Paper deep dive
Beyond Single-Pathogen Forecasts: Predicting Shared Respiratory Hospital Burden from Influenza, COVID-19, and RSV
Gaurob Saha, MinJae Woo, Lior Rennert
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Co-circulating influenza, COVID-19, and respiratory syncytial virus (RSV) create a shared hospital-admission burden, because beds, staff, testing capacity, and respiratory support are allocated to total demand rather than to one pathogen at a time. This makes respiratory forecasting a target-construction problem as well as a modeling problem: disease-specific forecasts preserve attribution, whereas hospitals also need a single planning signal for combined burden. We developed a supervised 1- to 4-week-ahead framework for weekly admissions and compared two total-burden strategies: summing independently generated pathogen-specific forecasts and forecasting the combined admission target directly. Evaluation included retrospective held-out testing and prospective operational forecasts issued before the corresponding admission weeks were observed, with performance summarized using MAE and Percent Agreement (PA; the smaller of model-generated and observed counts divided by the larger). In retrospective total-burden evaluation, the combined-burden model had PA values of 0.85, 0.78, 0.74, and 0.76 and MAE values of 80.63, 200.09, 183.97, and 156.13 from week 1 to week 4 ahead, compared with summed-forecast PA values of 0.85, 0.80, 0.69, and 0.69 and MAE values of 100.14, 200.06, 205.65, and 217.57. Prospectively, the combined-burden model had higher PA and lower MAE at all four horizons, with the largest separations at weeks 3 and 4 ahead. These results support layered respiratory forecasting: pathogen-specific forecasts remain necessary for etiologic interpretation, while combined-burden forecasts better align with hospital preparedness and shared-capacity planning.
Tags
Links
- Source: https://arxiv.org/abs/2610.16074v1
- Canonical: https://arxiv.org/abs/2610.16074v1
Trouble viewing inline? Open PDF directly â
Full Text
73,993 characters extracted from source content.
Expand or collapse full text
Beyond Single-Pathogen Forecasts: Predicting Shared Respiratory Hospital Burden from Influenza, COVID-19, and RSV Gaurob Saha, Minjae Woo, Lior Rennert Department of Public Health Sciences and Center for Public Health Modeling and Response, Clemson University, Clemson, SC 29601, USA E-mail addresses: gsaha@clemson.edu, mwoo@clemson.edu, liorr@clemson.edu Abstract Co-circulating influenza, COVID-19, and respiratory syncytial virus (RSV) create a shared hospital-admission burden, because beds, staff, testing capacity, and respiratory support are allocated to total demand rather than to one pathogen at a time. This makes respiratory forecasting a target-construction problem as well as a modeling problem: disease-specific forecasts preserve attribution, whereas hospitals also need a single planning signal for combined burden. We developed a supervised 1- to 4-week-ahead framework for weekly admissions and compared two total-burden strategies: summing independently generated pathogen-specific forecasts and forecasting the combined admission target directly. Evaluation included retrospective held-out testing and prospective operational forecasts issued before the corresponding admission weeks were observed, with performance summarized using MAE and Percent Agreement (PA; the smaller of model-generated and observed counts divided by the larger). In retrospective total-burden evaluation, the combined-burden model had PA values of 0.85, 0.78, 0.74, and 0.76 and MAE values of 80.63, 200.09, 183.97, and 156.13 from week 1 to week 4 ahead, compared with summed-forecast PA values of 0.85, 0.80, 0.69, and 0.69 and MAE values of 100.14, 200.06, 205.65, and 217.57. Prospectively, the combined-burden model had higher PA and lower MAE at all four horizons, with the largest separations at weeks 3 and 4 ahead. These results support layered respiratory forecasting: pathogen-specific forecasts remain necessary for etiologic interpretation, while combined-burden forecasts better align with hospital preparedness and shared-capacity planning. Keywords: Respiratory hospital burden; Influenza; COVID-19; RSV; Multi-pathogen forecasting; Hospital preparedness. 1 Introduction Respiratory viruses are a major cause of seasonal hospital burden, with influenza, coronavirus disease 2019 (COVID-19), and respiratory syncytial virus (RSV) contributing substantially to emergency department visits, inpatient admissions, intensive care utilization, and mortality (Bi et al., 2025; Dewey et al., 2025; Mellor et al., 2025; Yang et al., 2024; Gebremedhin et al., 2022). During the 2023â2024 respiratory season in the United States, these three pathogens accounted for more than one million hospitalizations, with combined daily admissions exceeding 10,000 during the seasonal peak (Bi et al., 2025). Such surges place considerable pressure on healthcare systems by increasing demand for inpatient beds, intensive care units, ventilators, diagnostic testing, healthcare personnel, and other hospital resources (Mellor et al., 2025; Osthus, 2022; Gebremedhin et al., 2022). Because influenza, COVID-19, and RSV frequently circulate during the same respiratory season, they compete for these shared resources, requiring hospitals to manage the overall respiratory disease burden rather than each pathogen independently (Bi et al., 2025; Dewey et al., 2025; Mellor et al., 2025). Numerous forecasting models have been developed to estimate short-term hospital admissions for influenza, COVID-19, and, more recently, RSV (Ray et al., 2025; Klein et al., 2023; Mellor et al., 2025; Yang et al., 2024). These studies have demonstrated the utility of statistical models, machine learning methods, ensemble approaches, and multi-source surveillance data for supporting near-term public health planning and hospital preparedness (Osthus, 2022; Ray et al., 2025; Park et al., 2024; Yoon et al., 2026; Mellor et al., 2025). Despite these advances, respiratory disease forecasting continues to be largely pathogen-specific. Influenza, COVID-19, and RSV are typically modeled as separate prediction problems, each developed using its own surveillance data, preprocessing strategy, feature engineering procedure, model optimization, calibration, and evaluation framework (Osthus, 2022; Ray et al., 2025; Klein et al., 2023; Yang et al., 2024). While this approach may provide useful forecasts for individual diseases, it implicitly treats respiratory pathogens as independent processes, even though they coexist within the same healthcare system. It also does not align with the way respiratory-season burden is observed or managed in hospitals, where multiple pathogens collectively determine demand for the same beds, staff, diagnostic capacity, and respiratory support (Bi et al., 2025; Dewey et al., 2025; Mellor et al., 2025). This pathogen-specific forecasting paradigm has important limitations. Influenza, COVID-19, and RSV do not circulate in isolation, and their observed hospital burden may be shaped by overlapping symptoms, testing behavior, healthcare-seeking patterns, population immunity, vaccination patterns, and concurrent respiratory virus activity (Bi et al., 2025; Dewey et al., 2025; Mellor et al., 2025; Gebremedhin et al., 2022; Yang et al., 2024). In our surveillance data, the relative contribution of each pathogen changed substantially across respiratory seasons, with COVID-19 dominating earlier periods, influenza becoming the largest contributor during selected winter peaks, and RSV adding a distinct seasonal component to the overall hospitalization burden (Fig. 1). This changing composition illustrates why a model trained only on one pathogen remains unaware of the contemporaneous respiratory disease environment in which that pathogen occurs. Consequently, simply adding predictions from independently developed pathogen-specific models is not equivalent to forecasting total respiratory burden; the summed estimate inherits separate model assumptions, separate calibration behavior, and separate error structures while ignoring information contained in the collective activity of respiratory pathogens. This creates an urgent gap in understanding how discrepancies between pathogen-specific forecasts and multi-pathogen respiratory burden forecasts may lead hospitals to overestimate or underestimate the resources needed for near-term respiratory-season planning (Mellor et al., 2025; Murph et al., 2026). Figure 1: Weekly respiratory hospital admissions by pathogen across respiratory seasons. The figure shows the changing relative contributions of COVID-19, influenza, and RSV to the overall respiratory hospitalization burden, motivating a unified multi-pathogen forecasting framework. In this paper, we describe a unified multi-pathogen forecasting framework for short-term respiratory hospital admissions, with the purpose of evaluating whether respiratory burden is better estimated using a single forecasting framework than by aggregating independently developed pathogen-specific models. The presented analysis compares these two operational forecasting paradigms through retrospective evaluation and prospective deployment across one- to four-week forecasting horizons. Beyond comparing forecasting accuracy, this study examines whether a unified forecasting strategy provides a more representative estimate of respiratory disease burden while enabling a single operational workflow for influenza, COVID-19, and RSV. The proposed framework has the potential to support hospital preparedness by improving estimation of overall respiratory demand, simplifying operational forecasting, and providing a scalable foundation for future multi-pathogen surveillance and response (Osthus, 2022; Mellor et al., 2025; Murph et al., 2026). 