key: cord-0101239-mggudlp5
authors: Pandianchery, Mredulraj S.; E.A, Gopalakrishnan; Sowmya, V; Ravi, Vinayakumar; K.P, Soman
title: Explainable AI Framework for COVID-19 Prediction in Different Provinces of India
date: 2022-01-12
journal: nan
DOI: nan
sha: d968d9e13a487c9e800ab7296ea39222105191db
doc_id: 101239
cord_uid: mggudlp5

In 2020, covid-19 virus had reached more than 200 countries. Till December 20th 2021, 221 nations in the world had collectively reported 275M confirmed cases of covid-19&total death toll of 5.37M. Many countries which include United States, India, Brazil, United Kingdom, Russia etc were badly affected by covid-19 pandemic due to the large population. The total confirmed cases reported in this country are 51.7M, 34.7M, 22.2M, 11.3M, 10.2M respectively till December 20, 2021. This pandemic can be controlled with the help of precautionary steps by government&civilians of the country. The early prediction of covid-19 cases helps to track the transmission dynamics&alert the government to take the necessary precautions. Recurrent Deep learning algorithms is a data driven model which plays a key role to capture the patterns present in time series data. In many literatures, the Recurrent Neural Network (RNN) based model are proposed for the efficient prediction of COVID-19 cases for different provinces. The study in the literature doesnt involve the interpretation of the model behavior&robustness. In this study, The LSTM model is proposed for the efficient prediction of active cases in each provinces of India. The active cases dataset for each province in India is taken from John Hopkins publicly available dataset for the duration from 10th June, 2020 to 4th August, 2021. The proposed LSTM model is trained on one state i.e., Maharashtra and tested for rest of the provinces in India. The concept of Explainable AI is involved in this study for the better interpretation&understanding of the model behavior. The proposed model is used to forecast the active cases in India from 16th December, 2021 to 5th March, 2022. It is notated that there will be a emergence of third wave on January, 2022 in India.

I N December 2019, there was an outbreak with unknown cause in the Wuhan city, Hubei province of China. The infection was named as 2019-nCoV on 12th January, 2020 & the genome sequences were send to World Health Organization (WHO). On 11th March 2020, WHO had declared covid-19 outbreak as global pandemic [1] .

Artificial intelligence (AI) provides improved methodologies to mitigate the effect of COVID-19 pandemic [12] . The deep learning tools provided by AI are one of the main highlights of improved methodology. Deep learning approaches Mredulraj S. Pandianchery, Gopalakrishnan E.A, Sowmya V and Soman K.P are with Centre for Computational Engineering and Networking (CEN), Amrita School of Engineering, Amrita Vishwa Vidyapeetham, Coimbatore, India. e-mail: mredulraj383@gmail.com, ea gopalakrishnan@cb.amrita.edu, kp soman@amrita.edu, v sowmya@cb.amrita.edu Vinayakumar Ravi is with Center for Artificial Intelligence, Prince Mohammad Bin Fahd University, Khobar, Saudi Arabia, e-mail: vravi@pmu.edu.sa succeeded to discover the pattern present in large datasets & had brought breakthrough in visual recognition, speech and video processing [2] . Deep learning approaches plays a prominent role in detecting COVID-19 infections among the infected people & the early prediction of COVID-19 cases. Deep Learning architectures are used for the detection of COVID-19 infection by the classification of chest radiography images [3] . They are also used for different real time applications such as face mask detection & social distance identification in public places which are the most important control measures for the mitigation of COVID-19 [4] , [5] . Before the implementation of these containment measures in public places, the government forecast the transmission trend of covid-19 among the population. The early prediction of covid-19 plays a key role to control the transmission dynamics of virus among the population. It helps the government to take the necessary precautions & alert the people by the implementation of different restriction in public places.

The early prediction of covid-19 cases can be done by the means of mathematical models & deep learning algorithms. The recurrent deep learning algorithms captures the pattern present in time series data. The forecasting of time series data plays a key role in different domain such as epidemics & finance [6] - [8] . Recurrent Neural Network (RNN) based models such as Long Short Term Memory (LSTM) & Gated Recurrent Unit (GRU) are capable to learn the dynamics in time series data more efficiently as compared to the traditional methods. From the beginning of pandemic, there was lots of study carried related to the transmission trend in different provinces of India [21] . Bahri [26] .

