Spencer James Fox
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Spencer James Fox has a total of 644 co-authors. Here are the top co-authors by number of shared publications:
- Lauren Ancel Meyers
Department of Integrative Biology, The University of Texas at Austin
30 shared publications - Remy Pasco
Biofrontiers Institute, University of Colorado at Boulder
11 shared publications - Matthew Biggerstaff
Centers for Disease Control and Prevention, Atlanta, GA, USA
9 shared publications Michael Lachmann
Santa Fe Institute, Santa Fe, USA
9 shared publications- Nicholas G. Reich
University of Massachusetts Amherst, Amherst, MA, USA
9 shared publications - Bryan Lewis
University of Virginia, Charlottesville, VA, USA
8 shared publications
Latest Publications
- Open Access
- 18 November 2026
- Published by National Academy of Sciences in Proceedings of the National Academy of Sciences of the United States of America
- Vol. 123 (47)
Local influenza forecasts outperform state-level forecasts in the United States
Accurate outbreak forecasts are critical for timely and effective public health response. In the United States, however, most forecasts are produced at the state level, which can mask substantial substate heterogeneity and limit their utility for local planning. We generated and evaluated forecasts of the percentage of Emergency Department visits attributable to influenza across 173 large metropolitan Health Service Areas (HSAs) using a gradient boosting quantile regression model and compared their accuracy to forecasts derived from state-level data alone. At a 1-wk, 2-wk, and 3-wk horizon, local forecasts outperformed state-based forecasts in 98.8%, 90.8%, and 78.6% of HSAs, respectively, achieving mean weighted interval scores that were on average a 39.2% lower (95% range: 5.9 to 76.7%), 19.6% lower (−6.3 to 59.5%), and 11.4% lower (−11.7 to 44.9%), respectively. The performance advantage of local forecasting was strongest in HSAs representing a smaller share of their state’s population and increased with the proportion of the HSA population living in urban areas and the number of metropolitan areas within a state. These results, based on an analysis of HSAs with populations greater than 250,000, demonstrate that fine-scale modeling can substantially improve forecast accuracy and highlight the potential value of local forecasts for outbreak preparedness and response.
- 18 November 2026
- Published by National Academy of Sciences in Proceedings of the National Academy of Sciences of the United States of America
- Vol. 123 (47)
Open AccessLocal influenza forecasts outperform state-level forecasts in the United States
Accurate outbreak forecasts are critical for timely and effective public health response. In the United States, however, most forecasts are produced at the state level, which can mask substantial substate heterogeneity and limit their utility for local planning. We generated and evaluated forecasts of the percentage of Emergency Department visits attributable to influenza across 173 large metropolitan Health Service Areas (HSAs) using a gradient boosting quantile regression model and compared their accuracy to forecasts derived from state-level data alone. At a 1-wk, 2-wk, and 3-wk horizon, local forecasts outperformed state-based forecasts in 98.8%, 90.8%, and 78.6% of HSAs, respectively, achieving mean weighted interval scores that were on average a 39.2% lower (95% range: 5.9 to 76.7%), 19.6% lower (−6.3 to 59.5%), and 11.4% lower (−11.7 to 44.9%), respectively. The performance advantage of local forecasting was strongest in HSAs representing a smaller share of their state’s population and increased with the proportion of the HSA population living in urban areas and the number of metropolitan areas within a state. These results, based on an analysis of HSAs with populations greater than 250,000, demonstrate that fine-scale modeling can substantially improve forecast accuracy and highlight the potential value of local forecasts for outbreak preparedness and response.