2 Methodology The methodological workflow was organized as a sequential forecasting pipeline. Weekly surveillance and hospital-admission data were first harmonized across sources, then used to construct disease-specific and combined respiratory-burden targets. For each target and horizon, the pipeline applied horizon shifting, feature selection, feature engineering, model tuning, adaptive forecast adjustment, and benchmark comparison before retrospective and prospective scoring. 2.1 Data Sources Weekly respiratory surveillance and hospital admission data were assembled separately for influenza, COVID-19, and RSV over a common experimental period from 2020-09-05 through 2026-06-13. These weekly aggregate data supported two related forecasting targets: disease-specific admissions and the aligned combined respiratory admission burden used for hospital-preparedness planning. Three source types were used. CDC NHSN served as the primary source for weekly admission targets and broader statewide hospital-burden indicators. MUSC and PRISMA supplied local health-system surveillance predictors. South Carolina Revenue and Fiscal Affairs (RFA) RSV hospitalization data were used only to bridge the historical RSV target before CDC RSV admissions were consistently available. This role-based harmonization was needed because no single stream provided complete disease-specific admission targets and local surveillance signals across the full study period. For influenza and COVID-19, the modeling target was the CDC NHSN weekly total admission count for the corresponding disease. For RSV, CDC NHSN weekly total admission count was used once the CDC RSV target was available for the modeling period. Before that point, the CDC RSV target was missing or not consistently usable for training, so the RFA weekly hospitalization series was used to extend the historical RSV target. Within the study window, RFA weekly hospitalizations contributed historical RSV target weeks from 2020-09-05 through 2024-09-21 after conversion to the CDC admission scale. CDC Total RSV Admissions was used for RSV target weeks from 2024-09-28 onward, including the test weeks. The MUSC and PRISMA files provided the same local weekly indicator types for each disease: disease-related diagnoses, emergency-department visits, tests performed, positive tests, vaccines recorded during the reporting week, season-to-date vaccinations, and total disease-related clinical encounters. Additional CDC predictors provided broader hospital-burden context when selected by the pipeline, including admission rates per 100,000 population, pediatric admission and hospitalization counts, ICU hospitalization counts, bed-occupancy percentages, and cumulative seasonal admission counts. This source mix is motivated by prior work showing that local health-system and other near-real-time surveillance streams can provide actionable signals for infectious disease burden when broader public health datasets are delayed or too coarse for local planning (Ahammed et al., 2025; Tabassum et al., 2026; Klein et al., 2023). 2.2 Data Preprocessing The preprocessing workflow harmonizes the source files before modeling. The disease-specific data sources are merged on week: D=(DMUSCâWeekDPRISMA)âWeekDCDC.D= (D_MUSC _WeekD_PRISMA ) _WeekD_CDC. (1) For RSV, the MUSC, PRISMA, and CDC files remained the feature sources in the merged modeling dataset. The RFA file was linked by Week only to construct the historical RSV target before the CDC RSV target period, after converting RFA weekly hospitalizations to the CDC admission scale. The RFA-to-CDC conversion was developed using only the weeks when both CDC weekly RSV admissions and RFA weekly RSV hospitalizations were available. This shared RFAâCDC overlap spanned 2024-09-28 through 2025-02-15 and provided 21 paired weekly observations. Four candidate conversion rules were evaluated in this overlap period. Each rule converted RFA weekly RSV hospitalizations into a CDC-scale RSV admission value. Simple ratio scaling used a single RFA-to-CDC multiplier; threshold-ratio scaling first mapped low RFA values to 0 and then applied a multiplier to the remaining values; RFA-only linear regression estimated CDC admissions from RFA hospitalizations alone; and the RFA-plus-indicators regression added selected PRISMA RSV inpatient-hospitalization, emergency-department, and positive-test indicators to the RFA value. Because the overlap period was short, the comparison used leave-one-out cross-validation across the 21 overlap weeks. In each fold, one overlap week was withheld, each candidate conversion rule was estimated from the other 20 overlap weeks, and the withheld RFA value was converted to the CDC scale and compared with the observed CDC RSV admission value for that same week. The bridge methods were compared using Percent Agreement (PA), because the conversion step was intended to preserve proportional agreement between the reconstructed RSV target and the CDC admission scale rather than optimize absolute error alone. In general, PA measures relative closeness by comparing the smaller value in a predicted-observed pair with the larger value, so values closer to 1 indicate stronger agreement. For each withheld overlap week, PA was calculated as the smaller of the observed CDC value and converted RFA value divided by the larger value. The leave-one-out PA values were 0.630 for simple ratio scaling, 0.761 for threshold-ratio scaling, 0.633 for RFA-only regression, and 0.697 for RFA plus selected indicators. Based on this criterion, the threshold-ratio bridge was selected because it had the highest PA and provided a transparent single-series conversion rule for historical RFA admissions. The selected bridge was fixed before retrospective and prospective evaluation. After selection, the final threshold and scaling values were calculated from the full overlap period. The overlap pattern showed that a simple multiplication rule would not be adequate for low RFA values. During the first three overlap weeks, CDC RSV admissions remained 0 while RFA reported 65, 69, and 74 weekly hospitalizations. The first positive CDC RSV admission value occurred when RFA reached 103. The zero threshold was therefore set at the midpoint between the largest RFA value paired with a CDC value of 0 and the smallest RFA value paired with a positive CDC value: (74 + 103)/2 = 88.5. y~t(RSV)=0,rt<88.5,0.1969157929ârt,rtâ„88.5, y_t^(RSV)= cases0,&r_t<88.5,\\ 0.1969157929\,r_t,&r_tâ„ 88.5, cases where rtr_t is the RFA weekly RSV hospitalization count and y~t(RSV) y_t^(RSV) is the converted CDC-scale RSV admission target. The scaling ratio was estimated from the 18 overlap weeks in which both CDC and RFA values were positive. For each positive overlap week, the weekly ratio was calculated as CDC weekly RSV admissions divided by RFA weekly RSV hospitalizations. The ratios were then sorted, the single lowest and single highest ratios were removed, and the remaining 16 ratios were averaged. This 10% trimmed mean was used to reduce sensitivity to unusually high or low overlap weeks. The converted RFA target was used only before 2024-09-28; CDC Total RSV Admissions was used once available and for all retrospective and prospective scoring. After the disease-specific sources were merged and the active disease target was constructed, including the RFA-to-CDC conversion for RSV, the remaining preprocessing steps were applied uniformly across diseases. All non-date fields were converted to numeric values after standardizing common missing-value tokens, including empty strings, null, and dash placeholders. The merged dataset was then sorted by Week. Rows without an observed target value for the active disease were removed from the modeling frame. 2.3 Forecast Target Construction and Burden Paradigms This study treats target construction as part of the forecasting question. Single-pathogen targets preserve etiologic attribution for influenza, COVID-19, and RSV. Summed forecasts estimate total burden only after separate pathogen-specific predictions have been generated. The combined-burden paradigm instead defines total respiratory admissions as the hospital planning target before model fitting and evaluation. This distinction is central because disease-specific interpretation and shared-capacity planning answer related but different operational questions. In retrospective evaluation, both paradigms are scored on out-of-sample test weeks for which future admission targets are already observed. In prospective forecasting, the same target definitions are used to generate forecasts from the most recent available predictor rows before the future admission counts are observed and scored. 