From the literature. It is evident that there are various researches for the efficient prediction of COVID-19 cases. In the literature review, the proposed model is used for the early prediction for different provinces individually i.e., training and testing the model on same parent provinces. The study doesn't involve whether the proposed model is robust, reliable or interpretable for different stakeholders which include Data Scientist, Product Owners, Regulatory entities, Medical Doctors & Executive Board Member etc. It is very critical to understand the model behaviour in different tasks such as explanation of predictions in-order to support the decisionmaking process and debugging the unexpected behaviour of the model. Due to the availability of large databases, well developed methodologies and good computational power, deep learning algorithms perform well on complex tasks. Due to non-linear structure in deep learning models, the decision taken by the neurons are considered to be black box. There is lack of transparency in understanding the model behavior. The requirements of transparency and trust are more crucial in critical applications. Explainable AI is a term which helps to make the results from AI systems more understandable to the stakeholders. Explainable AI contributes to "Responsible AI" considering the accountability, responsibility and transparency [9] , [18] . Explainable AI helps to detect the biases existing in the model or data and the weakness existing in the AI system. The detection of improper classification or wrong prediction helps to improve the AI models [10] , [11] , [17] . The explainable AI approach plays a key role in the medical domain. The explainable AI approach has been used for heart rate variability in ECG signal. The study involves of different deep learning architecture such as LSTM, GRU, CNN and RSCNN trained on one tachycardia disease and tested on different tachycardia diseases. As all tachycardia disease has fast beat rhythm as common characteristic, it will allow to check the robustness & reliability in the model [19] .

Motivated by the importance of "Explainable AI", this article is unique from few perspective: 1) India is a country with 36 provinces which includes 28 states & 8 union territories. This study had proposed a LSTM model for active cases per day prediction trained on single state with favourable dynamics & tested for rest of the provinces. This approach helps to understand the robustness present in behaviour of the model. 2) The study also involves the visualization & interpretation of the biased behaviour detected in model prediction.

In India, the first case of covid-19 was reported on 27th January, 2020 in Thrissur district of Kerala [27] . During covid-19 outbreak, WHO has acknowledged that the entire world had gazed their attention towards the strategies to be taken by Indian government in-order to control the covid-19 outbreak [28] . The nationwide lockdown was implemented on 24th March, 2020 in India as a precautionary step to control the pandemic [27] . In the absence of vaccine, non pharmacological intervention i.e. social distancing was the only way to control the pandemic. The sudden surge of covid-19 positive cases in hospital has brought the crises in the healthcare operations during the 1st wave of pandemic. The covid-19 outbreak has brought the challenge in the requirements of hospital beds, ventilators, testing services & front-line workers. It was a great challenge for the country to improve the primary health care system, as 65-68% of the population lives in the rural areas [29] . The forecasting of covid-19 cases was necessary to capture the transmission dynamics in the different provinces of India. It helps the government to predict the requirements of healthcare facilities & to make the decision for the implementation of precautionary measures which include lockdown & quarantine [30] - [32] .

Time Series Analysis is defined as an analysis of orderly sequenced value of the variable which is sampled at the specific interval of time. It captures the evolution of the variable with respect to time & helps to forecast the future data points [33] . The analysis of time series detect the regularities present in the data which include seasonality & cyclicity. The different techniques are used for time series forecasting which include regression method, auto regression method, mathematical modelling & machining learning methods [34] - [36] . The extreme events in dynamical systems are responsible for the sudden transitions in the time series data. This kind of dynamics in the data can be captured with deep learning methods [37] . This study focuses in the implementation of deep learning based model for the prediction of active cases per day in different province of India.

Conventional machine learning methods has limitation in its ability to convert the raw data into suitable representation. In the development of machine learning models, domain expertise is needed for the feature extraction from raw data. The representation based learning helps the machine to automatically discover the patterns present in the raw data. Deep learning methods is a representation based learning method which extracts the intricate structure present in the data with multiple levels of abstraction. Deep learning methods had achieved state of art accuracies in different fields which include visual recognition, speech recognition & natural language processing etc. For time series data, the hand crafted features are more expensive. The recurrent neural networks in deep learning methods helps to capture the past information present in the sequence & predicts the information in future time steps [2] , [38] .

1) Recurrent Neural Network (RNN): Recurrent Neural Network captures the pattern present in the data which varies with respect to time. RNN is a structured deep learning model with feedback loop which allows to do the forecasting. In feed forward neural network, there is a unidirectional flow from input layer to the output layer. In RNN, the output from the layer is added with input and fed back into the same layer. RNN takes the sequential input and gives the sequential output from the model. So, RNN plays a key role in sequential modelling with a different variety of applications [23] . The drawback in RNN is that it suffers from the problem of vansishing gradient & exploding gradient.

