Artificial Intelligence Forecasting of Covid-19 in China

METHODS

• The study used a modified stacked auto-encoder to model transmission dynamics and forecast confirmed COVID-19 cases in China in real time.Source: METHODS · p. 2
• WHO data collected from January 11 to February 27, 2020 were used for the forecasting.Source: METHODS · p. 2

RESULTS

• Forecasted cumulative-case curves covered Jan 20, 2020 to April 20, 2020; average errors were 1.64% (6-step), 2.27% (7-step), 2.14% (8-step), 2.08% (9-step), and 0.73% (10-step).Source: RESULTS · p. 2
• Province/city plateau entry estimates ranged from Jan 21 to April 19, 2020, and the 34 provinces/cities were grouped into 9 clusters.Source: RESULTS · p. 2

CONCLUSIONS

• The authors described the AI-based forecasting accuracy as high and predicted that the COVID-19 epidemic would be over by the middle of April.Source: CONCLUSIONS · p. 2
• Accurate forecasts across Chinese provinces and cities were conditional on reliable data and the absence of second transmissions.Source: CONCLUSIONS · p. 3
• The authors presented AI-inspired methods as a tool for supporting public health planning and policymaking.Source: CONCLUSIONS · p. 3

Introduction

• Epidemiological models, including SEIR, have been developed to analyze Covid-19 transmission, but they require parameters and assumptions that were not readily available in real time.The introduction notes that analyses using hypothesized parameters may not fit the data well, potentially limiting forecast accuracy.Source: Introduction · p. 4
• The authors develop an AI-based method to forecast new and cumulative confirmed cases across China and support public health planning and policymaking.They also aim to forecast transmission trends and plateaus and cluster provinces and cities by transmission patterns.Source: Introduction · p. 4
• The analysis uses surveillance data on confirmed Covid-19 cases in China.The supplied passage specifies that these data extend up to a stated cutoff, but the cutoff is incomplete in the excerpt.Source: Introduction · p. 4

Data Sources

• Confirmed COVID-19 case data for January 11–20, 2020, and January 21–February 27, 2020, came from Surging News Network and WHO situation reports, respectively.Source: Data Sources · p. 5
• Confirmed cases were counted from January 21, 2020 to February 13, 2020; after February 14, 2020, the count was the sum of clinical-confirmed and lab-confirmed cases.Source: Data Sources · p. 5
• For Hubei Province, China, the reported figures on February 14 were 15,384 and 36,602, respectively; confirmed-case counts before February 14, 2020, were adjusted by a formula.Source: Data Sources · p. 5
• The dataset covered total and new confirmed cases for China, 31 mainland provinces/cities, and Hong Kong, Macau, and Taiwan.The data included accumulated and new case counts across these geographic levels.Source: Data Sources · p. 5
• The data were arranged in a matrix whose rows represented China and each province/city and whose columns represented daily new confirmed cases.Source: Data Sources · p. 5
• Each province/city was represented as a time series of daily new cases starting January 11, 2020; the described matrix Z had 34 rows and m columns.Source: Data Sources · p. 5

Modified Auto-encoder for Modeling Time Series

• The MAE forecast cumulative and new confirmed cases using an 8-32-4-1 node architecture across its input, latent and output layers.The final forecast was produced for each of 34 provinces/cities.Source: Modified Auto-encoder for Modeling… · p. 6
• Training used 128 samples, each formed from an 8-day time-series segment beginning at a randomly selected start day.Each segment comprised the selected start day and its seven successive days.Source: Modified Auto-encoder for Modeling… · p. 6
• The MAE used weighted loss terms; backpropagation estimated weights and bias, and after five training repetitions, the average forecast for i = 1,...,34 was taken as the final forecast.Weights were 1 when j was in [1, 12] and 2 when j was in [13, 24].Source: Modified Auto-encoder for Modeling… · p. 6

Forecasting Procedures

• The trained MAE was used to forecast future confirmed cases for each province/city, with its output representing forecasted new cases.The model’s output was defined using parameters estimated during training.Source: Forecasting Procedures · p. 7
• The forecasting input was a 34 × 8 matrix of prior daily case values, normalized using each province/city’s row average.The target was the number of new confirmed cases on the day being forecast.Source: Forecasting Procedures · p. 7
• Recursive multiple-step forecasting repeatedly applied the one-step model, using each preceding step’s prediction as input for the next step.For example, the predicted new cases for the next day were used as an observational input to forecast day 2.Source: Forecasting Procedures · p. 7
• The province/city forecasts of new confirmed cases were summed to predict the total number of new confirmed cases.The text specifies summing the final forecasted number for each province/city.Source: Forecasting Procedures · p. 7

