📄 medicine

Anticipating the Next Week: Calibrated Prediction of State-Level Respiratory Illness Escalation from Public Surveillance Data

This study presents a transparent benchmark for predicting state-level transitions from moderate to high respiratory illness activity one week in advance, demonstrating that a simple, calibrated logistic regression model using public surveillance data can significantly reduce alert burden while retaining most short-horizon escalations.

Original authors: Tianyi Yu

Published 2026-08-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tianyi Yu

Original paper licensed under CC BY 4.0 (https://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

Every winter, public health officials watch a quiet but critical dashboard. It tracks how many people in each state are visiting doctors with fever, cough, and sore throat—symptoms that signal the spread of flu and other respiratory viruses. These reports are vital, but they often tell a story of what has already happened. The real challenge for health teams is not just to describe the current situation, but to spot the places that are about to get worse. Imagine a weather forecaster who can tell you exactly which towns will see a sudden storm tomorrow, allowing them to prepare before the rain starts. For respiratory viruses, that kind of early warning could mean the difference between a manageable week and an overwhelmed hospital system. The goal is to identify which areas are currently experiencing a moderate level of illness but are on the very edge of tipping into a high-risk zone in the coming week.

A new study by Tianyi Yu at King's College London tackles this specific problem using a method that prioritizes clarity and reliability over complexity. The researcher analyzed years of public data from the United States, combining reports on outpatient doctor visits with laboratory test results for flu, RSV, and SARS-CoV-2. The objective was not to predict the entire season or diagnose individual patients, but to answer a narrow, practical question: among the states that are already seeing a moderate amount of sickness, which ones are most likely to cross the line into high activity in the next seven days? To do this, the team built a computer model that learned from past patterns. They taught the system to recognize the subtle signs that precede a surge, such as how close a state is to its own historical peak, how quickly the number of sick visits is rising, and what other respiratory viruses are circulating in the region.

The study focused on a massive dataset covering thousands of weeks across fifty-one jurisdictions. The researchers split this data into two parts: a training set to teach the model and a separate, later set to test it, ensuring the model could handle new, unseen weeks without memorizing the past. They tested several different approaches, from simple rules to complex machine learning algorithms. The results showed that a straightforward, transparent statistical model performed the best. This model successfully identified nearly sixty percent of the upcoming surges while keeping the number of false alarms low. In contrast, a simpler strategy that would have flagged every single moderate state as a warning would have generated hundreds of unnecessary alerts, overwhelming health teams with noise. The winning model issued far fewer warnings, yet it still caught the majority of the dangerous escalations.

What makes this finding significant is not just the accuracy, but the balance it strikes. The model learned that the most important signal was how close a state was to its own specific high-activity threshold, combined with the time of year and recent trends in sickness. It also used data on other viruses, like RSV and the virus that causes COVID-19, as context clues. If a region was seeing a rise in these other viruses, it increased the likelihood that flu activity would spike soon. The researchers were careful to calibrate the model so that its predictions were honest probabilities, not just guesses. They found that the model's confidence levels matched reality, meaning when it predicted a high risk, that risk was real. This calibration is crucial because public health officials need to know not just which states to watch, but how urgently to watch them.

The study also highlighted the limits of what public data can tell us. The model works with broad, state-level numbers and cannot see individual patients, their ages, or their specific health conditions. It is a tool for triage, designed to help officials decide where to look first, not a tool for making medical decisions for specific people. The researchers were explicit that this is a benchmark for future research, not a rule that should be used immediately to trigger emergency responses without further testing. They emphasized that while the model performed well on historical data, it has not yet been tested in real-time as a live alert system. The true test will come when such a tool is used prospectively, in the future, to see if it can maintain its accuracy when the world changes in unexpected ways.

Ultimately, this work offers a clear path forward for respiratory surveillance. It demonstrates that simple, well-calibrated models can be more useful than complex, opaque ones when the goal is to manage limited resources. By reducing the number of false alarms while still catching the real threats, such a system could allow health teams to focus their attention where it is needed most. The study serves as a transparent example of how to build, test, and evaluate prediction tools in public health, ensuring that the methods are as robust as the data they rely on. It reminds us that in the race against spreading viruses, the best tool is often the one that is clear enough to be trusted and simple enough to be understood.

Technical Summary: Calibrated Prediction of State-Level Respiratory Illness Escalation

Problem Definition
Respiratory-virus surveillance systems, such as the CDC's FluView and NREVSS, effectively describe recent activity levels but are less optimized for triaging jurisdictions that are already at "moderate" activity levels. Public health teams require early, credible signals to identify which moderate jurisdictions are likely to cross into "high" activity in the immediate future (the next reporting week). The specific problem addressed is a short-horizon escalation prediction: among state-weeks currently at moderate Influenza-Like Illness (ILI) activity, which ones will transition to high activity in week $t+1$ ? This is distinct from national seasonal peak forecasting or pathogen-specific diagnosis; it is a prioritization task for resource allocation and situational awareness.

