1 Introduction

Due to the COVID-19 epidemic, the year 2020 will be remembered as a devastating year for humanity. The foremost case of COVID-19 had been seen in China's Wuhan city and detected as undefined etiology. This first case came in December 2019, and within a month, it took the whole world into its trap (Cohen 2020; Huang et al. 2020). In the second week of March, the WHO (World Health Organization) had announced COVID-19 as a pandemic disease (Coronavirus (COVID-19). Available online 2020; World Health Organization 2005; Sohrabi et al. 2020). Initially, it was acknowledged as the Wuhan virus, but later on, it was named COVID-19 or 2019-nCov or Novel coronavirus (“WHO 2020; World Health Organization 2020). As per the current stats by WHO, 10,764,786 confirm cases, 518,120 deaths, 4,288,454 active cases, and 5,958,212 recovered cases, have been come up to July 1, 2020 (Coronavirus disease 2020a). The epidemic situation causes severe social and economic impacts on the country. All the nations are taking preventive action, such as travel restraints, event deferrals, quarantines, testing, social distancing, and soft and tough lockdowns, to save their people's lives (Acter et al. 2020). Suppose we talk of India's perspective, which is a developing country and the second most populated country after China. So, the chance of COVID-19 is more as compared to other countries. India's government has taken various initiatives (i.e., lockdown 1–4, and unlock 1–2) to deal with the COVID-19 epidemic. As per the WHO's current COVID-19 stats for India, 605,220 confirm cases, 17,848 deaths, 227,476 active cases, and 359,896 recovered cases, have been come up to July 1, 2020 (Coronavirus disease 2020a).

If we talk about medical approaches, various clinical trials are underway across the world. Some of them are the fourth (final) phase of their trial, but to date, no antiviral treatments and vaccines are obtainable for COVID-19 (Ketu and Mishra 2020a). Initially, the new vaccine for any novel disease was taking approximately 5–10 years to be on the market. But due to the latest technologies, the fast traction of the vaccine is possible and can usually take around 18–24 months, to come into the market (Grenfell and Drew 2020). In the current scenario, any successful trials may take a minimum of 6–8 months to reach into the production line (if we follow parallel production) and may take a further 6–12 months to be in the market across the world. So, we have to follow preventive measures until the vaccine has come into the market.

In the last couple of years, substantial progress has been seen in the deep learning method in various application areas. If we talk about the deep learning techniques and methods to deal with real-life problems such as time series forecasting that have been successfully applied in various application areas (Ai et al. 2019; Li et al. 2019; Zheng et al. 2019; Zou and Xia 2019). These models are capable enough to deal with the chaotic and noisy nature and lead to more reliable and accurate forecasts. The LSTM (Long Short-term Memory) and CNN (Convolutional Neural Networks) networks are the most prevalent, well-established, proficient, and broadly used deep learning models (Fawaz et al. 2019). The primary aim of using these models in the problem solving of time series paradigms is that the LSTM models are capable of obtaining the sequence outline more efficiently, whereas the CNN models are capable of filtering the input noise and capturing valuable features. The standard CNNs have been developed to deal with spatial autocorrelation data, but they are less capable of handling the long and complex temporal dependencies (Bengio et al. 2013). On the other hand, the LSTM is capable of handling the temporal dependencies and uses the features of the training dataset. Therefore, such a time series model who utilizes these two deep learning models' benefits can improve the overall prediction results.

Figure 1 illustrates the current situation of COVID-19 epidemic in India with the bar chart. In Fig. 1, the y-axis represents the number of infected peoples, and the x-axis represents the aggregate number of recovered cases, the aggregate number of infected cases, the aggregate number of deaths, and the aggregate number of active cases (Coronavirus disease 2020a).