- Open Access
- 4 November 2026
- Published by JMIR Publications Inc. in JMIR Public Health and Surveillance
- Vol. 12, e97405
Community Risk, Mobility Patterns, and COVID-19 Hospitalization Risk in Texas: ZIP Code–Level Ecological Longitudinal Study
Background: The COVID-19 pandemic revealed substantial disparities in health outcomes across US communities. High-resolution mobility data offer new opportunities for real-time surveillance of behavioral responses, yet less is known about how precautionary behavior varied within cities and how these differences were associated with subsequent COVID-19 hospitalization risk. Objective: This study assessed how community-level risk shapes heterogeneity in precautionary mobility behavior within cities and whether these differences are associated with subsequent COVID-19 hospitalization risk. Methods: We analyzed anonymized mobile device data from 623 ZIP codes in the 20 most populous counties in Texas and COVID-19 hospitalization records comprising 65,474 hospital admissions from 620 ZIP codes (February-July 2020). We constructed a ZIP code–level measure of precautionary mobility reduction (PMR) and linked it to a composite community risk score derived from US Centers for Disease Control and Prevention social vulnerability indicators. Generalized additive mixed models estimated associations among community risk, mobility patterns, and weekly hospitalization rates, adjusting for epidemic timing, population size, and spatial clustering. Results: Mobility declined by approximately 73% relative to baseline following statewide stay-at-home orders but varied substantially across ZIP codes. On April 2, 2020, PMR was negatively associated with community risk (slope −0.16, 95% CI −0.19 to −0.13; P <.001). At peak divergence, high-risk ZIP codes reduced mobility by 12 to 13 percentage points less than low-risk ZIP codes (95% CI 11 to 15). Greater PMR was associated with lower subsequent hospitalization risk (incidence rate ratio [IRR] 0.41, 95% CI 0.31 to 0.53; P <.001), whereas higher community risk was associated with increased hospitalization risk (IRR 3.03, 95% CI 2.57 to 3.58; P <.001). The protective association of mobility reduction was attenuated in higher-risk communities (interaction IRR 2.79, 95% CI 2.08 to 3.73; P <.001). Conclusions: Mobility data can support real-time surveillance of behavioral responses during epidemics. Within-city differences in mobility reduction were associated with subsequent hospitalization risk, but their protective association was weaker in higher-risk communities. Integrating digital mobility data with community risk metrics may improve targeted and equitable public health responses.
- 4 November 2026
- Published by JMIR Publications Inc. in JMIR Public Health and Surveillance
- Vol. 12, e97405
Open AccessCommunity Risk, Mobility Patterns, and COVID-19 Hospitalization Risk in Texas: ZIP Code–Level Ecological Longitudinal Study
Background: The COVID-19 pandemic revealed substantial disparities in health outcomes across US communities. High-resolution mobility data offer new opportunities for real-time surveillance of behavioral responses, yet less is known about how precautionary behavior varied within cities and how these differences were associated with subsequent COVID-19 hospitalization risk. Objective: This study assessed how community-level risk shapes heterogeneity in precautionary mobility behavior within cities and whether these differences are associated with subsequent COVID-19 hospitalization risk. Methods: We analyzed anonymized mobile device data from 623 ZIP codes in the 20 most populous counties in Texas and COVID-19 hospitalization records comprising 65,474 hospital admissions from 620 ZIP codes (February-July 2020). We constructed a ZIP code–level measure of precautionary mobility reduction (PMR) and linked it to a composite community risk score derived from US Centers for Disease Control and Prevention social vulnerability indicators. Generalized additive mixed models estimated associations among community risk, mobility patterns, and weekly hospitalization rates, adjusting for epidemic timing, population size, and spatial clustering. Results: Mobility declined by approximately 73% relative to baseline following statewide stay-at-home orders but varied substantially across ZIP codes. On April 2, 2020, PMR was negatively associated with community risk (slope −0.16, 95% CI −0.19 to −0.13; P <.001). At peak divergence, high-risk ZIP codes reduced mobility by 12 to 13 percentage points less than low-risk ZIP codes (95% CI 11 to 15). Greater PMR was associated with lower subsequent hospitalization risk (incidence rate ratio [IRR] 0.41, 95% CI 0.31 to 0.53; P <.001), whereas higher community risk was associated with increased hospitalization risk (IRR 3.03, 95% CI 2.57 to 3.58; P <.001). The protective association of mobility reduction was attenuated in higher-risk communities (interaction IRR 2.79, 95% CI 2.08 to 3.73; P <.001). Conclusions: Mobility data can support real-time surveillance of behavioral responses during epidemics. Within-city differences in mobility reduction were associated with subsequent hospitalization risk, but their protective association was weaker in higher-risk communities. Integrating digital mobility data with community risk metrics may improve targeted and equitable public health responses.