Methodology
The study employs a retrospective prediction-modelling framework using public aggregate data from the U.S.

Data Sources: The analysis combines CDC ILINet outpatient ILI data (accessed via Delphi Epidata FluView) and CDC NREVSS regional nucleic-acid test positivity data for RSV and SARS-CoV-2.
Unit of Analysis: State-weeks.
Target Variable: A binary outcome indicating whether a state with moderate ILI activity in week $t$ crosses its state-specific high-activity threshold in week $t+1$ .
Threshold Definition: State-specific thresholds for "moderate" and "high" activity were derived exclusively from the development period (2019–2024) using the median and 75th percentile of weighted ILI, respectively.
Predictors: Features included current and lagged ILI metrics (1–3 weeks), ILI change rates, proximity to the high-activity threshold, reporting volumes (provider/patient counts), regional RSV and SARS-CoV-2 positivity, seasonal timing (sine/cosine), and geographic identifiers.
Validation Design: To prevent temporal leakage and over-optimism, the study utilized a strict chronological split.
- Development Period: 2019-10-05 to 2024-09-28.
- Holdout Period: 2024-10-05 to 2026-05-09 (evaluated only once after model selection).
- Model Selection: Hyperparameters and alert thresholds were tuned via repeated rolling-origin cross-validation within the development set.
- Uncertainty: Performance uncertainty was quantified using state-cluster bootstrap confidence intervals (1,000 resamples) to preserve the clustering structure of state-weeks.
Models Compared: The study evaluated an "alert-all-moderate" baseline, a tuned threshold rule, regularized logistic regression, random forest, and histogram gradient boosting.
Calibration: Probability calibration and decision-threshold selection were performed using development data only to ensure the output probabilities reflected observed risks.

Key Results
The modelling cohort consisted of 4,232 moderate state-weeks from 51 jurisdictions. The holdout period contained 943 state-weeks, including 127 transitions from moderate to high activity.

Model Performance: The selected regularized logistic regression model ( $C=0.1$ $C = 0.1$ ) outperformed other candidates on the holdout set.
- AUROC: 0.877 (95% CI: 0.827–0.921).
- Average Precision: 0.529 (95% CI: 0.460–0.618).
- F1 Score: 0.594.
- Precision/Recall: 0.589 / 0.598.
- Specificity: 0.935.
Alert Efficiency: At a calibrated threshold of 0.308, the model issued 129 alerts, successfully identifying 76 of the 127 transitions (recall ~60%). In contrast, a baseline strategy of alerting all moderate state-weeks would have required 943 alerts to catch all 127 transitions, generating 816 false positives.
Calibration: The model demonstrated strong calibration (Brier score 0.082; slope 1.013), with mean predicted risk (0.127) closely matching the observed event rate (0.135).
Feature Importance: Permutation importance analysis indicated that the most influential predictors were the current ILI level relative to the state's high-activity threshold, seasonal timing, recent ILI history, reporting volume, and lagged RSV/SARS-CoV-2 positivity.
Subgroup Analysis: Performance varied slightly across HHS regions and reporting-volume tertiles, with recall gaps of up to 0.434 in specific regions, highlighting transportability considerations.

Significance and Claims
The paper positions its contribution not as a deployed public health decision rule, but as a transparent benchmark and reproducible framework for respiratory-surveillance triage research.

Methodological Rigor: The study emphasizes the importance of separating development from evaluation, using temporal cross-validation, reporting bootstrap uncertainty, assessing calibration, and explicitly stating the absence of external validation.
Operational Utility: The results demonstrate that a simple, calibrated linear model can significantly reduce the alert burden for surveillance teams (reducing alerts from 943 to 129) while retaining the majority of short-horizon escalations.
Transparency: The author argues that the interpretability of the logistic regression model is a feature, not a bug, allowing public health officials to inspect, recalibrate, and understand the drivers of alerts (e.g., proximity to thresholds) in a way that complex "black box" ensembles may not.
Limitations Acknowledged: The paper explicitly notes that the outcome is a surveillance proxy (ILI transition) rather than a clinical diagnosis or hospitalization surge. It acknowledges the lack of external validation in a different surveillance ecosystem or country and the inability to assess patient-level fairness due to the aggregate nature of the data.

In conclusion, the study provides a validated, leakage-controlled workflow that improves one-week-ahead triage of respiratory illness escalation, offering a rigorous template for future surveillance research that prioritizes calibration, uncertainty quantification, and operational relevance over headline accuracy.

Technical Summary: Calibrated Prediction of State-Level Respiratory Illness Escalation

More like this