2) Long Short Term Memory (LSTM): Long Short Term Memory provides gating mechanism which helps to overcome the inefficiency of long term dependencies. LSTM is consist of three important gates which include Input Gate, Forget Gate and Output Gate. The Input Gate in LSTM helps to retain the information in current cell state. The Forget Gate plays a key role in determining which information has to be removed from the previous LSTM cell. The output from LSTM cell is determined by the Output Gate with the combination of cell value, current input and output from the previous iteration [24] . The gating mechanism helps in retaining the relevant information for the next cell and helps to sustain the long term dependency.

3 [39] , [40] . GRU does not contain current cell state which is present in LSTM. The information present in current cell state is directly fed in to the hidden state of GRU.

Now a days, there are lots of development in industrial & research domain with the help of deep learning. Deep learning had succeeded in solving the critical problems. But the wrong interpretation from the deep learning models is very crucial in different critical domains. So, it is necessary to understand the model behaviour in deep learning approach. The questions always arises in the research with respect to the failure of the model and also how to leverage the performance of the model. The term "Explainable Artificial Intelligence" helps to do the interpretation and understanding from the analysis of the deep learning models. It helps to interpret the decision made in the black box of the neurons. It helps the ML community to have a good interpretability and explainability of ML algorithms [25] . Explainable AI helps to detect the biased decision taken by the deep learning model and also interpret the model behaviour which improve the performance in various field of AI.

The development of prediction model & experiments are carried out in Python platform. Different recurrent deep learning architecture are compared for the efficient prediction of covid-19 cases.

The dataset for 28 states and 8 union territories of India has been extracted from John Hopkin's publicly available dataset [22] . The dataset for each province in India is consist of active cases per day, case fatality, cumulative confirmed cases, cumulative death cases, incident rate and cumulative recovered cases. The dataset extracted is for the duration from 10th June, 2020 to 4th August, 2021. Active cases per day is calculated based on cumulative confirmed, deceased and recovered cases. The study involves data which includes 1st wave & 2nd wave of the COVID-19 pandemic. Many states in India were affected badly during covid-19 pandemic. Maharashtra, Kerala, Karnataka & Tamil Nadu had reported 20%, 11%, 9% & 8% of cumulative confirmed cases in India respectively. Maharashtra, Karnataka, Tamil Nadu & Delhi had reported 31%, 9%, 8% & 6% of cumulative deceased cases in India respectively. Maharashtra is one among the province in India which was badly affected during the pandemic & also reported maximum active cases per day up to 7,01,614 cases during 2nd wave of pandemic.

In this study, deep learning architecture is trained on one province which had reported large number of active cases per day with good dynamics & this pretrained model is tested for rest of the provinces. This approach helps to understand the robustness in the model and it will also help to improve the results obtained based on the better understanding of the model behaviour. The study also focuses on the comparison between two approaches as follows 1) Proposed Approach-1 : To train the model on state with favourable dynamics & test for rest of the provinces. 2) Proposed Approach-2 : To train & test the model on same parent province. As per the data exploration, Maharashtra is one of the states which was badly affected during the 1st and 2nd wave of pandemic in India. Due to the high population density, the covid-19 pandemic had shown the worst effect in Maharashtra. So, 

where X max represents the maximum value of the data & X min is minimum value of the data. where p i represents the predicted value, a i is actual value & n is total number of observations. The research framework for training & testing the model for the efficient prediction of active cases per day is shown in Fig. 1. The pretrained model is also tested on the data of active cases per day for India & forecasted for 100 days ahead. Based on comparison with other deep learning architecture in Table I as in proposed approach-2 for majority of the provinces in India. From the comparison, it has been notated that proposed approach-1 had performed better as compared to proposed approach-2 for some provinces which include Chhattisgarh, Gujarat, Haryana, Karnataka, Kerala, Tamil Nadu, Telangana, Uttar Pradesh, Uttarakhand & Delhi with difference in MAE greater than 1000. The MAE calculated for the prediction of active cases per day for these ten states in proposed approach-1&2 is shown in Table II . In both the approaches, the LSTM model had incorporated one LSTM layer with 150 hidden units, "tanh" as input/output activation function & "hard sigmoid" as recurrent activation function. So, the study had focussed in the visualization of the output from the LSTM layer. The output from the LSTM layer is elementwise multiplication of output from tanh function (input/output activation function) & hard sigmoid function (recurrent activation function). At the input of the model, the input step size is 1 x 8 which is mapped to the vector dimension of 1 x 150 at the output of the LSTM layer. For the prediction of 203 days in test data, there will be 203 x 150 dimensional matrix at the output of the LSTM layer. LSTM model. In proposed approach-2, the large value of absolute error between actual & predicted data has been notated between 75th and 203rd day of prediction while the model is tested. So, the output value of LSTM layer from 75th to 203rd day of prediction is considered for further study of proposed approach-1&2. It is difficult to differentiate the heatmaps of proposed approach-1&2 visually. To differentiate the heatmaps, the maximum & minimum output value from LSTM layer for each hidden unit is extracted. These two extreme output values have been compared with the actual data to check whether the LSTM layer output had captured the pattern present in the actual data. From the Fig. 8 , it is evident that proposed approach-1 had captured the pattern in data in a better way as compared to the approach-2 for Kerala. The pattern present after the peak value in the actual test data has been captured well in maximum activation output as compared to minimum activation output. The pattern present before the peak value in the actual test data has been captured well in minimum output. The LSTM model in proposed approach-2 had not captured the trend up to the peak value present in actual test data which lead to the large value of MAE between actual & predicted data from 75th to 203rd day.

