Source-linked AI summary
Neural network based country wise risk prediction of COVID-19
Ratnabali Pal, Arif Ahmed Sekh, Samarjit Kar, Dilip K. Prasad
TL;DR
Limited data and uncertain, varying country-level COVID-19 trends complicate risk prediction. The paper uses Bayesian-optimized shallow LSTMs with fuzzy rules for country-specific risk classification, reporting that shallow models perform better than deep networks and that weather data do not improve performance.
Problem
Small datasets and uncertain, country-varying trends challenge AI methods for predicting country-level COVID-19 risk.
Source: §1 · pp. 1–2Method
The paper combines country-specific shallow LSTMs optimized with Bayesian optimization, trend data, and fuzzy rules to classify risk.
Source: §3.2 · p. 5, §3.4 · p. 6, §5 · p. 14Results
The authors report that shallow networks perform better than deep neural networks and that prediction performs similarly without weather data.
Source: §5 · p. 14Takeaways & Limitations
The method may be useful for predicting long-duration epidemic risk at the country level.
Source: §5 · p. 14
Abstract
from arXiv · showhide
The recent worldwide outbreak of the novel coronavirus (COVID-19) has opened up new challenges to the research community. Artificial intelligence (AI) driven methods can be useful to predict the parameters, risks, and effects of such an epidemic. Such predictions can be helpful to control and prevent the spread of such diseases. The main challenges of applying AI is the small volume of data and the uncertain nature. Here, we propose a shallow long short-term memory (LSTM) based neural network to predict the risk category of a country. We have used a Bayesian optimization framework to optimize and automatically design country-specific networks. The results show that the proposed pipeline outperforms state-of-the-art methods for data of 180 countries and can be a useful tool for such risk categorization. We have also experimented with the trend data and weather data combined for the prediction. The outcome shows that the weather does not have a significant role. The tool can be used to predict long-duration outbreak of such an epidemic such that we can take preventive steps earlier
1. Introduction
The introduction frames country-level COVID-19 risk forecasting as useful for planning and prevention, but difficult because data are limited and trends vary across countries. It proposes shallow, country-specific LSTMs optimized with Bayesian optimization, alongside an examination of weather data.
Motivation
- Forecasting can support planning, decision-making, and prevention, while early epidemic predictions can help governments and health-care departments respond in time.Source: §1 · pp. 1–2
Challenges
- Limited training data and uncertain, country-varying trends challenge accurate AI forecasting and make generic tools and deep neural networks unsuitable, according to the authors.The authors identify these challenges as motivation for shallow, country-specific optimized networks.Source: §1 · p. 2
Weather data
- The introduction motivates systematic analysis of worldwide weather parameters because prior work has claimed a role for weather in viral outbreaks.Source: §1 · p. 2
Proposed approach
- The proposed approach combines local trends, a shallow LSTM, a fuzzy-rule system, and Bayesian optimization of country-specific networks.Source: §1 · p. 2
2. Related Works
The related-work section reviews AI applications in epidemic research, COVID-19 studies, and multivariate regression. It positions the paper’s country-optimized shallow RNN and fuzzy-rule inference as a response to limited data, local trends, and questions about weather’s role.
AI based epidemic researches
- Prior epidemic-AI work spans neural networks, machine learning, surveillance, and combined statistical and machine-learning approaches to forecast future growth.Source: §2 · pp. 2–3
Researches on COVID-19
- COVID-19 studies reviewed include patient detection and risk prediction, epidemic forecasting, multimodal AI tools, and models estimating cases or climate-related projections.Source: §2 · p. 3
Multivariate Regression in AI
- The reviewed multivariate-regression and time-series work applies forecasting across business and health, alongside CNN, LSTM, GRU, and deep-neural-network methods.Source: §2 · p. 3
Gap bridged by our method
- The paper addresses long-term country risk prediction with a dynamic shallow RNN optimized per country and fuzzy-rule inference, reporting that prediction does not depend on weather.The authors report similar or better model performance without weather data.Source: §2 · p. 3
3. Proposed Model
The proposed model combines country-specific shallow LSTMs, optimized over each country's trend data, with fuzzy rules that classify risk from predicted rates. Its design spans recurrent sequence modeling, hyperparameter search, and risk categorization.