2.3.1 Single-Pathogen Target Construction For the single-pathogen paradigm, each disease was modeled as a separate weekly admission burden. These targets support pathogen attribution and allow influenza, COVID-19, and RSV forecasts to be evaluated individually. Let yt(d)y_t^(d) denote the observed admission count at week t for disease dâInfluenza,COVIDâ-â19,RSVdâ\Influenza,COVID -19,RSV\. The disease-specific targets were yt(Influenza),yt(COVIDâ-â19),yt(RSV).y_t^(Influenza), y_t^(COVID -19), y_t^(RSV). (2) For influenza and COVID-19, these targets were the corresponding CDC NHSN weekly admission counts. For RSV, the target combined the converted RFA-scale historical values before the CDC RSV target period with CDC NHSN RSV admissions once the CDC target was available. Each disease-specific model used its own target, selected predictors, engineered features, and horizon-specific training data. The resulting disease-specific forecasts can be evaluated separately or summed after prediction to obtain an aggregate estimate from independently modeled pathogens: y^t,h(sum)=y^t,h(Influenza)+y^t,h(COVIDâ-â19)+y^t,h(RSV). y_t,h^(sum)= y_t,h^(Influenza)+ y_t,h^(COVID -19)+ y_t,h^(RSV). (3) 2.3.2 Multi-Pathogen Target Construction For the multi-pathogen paradigm, the weekly respiratory burden target was constructed by aligning the disease-specific admission targets by week and summing them: yt(Combined) y_t^(Combined) =yt(Influenza)+yt(COVIDâ-â19) =y_t^(Influenza)+y_t^(COVID -19) (4) +yt(RSV). +y_t^(RSV). This target represents the total weekly hospital admission burden attributable to the three respiratory pathogens in the study. In contrast to the summed single-pathogen forecast, the multi-pathogen target construction defines the shared hospital-burden signal before horizon shifting, model fitting, and evaluation. 2.4 Performance Evaluation Performance was evaluated using point-prediction metrics applied consistently across retrospective estimation and prospective forecasting. Because the target variables are weekly admission counts, all model outputs were clipped to be nonnegative before scoring: y~i=maxâĄy^i,0, y_i= \ y_i,0\, (5) where y^i y_i is the raw predicted value and y~i y_i is the scored prediction. For an evaluated set q,hS_q,h of observations for target q and horizon h, mean absolute error (MAE) was defined as MAEq,h=1|q,h|ââiâq,h|yiây~i|.MAE_q,h= 1|S_q,h| _i _q,h|y_i- y_i|. (6) MAE is measured in admissions, so lower values indicate smaller absolute prediction errors. Percent Agreement (PA) was used as a scale-free measure of proportional closeness between observed and predicted admissions: PAq,h=1|q,h|ââiâq,hAi,PA_q,h= 1|S_q,h| _i _q,hA_i, (7) where Ai=1,yi=0âandây~i=0,minâĄ(yi,y~i)maxâĄ(yi,y~i),otherwise.A_i= cases1,&y_i=0\ and\ y_i=0,\\ (y_i, y_i) (y_i, y_i),&otherwise. cases (8) PA ranges from 0 to 1, with 1 indicating exact agreement and higher values indicating stronger proportional agreement. Thus, PA captures proportional agreement, whereas MAE captures admission-count error directly relevant to staffing and capacity planning. Metrics were computed separately by target and horizon. For total-burden evaluation, summed single-pathogen predictions were first added across influenza, COVID-19, and RSV and then compared with the corresponding summed observed admissions; the combined-burden model was scored against the observed combined respiratory-burden target. All performance metrics were calculated from point predictions; uncertainty intervals or full predictive distributions were not evaluated in this study. 2.5 Horizon-Specific Modeling Workflow For near-term respiratory-burden planning, each admission target was converted into horizon-specific prediction tasks for influenza, COVID-19, RSV, and the combined respiratory-burden target across four forecast horizons. Here, T denotes the forecast origin; T+1T+1 through T+4T+4 denote forecasts 1 through 4 weeks ahead, respectively. The horizon-specific shifted-label construction was applied separately within each paradigm. In the single-pathogen paradigm, separate shifted targets were created for influenza, COVID-19, and RSV. In the multi-pathogen paradigm, the aligned total respiratory burden target was shifted directly. Let yt(q)y_t^(q) denote the observed weekly hospital burden value for target q, where qâInfluenza,COVIDâ-â19,RSV,Combinedqâ\Influenza,COVID -19,RSV,Combined\. The label for target q and horizon h is yt,h(q)=yt+h(q).y_t,h^(q)=y_t+h^(q). (9) In implementation, this corresponds to shifting each target column upward by h weeks so that predictors at week t are paired with admissions at week t+ht+h. Here, Target represents either an active disease target: Total Influenza Admissions, Total COVID-19 Admissions, or Total RSV Admissions, or the aligned total respiratory burden target used in the multi-pathogen paradigm. This construction aligns each row of predictors at week t with the future burden value at week t+ht+h. Using one shifted-label rule across all targets keeps disease-specific and combined-burden results comparable at each horizon. It also preserves the forecast-time information set by pairing predictors from week t or earlier with future admissions. The shifted-label rule is handled differently for retrospective testing and prospective forecasting. During retrospective model training and testing, a row can be used only if its future target has already been observed. Therefore, after shifting the target, the T+1 dataset excludes the final one row without an observed week-t+1t+1 admission value, the T+2 dataset excludes the final two rows without observed week-t+2t+2 values, and the same logic applies to T+3 and T+4. These unlabeled end-of-series rows were excluded from retrospective training and scoring; scoring used only labeled rows in the chronological test segment after the fixed temporal gap. For prospective forecasting, the most recent row with a complete predictor vector was retained as the forecast-origin row, and the trained horizon-specific models were applied to this row to generate forecasts before the target admission weeks were observed. After the corresponding admission counts became available, the prospective forecasts were scored using the same evaluation metrics. 2.5.1 Feature Selection Feature selection is horizon-specific because the association between predictors and future admissions may change as the forecast horizon increases. For each forecasting target and horizon, the data with shifted targets are partitioned chronologically into a training segment, a fixed temporal gap, and a test segment: Dh=Dh,trainâ|Dgap|âDh,test.D_h=D_h,train\;|\;D_gap\;|\;D_h,test. (10) The fixed temporal gap corresponds to the skipped calendar weeks between the training and retrospective test segments, not to the forecast horizon h. These rows are excluded from both model training and retrospective testing to avoid temporal leakage. Predictors with any missing values in the training segment are removed based on the training set only, not the gap or test periods. Candidate features are then screened using mutual information, followed by variance inflation factor filtering. Mutual information (MI) score. Mutual information was used to rank candidate predictors according to their nonlinear dependence with the horizon-specific target. For candidate predictor XjX_j and target YhY_h, mutual information is defined as IâĄ(Xj,Yh)=â«âĄpâĄ(xj,yh)âlogâĄ(pâĄ(xj,yh)pâĄ(xj)âpâ(yh))âdâxjâdâyh.I(X_j;Y_h)= p(x_j,y_h) ( p(x_j,y_h)p(x_j)p(y_h) )dx_jdy_h. (11) In the implementation, MI is estimated with mutual info regression using only the training data. The top 40 predictors are retained for collinearity screening. In addition, the pipeline always appends these local hospital indicators to the selected feature list: Weekly Inpatient Hospitalizations MUSC and Weekly Positive Tests PRISMA. These forced-keep predictors, together with the top 10 MI-ranked predictors, are protected during VIF filtering. This design preserves clinically meaningful local signals even when feature rankings vary across small weekly training samples. More broadly, local and multi-source infectious disease forecasting studies emphasize the value of combining data-driven modeling with clinically relevant surveillance signals (Ahammed et al., 2025; Wu et al., 2025; Hossain et al., 2025). Variance inflation factor (VIF). Following MI screening, variance inflation factor filtering is used to reduce multicollinearity among retained features. For predictor XjX_j, let Rj2R_j^2 be the coefficient of determination obtained by regressing XjX_j on the remaining selected predictors. The VIF is VIFj=11âRj2.VIF_j= 11-R_j^2. (12) High VIF values indicate that a feature is strongly explained by other selected predictors. The pipeline uses a VIF threshold of 100. Features exceeding this threshold are removed iteratively, except for the protected top 10 MI-ranked predictors and the forced-keep predictors. When multiple removable features exceed the threshold, the implementation removes the feature with the lowest MI score first; if multiple removable features have similarly low MI scores, the higher-VIF feature is removed first. Although conventional VIF screening often uses lower cutoffs such as 5 or 10, these cutoffs are heuristic rather than universal (OâBrien, 2007). Preliminary sensitivity checks using thresholds of 5, 10, 15, 20, 25, and 50 retained substantially fewer predictors and reduced validation PA. The threshold of 100 was therefore retained to remove only extreme redundancy before tree-ensemble prediction, where collinearity mainly affects redundant split choices rather than coefficient interpretation. This horizon-specific MI and VIF process means that different numbers of predictors can be selected for T+1, T+2, T+3, and T+4, as shown in Table 1 for the disease-specific models and Table 2 for the combined-burden model. The resulting feature set is then passed to feature engineering. 