Fig. 1
figure 1

Current condition of COVID-19 in India

Full size image

In this epidemic situation, any help from the algorithmic perspective or a clinical standpoint is a precious but novel task. If we talk about the algorithmic perspective, forecasting can play a vibrant role in dealing with such critical situations and give an idea about the exposer so that the government can make plans accordingly and diminish the impact of the infections (Ketu and Mishra 2020b). The primary aim of this article is to develop such a model that can efficiently and correctly predict the COVID-19 epidemic. For this purpose, we have suggested a deep learning-based hybrid forecasting model. The proposed model uses convolutional layers, to extract meaningful information and learn from a given time series dataset. It is also enriched with the LSTM layer's capability, which means it can identify long-term and short-term dependencies. The recommended model has been compared with the other time series forecasting models to determine its rightness and appropriateness. From the experimental analysis, it is also clear that the use of extra convolutional layers with the LSTM layer may increase the performance of the forecasting model. Apart from this, the deep insides on the current status of medical resource availability across India have been discussed.

The primary purpose of this study is:

  • To establish such a model that can forecast the COVID-19 epidemic across India with greater exactness.

  • To give a deep-leaning-based solution for economic and social prosperity.

  • To find out the current status of medical resource availability across states of India.

This article is systematized as follows: In Sect. 2, a brief discussion about the latest literature based on algorithmic approaches used in COVID-19 forecasting is presented. In Sect. 3, we have briefly described the materials and methods, such as dataset description, proposed methodology, and statistical parameters. The statistical analysis-based prediction results are presented in Sect. 4. The detailed discussion about the forecasting results, along with the present status of medical resource availability across India, are displayed in Sect. 5. In Sect. 6, the summarization of our research findings and future directions is discussed.

2 Related work

In a couple of decades, substantial growth has been seen in the arena of information technology (IT). Resultantly, the application-centric algorithmic approaches are being developed to tackle the current trends. If we talk about the healthcare domain, various algorithmic solutions have been seen in disease detection and prevention. In the present scenario, we are facing a pandemic situation due to COVID-19. So, there is a need for such algorithmic approaches that can minimize the impact of COVID-19 across the globe.

The PIBA (Patient Information-Based Algorithm) was introduced by Wang et al. (Wang et al. 2020) to estimate deaths caused by COVID-19 in china. The authors had predicted the overall death rate outside Wuhan and Hubei, which was lies between 0.75 and 3%. The death rate of Wuhan and Hubei was predicted to 13%. They had also quoted that the death rate could vary in different temperatures and climates. Gupta et al. (Gupta et al. 2020) had established a relationship between the COVID-19 cases and temperature. They had forecasted the results for India based on the outbreak analysis of the USA. Based on the analysis result for India, they had quoted that, in summers, the massive reduction would be seen in the number of new cases. ARIMA and Sutte ARIMA model was used by Ahmar and Val (Ahmar and Val 2020) for forecasting the COVID-19 epidemic, and the stock market. Spain's COVID-19 and stock market dataset had been used for this analysis.

Similarly, the ARIMA model-based COVID -19 prediction of Italy, France, and Spain had been conducted by Ceylan (Ceylan 2020). The forecasting of the COVID-19 outbreak for Italy, China, and France had been done by Fanelli and Piazzain (Fanelli and Piazza 2020). They had forecasted the ventilation units and the number of COVID-19 cases by dividing the whole population into four parts, such as infected, susceptible, dead, and recovered. The LSTM model-based estimation of the COVID-19 end date for Canada had been performed by Reddy and Zhang (Chimmula and Zhang 2020) with an accurateness of 93.4% (short-term) and 92.3% (long-term). Sujatha and Chatterjee (Sujatha and Chatterjee 2020) used multilayer perceptron, vector autoregression, and linear regression models on the Kaggle dataset to forecast the COVID-19 epidemic in India. Elmousalami and Hassanien (Elmousalami and Hassanien 2020) had explored the correction among the day level gauging models with the help of numerical detailing and time series model. A blockchain-based framework was introduced by Torky and Hassanien (Torky and Hassanien 2020) to distinguish and confirm the ambiguous contaminated structures of COVID-19 infection. The GSA (Gravitational Search Algorithm) optimization-based hybrid deep learning model for the identification of COVID-19 had been proposed by Ezzat and Ella (Ezzat and Ella 2020).