- Open Access
- 14 October 2026
- Published by Springer Nature in Communications Medicine
A hierarchical model for forecasting seasonal respiratory epidemics across pathogens and regions
- 14 October 2026
- Published by Springer Nature in Communications Medicine
Open AccessA hierarchical model for forecasting seasonal respiratory epidemics across pathogens and regions
- Open Access
Multi-season evaluation and analysis of categorical trend forecasts of influenza hospital admissions in the United States
Forecasting influenza hospitalizations informs public health preparedness, yet questions remain about which types of forecasts best guide action. We evaluate categorical trend forecasts, which communicate probabilities of upcoming increases or decreases in epidemic trajectories, submitted to CDC’s FluSight Forecasting Challenge between Fall-2024 and Spring-2026. Teams submitted probability distributions over five categories describing direction and magnitude of week-over-week changes in laboratory-confirmed influenza hospital admissions. We assessed performance using Ranked Probability Skill Score, Brier Skill Score, and measures of forecast-observation agreement. Most models outperformed an equal-probability baseline; the FluSight ensemble ranked among the top three in the 2024–25 and 2025–26 seasons. Forecasts were most accurate during stable periods and least during periods of rapid change, with most models underestimating observed trends. Conclusions were robust to choice of scoring metric and reference model. These results support categorical trend ensembles as an approach to communicating infectious disease forecasts that may inform public health decision-making.
Open AccessMulti-season evaluation and analysis of categorical trend forecasts of influenza hospital admissions in the United States
Forecasting influenza hospitalizations informs public health preparedness, yet questions remain about which types of forecasts best guide action. We evaluate categorical trend forecasts, which communicate probabilities of upcoming increases or decreases in epidemic trajectories, submitted to CDC’s FluSight Forecasting Challenge between Fall-2024 and Spring-2026. Teams submitted probability distributions over five categories describing direction and magnitude of week-over-week changes in laboratory-confirmed influenza hospital admissions. We assessed performance using Ranked Probability Skill Score, Brier Skill Score, and measures of forecast-observation agreement. Most models outperformed an equal-probability baseline; the FluSight ensemble ranked among the top three in the 2024–25 and 2025–26 seasons. Forecasts were most accurate during stable periods and least during periods of rapid change, with most models underestimating observed trends. Conclusions were robust to choice of scoring metric and reference model. These results support categorical trend ensembles as an approach to communicating infectious disease forecasts that may inform public health decision-making.
- Open Access
Adaptive multi-model ensembles for improved epidemic projections and decision support
In recent years, the use of multi-model ensemble projections in infectious disease modeling has become an established methodological approach to account for and integrate across uncertainties and structural differences present in individual models. However, the creation of long-term ensemble projections through these coordinated efforts is resource-intensive, demanding the input of multiple research teams and substantial computational power. This typically limits the ability to refine projections, update the selection of plausible epidemic trajectories, or expand the number of scenarios that can be assessed, even as new empirical data become available. To address this challenge, we define an adaptive ensemble approach that, analogously to a multi-model particle filtering method, dynamically selects individual model trajectories based on observed data throughout the epidemic projection period. We demonstrate the effectiveness of this methodology using the U.S. Flu Scenario Modeling Hub (SMH) projections for influenza hospitalizations in the United States during the 2023-2024 and 2024-2025 winter seasons. Our findings show that the adaptive ensemble yields improved predictive accuracy with respect to the original SMH ensemble projections across several scoring rules and geographical resolutions. Furthermore, the adaptive ensemble approach offers two additional applications: i) the dynamic assignment of posterior probabilities to epidemic scenarios, identifying the most plausible scenario, and representing how reality is captured by a combination of scenarios, and ii) the potential use for short-term forecasting. The adaptive ensemble approach is able to identify the most likely scenarios for the 2023-2024 and 2024-2025 U.S. influenza seasons, even in the early stages of the epidemic. It outperforms, retrospectively, a baseline model in short-term forecasting of influenza hospitalizations in the United States during the two seasons across various horizons and scoring rules, showing potential to contribute to real-time collaborative forecasting challenges such as CDC’s FluSight. The proposed approach offers an efficient or low-resource strategy to increase the impact of multi-model epidemic projections by providing real-time support to modeling teams, public health authorities, and decision-makers.
