Source-linked AI summary
Time series forecasting of new cases and new deaths rate for COVID-19 using deep learning methods
Nooshin Ayoobi, Danial Sharifrazi, Roohallah Alizadehsani, Afshin Shoeibi, Juan M. Gorriz, Hossein Moosaei, Abbas Khosravi, Saeid Nahavandi, Abdoulmohammad Gholamzadeh Chofreh, Feybi Ariani Goni, Jiri Jaromir Klemes, Amir Mosavi
TL;DR
Forecasting COVID-19 cases and deaths may help organizations plan for future costs and facilities, while real-time forecasting motivates testing less frequent prediction intervals. The paper evaluates six deep-learning models on WHO data for Australia and Iran, and reports that bidirectional models usually outperform non-bidirectional counterparts. Its stated limitation is that changing time-series characteristics require retraining on newly observed data to maintain accuracy.
Problem
The study asks whether COVID-19 new-case and death-rate forecasts can be made one, three, and seven days ahead, including at less frequent intervals that may reduce computational complexity.
Source: §1 · pp. 2–3Method
The study compares LSTM, Conv-LSTM, and GRU with their bidirectional extensions using WHO data to forecast cases and deaths in Australia and Iran.
Source: §1 · p. 3, §7 · p. 21Results
Across most experiments, bidirectional methods achieved better results than the other methods.
Source: §7 · p. 21, §8 · p. 22Takeaways & Limitations
Forecasts based on observed data may provide health-crisis management centers with information for near-term preparation.
Source: §7 · p. 21Changing time-series characteristics require retraining on newly observed data to keep the models accurate.
Source: §7 · p. 21
Abstract
from arXiv · showhide
The first known case of Coronavirus disease 2019 (COVID-19) was identified in December 2019. It has spread worldwide, leading to an ongoing pandemic, imposed restrictions and costs to many countries. Predicting the number of new cases and deaths during this period can be a useful step in predicting the costs and facilities required in the future. The purpose of this study is to predict new cases and deaths rate one, three and seven-day ahead during the next 100 days. The motivation for predicting every n days (instead of just every day) is the investigation of the possibility of computational cost reduction and still achieving reasonable performance. Such a scenario may be encountered in real-time forecasting of time series. Six different deep learning methods are examined on the data adopted from the WHO website. Three methods are LSTM, Convolutional LSTM, and GRU. The bidirectional extension is then considered for each method to forecast the rate of new cases and new deaths in Australia and Iran countries. This study is novel as it carries out a comprehensive evaluation of the aforementioned three deep learning methods and their bidirectional extensions to perform prediction on COVID-19 new cases and new death rate time series. To the best of our knowledge, this is the first time that Bi-GRU and Bi-Conv-LSTM models are used for prediction on COVID-19 new cases and new deaths time series. The evaluation of the methods is presented in the form of graphs and Friedman statistical test. The results show that the bidirectional models have lower errors than other models. A several error evaluation metrics are presented to compare all models, and finally, the superiority of bidirectional methods is determined. This research could be useful for organisations working against COVID-19 and determining their long-term plans.
1. Introduction
The study forecasts COVID-19 new cases and deaths using deep-learning models, comparing standard and bidirectional architectures. It also tests whether forecasting every three or seven days can reduce computational complexity while retaining reasonable performance.
1. Introduction
- Forecasts target new cases and deaths, motivated by the value of estimating COVID-19 prevalence for pandemic control and planning.The introduction notes that infections may remain asymptomatic for several days.Source: §1 · p. 2
- The study compares LSTM, Conv-LSTM, and GRU with their bidirectional extensions for daily, three-day, and seven-day forecasting.The authors frame the longer intervals as a way to investigate computational savings while maintaining reasonable performance.Source: §1 · p. 3
2. Related works
Related work spans statistical, machine-learning, and mathematical models of COVID-19 spread and mortality. The paper distinguishes its data-driven forecasting focus from studies modeling disease mechanisms, interventions, or transmission dynamics.