The LSTM (Long short-term Memory Network) is one of the powerful models used in the time series prediction model (Hochreiter and Schmidhuber 1997; Gers et al. 1999). The LSTM model was firstly used by Li and Cao (Li and Cao 2018) for the forecasting of tourist drift. They have also quoted that the forecasting accuracy of the LSTM model could be improved as compared to BPNN and ARIMA, particularly in the long term. The Jaya-LSTM-based hybrid prediction model was proposed by Khalid et al. to predict the electricity demand and price (Khalid et al. 2020). The Jaya algorithm was used to optimize the hyper-parameters. Xue et al. carried out LSTM deep neural network-based analysis on the short-term forecasting of the financial market in the year 2020 (Yan et al. 2021). The experimental result establishes better forecasting accuracy as compared to other models (BP neural network, RNN, and LSTM) with the effective forecasting of the time series dataset.

CNN is one of the deep learning models, which is widely used for image recognition purposes (Chua and Roska 1993; Arena et al. 1998). However, in the year 2019, the experimental evaluation carried out by Cao and Wang shows that the CNN (Convolutional Neural Network) is also capable of dealing with the time series paradigms and also suitable for the prediction task (Cao and Wang 2019). Although the prediction accuracy of CNN alone is relatively low, it achieves better propagation accuracy if it is used in hybrid paradigms. In the year 2020, CNN-, MLP-, and LSTM-based forecasting of the stock price on four communal sector companies of the USA was performed by Kamalov (Alibašić et al. 2019). This experimental evaluation established better forecasting results than other state-of-the-art studies.

The CNN-LSTM hybrid model for travel time prediction was proposed by Wei et al. (Wei et al. 2018). Another Tensor-CNN-LSTM (TCL)-based framework for the prediction of the travel time was proposed by Shen et al. (Shen et al. 2019). The CNN-LSTM-based hybrid model for the energy consumption of the house was proposed by Kim and Cho (Kim and Cho 2019). The CNN-LSTM model for predicting the intensity of typhoons was proposed by Chen et al. (Chen et al. 2019). Another CNN-LSTM-based model for the prediction of particulate matter (PM2.5) was proposed by Huang and Kuo (Huang and Kuo 2018). It is clear from these state-of-the-art studies that the CNN-LSTM hybrid model can be applied to the time series prediction with excellent performance.

From this literature, it is clear that a lot of research in the COVID-19 data analytics and prediction has been performed, but there is still a lot of scope in evolving and testing the deep learning-based forecasting models. Along these lines, with the assistance of a productive and viable model, we can lessen the exposer of the COVID-19 and make the appropriate plan accordingly. It will also be beneficial for both social and economic types of factors.

3 Materials and methods

In this segment, we will confer the materials and methods that are being utilized in the progression of outcome findings. This segment consists of three subsections. The first subsection deliberates about the dataset. In the second subsection, we have discussed the mathematical modeling and architecture of the proposed forecasting model. In the last subsection, the statistical parameters and their formulation have been discussed.

3.1 Data

In this analysis, the data from the Arogya Setu App and the MoHFW (Ministry of Health and Family Welfare), GoI (Government of India), have been extracted on a day-to-day basis. The records used for this study are from the period 30/01/2020 to 10/6/2020. The data obtained from these data resources include the date, state/union territory name, number of active cases, number of discharged/cured/ migrated, number of deaths, and the total number of confirmed cases (Ministry of Health and Family Welfare Government of India. 2020; Aarogya Setu App 2020; Coronavirus disease 2020b). In the course of the result finding, all twenty-nine affected states have been taken into consideration. As we know, the situation due to the COVID-19 epidemic in India is getting worst day by day. The number of new cases is also breaking the record of the previous day. Now the virus is in its community-level spread, which means the upcoming weeks will be very crucial and tough.