2.5.2 Feature Engineering As shown in Figure 1, weekly admissions exhibited respiratory-burden dynamics, including recurrent seasonal peaks, pathogen-specific timing, and short-term persistence. These patterns motivated lagged, momentum, smoothing, and seasonal predictors designed to capture recent burden, trend, smoothed trajectory, and annual timing. This framing treats feature engineering as a way to represent respiratory-burden dynamics in the surveillance series, rather than as a purely algorithmic expansion of predictors. Lagged values and rolling means represent recent burden and smoothed trajectory, percentage changes, ratios, and first- and second-order differences represent local trend and short-term momentum, and Fourier terms represent annual seasonality. This step is applied after MI and VIF screening, so the feature expansion is performed only on the target- and horizon-specific predictors retained for that model. Consequently, the number of engineered features can differ across T+1, T+2, T+3, and T+4, as reported in Table 1. The retained surveillance and hospital variables are treated as exogenous predictors. The observed admission history is kept separately so that past admission values can be used, while the shifted future admission label is kept only as the outcome to be predicted. This target label is never engineered as a predictor. These feature families mirror common signals used in respiratory forecasting, where recent observed burden, lagged surveillance indicators, and seasonal terms often explain different parts of the trajectory (Osthus, 2022; Ray et al., 2025; Morris et al., 2023). For each retained exogenous predictor xtx_t observed at week t, the implementation first creates three backward-looking lagged values: xtâk,k=1,2,3.x_t-k, k=1,2,3. (13) These lagged values allow the model to use the recent history of diagnoses, visits, tests, positive tests, hospital indicators, and CDC surveillance variables rather than the current week alone. Short-term relative changes are encoded with safe percentage-change features: Î1%(xt) ^\%_1(x_t) =xtâxtâ1xtâ1, = x_t-x_t-1x_t-1, (14) Î2%(xt) ^\%_2(x_t) =xtâ1âxtâ2xtâ2. = x_t-1-x_t-2x_t-2. (15) In the implementation, these percentage-change features are protected against zero denominators. If the previous value is missing, the engineered change is treated as missing; if the previous value is zero, the percentage change is set to 0.0 rather than allowing an infinite value. The corresponding ratio features are r1â(xt) r_1(x_t) =xtxtâ1, = x_tx_t-1, (16) r2â(xt) r_2(x_t) =xtâ1xtâ2. = x_t-1x_t-2. (17) These ratio features are also computed with a safe denominator rule: a missing previous value remains missing, whereas a zero previous value is assigned a neutral ratio of 1.0. To smooth short-term fluctuations, the model includes right-aligned rolling means: RM2,t _2,t =(xt+xtâ1)/2, =(x_t+x_t-1)/2, (18) RM4,t _4,t =(xt+xtâ1+xtâ2+xtâ3)/4. =(x_t+x_t-1+x_t-2+x_t-3)/4. (19) Because the rolling windows are right-aligned, they use only week t and earlier values. The first difference is, st=xtâxtâ1,s_t=x_t-x_t-1, (20) and the second difference is, at=(xtâxtâ1)â(xtâ1âxtâ2).a_t=(x_t-x_t-1)-(x_t-1-x_t-2). (21) These transformations are applied to every retained exogenous predictor and are also applied separately to the observed history of the active target. In the single-pathogen paradigm, the active target is the corresponding disease admission series; for example, an RSV run may use Total RSV Admissions and its lags, percentage changes, ratios, rolling means, first difference, and second difference, but not Total RSV Admissions T+h as an input feature. In the multi-pathogen paradigm, the same rule is applied to the observed combined admission history, but not to Combined Admissions T+h. Therefore, the feature set uses week t and prior information to predict week t+ht+h, while the shifted target is used only as the supervised learning label. Seasonality is represented by Fourier terms derived from ISO week number: St S_t =sinâĄ(2âÏâ weekNumt52), = ( 2Ï·weekNum_t52 ), (22) Ct C_t =cosâĄ(2âÏâ weekNumt52). = ( 2Ï·weekNum_t52 ). (23) These seasonal terms provide a compact annual cycle without requiring the model to estimate a separate coefficient for each week of the year. After feature engineering, infinite values are converted to missing values. The engineered data are then separated into two related objects for each target-horizon pair. For retrospective training and testing, the labeled modeling frame keeps only rows that have both a complete engineered predictor vector and a known shifted target for that horizon; rows with missing engineered predictors or unknown future labels are excluded. This separation keeps the retrospective and prospective datasets aligned in their predictor definitions while allowing their target-availability requirements to differ. It also prevents information from future admission weeks from entering the engineered predictor matrix before forecasts are issued. For prospective forecasting, the future target is not required because it is the value being predicted. The pipeline therefore also retains the most recent row with complete engineered predictors as the forecast-origin row, even though its T+1, T+2, T+3, or T+4 admission value may not yet be observed. Table 1: Disease-specific selected and engineered feature counts from the model runs. Target Horizon Selected Features Engineered Features Influenza T+1 20 242 Influenza T+2 18 230 Influenza T+3 19 242 Influenza T+4 17 218 COVID-19 T+1 20 242 COVID-19 T+2 20 242 COVID-19 T+3 18 218 COVID-19 T+4 19 230 RSV T+1 21 254 RSV T+2 19 242 RSV T+3 16 206 RSV T+4 19 242 Table 2: Combined-target selected and engineered feature counts from the model runs. Target Horizon Selected Features Engineered Features Combined T+1 28 280 Combined T+2 28 280 Combined T+3 28 280 Combined T+4 26 256 2.5.3 Model Design and Training XGBoost Model To isolate the role of forecast-target construction, each targetâhorizon pair was modeled with a separate standard XGBoost regressor (Chen and Guestrin, 2016). The same model form was used for the three single-pathogen targets and for the combined respiratory-burden target, with separate fitted models for T+1 through T+4. XGBoost is a gradient-boosted tree model for continuous outcomes; here, the outcome is the future weekly admission count for the selected target and horizon. It was used as an operational supervised prediction model for weekly respiratory-admission burden, not as a mechanistic model of pathogen transmission. The primary machine-learning formulation was multivariate supervised time-series prediction: each target was predicted from its own recent admission history together with contemporaneously available local, clinical, and surveillance predictors. For target q, week t, and horizon h, the prediction is the sum of the contributions from all regression trees: y^t,h(q)=âm=1Mfm(q,h)â(t), y_t,h^(q)= _m=1^Mf_m^(q,h)(x_t), (24) where tx_t is the engineered feature vector and fm(q,h)f_m^(q,h) denotes the mmth regression tree for target q and horizon h. Model settings are selected through the training procedure described below. Model Training After feature engineering, each target- and horizon-specific labeled dataset was kept in chronological order and divided into an outer training segment, a temporal gap, and a retrospective test segment: Dmodel=Dtrainâtime-series validation insideâ|Dgap|âDtest.D_model= D_train_time-series validation inside\;|\;D_gap\;|\;D_test. (25) DtrainD_train contains the earlier labeled weeks used for model training. DtrainD_train spanned 2020-09-05 through 2025-10-11. DgapD_gap contains the skipped weeks between the development and test periods; this gap spanned 2025-10-18 through 2025-11-08. The gap was included to preserve temporal separation, prevent leakage from observations immediately adjacent to the retrospective test period, and ensure that model selection used only information available before the held-out evaluation period. DtestD_test contains later weeks held out for retrospective