2. Related works
- Pinter et al. (19) used MLP-ICA on Hungary data for a nine-day horizon, while Burke et al. (6) compared four ML methods and reported their method performed best among similar methods.Other forecasting examples include Polynomial Regression, Bayesian Edge, and LSTM (22), and a deep-learning comparison in which VAE outperformed other models (23).Source: §2 · pp. 3–4
- Other cited studies include SEIR-type modeling of protective measures (24), solvability and reproduction-number analysis (25), and age-compartment stochastic modeling of safety protocols (26).These works report roles for health strategies and safety protocols in containment.Source: §2 · p. 4
- Danane et al. (27) modeled COVID-19 with a stochastic SIQR system, whereas this study uses deep learning for prediction; their simulations used Morocco data, while this study uses Iran and Australia data.Danane et al. estimated infection and recovery rates from Morocco cases (29).Source: §2 · pp. 4–5
- The paper contrasts its neural-network forecasting with fractional-calculus studies (31-33), which it describes as mathematically grounded but harder for general readers to grasp and implement.Those studies examined spread dynamics, including reported and unreported cases.Source: §2 · p. 5
- Unlike Boudaoui et al. (34), whose simulations use a fixed mathematical model, this study trains and predicts from observed population data to adapt to changing dynamics.Boudaoui et al. reported lower infective peaks with reduced contact rates and increased isolation or hospitalization.Source: §2 · p. 5
- Other cited mathematical studies examine interventions, age, and transmission: examples include NPIs (36), quarantine without lockdown (37), reservoir-to-human transfer (38), and age-related mortality (21).The paper positions its work as forecasting mortality and spread rather than devising prevention strategies or modeling these mechanisms.Source: §2 · pp. 3–6
3. Deep Learning and its variations
The paper introduces recurrent neural-network architectures used for time-series prediction, including LSTM, its bidirectional and convolutional variants, and GRU. Their described distinctions concern sequence direction, data dimensionality, and gating.
3. Deep Learning and its variations
- Deep learning is presented as a machine-learning approach based on artificial neural networks and is used here to predict COVID-19 time series.The paper introduces LSTM, Bi-LSTM, Conv-LSTM, and GRU as methods for time-series prediction.Source: §3 · p. 6
- LSTM uses a repeating cell to retain sequence information through input, output, and forget gates.The forget gate controls what cell-state information is discarded, the input gate controls new information, and the output gate selects what reaches the cell output.Source: §3 · pp. 6–7
- Bi-LSTM trains two LSTMs on the input sequence, one in its original order and one in reverse.The passage states that Bi-LSTM can learn faster than traditional LSTM.Source: §3 · p. 7
- Conv-LSTM replaces LSTM gate layers with convolutional layers to handle multidimensional data and encode spatiotemporal information in its memory cell.The passage also describes convolution operators being supplanted by an LSTM memory cell to identify information to remember or forget.Source: §3 · p. 7
- GRU is an RNN variant with two gates: an update gate for passing past information forward and a reset gate for deciding how much past information to discard.The passage contrasts its two gates with LSTM's three.Source: §3 · p. 7
4. Dataset Description
The study uses WHO COVID-19 reports for Australia and Iran to forecast new-case and new-death rates. It selects cumulative and new case and death series from the reported dataset.
4. Dataset Description
- The dataset comes from WHO reports and includes country and date fields alongside cumulative case and death counts.The study uses the “New Cases”, “Cumulative Cases”, “New Deaths”, and “Cumulative Deaths” series for forecasting in Australia and Iran.Source: §4 · p. 7, Table 1 · p. 7
5. Proposed Method
The study trains six neural-network models to forecast new-case and new-death rates in Australia and Iran at one-, three-, and seven-day horizons. Training uses WHO time series, with separate model runs for each horizon.
5. Proposed Method
- The six predictors are LSTM, Conv-LSTM, GRU, and their bidirectional versions, each with three hidden layers of 50 neurons.All layers use ReLU activation; training uses MSE loss and the Adam optimiser for 200 epochs, with a 0.001 learning rate.Source: §5 · p. 8
- The input time series are from the WHO database for Australia and Iran, with approximately 70% of data used for training and the remainder for testing.About 20% of the training data were reserved for validation.Source: §5 · p. 8
- As the forecasting horizon extends from one to three to seven days, the error rate increases.The models were trained separately to predict each horizon for both new cases and new deaths.Source: §5 · p. 8
- Training was repeated for one-, three-, and seven-day predictions, for both new-case and new-death time series.Figure 1 illustrates the proposed method’s high-level steps.Source: §5 · p. 8, Figure 1 · p. 9
6. Experimental Results and Analysis
Across 12 datasets, Bi-GRU has the best average rank among six algorithms. Forecasting results vary by country, outcome, and horizon; LSTM-family models generally compare favorably with GRU, while longer horizons often bring larger errors.