For an improved understanding and seriousness of the COVID-19 outbreak across India, we have plotted the heatmap based on the confirmed COVID-19 cases (form 30/01/2020 to 10/6/2020) and presented in Fig. 2. The geographical location of the states has been used to visualize the epidemic situation due to COVID-19.

Fig. 2
figure 2

COVID-19 outbreak in India

Full size image

3.2 Proposed methodology

The contribution of this investigation is to build a deep learning-based hybrid prediction model for the forecasting of the COVID-19 epidemic across India. The fundamental purpose of convolutional layers is to extract meaningful evidence and learn from a given time series dataset, whereas the primary aim of LSTM networks is to identify long-term and short-term dependencies. This examination targets to integrate the advantages of these deep learning methods into our proposed model.

Thus, our proposed CNN-LSTM hybrid deep learning model is composed of two essential parts. The first part is composed of pooling and consensual layers in which complex scientific operations have been accomplished to build the input data's characteristics. The second component is composed of LSTM and dense layers to exploit the generated features.

The pooling and convolutional layers (Rawat and Wang 2017) are preprocessing layers that have been used for filtering and extracting fruitful information from the data. This refined data will be further used as an input in the network layer. We can say that convolutional layers apply the convolution operation among convolution kernels and input data for creating new features. The convolution kernel is nothing but a tiny window (compared to the input window) that carries the values of coefficients in the matrix form. This window applies the convolution operation to each subregion (patch) of the input matrix to determine the specified window input matrix. The outcome of these convolution processes is in the matrix form, which contains the attributes' value. This value of the attributes is derived by the dimension of the specified filter and coefficient values. Different convolution kernels can enhance the performance of the model because the convolved features that have been generated by the convolution operation are more fruitful than the original features.

The pooling layer is one of the subsampling techniques which has been used to extract the specified values based on the convolved features. The pooling layer's processing is similar to the convolutional layer, where the tiny sliding window is used. Convolved feature-based patch values are served as an input to the pooling layer, and the new values are received at the output layer. Thus, we can say that max-pooling is used to calculate the maximum costs of each patch, whereas average pooling is used to calculate each patch's average values. The pooling layer produces resultant, new matrices, or we can say that these new matrices are nothing but an abstract version of convolved features provided by a convolutional layer. The use of pooling operation can enhance the system's performance because any minor change in the input will not reflect in the output, or we can say small variations in input values will not modify the output values.

Our proposed CNN–LSTM hybrid deep learning model contains the two convolutional layers having the size of 32 filters and 64 filters, respectively, which is trailed by a pooling layer, an underlying LSTM layer (traditional LSTM Model), and an output layer. In Fig. 3, the architype of the proposed CNN–LSTM hybrid deep learning model has been depicted.

Fig. 3
figure 3

Proposed CNN-LSTM hybrid deep learning model

Full size image

The mathematical formulation of the CNN–LSTM hybrid deep learning model has been defined in the Eqs. 1, 2, 3, 4, 5, 6:

$$ \left( {{\mathbf{Input}} \, {\mathbf{Gate}}} \right)\;I_{t} = \sigma \left( {W_{I} x_{t} + W_{I} h_{t - 1} + b_{I} } \right) $$
(1)
$$ \left( {{\mathbf{Forget}} {\mathbf{Gate}}} \right)F_{t} = \sigma \left( {W_{F} x_{t} + W_{F} h_{t - 1} + b_{F} } \right) $$
(2)
$$ \left( {{\mathbf{Output}} {\mathbf{Gate}}} \right)O_{t} = \sigma \left( {W_{O} x_{t} + W_{O} h_{t - 1} + b_{O} } \right) $$
(3)

where \(\sigma\) = sigmoid function, \(b\) = voltage vectors, and \(W\) = weight matrices.

