Data-Driven Models Informed by Spatiotemporal Mobility Patterns for Understanding Infectious Disease Dynamics

Question

1. What is the proposed method for understanding disease dynamics and real-time surveillance?

2. How were risk indices developed in the study?

3. How can mobility-informed risk indices be calculated using individual mobility patterns and contact intensity?

4. How do logistic regression and random forest classifiers predict high-risk subdistricts?

Accepted Answer

The proposed method involves developing two mobility-informed risk indices to describe the risk of infectious disease transmission in space and time. These risk indices are combined with other relevant variables and used in statistical regression and machine learning models. The method can detect outbreaks caused by newly introduced acute human-to-human transmitted diseases at the early stage of the outbreak and predict short-term trends of transmission in community hotspots where populations have not yet acquired herd immunity. The method was tested using real-world data on COVID-19 outbreaks in Chinese cities and the United States, showing its effectiveness in identifying high-risk areas and assisting in the implementation of interventions to control disease spread. It also maintained a high level of performance for one-to four-week-ahead forecasts of the county-level COVID-19 prevalence in the United States, contributing to real-time surveillance of disease dynamics within the country.

Accepted Answer

Risk indices were developed by analyzing individuals' movements and contacts over space and time. This involved examining the patterns and interactions of people to identify potential risk factors for infectious disease transmission. By understanding the mobility and contact patterns, researchers were able to create risk indices that could serve as predictive variables for infectious disease dynamics. These indices were then combined with socio-economic, demographic, environmental, and epidemiological factors to enhance the accuracy of the statistical and machine learning models used in the study.

Accepted Answer

Mobility-informed risk indices can be calculated by examining individual mobility patterns and contact intensity. The case flow intensity (CFI) considers the location of initial cases and their movements across regions, such as subdistricts or counties. It quantifies regional infection risk by counting the cumulative number of initial cases that visited a region. The case transmission intensity (CTI) represents the risk introduced by both inter-regional movements and intra-regional contact with initial cases. It is based on the number of potential new infections resulting from the activity of initial cases. The CFI-based regional infection risk can be depicted by the hourly cumulative counts of initial cases, while the CTI risk index can be computed by the hourly cumulative counts of potential new infections due to contact with initial cases within a region. These indices are derived using an established travel network and detailed population flow data, as well as the locations of the first confirmed cases. Python programming language is used for data processing, calculation of mobility-informed risk indices, model training, and testing for statistical regression, including random forest models. The infection rate is given by the intra-regional transmission rate derived from the logged POI-based diversity index, which depicts neighborhood vibrancy and human activity. Overall, mobility-informed risk indices provide valuable insights into the transmission risks of infectious diseases by considering individual mobility patterns and contact intensity.

Accepted Answer

Logistic regression and random forest classifiers are used to predict high-risk subdistricts by utilizing CFI and CTI risk indices. These indices are computed from collected data on actual COVID-19 outbreaks in Beijing and Guangzhou. The sample data is used for training and tuning the models. The CFI and CTI risk indices are then inputted into the classifiers to determine which subdistricts are at risk of being affected. The fitted models are applied to predict the affected subdistricts in actual COVID-19 outbreaks, and the accuracy of the predictions is evaluated by comparing them to real-world data. This method helps in assessing outbreaks of newly emerging or emergent acute human-to-human transmitted diseases in a city, allowing for early intervention and prevention strategies.

Accepted Answer

In analyzing COVID-19 outbreaks in Beijing and Guangzhou, various data sources were utilized. For Beijing, information on affected subdistricts and the number of cases were obtained from press releases and daily pandemic notification reports released by the Beijing Municipal Health Commission. Population data for 2021 at a 100-meter resolution were obtained from WorldPop. Hourly two-day data from 11-12 June 2020 was used to capture people's movement between subdistricts. For Guangzhou, hourly inter-subdistrict population flow data from 21-22 May 2021 was used. POI data for 2020 was obtained from AMap Services. These data sources provided valuable insights into the mobility, demographics, and epidemiological outbreaks in the two cities.

Accepted Answer

The SEIR model was employed to simulate COVID-19 transmission across various outbreak scenarios in Beijing and Guangzhou. It utilized a travel network-based framework and simulated epidemiological data to develop predictive models for subdistricts. The model estimated the daily number of new cases in each subdistrict to determine the number of affected subdistricts throughout an outbreak. Different levels of transmissibility and random source locations were considered, generating a total of 90 sets of COVID-19 outbreak epidemic data for each city. The time-series changes in the number of daily cases were shown in Figure A1. Primary predictive variables included population, population density, number of POIs, POI density, population flow volume, and mobility-informed risk indices. The mobility-informed risk indices, CFI and CTI, were calculated based on hourly flow data and location of initial confirmed cases to determine initial stage risk levels. The spatiotemporal population, number of cases, and potential infections were used to obtain the early-stage CFI and CTI risks. This approach allowed for the simulation of COVID-19 propagation across subdistricts within a city.