evaluation. The outer gap and test rows are not used during model selection. Within DtrainD_train, Optuna with the Tree-structured Parzen Estimator sampler is used for hyperparameter tuning (Akiba et al., 2019; Bergstra et al., 2011). Tuning uses rolling-origin time-series validation. For validation split s, the training segment is divided into an earlier fitting block and a later validation block: Dtrain(s)=Dfit(s)|Dval(s).D_train^(s)=D_fit^(s)\;|\;D_val^(s). (26) For that split, the candidate model is trained only on Dfit(s)D_fit^(s) and scored only on Dval(s)D_val^(s). The validation block follows the fitting block in time, and a validation gap equal to the forecast horizon separates fitting rows from validation rows. For all diseases and horizons, tuning used a fixed random seed of 42, an Optuna TPE sampler with seed 42, 25 trials, and 10 rolling-origin validation splits. Optuna selected hyperparameters by maximizing validation PA aggregated across the chronological validation splits. This preserves temporal order and follows forecast evaluation practice emphasizing standardized horizons and retrospective assessment because forecast accuracy varies over time, geography, model class, and lead time. The development workflow therefore involved 16 targetâhorizon tuning runs: influenza, COVID-19, RSV, and the combined burden target across four forecast horizons. With 25 trials and 10 rolling-origin splits per run, hyperparameter search was the main computational cost, whereas prospective operation reused the selected settings and required only weekly data updating, feature construction, final model fitting, and prediction. This separation limited issue-time latency and clarifies the distinction between model development and real-time forecast generation (Pollett et al., 2021). For each target and horizon, the search space Î includes n estimators (tree count) from 200 to 800, max depth (tree depth) from 3 to 8, learning rate (η; step size) from 0.01 to 0.10, subsample (row sampling) and colsample bytree (feature sampling) from 0.60 to 1.00, min child weight (split size) from 1 to 10, gamma (split penalty) from 0 to 1, reg alpha (α; L1 penalty) from 0 to 1, and reg lambda (λ; L2 penalty) from 0 to 5. The Optuna-selected values are summarized for the disease-specific runs in Table 3 and for the combined-burden model in Table 4. Table 3: Optuna-selected XGBoost hyperparameters by disease and forecast horizon from the model runs. Target Horizon Trees Depth η Subsample Colsample Child Gamma α λ Influenza T+1 414 4 0.0588 0.656 0.921 1 0.987 0.772 0.994 Influenza T+2 537 7 0.0666 0.828 0.707 1 0.896 0.741 3.755 Influenza T+3 383 3 0.0716 0.776 0.649 5 0.034 0.909 1.294 Influenza T+4 256 7 0.0283 0.708 0.851 3 0.564 0.396 0.552 COVID-19 T+1 786 3 0.0432 0.949 0.751 3 0.222 0.010 0.368 COVID-19 T+2 286 3 0.0674 0.779 0.905 5 0.226 0.791 3.333 COVID-19 T+3 517 5 0.0701 0.874 0.640 7 0.272 0.701 1.218 COVID-19 T+4 707 3 0.0951 0.969 0.774 8 0.227 0.427 2.883 RSV T+1 675 8 0.0801 0.877 0.643 2 0.857 0.997 1.922 RSV T+2 509 6 0.0142 0.843 0.668 1 0.949 0.966 4.042 RSV T+3 696 6 0.0286 0.875 0.876 9 0.576 0.036 3.356 RSV T+4 668 5 0.0267 0.998 0.852 10 0.592 0.432 3.525 Table 4: Optuna-selected XGBoost hyperparameters for the combined respiratory-burden model. Target Horizon Trees Depth η Subsample Colsample Child Gamma α λ Combined T+1 349 8 0.0889 0.869 0.866 4 0.504 0.454 1.806 Combined T+2 305 5 0.0496 0.608 0.956 4 0.944 0.607 0.639 Combined T+3 326 3 0.0461 0.617 0.756 6 0.242 0.454 0.136 Combined T+4 294 4 0.0516 0.804 0.728 10 0.559 0.456 1.948 For a trial parameter set ΞâÎΞâ , predictions are made for the validation rows across the chronological splits and scored using the MAE and PA definitions in Section 2.4. The same nonnegative clipping rule defined in Section 2.4 was applied before validation scoring and final evaluation. Let hâÎT_hâ denote the configurations evaluated by Optuna for horizon h. The selected parameter set is the trial that maximizes validation PA. After tuning hyperparameters, the selected parameter set is used to fit a retrospective test model on the training segment for retrospective evaluation. For each forecasting target and horizon, the retrospective test model is trained using only DtrainD_train. The gap rows are skipped, so they are not used for fitting: Fhtest=TrainhâĄ(Dtrain).F^test_h=Train_h(D_train). (27) For prospective forecasting, hyperparameters are not re-tuned. The selected parameters are carried forward to fit a final model on all complete rows with known horizon-specific labels before the forecast origin: final,h=(t,yt,h): _final,h=\(x_t,y_t,h): tâcomplete, _t\ complete, (28) yt,hâknown, y_t,h\ known, t<torigin. t<t_origin\. Fhfinal=TrainhâĄ(final,h).F^final_h=Train_h(D_final,h). (29) This final model is applied to the latest complete feature row to generate the T+1, T+2, T+3, and T+4 point predictions. 2.6 Adaptive Prediction Generation For both retrospective estimation and prospective forecasting, the fitted XGBoost model first generates a raw prediction. For target q, horizon h, and feature vector tx_t, the raw prediction is y^q,t,hraw=Fq,hâ(t). y^raw_q,t,h=F_q,h(x_t). (30) For retrospective evaluation, raw predictions are generated for the held-out test weeks. For prospective forecasting, the same raw-prediction form is generated from the final horizon-specific model at the forecast origin. In both settings, the reported and scored value is the adaptive prediction rather than the unadjusted raw output. The adjustment uses only historical scored weeks available before the prediction being evaluated; held-out retrospective targets and future prospective targets are not used to estimate the adjustment before prediction. The historical actual-to-predicted ratio for target q and horizon h is Ïq,h=1|âq,h|ââtââq,hyq,t,h1|âq,h|ââtââq,hy^q,t,hraw, _q,h= 1|H_q,h| _t _q,hy_q,t,h 1|H_q,h| _t _q,h y^raw_q,t,h, (31) where âq,hH_q,h denotes the historical scored weeks used for the target- and horizon-specific adjustment. The adaptive prediction is then generated as y^q,t,hadapt=maxâĄ(y^q,t,hrawâ Ïq,hâ αq,h,0), y^adapt_q,t,h= ( y^raw_q,t,h· _q,h· _q,h,0 ), (32) where αq,h _q,h is a target- and horizon-specific smoothing constant with 0<αq,h<10< _q,h<1, chosen during retrospective workflow development based on PA and then fixed for retrospective and prospective scoring. Thus, the adaptive prediction rescales the raw output by Ïq,h _q,h and αq,h _q,h. 2.7 Baseline Models Baseline comparisons were defined to contextualize both the disease-specific models and the combined-burden model before interpreting the adaptive forecasts. These comparisons show how the raw XGBoost model performs relative to standard time-series and tree-ensemble references. For the combined target, the same comparator set evaluates the combined-burden XGBoost output relative to alternative model classes. Baseline or reference models are commonly used in infectious disease forecast evaluations to determine whether a proposed approach improves on simpler forecasting strategies and to make performance differences easier to interpret across horizons and settings (Reich et al., 2019; Cramer et al., 2022). Baseline models were fit separately for influenza, COVID-19, RSV, and the combined target at each forecast horizon. For target qâInfluenza,COVIDâ-â19,RSV,Combinedqâ\Influenza,COVID -19,RSV,Combined\ and horizon h, the comparison included Baseline XGBoost, SARIMA, Random Forest, and the adaptive XGBoost model. Baseline XGBoost was defined as the raw XGBoost prediction: y^t,hbaseâ-âXGB,(q)=Fh(q)â(t). y_t,h^base -XGB,(q)=F_h^(q)(x_t). (33) SARIMA was retained as a univariate seasonal time-series baseline. For each target, it forecasts future admissions from the historical target series rather than from the full engineered feature set: y^t,hSARIMA,(q)=Sh(q)â(yt(q),ytâ1(q),âŠ). y_t,h^SARIMA,(q)=S_h^(q)(y_t^(q),y_t-1^(q),âŠ). (34) Random Forest was used as a non-boosted tree-ensemble baseline. It uses the same horizon-specific feature matrix as the machine-learning models and averages predictions across regression trees: y^t,hRF,(q)=1Bââb=1BTb(q,h)â(t). y_t,h^RF,(q)= 1B _b=1^BT_b^(q,h)(x_t). (35) Together, these baselines compare the adaptive XGBoost forecasts with an unadjusted boosting model, a seasonal time-series model, and a standard tree-ensemble model for the target-specific forecasts. The disease-specific and combined-burden baseline results are reported in Tables 3.1 and 3.1. 3 Results This section reports forecast performance in three stages: retrospective disease-specific results for influenza, COVID-19, and RSV; retrospective comparison of summed disease-specific forecasts with the combined-burden model; and prospective comparison of the same total-burden forecasting paradigms. 