$$ \left( {{\mathbf{New}} {\mathbf{Memory}} {\mathbf{Cell}}} \right)c_{t} = W_{t} c_{t - 1} + I_{t} \tilde{c}_{t} $$
(4)
$$ \left( {{\mathbf{Final}} {\mathbf{Memory}} {\mathbf{Cell}}} \right)\tilde{c}_{t} = \tanh \left( {W_{C} x_{t} + W_{C} h_{t - 1} + b_{C} } \right) $$
(5)
$$ \left( {{\mathbf{Final}} {\mathbf{Output}}} \right)h_{t} = O_{t} \tanh \left( {c_{t} } \right) $$
(6)

3.2.1 Hyper-parameter for proposed CNN-LSTM model

The hyper-parameter that has been used by the proposed CNN-LSTM model for the forecasting of the Covid-19 outbreak is presented in Table 1.

Table 1 Hyper-parameter used in CNN-LSTM model
Full size table

3.3 Statistical analysis

For the evaluation of the prediction results, three performance evaluators such as mean absolute percentage error (MAPE), coefficient of determination (R2 Score), and root mean-squared error (RMSE) have been utilized. These performance evaluators have been used to measure the accuracy, suitability, and performance of prediction models. The mathematical formulation, of these performance evaluators, is presented in Eqs. 7, 8, and 9, respectively (Nagelkerke 1991).

$$ {\text{MAPE}} = \frac{1}{N}\mathop \sum \limits_{i = 1}^{N} \left| {{ }\frac{{Y_{i} - X_{i} }}{{Y_{i} }}} \right| \times 100{\text{\% }} $$
(7)

where \(N\) = number of predicted samples, \(Y_{i}\) = actual values, and \(X\) i = predicted values.

$$ R^{2} = 1 - \frac{{\mathop \sum \nolimits_{i = 1}^{N} \left( {\hat{x}_{i} - x_{i} } \right)^{2} }}{{\mathop \sum \nolimits_{i = 1}^{N} \left( {\hat{x}_{i} - y_{i} } \right)^{2} }} $$
(8)

where \(\left( {\hat{x}_{i} - x_{i} } \right)^{2}\) = squares of residuals, \(\left( {\hat{x}_{i} - y_{i} } \right)^{2}\) = squares of the total (sum), \(N\) = number of errors, \(\hat{x}_{i}\), \(x_{i}\) = observed values, and \(y_{i}\) = forecasted values.

$$ {\text{RMSE}} = \sqrt {\frac{1}{N}\mathop \sum \limits_{i = 1}^{N} \left( {\hat{x}_{i} - x_{i} } \right)^{2} } $$
(9)

where \(\left( {\hat{x}_{i} - x_{i} } \right)^{2}\) = squares of errors, \( N\) = number of errors, \(\hat{x}_{i}\) = observed values, and \(x_{i}\) = forecasted values.

4 Result

It is always a challenging but novel task to find out such a model, which can accurately predict the pandemic situation in time. So, there is a need for such forecasting models that can forecast the COVID-19 epidemic across India more accurately. The primary purpose of this investigation is to establish such a model that can forecast the COVID-19 epidemic across India, with greater exactness. Correct prediction of the epidemic is one of the essential needs because, without it, we can neither detect the seriousness of the pandemic nor make effective preventive policies. Thus, in this pandemic situation, the correct prediction will help determine the exposer of the COVID-19 outbreak. It will also help the government to formulate efficient and effective preventive policies. It can also helpful for diminishing the overall risk of the COVID-19 explosion.

All the experimental evaluations have been performed on India’s COVID-19 dataset with the help of two well-established time series prediction models, such as LSTM, and ARIMA, along with our proposed hybrid model. These models are shown in Fig. 4.

Fig. 4
figure 4

Forecasting models assessment a quick lookup

Full size image

The state-wise population distribution of India is shown in Fig. 5. The name of the states and population density has represented on the y-axis and x-axis, respectively (Census 2011).