6. Experimental Results and Analysis
- The study reports the first use of Bi-Conv-LSTM and Bi-GRU to predict COVID-19 new-case and death time series.The experimental section also reports results for LSTM, Conv-LSTM, GRU, and their bidirectional counterparts.Source: §6 · pp. 9–10
- The comparison evaluates six methods on one-, three-, and seven-day forecasts over 100 days in Australia and Iran using MSLE, MAPE, RMSLE, and explained variance.The reported 100-day errors compare predicted values with actual values.Source: §6 · p. 10, §6.1 · p. 10
6.1. Forecasting performance
- In Australia cases, Bi-GRU and LSTM perform best at one day, Conv-LSTM at three days, and Bi-Conv-LSTM at seven days; for deaths, LSTM significantly outperforms GRU at seven days.These comparisons are based on MAPE; the methods have approximately similar explained variance for cases.Source: §6.1 · p. 10, Figure 2 · p. 11, Figure 3 · p. 11
- For Iran cases and deaths, LSTM and its variations outperform GRU most of the time, especially at three- and seven-day horizons.The comparison is based on MAPE in Figures 4 and 5.Source: §6.1 · p. 11, Figure 4 · p. 12, Figure 5 · p. 12
- Longer prediction horizons often produce larger errors, although Bi-Conv-LSTM deviates badly from actual data in the one-day Iran-cases forecast.The paper illustrates horizon-related error growth with an Iran-cases prediction whose absolute error rises from 9,000 at one day to 59,000 at seven days.Source: Figure 5 · p. 12, §6.1 · p. 16
- The absolute-error histograms count predictions within error intervals; the authors note that large absolute errors can be modest relative to cumulative reported totals.For Iran one-day cases, GRU's absolute error of 9,000 is 2.6% of the true value of 345,000.Source: Figure 9 · p. 15
6.2. Statistical Analysis
- Bi-GRU has the best average rank, followed by LSTM, GRU, Bi-Conv-LSTM, Bi-LSTM, and Conv-LSTM across 12 datasets.The reported F-distribution degrees of freedom are (5,55), with a 0.05 critical value of 2.38.Source: Table 4 · p. 19, §6.2 · p. 19
7. Discussion
The discussion reports a broad comparison of six forecasting models, including models newly applied to COVID-19 time series, and examines whether less frequent forecasts can reduce computational complexity. Bidirectional methods performed better in most experiments, while the preferred model varied by country, forecast horizon, and outcome.
7. Discussion
- Predicting every three days approximately doubled RMSLE, while every-seven-day prediction incurred more than four times the daily error.The authors frame less frequent prediction as a possible computational-complexity reduction, with acceptable error depending on application requirements.Source: §7 · pp. 20–21
- The study presents a comprehensive evaluation of LSTM, Conv-LSTM, GRU, and their bidirectional extensions, with statistical comparison using the Friedman test.The authors report that Bi-GRU and Bi-Conv-LSTM had not previously been used to predict COVID-19 new cases and deaths time series.Source: §7 · pp. 20–21
- In most experiments, bidirectional methods achieved better results than the other methods.The comparison covered LSTM, GRU, Conv-LSTM, and their bidirectional counterparts on new cases and deaths in Australia and Iran.Source: §7 · p. 21
- For next-day deaths, Bi-GRU performed best in both countries; at three and seven days, the leading model differed by country.For three-day deaths, GRU led in Australia and Bi-Conv-LSTM in Iran; for seven-day deaths, LSTM led in Australia and Bi-Conv-LSTM in Iran.Source: §7 · p. 21
- For new cases, Australia’s leaders were LSTM and Bi-GRU at one day, Conv-LSTM at three days, and Bi-Conv-LSTM at seven days.In Iran, Bi-GRU led at one and three days, while Bi-Conv-LSTM led at seven days.Source: §7 · p. 21
- The authors identify changing time-series characteristics as a limitation, requiring retraining on newly observed data to keep models accurate.This retraining incurs additional cost.Source: §7 · p. 21
8. Conclusion
The study compares six models for forecasting new cases and deaths over the next 100 days at daily, three-day, and seven-day intervals. Bidirectional models generally outperform their non-bidirectional counterparts, and future work may explore Gaussian Process forecasting.
8. Conclusion
- Most of the time, bidirectional models outperformed their non-bidirectional counterparts in forecasts of new cases and deaths over the next 100 days.The comparison used daily, three-day, and seven-day predictions.Source: §8 · p. 22
- Future work may combine machine-learning and deep-learning methods, including Gaussian Processes that can provide prediction uncertainty.The authors suggest uncertainty could help determine an appropriate prediction horizon.Source: §8 · p. 22