Open AccessAdaptive multi-model ensembles for improved epidemic projections and decision support
In recent years, the use of multi-model ensemble projections in infectious disease modeling has become an established methodological approach to account for and integrate across uncertainties and structural differences present in individual models. However, the creation of long-term ensemble projections through these coordinated efforts is resource-intensive, demanding the input of multiple research teams and substantial computational power. This typically limits the ability to refine projections, update the selection of plausible epidemic trajectories, or expand the number of scenarios that can be assessed, even as new empirical data become available. To address this challenge, we define an adaptive ensemble approach that, analogously to a multi-model particle filtering method, dynamically selects individual model trajectories based on observed data throughout the epidemic projection period. We demonstrate the effectiveness of this methodology using the U.S. Flu Scenario Modeling Hub (SMH) projections for influenza hospitalizations in the United States during the 2023-2024 and 2024-2025 winter seasons. Our findings show that the adaptive ensemble yields improved predictive accuracy with respect to the original SMH ensemble projections across several scoring rules and geographical resolutions. Furthermore, the adaptive ensemble approach offers two additional applications: i) the dynamic assignment of posterior probabilities to epidemic scenarios, identifying the most plausible scenario, and representing how reality is captured by a combination of scenarios, and ii) the potential use for short-term forecasting. The adaptive ensemble approach is able to identify the most likely scenarios for the 2023-2024 and 2024-2025 U.S. influenza seasons, even in the early stages of the epidemic. It outperforms, retrospectively, a baseline model in short-term forecasting of influenza hospitalizations in the United States during the two seasons across various horizons and scoring rules, showing potential to contribute to real-time collaborative forecasting challenges such as CDC’s FluSight. The proposed approach offers an efficient or low-resource strategy to increase the impact of multi-model epidemic projections by providing real-time support to modeling teams, public health authorities, and decision-makers.
- Open Access
Local Influenza Forecasts Outperform State-Level Forecasts in the United States
Accurate outbreak forecasts are critical for timely and effective public health response. In the United States, however, most forecasts are produced at the state level, which can mask substantial sub-state heterogeneity and limit their utility for local planning. We generated and evaluated forecasts of the percentage of Emergency Department visits attributable to influenza across 173 large metropolitan Health Service Areas (HSAs) using a gradient boosting quantile regression (GBQR) model, and compared their accuracy to forecasts derived from state-level data alone. At a one-week, two-week and three-week horizon, local forecasts outperformed state-based forecasts in 98.8%, 90.8%, and 78.6% of HSAs, respectively, achieving mean weighted interval scores that were on average a 39.2% lower (95% range: 5.9% to 76.7%), 19.6% lower (−6.3% to 59.5%), and 11.4% lower (−11.7% to 44.9%), respectively. The performance advantage of local forecasting was strongest in HSAs representing a smaller share of their state’s population and increased with the proportion of the HSA population living in urban areas and the number of metropolitan areas within a state. These results, based on an analysis of HSAs with populations greater than 250,000, demonstrate that fine-scale modeling can substantially improve forecast accuracy and highlight the potential value of local forecasts for outbreak preparedness and response. Significance Statement: During severe influenza seasons and pandemics such as COVID-19, anticipating the timing and severity of local peaks is critical for managing healthcare surges and communicating risk—yet these dynamics can vary substantially across local areas within the same state. However, most forecasting efforts in the United States provide only state-or national-level projections. This study shows that local influenza forecasts are both feasible and consistently more accurate than state-level projections, particularly in metropolitan areas that make up a small share of large states with multiple metropolitan areas.
