Source-linked AI summary
Artificial Intelligence Forecasting of Covid-19 in China
Zixin Hu, Qiyang Ge, Shudi Li, Li Jin, Momiao Xiong
TL;DR
The paper addresses the need for real-time Covid-19 forecasting when epidemiological models require unavailable parameters and assumptions. It develops an AI-based modified auto-encoder approach using confirmed-case data and clustering, and predicts high forecasting accuracy with the epidemic ending by mid-April under reliable-data and no-second-transmission conditions.
Problem
Epidemiological models require parameters and assumptions that may not be readily available during an outbreak, limiting real-time forecasting for public health planning.
Method
The study uses a modified stacked auto-encoder to forecast confirmed cases and combines latent variables with clustering algorithms to analyze transmission structure across China.
Results
The authors report high accuracy and predict that Covid-19 epidemics in China will be over by the middle of April.
Takeaways & Limitations
The AI-inspired methods are presented as a tool for public health planning and policymaking.
Takeaways & Limitations
The forecast assumes reliable data and no second transmissions.
Abstract
from arXiv · showhide
BACKGROUND An alternative to epidemiological models for transmission dynamics of Covid-19 in China, we propose the artificial intelligence (AI)-inspired methods for real-time forecasting of Covid-19 to estimate the size, lengths and ending time of Covid-19 across China. METHODS We developed a modified stacked auto-encoder for modeling the transmission dynamics of the epidemics. We applied this model to real-time forecasting the confirmed cases of Covid-19 across China. The data were collected from January 11 to February 27, 2020 by WHO. We used the latent variables in the auto-encoder and clustering algorithms to group the provinces/cities for investigating the transmission structure. RESULTS We forecasted curves of cumulative confirmed cases of Covid-19 across China from Jan 20, 2020 to April 20, 2020. Using the multiple-step forecasting, the estimated average errors of 6-step, 7-step, 8-step, 9-step and 10-step forecasting were 1.64%, 2.27%, 2.14%, 2.08%, 0.73%, respectively. We predicted that the time points of the provinces/cities entering the plateau of the forecasted transmission dynamic curves varied, ranging from Jan 21 to April 19, 2020. The 34 provinces/cities were grouped into 9 clusters. CONCLUSIONS The accuracy of the AI-based methods for forecasting the trajectory of Covid-19 was high. We predicted that the epidemics of Covid-19 will be over by the middle of April. If the data are reliable and there are no second transmissions, we can accurately forecast the transmission dynamics of the Covid-19 across the provinces/cities in China. The AI-inspired methods are a powerful tool for helping public health planning and policymaking.
BACKGROUND
The study applies modified auto-encoders to forecast Covid-19 transmission across China, estimating cumulative-case trajectories, plateau timing, and province/city clusters. It reports high forecasting accuracy and projects the epidemic ending by mid-April under stated conditions.
- A modified stacked auto-encoder modeled Covid-19 transmission dynamics and forecast confirmed cases across China.
- Multiple-step forecasting errors were 1.64%, 2.27%, 2.14%, 2.08%, and 0.73% for 6-step through 10-step forecasts, respectively.
- Province/city plateau timing varied from Jan 21 to April 19, 2020, and 34 provinces/cities formed 9 clusters.
- The authors characterized the AI-based forecasting accuracy as high and predicted the epidemic would be over by the middle of April.
- Accurate forecasting was conditional on reliable data and the absence of second transmissions.
Introduction
Existing epidemiological models can analyze transmission dynamics but require parameters and assumptions that may not be readily available during an outbreak. The study therefore develops an AI-based real-time forecasting method to support planning and characterize provincial transmission patterns.
- Epidemiological models such as SEIR have been developed to analyze Covid-19 transmission dynamics.
- These models can estimate transmission dynamics, target resources, and evaluate intervention strategies.
- Their use requires parameters and depends on many assumptions.
- Real-time outbreak data were insufficient for readily estimating model parameters, so analyses often used hypothesized parameters that did not fit the data well.
- The study develops an AI-based method for real-time forecasting of new and cumulative confirmed cases across China.
- It also forecasts transmission trends and plateaus and groups provinces/cities into clusters according to dynamic transmission patterns.
Data Sources
The analysis uses confirmed Covid-19 case data from China, organized as province/city time series of daily new and accumulated cases. The dataset incorporates reporting sources and a stated adjustment to Hubei counts before February 14, 2020.
- The confirmed-case time series began on January 11, 2020 for each province/city.
- Reporting from January 21 to February 27, 2020 used data from the Surging News Network and WHO situation reports, respectively.
- WHO used laboratory-confirmed cases through February 13 and combined clinical and laboratory-confirmed cases after February 14.
- Hubei counts before February 14 were adjusted using the laboratory-confirmed count and a stated formula.
- The dataset included accumulated and new confirmed cases for China, 31 mainland provinces/cities, Hong Kong, Macau, and Taiwan.
- Data were organized in a matrix with China and province/city units as rows and daily new confirmed cases as columns.