3.1 Retrospective Estimation Retrospective out-of-sample performance was evaluated after the training and temporal-gap periods for held-out target weeks from 2025-11-15 through 2026-02-21 across T+1 through T+4 horizons, using the split shown in Figure 3.1. Results are reported first for disease-specific influenza, COVID-19, and RSV forecasts (Table 3.1) and the combined-burden model (Table 3.1), followed by the total-burden comparison between summed disease-specific forecasts and the combined-burden model (Tables 7 and 3.1; Figure ). This ordering separates disease-level prediction performance from total-burden estimation, making the single-pathogen and combined-target results interpretable before the two burden paradigms are compared. Figure 2: Training, gap, retrospective, and prospective data split used for respiratory-admission forecasting. Table 5: Retrospective results across disease-specific models. Disease Horizon Baseline XGBoost MAE SARIMA MAE Random Forest MAE Adaptive XGBoost MAE Baseline XGBoost PA SARIMA PA Random Forest PA Adaptive XGBoost PA Influenza T+1 88.84 92.01 57.16 74.64 0.80 0.75 0.83 0.84 Influenza T+2 190.88 209.70 142.90 189.58 0.77 0.51 0.69 0.75 Influenza T+3 164.33 286.84 152.86 176.53 0.63 0.40 0.69 0.62 Influenza T+4 194.53 338.12 158.58 191.29 0.58 0.28 0.67 0.62 COVID-19 T+1 18.66 26.00 18.39 20.70 0.79 0.66 0.78 0.75 COVID-19 T+2 21.53 34.48 29.53 17.92 0.76 0.57 0.71 0.80 COVID-19 T+3 23.41 37.31 32.73 24.88 0.73 0.54 0.70 0.73 COVID-19 T+4 27.51 36.80 36.15 19.60 0.69 0.54 0.67 0.79 RSV T+1 17.18 38.71 37.84 11.67 0.81 0.65 0.69 0.89 RSV T+2 29.83 40.54 38.05 21.57 0.75 0.65 0.70 0.81 RSV T+3 30.98 35.02 38.91 25.64 0.70 0.68 0.69 0.80 RSV T+4 21.20 89.73 33.13 26.85 0.67 0.53 0.73 0.79 Table 6: Retrospective results across combined-burden models. Disease Horizon Baseline XGBoost MAE SARIMA MAE Random Forest MAE Adaptive XGBoost MAE Baseline XGBoost PA SARIMA PA Random Forest PA Adaptive XGBoost PA Combined T+1 114.42 144.52 114.32 80.63 0.83 0.73 0.81 0.85 Combined T+2 228.97 251.86 216.66 200.09 0.67 0.76 0.78 0.78 Combined T+3 289.41 257.83 160.32 183.97 0.62 0.72 0.74 0.74 Combined T+4 291.27 140.57 164.56 156.13 0.61 0.74 0.74 0.76 Table 3.1 shows disease- and horizon-specific variation across comparator models. For influenza, Adaptive XGBoost PA was 0.84, 0.75, 0.62, and 0.62 from T+1 to T+4, with MAE values of 74.64, 189.58, 176.53, and 191.29. COVID-19 PA was 0.75, 0.80, 0.73, and 0.79, with MAE values of 20.70, 17.92, 24.88, and 19.60. RSV had PA values of 0.89, 0.81, 0.80, and 0.79 and MAE values of 11.67, 21.57, 25.64, and 26.85. Across Baseline XGBoost, SARIMA, Random Forest, and Adaptive XGBoost, no single method dominated both PA and MAE for every diseaseâhorizon combination. Table 3.1 shows the same comparator-model evaluation for the combined-burden target. Adaptive XGBoost reduced MAE relative to baseline XGBoost at all horizons, from 114.42 to 80.63 at T+1, 228.97 to 200.09 at T+2, 289.41 to 183.97 at T+3, and 291.27 to 156.13 at T+4. PA also increased from 0.83, 0.67, 0.62, and 0.61 under baseline XGBoost to 0.85, 0.78, 0.74, and 0.76 after adaptive adjustment. Among the comparator models, Random Forest had the lowest MAE at T+3 (160.32), while Random Forest and Adaptive XGBoost had the same PA at that horizon (0.74). Adaptive XGBoost improved MAE and PA relative to baseline XGBoost at all combined-target horizons, although Random Forest matched or exceeded it for selected T+3 metrics. For the total-burden comparison, âSummedâ refers to the post-modeling sum of the separate influenza, COVID-19, and RSV adaptive forecasts, while âCombinedâ refers to the model trained directly on the aligned total respiratory-admission target. Tables 7 and 3.1 report the retrospective total-burden scores, and Figure displays the same comparison across T+1 through T+4. Table 7: Retrospective results for summed disease-specific forecasts. Target Horizon Adaptive XGBoost MAE Adaptive XGBoost PA Summed T+1 100.14 0.85 Summed T+2 200.06 0.80 Summed T+3 205.65 0.69 Summed T+4 217.57 0.69 Table 8: Retrospective results for the combined-burden model. Target Horizon Adaptive XGBoost MAE Adaptive XGBoost PA Combined T+1 80.63 0.85 Combined T+2 200.09 0.78 Combined T+3 183.97 0.74 Combined T+4 156.13 0.76 Tables 7 and 3.1 show that the two total-burden paradigms were similar at T+1 and T+2 but separated at longer horizons. From T+1 through T+4, the summed forecasts had MAE values of 100.14, 200.06, 205.65, and 217.57, with PA values of 0.85, 0.80, 0.69, and 0.69. The combined-burden model had MAE values of 80.63, 200.09, 183.97, and 156.13, with PA values of 0.85, 0.78, 0.74, and 0.76. PA was identical at T+1, slightly higher for the summed forecasts at T+2, and higher for the combined-burden model at T+3 and T+4. The combined-burden model had lower MAE at T+1, T+3, and T+4, while T+2 MAE was nearly identical between the two approaches. Figure highlights the longer-horizon separation: at T+3 and T+4, combined PA was 0.74 and 0.76 versus 0.69 and 0.69 for the summed forecasts, with MAE of 183.97 and 156.13 versus 205.65 and 217.57. 3.2 Prospective Forecasting The prospective evaluation assessed operational forecasts issued at the Center for Public Health Modeling and Response before the target weeks were observed and scored after admission counts became available. Target timestamps ran from 2026-02-28 through 2026-06-13. The horizon-specific models and calibration settings selected during retrospective development were carried forward without re-tuning. As in the retrospective analysis, total burden was scored using two paradigms: summed disease-specific forecasts and the combined-burden model. Tables 9 and 10 report the prospective total-burden scores, and Figure displays the comparison across T+1 through T+4. Table 9: Prospective results for summed disease-specific forecasts. Target Horizon Adaptive XGBoost MAE Adaptive XGBoost PA Summed T+1 40.21 0.70 Summed T+2 42.46 0.69 Summed T+3 64.43 0.62 Summed T+4 43.47 0.66 Table 10: Prospective results for the combined-burden model. Target Horizon Adaptive XGBoost MAE Adaptive XGBoost PA Combined T+1 36.33 0.71 Combined T+2 37.57 0.70 Combined T+3 33.19 0.70 Combined T+4 21.97 0.76 Tables 9 and 10 show that the combined-burden model had higher PA and lower MAE than the summed forecasts at all four prospective horizons. From T+1 through T+4, the summed forecasts had MAE values of 40.21, 42.46, 64.43, and 43.47, with PA values of 0.70, 0.69, 0.62, and 0.66. The combined-burden model had MAE values of 36.33, 37.57, 33.19, and 21.97, with PA values of 0.71, 0.70, 0.70, and 0.76. Differences were small at T+1 and T+2 and larger at T+3 and T+4; the largest MAE separation occurred at T+3, while the largest PA separation occurred at T+4. Figure shows the same prospective pattern visually: at T+3 and T+4, combined MAE was 33.19 and 21.97 versus 64.43 and 43.47 for the summed forecasts, with PA of 0.70 and 0.76 versus 0.62 and 0.66. 4 Discussion This study evaluated a central operational forecasting question: whether short-term respiratory hospital burden is better estimated with a unified multi-pathogen forecasting framework or by independently developed pathogen-specific models. The results support the unified combined-burden approach when the endpoint is total admissions requiring beds, staff, testing capacity, and respiratory support. Disease-specific forecasts still characterize influenza, COVID-19, and RSV trajectories separately, but the operational total-burden signal is better treated as its own forecasting target. This distinction is more than a technical modeling choice. Forecast target definition determines what error the model is trained to minimize and what operational question the forecast is designed to answer. When independently developed pathogen-specific forecasts are later aggregated, the resulting total inherits the timing, calibration, and scale errors of three separately tuned models. A combined-burden forecast instead learns the aggregate admission trajectory directly as its own target. This study therefore contributes to the growing literature on real-time and multi-source respiratory forecasting by focusing on target construction as well as model class (Mellor et al., 2025; Ray et al., 2025; Meyer et al., 2025; Xu et al., 2025). The combined-burden target should therefore be interpreted as a decision-relevant admission signal, not as a mechanistic representation of influenza, COVID-19, and RSV transmission. 