Open AccessLocal Influenza Forecasts Outperform State-Level Forecasts in the United States
Accurate outbreak forecasts are critical for timely and effective public health response. In the United States, however, most forecasts are produced at the state level, which can mask substantial sub-state heterogeneity and limit their utility for local planning. We generated and evaluated forecasts of the percentage of Emergency Department visits attributable to influenza across 173 large metropolitan Health Service Areas (HSAs) using a gradient boosting quantile regression (GBQR) model, and compared their accuracy to forecasts derived from state-level data alone. At a one-week, two-week and three-week horizon, local forecasts outperformed state-based forecasts in 98.8%, 90.8%, and 78.6% of HSAs, respectively, achieving mean weighted interval scores that were on average a 39.2% lower (95% range: 5.9% to 76.7%), 19.6% lower (−6.3% to 59.5%), and 11.4% lower (−11.7% to 44.9%), respectively. The performance advantage of local forecasting was strongest in HSAs representing a smaller share of their state’s population and increased with the proportion of the HSA population living in urban areas and the number of metropolitan areas within a state. These results, based on an analysis of HSAs with populations greater than 250,000, demonstrate that fine-scale modeling can substantially improve forecast accuracy and highlight the potential value of local forecasts for outbreak preparedness and response. Significance Statement: During severe influenza seasons and pandemics such as COVID-19, anticipating the timing and severity of local peaks is critical for managing healthcare surges and communicating risk—yet these dynamics can vary substantially across local areas within the same state. However, most forecasting efforts in the United States provide only state-or national-level projections. This study shows that local influenza forecasts are both feasible and consistently more accurate than state-level projections, particularly in metropolitan areas that make up a small share of large states with multiple metropolitan areas.
- Open Access
Assessing the Impact of Timing and Coverage of United States COVID-19 Vaccination Campaigns: A Multi-Model Approach
Six years after its emergence, SARS-CoV-2 continues to have a substantial burden, however, the impact of vaccination and the optimal timing of its rollout remain uncertain. To explore these uncertainties, the US Scenario Modeling Hub convened its 19th round of ensemble projections for COVID-19 hospitalizations and deaths in the United States. Eight teams provided outcomes for each US state and nationally from April 2025 to April 2026 under five scenarios regarding vaccine recommendations and timing. We assessed recommendations with two eligibility scenarios (high-risk individuals only and all-eligible) and two timing scenarios (classic start: mid-August, earlier start: late June). These were crossed to create four scenarios and were compared against a counterfactual scenario with no vaccination. We found that compared to no vaccination, our ensemble projections estimated 90,000 (95% PI 53,000-126,000) hospitalizations averted in the high-risk and classic timing scenario across the US. Expanding coverage averted an additional 26,000 (95% PI 14,000-39,000) hospitalizations, which when coupled with earlier vaccination timing further reduced national hospitalizations by 15,000 (95% PI −3,000-33,000). These findings estimate significant benefits from a broad all-eligible vaccination recommendation, and suggest an additional benefit is likely to be gained from an earlier vaccination campaign.
Open AccessAssessing the Impact of Timing and Coverage of United States COVID-19 Vaccination Campaigns: A Multi-Model Approach
Six years after its emergence, SARS-CoV-2 continues to have a substantial burden, however, the impact of vaccination and the optimal timing of its rollout remain uncertain. To explore these uncertainties, the US Scenario Modeling Hub convened its 19th round of ensemble projections for COVID-19 hospitalizations and deaths in the United States. Eight teams provided outcomes for each US state and nationally from April 2025 to April 2026 under five scenarios regarding vaccine recommendations and timing. We assessed recommendations with two eligibility scenarios (high-risk individuals only and all-eligible) and two timing scenarios (classic start: mid-August, earlier start: late June). These were crossed to create four scenarios and were compared against a counterfactual scenario with no vaccination. We found that compared to no vaccination, our ensemble projections estimated 90,000 (95% PI 53,000-126,000) hospitalizations averted in the high-risk and classic timing scenario across the US. Expanding coverage averted an additional 26,000 (95% PI 14,000-39,000) hospitalizations, which when coupled with earlier vaccination timing further reduced national hospitalizations by 15,000 (95% PI −3,000-33,000). These findings estimate significant benefits from a broad all-eligible vaccination recommendation, and suggest an additional benefit is likely to be gained from an earlier vaccination campaign.