Fig. 5
figure 5

State-wise population of India

Full size image

Experimental assessment on the COVID-19 epidemic (total number of confirmed cases) across India over time series forecasting models is presented in Table 2. The experimental assessment has been accomplished on the twenty-nine affected states. The fundamental objective of this study is to find out the exactness and aptness of our proposed CNN-LSTM hybrid deep learning model. The experimental evaluation has been executed with the help of two well-established time series prediction models, such as LSTM and ARIMA, along with our proposed hybrid deep learning model. For evaluating the prediction results, three performance evaluators, such as MAPE (ought to be low), R2 Score (ought to be high), and RMSE (ought to be low), have been utilized. It is clear from the forecasting results of the COVID-19 epidemic, across the various states of India (29 states), that our proposed CNN-LSTM hybrid deep learning model performed exceptionally well throughout the experiment compared to the other well-grounded time series forecasting models. Our proposed model achieves minimal MAPE, highest R2 Score, and minimal RMSE under several selection settings (state-wise).

Table 2 Performance result of the forecasting algorithms (Confirmed Cases)
Full size table

5 Discussion

To find the results, we have taken data from the Arogya Setu App and MoHFW (Ministry of Health and Family Welfare), GoI (Government of India), on a day-to-day basis. Data used for this study are from the period 30/01/2020 to 10/6/2020. The data extracted from these data resources include the date, state/union territory name, number of active cases, number of discharged/cured/migrated, number of deaths, and the aggregate number of confirmed cases. In the process of result finding, all twenty-nine affected states have been taken into consideration.

The simulation code is implemented using Python 3.4 on a windows 10-based Dell workstation (64-bit Intel Xeon processor with 32 GB RAM and 3.60 GHz speed). The deployment of the deep learning algorithms has been done using the Keras (Géron 2019; Manaswi 2018) library at the front end and Theano (Brownlee 2016; Bahrampour et al. 2016) at the backend. The CNN-LSTM and LSTM models have been trained using adaptive moment estimation (ADAM) with a batch size of 64 and a mean-squared loss function for 100 epochs. The ADAM algorithm is used to take care of the learning steps in the training process. It certifies that the learning phase (in the training process) is scale-invariant compared to the parameter-gradient. In addition, we implement the same padding to ensure that no features are terminated during the conversion operations.

For the experimental analysis, three forecasting models (two traditional models with one proposed hybrid model), i.e., LSTM, ARIMA, and a proposed CNN-LSTM hybrid deep learning model, have been taken. The 60–40 evaluation criteria have been used, meaning that 60% of the dataset has participated in the training process, and the remaining 40% have taken in the testing process.

Statistical parameters, such as MAPE-, R2 Score-, and RMSE-based prediction results of time series prediction algorithms (LSTM, ARIMA, and proposed CNN-LSTM Hybrid Model), are shown in Figs. 6, 7, and 8, respectively. From the experimental outcomes, it is clear that the performance of our proposed model is much superior to the other forecasting models. It is more appropriate in predicting the COVID-19 epidemic across India.

Fig. 6
figure 6

MAPE-based prediction results of COVID-19 outbreak of India

Full size image
Fig. 7
figure 7

R2 Score-based prediction results of COVID-19 outbreak of India

Full size image
Fig. 8
figure 8

RMSE-based prediction results of COVID-19 outbreak of India

Full size image

The LSTM, ARIMA, and our proposed CNN-LSTM hybrid model-based forecasting results of the COVID-19 epidemic across India are presented in Fig. 9, via a line graph. The aggregate number of cases (target values) and the aggregate number of testing samples have been denoted on the y-axis and x-axis, respectively. In the process of result finding, twenty-nine affected states of India have been taken into consideration. The data of these states have been extracted, from the Arogya Setu App and MoHFW (Ministry of Health and Family Welfare), GoI (Government of India). The line chart has presented the experimental evaluation, and it is based on the comparison of actual values (real values) and predicted values (observed values). It is clear from the experimental result that the performance of our proposed CNN-LSTM hybrid model is much superior to the other time series prediction models (i.e., LSTM and ARIMA) for the forecasting of the COVID-19 epidemic across India.