In proposed approach-2 for Chhattisgarh, the large value of absolute error between actual & predicted data has been notated between 50th and 150th day of prediction as shown in Fig. 9 . The minimum & maximum output value from the LSTM layer is shown with actual trend of data in Fig. 13 for Chhattisgarh. So, the output value of LSTM layer from 50th to 150th day of prediction is considered for further study of proposed approach-1&2. The maximum & minimum output value had captured the trend in the actual test data in proposed approach-1&2. While in proposed approach-2, there is a lack of scalability in the two extreme value of output from LSTM layer as compared to the proposed approach-1 which is also reflected in the predicted value. For states which include Gujarat, Karnataka & Tamil Nadu. the maximum difference in actual & predicted data is notated between 75th to 175th day of prediction in proposed approach-2 as shown in Fig 10, 11 & 12 respectively . From the visualization of maximum & minimum output of LSTM layer for Gujarat in Fig. 14 , it is evident that there is right drift in the pattern captured in proposed approach-2 as compared to approach-1. The right drift in the trend of predicted value is notated for proposed approach-2 in Fig 10. For Karnataka, the trend in actual data is captured in both proposed approach-1 & 2. There is a slight deviation at the peak value of maximum & minimum output from LSTM layer in proposed approach-2 as shown in Fig.  15 . For Tamil Nadu, there is a deviation in the predicted data after the peak value for proposed approach-2 as shown in Fig.  12 . The same deviation of trend has been notated in maximum and minimum activation output in proposed approach-2 with respect to proposed approach-1 and actual trend of test data. The pretrained LSTM model had captured the trend of active cases per day in test data for 36 provinces in India. The visualization of maximum & minimum output from LSTM layer for proposed approach-1 had shown the robustness in the model to capture the different dynamics.

The proposed pretrained LSTM model is used for the forecasting of active cases per day in India. The data of active cases per day for India is collected from https://prsindia.org/covid-19/cases. The data for active cases in India from 9th February, 2021 to 15th December, 2021 is used to test the LSTM model. The MAE calculated between actual & predicted data for active cases in India is 19547.95. The forecasting of active cases per day is done from 16th December, 2021 to 5th March, 2022 as shown in Fig. 17 . As per the forecasting, there will be a rise in the active cases of covid-19 in January & February 2022. The active cases in India will spike more than 3,00,000 cases. The different variants of SARS-CoV-2 i.e. Beta, Delta & Omicron is a major threat to entire countries in the world [41] , [42] . As per the forecasting, there will be a emergence of third wave of pandemic in India from January, 2022. The severity of third wave will be less as compared to second wave of pandemic because of the massive vaccination drive in India. In India, three vaccines which include Covishield, Covaxin & Sputnik has administered more than 1.37 billion doses (Source: https://www.mohfw.gov.in). More than 0.82 billion & 0.54 billion population is inoculated with 1st & 2nd dose respectively. Apart from vaccination, it is necessary to follow the precautionary steps in daily life which include social distancing, use of face mask & hand-sanitizer to overcome this pandemic [43] . [44] , [45] . The health infrastructure plays a key role to control the pandemic. The forecast model incorporated with medical facilities will help to do the efficient prediction of active cases & requirements in health infrastructure. The another limitation in this article is that the proposed model is not linked with contact tracing of covid-19 infected people. This factors incorporated in the model will helps to efficiently capture the transmission dynamics of covid-19.