4.1 Retrospective Estimation Analysis The retrospective evaluation showed that forecast behavior differed by pathogen, forecast horizon, and metric (Table 3.1). RSV had the most stable disease-specific agreement across horizons, whereas influenza showed the clearest deterioration at longer horizons. COVID-19 did not follow a simple monotonic pattern, which is consistent with the broader challenge of forecasting a pathogen whose recent hospitalization signal has been shaped by changing immunity, testing, reporting, variants, and public behavior (Klein et al., 2023; Lopez et al., 2024). These differences matter because the aggregated pathogen-specific paradigm, when used for total-burden planning, is only as stable as its separate components. If one pathogen has a timing error, another has a scale error, and a third has more stable relative agreement, the aggregate forecast can improve through accidental cancellation or deteriorate through error accumulation. The comparison with baseline XGBoost, SARIMA, and Random Forest also shows why a single accuracy measure is not sufficient. Adaptive XGBoost was strongest for RSV agreement across the four horizons and improved selected COVID-19 horizons, whereas Random Forest produced lower influenza MAE at several horizons. This absence of a universal winner is common in epidemic forecasting, where models may differ in short-term tracking, peak timing, calibration, and scale sensitivity (Reich et al., 2019; Cramer et al., 2022). PA and MAE therefore serve complementary roles. PA describes proportional agreement between predicted and observed admissions, while MAE describes the average number of admissions missed. For public health use, both are needed: a model that follows the trajectory but misses the admission count can still mislead staffing decisions, while a model with low average error may still fail to track the timing of a surge. At the total-burden level, the retrospective results indicate that the aggregated pathogen-specific and combined-burden paradigms were most similar at the shortest horizons and separated more clearly at longer horizons (Tables 7 and 3.1; Figure ). The key interpretation is not simply that the combined-burden model had better scores at horizon three and horizon four. Figure makes this separation visible across both PA and MAE panels. Rather, the longer-horizon pattern suggests that the combined target preserved the collective respiratory admission signal more effectively when disease-specific forecasts became harder to synchronize. This is plausible because the combined-burden model does not require each pathogen-specific model to correctly estimate its own timing and magnitude before aggregation. It directly optimizes the quantity that is later evaluated. This interpretation should be treated carefully. The present analysis does not prove that error accumulation is the sole mechanism behind the summed modelâs weaker longer-horizon performance. Differences may also reflect the reconstructed RSV history, the selected features, horizon-specific calibration, the timing of respiratory waves, or the particular evaluation period. Still, the result is operationally meaningful because horizon three and horizon four forecasts are often the horizons at which hospitals make schedule, staffing, supply, and capacity decisions that cannot be changed immediately. 4.2 Prospective Performance Analysis The prospective evaluation is especially important because it tested the workflow under real-time conditions rather than only through retrospective simulation. Forecasts were issued before the target weeks were observed, using the model choices and calibration settings carried forward from retrospective development. This mirrors the operational setting emphasized in recent respiratory forecasting work, where timeliness, repeatability, and decision relevance are as important as retrospective fit (Mellor et al., 2025; Meyer et al., 2025; Ahammed et al., 2025). This prospective step therefore tested the aggregated pathogen-specific versus unified combined-burden comparison under the timing constraints that would govern routine forecast use. The prospective results showed a more consistent advantage for the combined-burden model than the retrospective results (Tables 9 and 10; Figure ). The combined-burden model had higher PA and lower MAE at every horizon, with the largest separation at T+3 and T+4. Figure shows the same prospective pattern visually, with clearer separation at the longer horizons. This pattern strengthens the practical argument for using a unified burden target rather than aggregating separate disease forecasts when the operational target is total respiratory burden: in real-time use, the combined target remained competitive at the shortest horizons and became more favorable at the horizons most relevant for advance planning. The prospective MAE values should not be interpreted as a simple function of forecast horizon. Absolute error also depends on the level and shape of admissions during the evaluated weeks, and the prospective period covered a different calendar interval than the retrospective analysis. For that reason, the most defensible comparison is within the same evaluation period and horizon. Under that comparison, the combined-burden model had more favorable point-forecast performance. 4.3 Public Health Relevance The public health relevance of this work is that it separates two operational needs that are often conflated. Disease-specific forecasts for influenza, COVID-19, and RSV help identify which pathogen is contributing to expected admissions, while the combined-burden forecast estimates the total respiratory pressure hospitals must manage. In this study, the combined-burden model provided the more consistent total-admission signal, especially at longer planning horizons. This distinction matters because staffing, bed management, testing capacity, respiratory-support planning, and regional coordination depend on total burden rather than pathogen-specific counts alone (Bi et al., 2025; Mellor et al., 2025; Brouwer et al., 2026). The framework therefore preserves pathogen-specific interpretation while adding a hospital-facing burden estimate for respiratory-season preparedness, particularly when influenza, COVID-19, and RSV overlap or peak at different times (Bi et al., 2025; Dewey et al., 2025; Mellor et al., 2025). The 1â4 week horizon also has practical meaning. One-week forecasts can support immediate situational awareness and short-term staffing adjustments. Forecasts three and four weeks ahead are less certain but more useful for actions that require lead time, such as schedule planning, supply ordering, surge protocols, and coordination with public health partners. The finding that the combined-burden model had more favorable PA and MAE at these longer planning horizons therefore has direct relevance for preparedness, even though the present analysis does not yet translate forecast errors into specific operational cost or capacity outcomes. 4.4 Future Extensions Several limitations define the next steps for this work. First, the evaluation used data from one geographic setting and a limited number of respiratory seasons. Generalizability should therefore be tested across additional states, health systems, hospital catchment areas, and reporting environments where population structure, health system coverage, testing practice, and admission coding may differ (Ahammed et al., 2025). Second, the analysis used point forecasts. This is useful for comparing PA and MAE, but operational decision-making also requires calibrated uncertainty intervals, especially when forecasts are used to prepare for high-consequence surges (Pollett et al., 2021; Brouwer et al., 2026). Third, the evaluation did not formally quantify whether score differences were statistically distinguishable across horizons, nor did it decompose summed-forecast error into pathogen-specific timing, magnitude, and calibration components. Future work should extend the framework in five directions. Longer prospective evaluation is needed to determine whether the combined-burden modelâs advantage persists across multiple respiratory seasons, off-season activity, repeated waves, and high-burden peaks. Probabilistic forecasts should be added so that hospitals can reason about plausible ranges rather than single values. Coherent hierarchical forecasting should be explored so that pathogen-specific forecasts and total-burden forecasts can be reconciled rather than treated as competing products. Additional data streams, including wastewater, syndromic, mobility, laboratory, and local health-system signals, may improve timeliness and robustness when traditional reporting is delayed or revised (Ray et al., 2025; Wu et al., 2025; Tabassum et al., 2026). Finally, the forecasts should be linked to decision thresholds, such as expected bed demand, staffing triggers, testing-capacity thresholds, or respiratory support readiness, so forecast performance can be evaluated in terms of actionability rather than accuracy alone. 