Prediction and analysis of Coronavirus Disease 2019

• The listed affiliation is Stanford University in Stanford, California.Source: §1 · p. 1
• The listed keywords are mathematical model, COVID-19, and epidemic prediction.Source: §1 · p. 1

1 Introduction

• By February 29, 2020, China reported 79,824 COVID-19 cases and 2,870 deaths; Wuhan accounted for 61.5% of cases and 76.5% of deaths.The authors describe infectious-disease prediction models as important for reducing harm and supporting prevention and control.Source: §1 · p. 2

1.1 Traditional infectious disease prediction model

• Traditional prediction uses differential-equation and time-series models, but fixed-population assumptions and curve-fitting dependence limit realism and use with irregular data.The discussion notes that differential-equation models ignore changing populations and individual differences, while time-series approaches rely on available data for curve fitting and parameter estimation.Source: §1.1 · pp. 2–3

1.2 Internet-based infectious disease prediction model

• Internet-based prediction can provide faster, real-time monitoring suited to large populations, but its sensitivity, spatial resolution, and accuracy need improvement.The approaches draw on search engines, social media, and other online information to track disease signals.Source: §1.2 · pp. 3–4
• Internet-based infectious-disease prediction models cannot replace traditional models and are presented only as an extension of them (16).Source: §1.2 · p. 4
• The study will validate Logistic, Bertalanffy, and Gompertz models on 2003 SARS data, then apply them to COVID-19 cases, deaths, and epidemic end times across Wuhan, mainland China, and non-Hubei areas.The models are used to predict development trends and analyze regional epidemic trends.Source: §1.2 · p. 4

1.3 Early Prediction Model of Infectious Diseases Based on Machine Learning

• Neural networks can model high-dimensional nonlinear input-output mappings, while traditional statistical methods may be less effective on complex data relations (17, 18).Machine learning learns useful information from data using algorithmic models.Source: §1.3 · p. 4
• Machine learning can identify species and environmental conditions associated with zoonotic spillover and predict disease reservoirs and carriers (19).Source: §1.3 · p. 4
• Machine learning aims to identify and quantify factors behind zoonotic outbreaks and visualize causal relationships, but early prediction models can be limited by local and global minima (20).Source: §1.3 · p. 4

2 Mathematical model

• Logistic, Bertalanffy, and Gompertz models were selected as relatively simple alternatives to more complex approaches that may overfit.The paper describes the selected models as consistent with epidemiological statistical laws and fits them using least squares.Source: §2 · p. 5
• The least squares method fits a curve by minimizing the sum of squared errors between the fitted function and observed data.The text says this optimization technique is used to obtain unknown data for the models.Source: §2 · p. 5

2.1 Model Selection

• The Logistic model is used to describe epidemic development, while the Bertalanffy and Gompertz models are selected to describe infectious-disease spread and relevant factors.The paper motivates the latter models by comparing infectious-disease development with growth in individuals and populations.Source: §2.1 · pp. 5–6
• For the three models, Qt represents cumulative confirmed cases or deaths, a is the predicted maximum, b and c are fitting coefficients, and t counts days since the first case.The text also identifies t0 as the time when the first case occurred.Source: §2.1 · pp. 5–6

2.2 Model Evaluation

• R² evaluates fitting ability using actual and predicted cumulative confirmed COVID-19 cases, with values closer to 1 indicating more accurate predictions.The equation compares predicted and actual values against the average of actual values.Source: §2.2 · p. 6

3 Fitting and analysis of SARS epidemic

• Because SARS and COVID-19 are both coronaviruses, the authors use SARS data to assess the models before applying them to COVID-19.They also treat the time when cumulative confirmed cases peak as approximately the epidemic end, using this rule to predict COVID-19's end time.Source: §3 · p. 6, §3.1 · p. 6

3.1 Number of Confirmed Cases

• All three models accurately predict cumulative confirmed SARS cases, with Logistic and Gompertz performing better than Bertalanffy.The cumulative-case fitting uses data after April 21, 2003.Source: §3.1 · p. 6, Figure 1 · p. 7

3.2 Death Toll

• Gompertz and Logistic models predict the SARS death toll well.Figure 2 presents the death-toll fits for Gompertz, Logistic, and Bertalanffy models.Source: §3.2 · p. 7, Figure 2 · p. 8

4.1 Number of Confirmed Cases

• All three models were judged to predict later-stage COVID-19 well; Logistic fit Wuhan best, while Gompertz fit outside Wuhan best.The paper also reports Logistic as better than the other two models in fitting all the data in Wuhan.Source: §4.1 · p. 8
• The 13,332-case increase was treated as a data mutation and its impact removed from the fitting, with the sudden increase considered in later fitting.The paper states that this mutation did not originate from the virus’s mechanism.Source: §4.1 · p. 8
• Predicted turning points for new confirmed cases are February 9 in Wuhan, February 6 in mainland China, and February 2 in non-Hubei areas.Figure 4 compares the three models’ predictions for new cases.Source: §4.1 · p. 11, Figure 4 · p. 12

4.2 Death Toll

• Three-model death-toll forecasts span 2,502–5,108 in Wuhan, 107–125 in non-Hubei areas, and 3,150–6,286 in mainland China.The authors suggest the differences may reflect factors affecting death rates, including improvements in treatment, equipment, and measures.Source: §4.2 · p. 14
• For later fits, the authors favor Logistic and estimate deaths at about 2,502 in Wuhan, 107 in non-Hubei areas, and 2,150 in mainland China.They consider later fitting results more important than earlier ones.Source: §4.2 · p. 14

5 Discussion

• Logistic fits were better for cumulative data in Wuhan and for newly confirmed cases in both Wuhan and non-Hubei areas, while Gompertz fit better in non-Hubei cumulative data.All three models were reported to predict newly confirmed cases well in early and late epidemic stages.Source: §5 · p. 15
• Improved medical resources, personnel capabilities, and treatment were considered likely to reduce mortality rapidly, with potentially smaller effects on cumulative confirmed cases.The discussion attributes the smaller expected case-count effect to the large number of confirmed cases.Source: §5 · pp. 15–16
• Predicted cumulative cases were 49852-57447 in Wuhan, 12972-13405 in Non-Hubei areas, and 80261-85140 in China mainland.Predicted total deaths were 2502-5108 in Wuhan, 107-125 in non-Hubei areas, and 3150-6286 in China mainland.Source: §5 · p. 16
• The epidemic was estimated to end in Wuhan at the end of April 2020 and in Non-Hubei areas at the end of March 2020.The China mainland endpoint was estimated between the end of March 2020 at the earliest and around April 2020 at the latest.Source: §5 · p. 16
• The forecasts assume reliable, stable prevention and control, no large-scale importation of cases into China, and no new serious epidemic caused by the virus.The paper also notes that predictions from different models vary substantially.Source: §5 · p. 16

6 Conclusion

• COVID-19 was estimated to end in late-April 2020 in Wuhan and before late-March 2020 in other areas.The discussion elsewhere specifies the non-Hubei estimate as the end of March 2020.Source: §6 · p. 18
• The cumulative number of confirmed COVID-19 cases is 49,852–57,447 in Wuhan and 12,972–13,405 in non-Hubei areas, respectively.Source: §6 · p. 18
• According to the current trend, the three models predicted a cumulative death toll of 2502-5108.Source: §6 · p. 18
• The epidemic inflection points for Wuhan, non-Hubei areas, and China mainland were estimated to fall around mid-February 2020.Source: §6 · p. 18
• Model predictions varied by parameter and region; Logistic fitting may be best overall, while Gompertz may fit better than Bertalanffy.Source: §6 · p. 18