Modified Auto-encoder for Modeling Time Series
The modified auto-encoder forecasts accumulated confirmed cases from normalized time-series segments. Its architecture expands from 8 input nodes to 32 and 4 latent-layer nodes, produces one output, and uses weighted loss with repeated training.
- Modified auto-encoders forecast accumulated and new confirmed Covid-19 cases from time-series data.
- Each 8-day time-series segment was treated as a sample, with 128 segments used for training.
- Segments were formed by randomly selecting a start day and its 7 successive days from the case matrix.
- The loss function compares observed and MAE-forecasted cases using weights that increase across later 12-day intervals.
- Backpropagation estimated MAE weights and biases, and five training repetitions were averaged into the final province/city forecast.
Forecasting Procedures
The trained MAE forecasts future confirmed cases using normalized recent case histories, with recursive multiple-step forecasting extending predictions across successive days.
- The trained MAE produces forecasts of future new confirmed Covid-19 cases.
- For each province/city, the model uses an 8-dimensional matrix of recent case values and its row-wise average for normalization.
- The MAE output is converted into one-step forecasts of new confirmed cases for each province/city.
- Recursive multiple-step forecasting feeds each preceding prediction back into the model to forecast the following time step.
- The predicted total number of new confirmed cases is obtained by summing the final forecasted new-case values for each province/city.
Clustering
Latent representations from the modified auto-encoder were transformed into province/city feature vectors and clustered to investigate transmission structure.
- The second latent-layer values of the MAE were extracted for each province/city.
- A 34 × 4 latent matrix was formed for each province/city, and its largest singular value was obtained by singular value decomposition.
- Five training runs produced five largest singular values for each province/city.
- Each feature vector combined the five singular values with outbreak timing, peak new cases, peak forecasted new cases, and forecasted accumulated cases.
- The k-means algorithm grouped the 34 province/city feature vectors into clusters.
Results
The forecasts closely tracked reported curves and projected heterogeneous provincial trajectories, while clustering revealed geographically and socioeconomically related transmission patterns.
- 5,236 new confirmed cases were forecasted at the February 5, 2020 peak, declining to zero on April 20.
- 83,401 cumulative confirmed cases were projected at the national plateau on April 20, 2020.
- Forecasting accuracy was reported as very high, although average errors did not strictly increase with longer forecast horizons because of data fluctuations.
- Forecasted cumulative-case curves covered 31 mainland provinces/cities plus Hong Kong, Macau, and Taiwan through April 20, 2020.
- Plateau entry times varied across provinces/cities: Xizang entered first on January 21, while Hubei was among the last around April 20 with 70,019 cases.
- Most provinces/cities were projected to enter the plateau around mid-March, with differing curve shapes indicating different transmission-dynamics patterns.
- 34 provinces/cities formed nine clusters, partially reflecting geography and contact patterns.
- Hubei formed its own cluster, while surrounding provinces formed another and Inner Mongolia, Xizang, and Qinghai had the lowest case counts in a distant-source cluster.
Discussion
The discussion presents modified auto-encoder methods as an alternative, data-driven approach for real-time Covid-19 forecasting and transmission-structure analysis across China. The forecasts estimated epidemic trajectories, plateau timing, and provincial or city clusters, while their intervention-analysis functions remained unexplored because of insufficient data.
- Modified auto-encoder methods provide an alternative to epidemiologic transmission models for real-time Covid-19 forecasting across China.
- Forecast accuracy and subsequent multiple-step forecasting were high, and longer training time improved forecasting performance.
- The forecasts estimated decreasing growth, epidemic lengths, and plateau timing for cumulative confirmed-case curves across 34 provinces/cities.
- The 34 provinces/cities were grouped into 9 clusters to investigate geographic and healthcare-resource patterns in transmission structure.
- The models can incorporate intervention information and simulate intervention impacts on outbreak size, severity, and ending time, but these functions were not explored because of limited data.
- The AI-based methods were presented as tools for tracking epidemic trajectories, assessing severity, predicting epidemic lengths, and supporting public-health planning.
Legend
The figure legend describes Covid-19 case time series and provincial or city transmission dynamics over time, along with a nine-cluster grouping.
- Figure 1 depicts the architecture of a MAE.
- The figures cover Covid-19 cases across China as a function of days from January 11, 2020 to April 20, 2020.
- The case time-series figure includes cumulative confirmed cases across 31 mainland provinces/cities and three other Chinese regions.
- The provinces/cities and other regions formed 9 clusters.
Supplementary Figure Legend
The supplementary figures show reported and fitted Covid-19 curves, forecasted cumulative cases, province/city groupings, and national forecasting errors across China.
- The reported and fitted cumulative confirmed-case curves for China span January 11 to February 27, 2020.
- The reported and forecasted national cumulative and new confirmed-case curves are presented across the forecasting period.
- Forecasted cumulative confirmed-case curves are shown for provinces and cities in China from January 11 to April 20, 2020.
- Cumulative confirmed-case time series across 31 mainland provinces/cities and three other Chinese regions formed 9 clusters.
- Table 1 reports errors of forecasting China’s national cumulative confirmed cases.
- The supplementary national curves distinguish reported cases from fitted cases using red and green curves, respectively.