Beyond Single-Pathogen Forecasts: Predicting Shared Respiratory Hospital Burden from Influenza, COVID-19, and RSV

1 Introduction

• Existing respiratory forecasts are largely pathogen-specific, with separate data, preprocessing, features, model optimization, calibration, and evaluation.This approach may provide useful disease-specific forecasts but treats pathogens as independent processes.Source: §1 · p. 2
• Hospitals manage the combined demand for beds, staff, testing, and respiratory support, rather than demand from each pathogen in isolation.The three viruses frequently circulate during the same respiratory season.Source: §1 · p. 2
• The changing contributions of COVID-19, influenza, and RSV across seasons make the total-burden forecast a distinct target from the sum of independently developed forecasts.The paper compares these approaches retrospectively and prospectively across one- to four-week horizons.Source: §1 · p. 2

2 Methodology

• The pipeline harmonizes weekly surveillance and admission data before constructing disease-specific and combined respiratory-burden targets.It then applies horizon shifting, feature selection and engineering, model tuning, adaptive forecast adjustment, and benchmark comparison before retrospective and prospective scoring.Source: §2 · pp. 2–3
• Figure 1 shows changing relative contributions of COVID-19, influenza, and RSV to overall respiratory hospitalization burden across seasons.The figure motivates a unified multi-pathogen forecasting framework.Source: Figure 1 · p. 3
• For each target and horizon, the workflow applies prediction preparation, model tuning, adjustment, and benchmark comparison before scoring.Evaluation includes retrospective and prospective scoring.Source: §2 · p. 3

2.1 Data Sources

• Weekly data for influenza, COVID-19, and RSV covered 2020-09-05 through 2026-06-13 and supported disease-specific and aligned combined respiratory-admission targets.The combined target was intended for hospital-preparedness planning.Source: §2.1 · p. 3
• CDC NHSN supplied primary weekly admission targets and broader statewide indicators, while MUSC and PRISMA supplied local health-system surveillance predictors.RFA RSV hospitalization data were used only to bridge the historical RSV target before CDC RSV admissions were consistently available.Source: §2.1 · p. 3
• RFA RSV hospitalizations supplied the historical target through 2024-09-21, and CDC RSV admissions supplied target weeks from 2024-09-28 onward, including test weeks.For influenza and COVID-19, CDC NHSN weekly total admissions were the modeling targets.Source: §2.1 · p. 3

2.2 Data Preprocessing

• RFA data were linked by week only to construct historical RSV targets before CDC RSV admissions became available; the other RSV files remained feature sources.After merging, non-date fields were converted to numeric values and rows without an observed active-disease target were removed.Source: §2.2 · p. 4
• The bridge comparison used leave-one-out cross-validation across 21 overlap weeks, estimating each conversion rule on 20 weeks and testing it on the withheld week.PA was chosen to assess proportional agreement with the CDC admission scale.Source: §2.2 · p. 4
• PA was 0.761 for threshold-ratio scaling, the highest among four tested RSV bridge rules, so that method was selected for historical target conversion.The alternatives scored 0.630 for simple ratio scaling, 0.633 for RFA-only regression, and 0.697 for RFA plus selected indicators; the selected bridge was fixed before evaluation.Source: §2.2 · p. 4
• The zero threshold was set to 88.5 RFA hospitalizations, the midpoint between 74 paired with zero CDC admissions and 103 paired with positive CDC admissions.A simple multiplication rule would not adequately handle the low RFA values.Source: §2.2 · p. 4
• The scaling ratio used a 10% trimmed mean of the 18 overlap-week ratios where both CDC and RFA values were positive.The lowest and highest ratios were removed before averaging; converted RFA targets were used only before 2024-09-28.Source: §2.2 · p. 4