4.5 Conclusion This study suggests that respiratory-hospital forecasting can benefit from treating total respiratory burden as a distinct forecasting target when the planning question concerns shared hospital capacity. Disease-specific forecasts provide the pathogen-level information needed to monitor influenza, COVID-19, and RSV separately and interpret burden composition. In this evaluation, however, the combined-burden model more consistently captured shared respiratory admission demand than the total obtained from separate disease-specific forecasts, with the difference most apparent at longer planning horizons. These findings support a two-output forecasting framework: disease-specific models identify respiratory-burden composition, while the combined-burden model estimates shared capacity pressure. Combined-burden forecasting may therefore provide a decision-aligned endpoint for respiratory-season planning and a scalable foundation for epidemic and pandemic preparedness when multiple pathogens create overlapping hospital pressure. CRediT Authorship Contribution Statement Gaurob Saha: Conceptualization, methodology, software, validation, formal analysis, investigation, visualization, project administration, writing â original draft, writing â review and editing. Minjae Woo: Writing â review and editing. Lior Rennert: Writing â review and editing. Ethics and Data Governance Statement This study used aggregate or de-identified respiratory-surveillance and hospital-utilization data. No individual-level identifiable information is reported. Data use followed the applicable institutional and data-provider approvals and restrictions. Declaration of Competing Interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Funding The authors acknowledge the Public Health Science department at Clemson University for providing facilities and support. Acknowledgements The authors acknowledge the data providers and collaborators supporting respiratory disease surveillance and forecasting. Data Availability The complete analytic dataset is not publicly available because it combines public CDC NHSN source data with restricted South Carolina RFA hospitalization data and internal MUSC and PRISMA surveillance indicators. Public CDC NHSN source components can be obtained from CDC data repositories. RFA hospitalization data may be available from South Carolina RFA subject to approval and data-use restrictions. MUSC and PRISMA indicators are not publicly redistributable by the authors. Analysis code is available from the corresponding authors upon reasonable request, subject to institutional approvals and data-use restrictions. References Ahammed et al. (2025) T. Ahammed, M. S. Hossain, C. McMahan, and L. Rennert Machine learning approaches for real-time ZIP code and county-level estimation of state-wide infectious disease hospitalizations using local health system data. Epidemics 51, p. 100823. External Links: Document Cited by: §2.1, §2.5.1, §4.2, §4.4. Akiba et al. (2019) T. Akiba, S. Sano, T. Yanase, T. Ohta, and M. Koyama Optuna: a next-generation hyperparameter optimization framework. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, p. 2623â2631. External Links: Document Cited by: §2.5.3. Bergstra et al. (2011) J. Bergstra, R. Bardenet, Y. Bengio, and B. Kegl Algorithms for hyper-parameter optimization. In Advances in Neural Information Processing Systems, Vol. 24. External Links: Link Cited by: §2.5.3. Bi et al. (2025) K. Bi, S. R. Bandekar, A. Bouchnita, S. J. Fox, and L. A. Meyers Annual hospitalizations for COVID-19, influenza, and respiratory syncytial virus, united states, 2023-2024. Emerging Infectious Diseases 31 (3), p. 636â639. External Links: Document Cited by: §1, §1, §1, §4.3. Brouwer et al. (2026) A. F. Brouwer, M. C. Eisenberg, N. E. Dean, H. Hochheiser, P. Huang, J. R. Coyle, and L. Rennert Infectious disease modeling for public health practice: projections, scenarios, and uncertainty in 3 phases of outbreak response. American Journal of Epidemiology. External Links: Document Cited by: §4.3, §4.4. Chen and Guestrin (2016) T. Chen and C. Guestrin XGBoost: a scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, p. 785â794. External Links: Document Cited by: §2.5.3. Cramer et al. (2022) E. Y. Cramer, E. L. Ray, V. K. Lopez, J. Bracher, A. Brennen, et al. Evaluation of individual and ensemble probabilistic forecasts of COVID-19 mortality in the us. Proceedings of the National Academy of Sciences 119 (15), p. e2113561119. External Links: Document Cited by: §2.7, §4.1. Dewey et al. (2025) G. Dewey, A. G. Meyer, R. Garrido Garcia, and M. Santillana Uncovering the post-pandemic timing of influenza, RSV, and COVID-19 driving seasonal influenza-like illness in the united states. medRxiv. Note: Preprint External Links: Document Cited by: §1, §1, §1, §4.3. Gebremedhin et al. (2022) A. T. Gebremedhin, A. B. Hogan, C. C. Blyth, K. Glass, and H. C. Moore Developing a prediction model to estimate the true burden of respiratory syncytial virus (RSV) in hospitalised children in western australia. Scientific Reports 12, p. 332. External Links: Document Cited by: §1, §1. Hossain et al. (2025) M. S. Hossain, R. Goyal, N. K. Martin, V. DeGruttola, T. Ahammed, C. McMahan, and L. Rennert Machine learning and probabilistic approaches for forecasting COVID-19 transmission and cases. medRxiv. Note: Preprint External Links: Document Cited by: §2.5.1. Klein et al. (2023) B. Klein, A. C. Zenteno, D. Joseph, M. Zahedi, M. Hu, M. S. Copenhaver, M. U. G. Kraemer, M. Chinazzi, M. Klompas, A. Vespignani, S. V. Scarpino, and H. Salmasian Forecasting hospital-level COVID-19 admissions using real-time mobility data. Communications Medicine 3, p. 25. External Links: Document Cited by: §1, §2.1, §4.1. Lopez et al. (2024) V. K. Lopez, E. Y. Cramer, R. Pagano, J. M. Drake, E. B. OâDea, et al. Challenges of COVID-19 case forecasting in the us, 2020-2021. PLOS Computational Biology 20 (5), p. e1011200. External Links: Document Cited by: §4.1. Mellor et al. (2025) J. Mellor, M. L. Tang, O. Jones, T. Ward, S. Riley, and S. R. Deeny Forecasting COVID-19, influenza, and RSV hospitalizations over winter 2023-4 in england. International Journal of Epidemiology 54 (3), p. dyaf066. External Links: Document Cited by: §1, §1, §1, §1, §4.2, §4.3, §4. Meyer et al. (2025) A. G. Meyer, F. Lu, L. Clemente, and M. Santillana A prospective real-time transfer learning approach to estimate influenza hospitalizations with limited data. Epidemics 50, p. 100816. External Links: Document Cited by: §4.2, §4. Morris et al. (2023) M. Morris, P. Hayes, I. J. Cox, and V. Lampos Neural network models for influenza forecasting with associated uncertainty using web search activity trends. PLOS Computational Biology 19 (8), p. e1011392. External Links: Document Cited by: §2.5.2. Murph et al. (2026) A. C. Murph, L. J. Beesley, G. C. Gibson, L. A. Castro, S. Y. Del Valle, and D. Osthus A disease-agnostic approach to ensemble learning for infectious disease forecasting. Nature Communications 17, p. 4255. External Links: Document Cited by: §1, §1. Osthus (2022) D. Osthus Fast and accurate influenza forecasting in the united states with Inferno. PLOS Computational Biology 18 (1), p. e1008651. External Links: Document Cited by: §1, §1, §1, §2.5.2. OâBrien (2007) R. M. OâBrien A caution regarding rules of thumb for variance inflation factors. Quality & Quantity 41, p. 673â690. External Links: Document Cited by: §2.5.1. Park et al. (2024) S. Park, J. Moon, S. Jung, S. Rho, and E. Hwang Explainable influenza forecasting scheme using DCC-based feature selection. Data & Knowledge Engineering 149, p. 102256. External Links: Document Cited by: §1. Pollett et al. (2021) S. Pollett, M. A. Johansson, N. G. Reich, D. Brett-Major, S. Y. Del Valle, S. Venkatramanan, R. Lowe, T. Porco, I. M. Berry, A. Deshpande, et al. Recommended reporting items for epidemic forecasting and prediction research: the EPIFORGE 2020 guidelines. PLOS Medicine 18 (10), p. e1003793. External Links: Document Cited by: §2.5.3, §4.4. Ray et al. (2025) E. L. Ray, Y. Wang, R. D. Wolfinger, and N. G. Reich Flusion: integrating multiple data sources for accurate influenza predictions. Epidemics 50, p. 100810. External Links: Document Cited by: §1, §2.5.2, §4.4, §4. Reich et al. (2019) N. G. Reich, L. C. Brooks, S. J. Fox, S. Kandula, C. J. McGowan, E. Moore, D. Osthus, E. L. Ray, A. Tushar, T. K. Yamana, M. Biggerstaff, M. A. Johansson, R. Rosenfeld, and J. Shaman A collaborative multiyear, multimodel assessment of seasonal influenza forecasting in the united states. Proceedings of the National Academy of Sciences 116 (8), p. 3146â3154. External Links: Document Cited by: §2.7, §4.1. Tabassum et al. (2026) N. Tabassum, M. M. Chowdhury, C. S. McMahan, S. Self, M. Isanovic, K. E. Correa Velez, S. C. Sellers, R. S. Norman, and L. Rennert Granular insights: a wastewater-based machine learning approach for localized COVID-19 hospitalization forecasting. Epidemics 55, p. 100907. External Links: Document Cited by: §2.1, §4.4. Wu et al. (2025) J. Wu, S. Tanim, M. Woo, T. Ahammed, A. M. Bleichrodt, and L. Rennert A deep learning approach for enhancing pandemic prediction: a retrospective evaluation of transformer neural networks and multi-source data fusion for infectious disease forecasting. Epidemics 53, p. 100865. External Links: Document Cited by: §2.5.1, §4.4. Xu et al. (2025) S. Xu, H. Du, E. Dong, X. Wang, L. Zhang, and L. M. Gardner An optimized geo-hierarchical ensemble model to forecast hospitalizations from respiratory viruses in the united states. Epidemics 53, p. 100869. External Links: Document Cited by: §4. Yang et al. (2024) C. Yang, J. Gao, L. Glass, A. Cross, and J. Sun Multi-faceted analysis and prediction for the outbreak of pediatric respiratory syncytial virus. Journal of the American Medical Informatics Association 31 (1), p. 198â208. External Links: Document Cited by: §1, §1, §1. Yoon et al. (2026) G. Yoon, A. Bleichrodt, G. Chowell, and S. Lee Enhancing Influenza-Like Illness forecasting: an ensemble approach combining mathematical and deep learning models amidst the COVID-19 pandemic. Epidemics 54, p. 100901. External Links: Document Cited by: §1.