Clustering

• For each province or city, the authors extracted second-layer latent variables, formed a 34 × 4 matrix, and obtained its largest singular value across five trainings.The five largest singular values were included in each province/city feature vector.Source: Clustering · p. 7
• K-means grouped the 34 province/city feature vectors, which combined latent singular values with forecasted outbreak timing and case-count characteristics.The features included the outbreak starting and forecasted end days, the day and number of new confirmed cases at their maximum, the largest forecasted number of new confirmed cases, and forecasted accumulated confirmed cases.Source: Clustering · p. 8

Results

• The forecast for China projected new confirmed cases to peak at 5,236 on February 5, 2020, then decline to zero on April 20, when potential cumulative confirmed cases reached a plateau of 83,401.Source: Results · p. 8
• The study reported that forecasting accuracy was very high; Table 1 showed that average forecasting errors did not strictly increase as the number of forecast steps increased, due to data fluctuations.Source: Results · p. 8
• Plateau entry was forecast to range from January 21 to April 20, with most provinces and cities reaching plateaus around mid-March and Hubei among the last.Xizang, Macao, and Qinghai were forecast to enter the plateau on January 21, January 26, and January 28, respectively; Hubei’s plateau was forecast at 70,019 around April 20.Source: Results · p. 9
• The 34 provinces and cities formed nine clusters with partially geographic patterns, including a distinct Hubei cluster and a cluster of surrounding provinces.Inner Mongolia, Xizang, and Qinghai formed a cluster with the lowest case counts and locations far from Covid-19 sources.Source: Results · p. 9
• Other clusters included Shanghai, Hong Kong, Beijing, Shandong, and Sichuan, while Fujian, Guangxi, Hebei, Jilin, and Tianjin grouped together despite not sharing a geographic area.The authors suggested the latter group might share economic ties with Wuhan, healthcare resources, and interventions.Source: Results · p. 10

Discussion

• Prediction and multiple-step forecasting accuracies were high, and forecasting improved with longer training time.The discussion reports no numerical accuracy values here.Source: Discussion · p. 10
• Across 34 provinces/cities, MAE estimated decreasing growth, epidemic lengths, and plateau timing, and predicted an end in mid-April if data were reliable and no second transmission occurred.The forecast covered the timing of decreases in new confirmed cases and plateaus in cumulative-case curves.Source: Discussion · p. 10
• Cluster analysis grouped 34 provinces/cities into 9 clusters to explore geographic patterns and healthcare resources.Source: Discussion · p. 10
• Although MAE can incorporate intervention information to assess outbreak size and end time, the study did not explore this function because data were lacking.Source: Discussion · p. 10
• With hypothesized initial outbreak sizes and known parameters and architecture, trained MAE can approximate dynamic processes to estimate future outbreak sizes and simulate intervention impacts on epidemic size and severity.Source: Discussion · pp. 10–11
• Complementing transmission-dynamics models, data-driven AI forecasting can track epidemic trajectories, assess severity, predict epidemic lengths, and support government and health-worker planning.Source: Discussion · p. 11

• Figure 1 depicts the architecture of a MAE.Source: Legend · p. 13
• Three other regions in China formed 9 clusters.Source: Legend · p. 13

• Figure 1 depicts the architecture of a MAE.Source: Figure 1 · p. 15
• Figure 2 shows national reported and forecasted cumulative and new confirmed-case curves from January 11, 2020 to April 20, 2020.Source: Figure 2 · p. 16
• Figure 3 shows forecasted cumulative confirmed-case curves for 34 provinces/cities from January 11, 2020 to April 20, 2020.Source: Figure 3 · p. 17
• Figure 4 presents nine clusters formed from features extracted by the MAE and cumulative confirmed-case time series across 31 mainland provinces/cities and three other regions.Source: Figure 4 · p. 18, Supplementary Figure Legend · p. 18
• Table 1 is titled as reporting errors for forecasts of national cumulative confirmed cases in China.Source: Table 1 · p. 19
• Figure S1 compares the national reported and fitted cumulative confirmed-case curves from January 11 to February 27, 2020, with reported cases in red and fitted cases in green.Source: Figure S1 · p. 20, Supplementary Figure Legend · p. 20