- Open Access
Basic Baseline model design choices can substantially influence performance in collaborative forecast hubs
Over the past decade, outbreak forecasting has become an increasingly used tool to assist public health decision-making during epidemics. Collaborative forecast hubs, where multiple teams submit predictions in real-time, are the gold standard for such efforts. For each hub, a Baseline model is used as a performance benchmark for other models. Although the Baseline is understood as a naïve forecast, its design is subjective, and the impact of model design decisions remains understudied. We evaluated how three Baseline specification decisions influence forecast performance on trend models that forecast based on historically observed dynamics: (1) the amount of historical data used for training, (2) whether the data are transformed, and (3) whether forecasts follow a flatline variant (constant predictions) or a drift variant (allowing a slope). Retrospective forecasts were generated for multiple years across four surveillance targets: COVID-19, influenza and RSV hospital admissions, and weighted influenza-like illness percentage. For wILI, we additionally compared trend baselines with a seasonal baseline model leveraging long-term historical patterns. Model specification significantly altered performance. The optimal performing model across targets was a flatline model that used the most recent 6-12 transformed observations. The optimal model outperforms the current standard Baseline used in many forecast hubs by an average of 9.6% (range: 3.7-12.9%) across forecast targets, and it outperformed the seasonal baseline model by 32.3% across nine influenza seasons. Our results demonstrate that subjective Baseline design decisions can materially influence forecast accuracy and, consequently, the perceived rankings of models within collaborative forecast hubs. Based on the varying approaches and their performance differences, these findings highlight the need for increased transparency in Baseline model specifications and support the routine inclusion of multiple benchmark models within collaborative forecast hubs. Author Summary: Forecasting the spread of infectious diseases is increasingly used to support public health decision-making during outbreaks. Many large forecasting efforts operate through collaborative forecast hubs, where multiple research teams submit predictions that are evaluated together. In these systems, a simple Baseline model serves as a benchmark against which more complex models are compared. However, the design of these Baseline models often involves subjective choices that have rarely been systematically studied. We examined how several common design decisions affect the performance of widely used Baseline approaches using multiple years of data for several respiratory diseases. We found that small changes in parameterization can substantially alter forecast accuracy and, consequently, the relative rankings of other models. Across all datasets, a single simple Baseline configuration consistently achieved the strongest performance, outperforming the most commonly used model. Our findings highlight the central and understudied role that Baseline models play in collaborative forecast hubs, and provide guidance to hub organizers regarding the appropriate design and communication of Baseline models.