2.3 Forecast Target Construction and Burden Paradigms

• Single-pathogen targets preserve etiologic attribution, whereas the combined-burden paradigm defines total respiratory admissions as the planning target before model fitting and evaluation.Retrospective forecasts use out-of-sample weeks with observed future targets; prospective forecasts are issued before future admission counts are observed.Source: §2.3 · pp. 4–5

2.3.1 Single-Pathogen Target Construction

• Single-pathogen targets model weekly admissions separately for influenza, COVID-19, and RSV, using disease-specific targets, predictors, features, and horizon-specific training data.Source: §2.3.1 · p. 5

2.3.2 Multi-Pathogen Target Construction

• The multi-pathogen target aligns disease-specific admission targets by week and sums them to represent total weekly admissions from the three pathogens before model fitting.Source: §2.3.2 · p. 5

• All model outputs were clipped to nonnegative values before scoring because the targets are weekly admission counts.Source: §2.4 · p. 5
• MAE measures admission-count error, with lower values indicating smaller errors; PA measures proportional closeness, with higher values up to 1 indicating stronger agreement.Source: §2.4 · p. 5
• Total-burden scoring compared summed pathogen predictions with summed observed admissions and scored the combined-burden model against the observed combined target.Uncertainty intervals and full predictive distributions were not evaluated.Source: §2.4 · p. 5

2.5 Horizon-Specific Modeling Workflow

• Each disease-specific or combined target is shifted to T+1–T+4; retrospective rows require observed labels, while prospective forecasts use the latest complete predictor row.Source: §2.5 · pp. 5–6

2.5.1 Feature Selection

• For each target and horizon, mutual-information ranking precedes iterative VIF filtering above 100, with the top 10 MI-ranked and forced local indicators protected.The top 40 MI candidates enter collinearity screening; tested lower thresholds retained fewer predictors and reduced validation PA.Source: §2.5.1 · pp. 6–7

2.5.2 Feature Engineering

• Retained predictors gain lags, changes, ratios, rolling means, differences, and Fourier seasonality; observed active-target history is transformed separately, never future target labels.The latest complete predictor row is retained for prospective forecasting even when its future admission label is unknown.Source: §2.5.2 · pp. 7–8

2.5.3 Model Design and Training

• Separate XGBoost regressors predict each target–horizon pair from recent admissions and available local, clinical, and surveillance predictors.Source: §2.5.3 · p. 8
• Retrospective testing separates training from held-out weeks with a temporal gap; prospective models reuse selected parameters and refit on eligible labeled rows.Hyperparameters are selected using chronological rolling-origin validation that maximizes validation PA.Source: §2.5.3 · pp. 8–9

2.6 Adaptive Prediction Generation

• For target q, horizon h, and feature vector x_t, the fitted model first generates a raw prediction.This applies to both retrospective evaluation and prospective forecasting.Source: §2.6 · p. 9
• The adjustment uses only historical scored weeks available before the prediction, excluding held-out retrospective and future prospective targets from adjustment estimation.Source: §2.6 · p. 9
• The target- and horizon-specific smoothing constant α is chosen during retrospective workflow development based on PA, then fixed for retrospective and prospective scoring.The adaptive prediction rescales the raw output using ρ and α.Source: §2.6 · p. 9

2.7 Baseline Models

• For each target and horizon, the comparison includes adaptive XGBoost, raw XGBoost, SARIMA, and Random Forest.Baseline models were fit separately for influenza, COVID-19, RSV, and the combined target.Source: §2.7 · p. 10
• Baseline XGBoost is the raw XGBoost prediction, providing an unadjusted boosting reference for adaptive forecasts.Source: §2.7 · p. 10
• SARIMA forecasts future admissions from each target's historical series rather than the full engineered feature set.Source: §2.7 · p. 10
• Random Forest uses the same horizon-specific feature matrix as the machine-learning models and averages predictions across regression trees.Source: §2.7 · p. 10

