "Sniper" approach was much better at telling them apart.

2. The Age Factor: The "Tiny vs. Grown-Up" Puzzle

The researchers noticed something interesting about who gets misidentified:

• Older Adults: Their symptoms are often distinct enough that the computer rules work well, no matter which net they use.
• Infants: This is the tricky part. Babies with severe illness are like chameleons; their symptoms can look like any of the three viruses. Even with the best rules, it's harder for the computer to tell if a sick baby has RSV, Flu, or COVID just by looking at their hospital notes. The risk of mixing them up is higher for infants than for adults.

3. Checking the Map: The "Weather Report" Test

How did they know their new rules worked? They compared their digital findings to the official weather reports (public surveillance data) that track how many people are sick each week.

• The Result: Their digital maps matched the real-world weather reports perfectly. The seasons for when these viruses peak looked the same in their data as they did in the real world.

The Big Takeaway

Think of this paper as creating a universal translator for hospital records.

In the past, if a patient wasn't tested for a specific virus, doctors and researchers often had to guess or ignore the data. Now, thanks to this study, we have a reliable set of instructions that can look at a patient's coded history and say, "Ah, this pattern looks 90% like RSV," even without a lab test result.

This is a game-changer because it allows scientists to study these viruses in real-time, understand how they affect different age groups, and plan better for the future—all without needing to dig up old test tubes, just by reading the digital stories already written in the NHS library.

Technical Summary

Based on the abstract provided, here is a detailed technical summary of the paper "Classifying and Differentiating Individuals with Respiratory Syncytial Virus, Influenza, and COVID-19 Cases in OpenSAFELY."

1. Problem Statement

The core challenge addressed by this research is the difficulty of accurately identifying and distinguishing between three major respiratory pathogens—Respiratory Syncytial Virus (RSV), Influenza, and COVID-19—using only coded electronic health records (EHRs).

• Data Limitation: In many clinical scenarios, specific testing information (e.g., PCR results or antigen tests) is missing from the records.
• Need for Phenotyping: There is a critical need for "computable phenotypes" (algorithms that define a disease state based on available data) to analyze health outcomes and epidemiological trends when direct diagnostic codes or test results are unavailable.
• Differentiation: Distinguishing between these viruses is difficult because they often present with similar symptoms and clinical codes, leading to potential misclassification.

2. Methodology

The study utilized a large-scale, secure data analytics approach to design and validate these phenotypes:

• Data Source & Platform: The research leveraged the OpenSAFELY secure analytics platform, which processes data from the NHS England electronic health records.
• Cohort & Timeframe: The analysis covered English coded health data spanning September 2016 to August 2024. This period is significant as it captures pre-pandemic, pandemic, and post-pandemic seasons for all three viruses.
• Phenotype Design: The team developed two distinct types of phenotypes for each virus:
  1. Specific Phenotypes: Algorithms designed to minimize false positives (high precision), likely requiring stricter criteria (e.g., specific test codes or severe clinical markers).
  2. Sensitive Phenotypes: Algorithms designed to minimize false negatives (high recall), casting a wider net to capture potential cases, including mild ones.
• Validation Strategy:
  • Internal Comparison: The specific and sensitive phenotypes were compared against each other to assess trade-offs in classification.
  • External Validation: The resulting case counts and seasonal patterns were benchmarked against publicly available surveillance data to ensure epidemiological validity.

3. Key Contributions

• Development of Robust Phenotypes: The paper provides validated, computable definitions for RSV, Influenza, and COVID-19 specifically tailored for the English NHS EHR context.
• Handling Missing Test Data: The study offers a practical solution for classifying viral infections in the absence of testing information, a common scenario in retrospective EHR studies.
• Granular Risk Analysis: It moves beyond simple case counting to analyze how misclassification risks vary by disease severity (mild vs. severe) and patient demographics (infants vs. older adults).

4. Key Results

• Epidemiological Validity: Cases identified by both the specific and sensitive phenotypes demonstrated seasonal patterns consistent with public surveillance data, confirming that the phenotypes accurately capture the temporal dynamics of these viruses.
• Trade-off in Classification:
  • Mild Cases: Sensitive phenotypes resulted in a higher risk of misclassification compared to specific phenotypes. This suggests that while sensitive algorithms catch more cases, they are more prone to including non-cases or misidentifying the specific virus in mild presentations.
  • Severe Cases: Misclassification risks were found to be higher in infants than in older adults. This disparity persisted regardless of whether the specific or sensitive phenotype was used, indicating that coding practices or clinical presentations in infants are inherently more difficult to distinguish from other respiratory conditions.
• Differentiation Capability: The study successfully demonstrated that these three viruses can be differentiated within coded data, though with varying degrees of confidence depending on the phenotype strictness and patient age.

5. Significance

• Research Utility: These phenotypes enable researchers to conduct large-scale retrospective studies on respiratory virus outcomes (e.g., mortality, hospitalization, comorbidities) using routine EHR data, even when specific test results are missing.
• Public Health Surveillance: The ability to generate accurate case counts from EHRs complements traditional surveillance systems, potentially offering faster or more granular insights into viral spread and burden.
• Clinical Insight: The finding that misclassification is higher in infants highlights a specific area where clinical coding or diagnostic protocols may need improvement to better distinguish between RSV, Influenza, and COVID-19 in pediatric populations.
• Scalability: By utilizing the OpenSAFELY platform, this methodology is scalable and reproducible for other health systems with similar EHR structures, advancing the field of digital epidemiology.