VI. CONCLUSION In this study, the proposed pretrained LSTM model is used for the prediction of active cases in different province of India. The different deep learning architecture which include SimpleRNN, LSTM, GRU, Stacked RNN, Stacked LSTM & Stacked GRU are trained & tested on data of Maharashtra. The detailed comparison has been done between these architecture based on different input window size & number of hidden units. The LSTM model has outperformed the other deep learning architectures on the dataset of Maharashtra. The proposed LSTM model trained on data of Maharashtra had captured the transmission dynamic for rest of the province in India. Under the umbrella of Explainable AI (XAI), this kind of approach tests the robustness present in the model to capture the different dynamics. The proposed approach-1 has given the better results as compared to proposed approach-2. For Lakshadweep, the first case was reported on 18th January, 2021. It was notated that number of training days has to be increased in proposed approach-2 as there is negligible dynamics in data. This kind of problem can be overcomed by proposed approach-1. The study had also focused on the visualization of maximum & minimum output value from LSTM layer for few states which include Kerala, Chhattisgarh, Gujarat, Karnataka & Tamil Nadu. From the visualization, it is evident that the proposed approach-1 had succeeded to generalize the model inorder to capture the different dynamics of transmission. The pretrained LSTM model is used to forecast the active cases in India from 16th December, 2021 to 5th March, 2022. The forecasted data had shown that there will be a emergence of third wave by January, 2022 in India. The forecasting of active cases helps the Indian government to implement the necessary restrictions & carry out the covid-19 testing in airports, bus & railway stations.

Epidemiology of COVID-19

Deep learning

Detection of COVID-19 Using X-ray Image Classification

Face Mask Detection by using Optimistic Convolutional Neural Network

Social Distance Identification Using Optimized Faster Region-Based Convolutional Neural Network

Deep learning methods for forecasting COVID-19 time-Series data: A Comparative study

Predictions for COVID-19 with deep learning models of LSTM, GRU and Bi-LSTM

Stock price prediction using LSTM, RNN and CNNsliding window model

Peeking Inside the Black-Box: A Survey on Explainable Artificial Intelligence (XAI)

Explainable AI: interpreting, explaining and visualizing deep learning

Explainable artificial intelligence: Understanding, visualizing and interpreting deep learning models

Use of AI, Robotics, and Modern Tools to Fight Covid-19

Deep Learning for COVID-19 prediction

COVID-19 Time Series Forecasting of Daily Cases, Deaths Caused and Recovered Cases using Long Short Term Memory Networks

COVID-19 prediction using LSTM algorithm: GCC case study

Predicting and analyzing the COVID-19 epidemic in China: Based on SEIRD, LSTM and GWR models

Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI

Responsible artificial intelligence: designing AI for human values

Explainable artificial intelligence for heart rate variability in ECG signal

Prediction and analysis of COVID-19 positive cases using deep learning models: A descriptive case study of India

Lessons learnt during the first 100 days of COVID-19 pandemic in India

An interactive web-based dashboard to track COVID-19 in real time

Predictive maintenance of oil and gas equipment using recurrent neural network

A review on the long short-term memory model

A Survey on Explainable Artificial Intelligence (XAI): Toward Medical XAI

Forecasting of covid-19 using deep layer recurrent neural networks (rnns) with gated recurrent units (grus) and long short-term memory (lstm) cells

First confirmed case of COVID-19 infection in India: A case report

A comprehensive analysis of COVID-19 outbreak situation in India

COVID-19: Challenges and its consequences for rural health care in India

Modeling and forecasting the COVID-19 pandemic in India

Trend Analysis and Forecasting of COVID-19 outbreak in India

Assessment of lockdown effect in some states and overall India: A predictive mathematical study on COVID-19 outbreak

Decision trees and genetic algorithms for condition monitoring forecasting of aircraft air conditioning

Estimation of COVID-19 prevalence in Italy

Prediction of COVID-19 corona virus pandemic based on time series data using Support Vector Machine

A model based study on the dynamics of COVID-19: Prediction and control

Healthcare Operations and Black Swan Event for COVID-19 Pandemic: A Predictive Analytics

Deep learning for time-series analysis

Deep learning with gated recurrent unit networks for financial sequence predictions

Gate-variants of Gated Recurrent Unit (GRU) neural networks

Increased risk of SARS-CoV-2 reinfection associated with emergence of the Omicron variant in South Africa

Omicron SARS-CoV-2 variant: a new chapter in the COVID-19 pandemic

The benefits and costs of using social distancing to flatten the curve for COVID-19

A Supply Chain Resilience Capability Framework and Process for Mitigating the COVID-19 Pandemic Disruption

Roles of Innovation Leadership on Using Big Data Analytics to Establish Resilient Healthcare Supply Chains to Combat the COVID-19 Pandemic: A Multimethodological Study