Accepted Answer

Sensitivity and specificity are key metrics used to assess the performance of logistic regression and random forest classifiers. Sensitivity, also known as the true positive rate, is calculated as the ratio of correctly estimated affected subdistricts to the total number of actual affected subdistricts. It measures the proportion of actual positives that are correctly identified by the model. Specificity, on the other hand, is calculated as the ratio of correctly estimated unaffected subdistricts to the total number of actual unaffected subdistricts. It measures the proportion of actual negatives that are correctly identified by the model. Both sensitivity and specificity are derived from the confusion matrix, a two-by-two table generated by a binary classifier, which presents four possible outcomes: true positive, false positive, true negative, and false negative. These metrics are crucial in evaluating the accuracy and reliability of the prediction models in identifying affected and unaffected subdistricts.

Accepted Answer

Regression models utilize mobility-informed risk indices, such as CFI and CTI, to predict COVID-19 cases. These indices are computed with temporal lags based on daily population flows and reported weekly cases. The models, including elastic net and random forest regression, are trained using log-transformed incidence rates as the target variable. They incorporate multiple base predictive variables from previous studies. The models generate forecasts of weekly case increases, which are evaluated against actual confirmed cases to assess predictive performance.

Accepted Answer

The source of county-level daily COVID-19 cases data is USA Facts (usafacts.org/visualizations/coronavirus-covid-19-spread-map, accessed on 15 July 2022). USA Facts is a reputable non-profit organization that provides data on government tax revenues, expenditures, and outcomes. The data on COVID-19 cases from USA Facts have been widely used in previous studies on COVID-19 spread characteristics [43, 44].

Accepted Answer

The rationale behind using the log-transformed target variable, as opposed to directly predicting the number of weekly new cases, was to minimize skewness and, more importantly, reduce the sensitivity of the models to the population of counties. This approach helps in achieving more accurate and reliable predictions by addressing the distributional characteristics of the data and reducing the impact of extreme values.

Accepted Answer

The proposed mobility-informed indices, logistic regression, and random forest classifiers outperformed the SEIR epidemiological model in predicting affected subdistricts during COVID-19 outbreaks in Beijing and Guangzhou. The SEIR model failed to identify 24 out of 52 affected subdistricts, resulting in a sensitivity of only 54%. In contrast, the proposed models demonstrated a range of 50-90% accuracy in identifying affected subdistricts, with logistic regression and random forest classifiers achieving high sensitivity and specificity rates. The logistic regression model successfully identified 87% of affected subdistricts in Guangzhou, while the random forest classifier outperformed it in Beijing. The proposed models also achieved an Area Under the Curve (AUC) value exceeding 0.9 in the Receiver Operating Characteristic (ROC) analysis for Beijing, indicating their superior performance in predicting affected subdistricts compared to the SEIR model.

Accepted Answer

The proposed models, which incorporate CFI and CTI risk indices, led to a decrease in MAE for most forecast dates compared to the reference studies. Specifically, the proposed method using predictive variables in REF1 (Proposed1) had a lower MAE than REF1 for one-week-ahead forecasts on 28 and 37 of 39 dates using the elastic net and random forest models, respectively. On average, REF1 had a higher MAE for four-weeks-ahead forecasts on 20 and 24 dates using the two models. The incorporation of a regression forecasting model, in combination with CFI and CTI related risk indices, led to a decrease in MAE of REF2 for forecasts ranging from one to four weeks in advance. Specifically, the utilization of elastic net regression resulted in MAE reductions for 9, 9, 7, and 7 out of the 11 dates, respectively. Similarly, employing random forest regression in conjunction with CFI and CTI related risk indices resulted in MAE reductions of 9, 7, 7, and 7 out of the 11 dates for forecasts spanning one to four weeks ahead. The Proposed1 method improved R 2 for one-to four-weeks-ahead forecasts on 35, 31, 31, and 31 dates compared to REF1 when using random forest regression. Additionally, Proposed2 improved R 2 on 8, 9, 8, and 7 dates against REF2. The Proposed1 method achieved an R 2 higher than 0.5 for one-to four-weeks-ahead forecasts on 39 dates, and the Proposed2 method achieved an R 2 higher than 0.8 for the forecasts on 11 dates. The proposed models incorporating CFI and CTI risk indices demonstrated improved forecasting performance compared to the reference studies.

Accepted Answer

The use of random forest regression on the same forecast dates revealed that while the R 2 of REF 2 was higher, its MAE was generally greater when compared to REF 1. The average increase in MAE from REF 1 to REF 2 was -1.6, 23.8, 37.8, and 33.6 for one-to four-weeks-ahead forecasts, respectively. R 2 increased by 0.04, 0.04, 0.03, and 0.07. This indicates that REF2 had a higher R 2 but a greater MAE compared to REF1.