Fig. 9
figure 9figure 9figure 9figure 9figure 9figure 9figure 9figure 9

COVID-19 epidemic analysis (for India) utilizing ARIMA, LSTM, and proposed CNN-LSTM hybrid model

Full size image

5.1 Current status of resource availability in India

As we know, the situation due to the COVID-19 outbreak in India is getting worst day by day. The number of fresh cases is breaking the record of the previous day. Now the virus is in its community-level spread, which means in the upcoming weeks, there will be a need for various healthcare resources to deal with this pandemic situation. If we talk about healthcare resources, such as hospitals and doctors in India, they are less compared to the total population. It is a serious matter of concern. So, India's government is taking various safety measures to slow down the COVID-19 growth rate.

In this section, we are going to talk about the current status of resource availability in India to deal with the COVID-19 outbreak. For this study, the data from the website of MoHFW (Ministry of Health and Family Welfare) has been taken. From there, we have extracted the information regarding the hospitals and the number of beds across India. This discussion is based on the Health Management Information System (HMIS) and National Health Profile (NHP18) (Hospitals in the Country. 2020). The Health Management Information System (HMIS) consists of the community health centers, district hospitals, primary health centers, sub-district hospitals, and the total number of public health facilities under HMIS. The National Health Profile (NHP18) consists of rural hospitals, urban hospitals, and the total number of public health facilities under NHP18. The state-wise public health facilities of the Health Management Information System (HMIS) and National Health Profile (NHP18) are shown in Fig. 10a and b, respectively.

Fig. 10
figure 10

The state-wise public health facilities in India a HMIS b NHP18

Full size image

Figure 11 shows the aggregate number of beds across India, consisting of public beds and urban beds. The total number of 37,725 Health Management Information System (HMIS) exists across India, consisting of 5568 community health centers, 1003 district hospitals, 29,899 primary health centers, and 1255 sub-district hospitals. Similarly, 23,582 National Health Profile (NHP18) exists across India, consisting of 19,810 rural hospitals and 3772 Urban Hospitals (Chua and Roska 1993). Figure 12b and b shows the public health facilities by HMIS and public health facilities by NHP18.

Fig. 11
figure 11

Aggregate number of beds in India

Full size image
Fig. 12
figure 12

Public health facilities a HMIS and b NHP18

Full size image

6 Conclusions

In the epidemic situation, any help from the algorithmic perspective or the clinical perspective is a precious but novel task. If we talk about the algorithmic perspective, the correct forecasting can play a vital role in dealing with such a critical situation and also give an idea about the exposure. Thus, the public authority can make arrangements in like manner and diminish the general effect of the illnesses. The main role of this investigation is to build up a particular model that can gauge the COVID-19 flare-up across India, with more noteworthy precision. In this paper, a CNN-LSTM hybrid forecasting model has been proposed, which can precisely foresee the COVID-19 episode across India contrasted with other conventional models (i.e., LSTM and ARIMA). The proposed model uses convolutional layers to extract meaningful information and learn from a given time series dataset. It is also enriched with the capability of the LSTM layer, such as categorizing long-term and short-term dependencies. The trial assessment has been performed by utilizing the three factual measures, like MAPE (ought to be low), R2 Score (ought to be high), and RMSE (ought to be low). The statistical measure-based experimental evaluation has been performed to assess the performance and suitability of the recommended model over the other well-grounded time series forecasting models. It is clear from the prediction outcomes of the COVID-19 epidemic, across the various states of India (29 states) that our proposed CNN-LSTM hybrid deep learning model performed exceptionally well throughout the experiment as compared to the other well-grounded time series forecasting models. Our proposed model achieves minimal MAPE, highest R2 Score, and minimal RMSE under several selection settings (state-wise). From the experimental analysis, it is also clear that the use of extra convolutional layers with the LSTM layer may increase the performance of the forecasting model. Apart from this, the deep insides on the current status of medical resource availability across India have been discussed.

In the future, we will do the testing of this proposed algorithm on the other time series datasets to determine suitability and correctness.