3.1. Background
- An LSTM uses gates and memory cells to model sequence dependencies across different time scales.The forget, input, and output gates regulate information passed through the cell state and onward to hidden layers.Source: §3.1 · pp. 4–5, Figure 3 · p. 5
3.2. Search Space Definition
- Each country has distinct trends, so the method defines a shared set of input concerns and searches country-specific network hyperparameters.The inputs are cases, deaths, and recoveries; active cases are calculated as κ − (ρ + δ).Source: §3.2 · p. 5, Table 1 · p. 5
3. Proposed Model
- Country-specific shallow LSTMs predict case, death, and recovery counts, which feed fuzzy rules for country risk classification.The framework has four modules: search-space definition, network search, local-trend prediction, and fuzzy rule-based risk assessment.Source: §3 · p. 4, §3.3 · p. 6, §3.4 · p. 6
3.3. Network Search
- Bayesian optimization searches hyperparameters by estimating candidate error from previous outcomes, then uses the selected network to predict country-level counts.Validation RMSE is used to choose the architecture.Source: §3.3 · pp. 5–6
3.4. Fuzzy Rule-Based Risk Categorization
- Predicted deaths, cases, and recoveries are converted into death, new-case, and recovery rates, then Gaussian fuzzy membership functions and rules assign risk.The three categories are high risk, medium risk, and recovering.Source: §3.4 · p. 6, Figure 4 · p. 6, Table 2 · p. 7
4. Results and Discussion
The experiments assess country-specific network optimization, trend and risk prediction, and weather features. Across these analyses, the authors report that shallow networks perform better than deep networks and that prediction performance is similar with and without weather data.
4.1. Dataset
- The dataset spans 22 January to 2 August 2020 and combines country-level confirmed, recovered, and death cases with weather measurements summarized by city-level means and standard deviations.The weather variables include humidity, dew, ozone, perception, maximum and minimum temperature, and UV.Source: §4.1 · p. 7
4.2. Network Optimization
- Bayesian optimization generated country-specific LSTMs with different architectures and settings; the USA model used three hidden layers of 171 nodes, learning rate 0.0002, zero dropout, batch size 6, and ReLU.Across the optimization runs, most networks had a few layers and hidden units, used ReLU, and selected zero dropout; optimization minimized validation RMSE.Source: §4.2 · pp. 7–8, Figure 7 · p. 9
4.3. Training
- Each country-specific network was trained on its own case data, validated on active cases, and then used to predict deaths, recoveries, and total cases.The final 25 days were reserved for testing, and training used early stopping; Figure 8 shows USA training loss and active-case validation predictions.Source: §4.3 · p. 9, Figure 8 · p. 10
4.4. Prediction Accuracy
- Linear, lasso, and ridge regression performed very poorly on the small dataset, deep networks also failed, and Bayesian-optimized shallow GRU performed closer to the proposed method.Table 3 summarizes average RMSE for 25-day predictions of active, recovered, and death cases; Figure 9 shows USA trend predictions.Source: §4.4 · p. 10, Table 3 · p. 10, Figure 9 · p. 11
4.5. Risk Classification Accuracy
- 77.6% average accuracy was achieved across 180 country-specific datasets; MR prediction was relatively less accurate because of incorrect trend-rate prediction.Figure 10 presents the four-class confusion matrix for 25-day-ahead prediction.Source: §4.5 · p. 11, Figure 10 · p. 12
4.6. Implementation and Computational Cost
- The experiments used about 72 computational hours for feature selection, network optimization, training, and evaluation on a CPU and GPU workstation.The implementation used Python, scikit-optimize, TensorFlow, and simple if-then fuzzy rules.Source: §4.6 · p. 12
4.7. The Effect of Weather
- Prediction performance was almost similar with and without weather data; weather parameters were not significantly correlated with active cases across countries.The analysis compared selected features, combined weather features, and features without weather; Figure 12 describes average RMSE for total cases, recoveries, and deaths.Source: §4.7 · p. 13, Figure 12 · p. 14, Figure 11 · p. 14
5. Conclusions
The paper concludes that Bayesian-optimized shallow LSTMs can predict country-specific COVID-19 risk using trend data, with similar prediction performance when weather data are included or omitted. It also identifies long-duration epidemic-risk prediction and additional data modalities as directions for use and future work.
5. Conclusions
- The authors conclude that shallow networks perform better than deep networks when their country-specific optimized LSTMs are considered together.They describe the approach as suitable for small, uncertain datasets.Source: §5 · p. 14
- The model performs similarly with and without weather data, according to the authors’ analysis.Source: §5 · p. 14
- The authors identify long-duration epidemic-risk prediction as a potential use and propose exploring flight, traveler, business, and tourist data in future work.They also note possible use in predicting epidemics’ economic effects.Source: §5 · pp. 14–15