Accepted Answer

The comparison of forecasting performance between two reference studies using mean absolute error (MAE) and R-square (R2) was conducted by predicting weekly increases in the number of COVID-19 cases in U.S. counties for one to four weeks ahead. The MAE and R2 values were calculated for the estimated and actual number of cases. The proposed method was compared to two reference studies, REF1 and REF2, respectively. The MAE and R2 differences between REF1 and REF2 for one-to four-weeks-ahead forecasts using random forest regression were analyzed. The MAE and R2 differences between the proposed method and reference study using random forest regression were also examined. The results showed that the proposed method consistently included several CFI and CTI related risk variables, such as IN_CFI_T and IN_CTI_T, which enhanced the precision of COVID-19 forecasting models. The models that incorporated CFI and CTI indices demonstrated a reduced MAE and an increased R2 compared to the reference experiments. The impact of including CFI and CTI was more significant in reducing MAE and increasing R2 for REF1 than REF2. Overall, the comparison of MAE and R2 highlighted the importance of including CFI and CTI risk indices in COVID-19 forecasting models.

Accepted Answer

Mobility-informed risk indices, such as CFI and CTI, can significantly enhance the spatiotemporal prediction of the prevalence of infectious diseases by accounting for the physics of disease spread. These indices provide valuable supplemental physical information that is not heavily relied upon in traditional SEIR models. By incorporating CFI and CTI risk indices, which reflect the spatial risk in the two weeks following the forecast date, the accuracy of disease prevalence predictions can be improved, particularly when other predictive variables have limited physics-related information. The addition of mobility-informed indices has been shown to improve forecasting performance, especially for one-and two-weeks-ahead forecasts. However, the improvement slightly reduces as the forecasting horizon extends to three-and four-weeks-ahead forecasts. Despite the limitations of the proposed method, such as the need for a training set generated by SEIR model simulations and the potential influence of multicollinearity among variables, the incorporation of mobility-informed risk indices has proven to be essential in ensuring the accuracy of disease prevalence predictions.

Accepted Answer

Predictive models are crucial for timely interventions in mitigating and monitoring disease outbreaks. A data-driven approach enables rapid predictions, enhancing understanding of disease dynamics. Mobility-informed risk indices provide accurate risk portrayal, improving prediction accuracy and physical consistency. While SEIR models are widely used, machine learning and statistical regression models also show potential in complex data scenarios. Overall, a data-driven approach informed by physical information is promising for detecting and responding to infectious disease outbreaks and epidemics.

Data-Driven Models Informed by Spatiotemporal Mobility Patterns for Understanding Infectious Disease Dynamics

Chat with Paper

AI Agents for this Paper

Most frequently asked questions

1. What is the proposed method for understanding disease dynamics and real-time surveillance?

2. How were risk indices developed in the study?

3. How can mobility-informed risk indices be calculated using individual mobility patterns and contact intensity?

4. How do logistic regression and random forest classifiers predict high-risk subdistricts?

5. What data sources were used for analyzing COVID-19 outbreaks in Beijing and Guangzhou?

6. How was the SEIR model used in simulating COVID-19 transmission?

7. How are sensitivity and specificity calculated in logistic regression and random forest classifiers?

8. How do regression models predict COVID-19 cases using mobility-informed risk indices?

9. What is the source of county-level daily COVID-19 cases data?

10. What is the rationale behind using the log-transformed target variable in predicting weekly new cases of COVID-19?

11. How do proposed models compare to SEIR in predicting affected subdistricts during COVID-19 outbreaks?

12. How did the proposed models incorporating CFI and CTI risk indices impact forecasting performance compared to reference studies?

13. How did REF2 compare to REF1 in terms of MAE and R 2?

14. How do MAE and R2 compare in forecasting performance?

15. How can mobility-informed risk indices improve the accuracy of disease prevalence predictions?

16. Importance of predictive models in infectious disease control?

Citations

Predicting People’s Concentration and Movements in a Smart City

References

Random Forests

Regularization and variable selection via the elastic net

Early Transmission Dynamics in Wuhan, China, of Novel Coronavirus-Infected Pneumonia.

The Mathematics of Infectious Diseases

Temporal dynamics in viral shedding and transmissibility of COVID-19.

Related Papers (5)

An outbreak of foodborne giardiasis.

Monitoring the incidence and causes of disease potentially transmitted by food in Australia: Annual report of the OzFoodNet network, 2017.

Norovirus outbreaks: a systematic review of commonly implicated transmission routes and vehicles.

A scoping review of risk factors and transmission routes associated with human giardiasis outbreaks in high-income settings

Get the News Out Loudly and Quickly: Modeling the Influence of the Media on Limiting Infectious Disease Outbreaks