Open AccessBasic Baseline model design choices can substantially influence performance in collaborative forecast hubs
Over the past decade, outbreak forecasting has become an increasingly used tool to assist public health decision-making during epidemics. Collaborative forecast hubs, where multiple teams submit predictions in real-time, are the gold standard for such efforts. For each hub, a Baseline model is used as a performance benchmark for other models. Although the Baseline is understood as a naïve forecast, its design is subjective, and the impact of model design decisions remains understudied. We evaluated how three Baseline specification decisions influence forecast performance on trend models that forecast based on historically observed dynamics: (1) the amount of historical data used for training, (2) whether the data are transformed, and (3) whether forecasts follow a flatline variant (constant predictions) or a drift variant (allowing a slope). Retrospective forecasts were generated for multiple years across four surveillance targets: COVID-19, influenza and RSV hospital admissions, and weighted influenza-like illness percentage. For wILI, we additionally compared trend baselines with a seasonal baseline model leveraging long-term historical patterns. Model specification significantly altered performance. The optimal performing model across targets was a flatline model that used the most recent 6-12 transformed observations. The optimal model outperforms the current standard Baseline used in many forecast hubs by an average of 9.6% (range: 3.7-12.9%) across forecast targets, and it outperformed the seasonal baseline model by 32.3% across nine influenza seasons. Our results demonstrate that subjective Baseline design decisions can materially influence forecast accuracy and, consequently, the perceived rankings of models within collaborative forecast hubs. Based on the varying approaches and their performance differences, these findings highlight the need for increased transparency in Baseline model specifications and support the routine inclusion of multiple benchmark models within collaborative forecast hubs. Author Summary: Forecasting the spread of infectious diseases is increasingly used to support public health decision-making during outbreaks. Many large forecasting efforts operate through collaborative forecast hubs, where multiple research teams submit predictions that are evaluated together. In these systems, a simple Baseline model serves as a benchmark against which more complex models are compared. However, the design of these Baseline models often involves subjective choices that have rarely been systematically studied. We examined how several common design decisions affect the performance of widely used Baseline approaches using multiple years of data for several respiratory diseases. We found that small changes in parameterization can substantially alter forecast accuracy and, consequently, the relative rankings of other models. Across all datasets, a single simple Baseline configuration consistently achieved the strongest performance, outperforming the most commonly used model. Our findings highlight the central and understudied role that Baseline models play in collaborative forecast hubs, and provide guidance to hub organizers regarding the appropriate design and communication of Baseline models.
- 26 November 2025
- Published by The Royal Society in Journal of The Royal Society Interface
- Vol. 22 (232) , 20250379
Physics-informed deep learning for infectious disease forecasting
Accurate forecasting of contagious illnesses has become increasingly important to public health policymaking and better prediction could prevent the loss of millions of lives. To better prepare for future pandemics, it is essential to improve forecasting methods and capabilities. In this work, we implement physics-informed neural networks (PINNs), a popular tool in the area of scientific machine learning, to perform infectious disease forecasting. The used PINNs model incorporates dynamical systems representations of disease transmission into the loss function, thereby assimilating epidemiological theory and data using neural networks. Our approach is designed to prevent model overfitting, which often occurs when training deep-learning models with observation data alone. In addition, we use an additional sub-network to account for mobility, cumulative vaccine doses and other covariates that influence the transmission rate, a key parameter in the compartmental model. To demonstrate the capability of the proposed model, we examine the performance of the model using state-level COVID-19 data in California. Our simulation results show that predictions of the PINNs model on the number of cases, deaths and hospitalizations are consistent with existing benchmarks. In particular, the PINNs model outperforms naive baseline forecasts and various sequence deep-learning models, such as recurrent neural networks, long short-term memory networks, gated recurrent units and transformer models. We also show that the performance of the PINNs model is comparable with that of a sophisticated Gaussian infection state forecasting model that combines the compartmental model, a data observation model and a regression model for inferring parameters in the compartmental model. Nonetheless, the PINNs model offers a simpler structure and is easier to implement. In summary, we perform a systematic study of the predictive capability of the PINNs model in forecasting the dynamics of infectious diseases and our results showcase the potential of the proposed model as an efficient computational tool to enhance the current capacity of infectious disease forecasting.