3.1 Retrospective Estimation

• The retrospective test evaluated held-out target weeks from 2025-11-15 through 2026-02-21 across T+1 to T+4 after training and a temporal gap.Figure 2 depicts the training, gap, retrospective, and prospective split.Source: §3.1 · p. 10, Figure 2 · p. 11
• Retrospective Adaptive XGBoost PA varied by disease: influenza was 0.84, 0.75, 0.62, and 0.62; COVID-19 was 0.75, 0.80, 0.73, and 0.79; RSV was 0.89, 0.81, 0.80, and 0.79 from T+1 to T+4.MAE also varied by disease and horizon; Table 5 compares disease-specific models.Source: Table 5 · p. 11, §3.1 · p. 11
• For the combined target, adaptive XGBoost reduced MAE relative to baseline XGBoost at all four retrospective horizons, while Random Forest had lower MAE at T+3.Random Forest and adaptive XGBoost had the same PA at T+3; Table 6 presents the combined-target comparator results.Source: Table 6 · p. 11, §3.1 · pp. 11–12
• At T+3/T+4 retrospectively, Combined PA was 0.74/0.76 versus Summed 0.69/0.69, and Combined MAE was 183.97/156.13 versus Summed 205.65/217.57.The approaches were similar at T+1 and T+2; the longer-horizon separation is shown in Figure 3.Source: Figure 3 · p. 12, §3.1 · p. 12, Table 7 · p. 12

3.2 Prospective Forecasting

• Across prospective horizons, Combined had higher PA and lower MAE; at T+4, PA was 0.76 versus 0.66 and MAE was 21.97 versus 43.47.Differences were larger at T+3 and T+4; Figure 4 displays the comparison.Source: Figure 4 · p. 13, §3.2 · p. 13, Table 9 · p. 13

4 Discussion

• The combined-burden target directly learns aggregate admissions, whereas summed forecasts inherit timing, calibration, and scale errors from separately tuned pathogen models.Disease-specific forecasts retain etiologic interpretation; the combined target is a decision-relevant admission signal, not a mechanistic representation of transmission.Source: §4 · p. 14

4.1 Retrospective Estimation Analysis

• Retrospective performance varied by pathogen, horizon, and metric; adaptive XGBoost led RSV agreement, while Random Forest lowered influenza MAE at several horizons.PA and MAE capture complementary aspects of performance: proportional agreement and average admissions missed.Source: §4.1 · p. 14
• Retrospectively, the combined-burden and aggregated pathogen-specific approaches were most similar at short horizons and separated more clearly at longer horizons.The authors caution that error accumulation alone does not explain the longer-horizon difference; reconstructed RSV history, features, calibration, respiratory-wave timing, and evaluation period may also contribute.Source: §4.1 · p. 14

4.2 Prospective Performance Analysis

• Prospectively, the combined-burden model had higher PA and lower MAE at every horizon, with the largest separation at T+3 and T+4.Forecasts were issued before the corresponding target weeks were observed.Source: §4.2 · pp. 14–15

4.3 Public Health Relevance

• Pathogen-specific forecasts describe burden composition, while combined-burden forecasts estimate shared hospital pressure for staffing, beds, testing, and respiratory support.Three- and four-week forecasts can inform actions requiring lead time, including schedule planning, supply ordering, surge protocols, and coordination.Source: §4.3 · p. 15

4.4 Future Extensions

• The evaluation used one geographic setting and few respiratory seasons, point forecasts, and no formal test of whether score differences were statistically distinguishable.Future work includes longer prospective evaluation, probabilistic and coherent hierarchical forecasts, additional data streams, and decision-threshold evaluation.Source: §4.4 · p. 15

4.5 Conclusion

• The authors conclude that a two-output framework can preserve pathogen-level composition information while estimating shared capacity pressure with a distinct total-burden target.The combined-burden model more consistently captured shared admission demand than summed disease-specific forecasts, especially at longer planning horizons.Source: §4.5 · p. 16

Funding

• The authors acknowledge the Public Health Science department at Clemson University for providing facilities and support.Source: Funding · p. 16