- 26 November 2025
- Published by The Royal Society in Journal of The Royal Society Interface
- Vol. 22 (232) , 20250379
Physics-informed deep learning for infectious disease forecasting
Accurate forecasting of contagious illnesses has become increasingly important to public health policymaking and better prediction could prevent the loss of millions of lives. To better prepare for future pandemics, it is essential to improve forecasting methods and capabilities. In this work, we implement physics-informed neural networks (PINNs), a popular tool in the area of scientific machine learning, to perform infectious disease forecasting. The used PINNs model incorporates dynamical systems representations of disease transmission into the loss function, thereby assimilating epidemiological theory and data using neural networks. Our approach is designed to prevent model overfitting, which often occurs when training deep-learning models with observation data alone. In addition, we use an additional sub-network to account for mobility, cumulative vaccine doses and other covariates that influence the transmission rate, a key parameter in the compartmental model. To demonstrate the capability of the proposed model, we examine the performance of the model using state-level COVID-19 data in California. Our simulation results show that predictions of the PINNs model on the number of cases, deaths and hospitalizations are consistent with existing benchmarks. In particular, the PINNs model outperforms naive baseline forecasts and various sequence deep-learning models, such as recurrent neural networks, long short-term memory networks, gated recurrent units and transformer models. We also show that the performance of the PINNs model is comparable with that of a sophisticated Gaussian infection state forecasting model that combines the compartmental model, a data observation model and a regression model for inferring parameters in the compartmental model. Nonetheless, the PINNs model offers a simpler structure and is easier to implement. In summary, we perform a systematic study of the predictive capability of the PINNs model in forecasting the dynamics of infectious diseases and our results showcase the potential of the proposed model as an efficient computational tool to enhance the current capacity of infectious disease forecasting.
- Open Access
- 24 October 2025
- Published by National Academy of Sciences in Proceedings of the National Academy of Sciences
- Vol. 122 (43)
Improving outbreak forecasts through model augmentation
Accurate forecasts of disease outbreaks are critical for effective public health responses, management of healthcare surge capacity, and communication of public risk. There are a growing number of powerful forecasting methods that fall into two broad categories-empirical models that extrapolate from historical data, and mechanistic models based on fixed epidemiological assumptions. However, these methods often underperform precisely when reliable predictions are most urgently needed-during periods of rapid epidemic escalation. Here, we introduce epimodulation, a hybrid approach that integrates fundamental epidemiological principles into existing predictive models to enhance forecasting accuracy, especially around epidemic peaks. When applied to empirical and machine learning forecasting methods (Autoregressive Integrated Moving Average, Holt-Winters, gradient-boosting machines, Prophet, and spline models), epimodulation improved overall prediction accuracy by an average of 12.3% (range: 8.5 to 18.7%) for COVID-19 hospital admissions and by 32.9% (range: 24.2 to 43.7%) for influenza hospital admissions; accuracy during epidemic peaks improved even further, by an average of 27.9% and 43.8%, respectively. Epimodulation also substantially enhanced the performance of complex forecasting methods, including the COVID-19 Forecast Hub ensemble model, demonstrating its broad utility in improving forecast reliability at critical moments in disease outbreaks.
- 24 October 2025
- Published by National Academy of Sciences in Proceedings of the National Academy of Sciences
- Vol. 122 (43)
Open AccessImproving outbreak forecasts through model augmentation
Accurate forecasts of disease outbreaks are critical for effective public health responses, management of healthcare surge capacity, and communication of public risk. There are a growing number of powerful forecasting methods that fall into two broad categories-empirical models that extrapolate from historical data, and mechanistic models based on fixed epidemiological assumptions. However, these methods often underperform precisely when reliable predictions are most urgently needed-during periods of rapid epidemic escalation. Here, we introduce epimodulation, a hybrid approach that integrates fundamental epidemiological principles into existing predictive models to enhance forecasting accuracy, especially around epidemic peaks. When applied to empirical and machine learning forecasting methods (Autoregressive Integrated Moving Average, Holt-Winters, gradient-boosting machines, Prophet, and spline models), epimodulation improved overall prediction accuracy by an average of 12.3% (range: 8.5 to 18.7%) for COVID-19 hospital admissions and by 32.9% (range: 24.2 to 43.7%) for influenza hospital admissions; accuracy during epidemic peaks improved even further, by an average of 27.9% and 43.8%, respectively. Epimodulation also substantially enhanced the performance of complex forecasting methods, including the COVID-19 Forecast Hub ensemble model, demonstrating its broad utility in improving forecast reliability at critical moments in disease outbreaks.