An Accurate Prediction of Disease Using Trapezoidal Long Short Term Recurrent Neural Network Model in Big Data
K.Tamilselvi1, Dr. K.Ramesh Kumar2
1Research Scholar, Research and Development, Bharathiar University, Coimbatore - India
2 Research Guide, Bharathiar University, Coimbatore - India [email protected]1, [email protected]2
Abstract
;[„.The eruption of Coronavirus 2019 (COVID‐19)has crashedday to day lives across the globe.The positive case count is increasing and India is in the top most position among the world countries.
Human beings undergo chronic diseases with noidentification in time,thattransports increased load of disease to the society. This paper createsprognosticrepresentationwhich canforecast the positive count with increased accuracy. Regression‐based, Decision tree‐based, and Random forest‐based models are built on the data from China and are authenticated on sample of India. This is effectual and iscapablefor predicting the positive count with decreasedmistakes in future. This papergives an idea ofinfectionthreatforecast method toevaluates methodically furtherinfection threat for patients based on their present medical records using Trapezoidal Long Short Term Recurrent Neural Network model(TRAP-LSTM).This TRAP-LSTM follows multilayer structure with aggregation function and n–gram masking. The proposed TRAP-LSTM is compared with two state of art methods such as, Long Short Term Recurrent Neural Network (LST-RNN) and Gated Recurrent Units (GRU)interms of accuracy, precision, recall and harmonic score and hence the proposed method achieves90.2% of precision, 81.8% of recall,70.2% of accuracy and 46.4% of hamming score.
Keywords- Neural network, COVID-19, disease prediction, accuracy, big data
1. Introduction
Thetechnological development causes the computer to create immensedata. Furthermore, such progressions and developments in the clinical DBMS produce enormous of clinical data. Medical services industry comprises huge and delicate data. This data needs to dealt with very caution to get profited by it. Admittance to prediction models is fundamental to acquire knowledge toextend probably and results indiseases which are infectious. State Admisstration and other authoritative organization depend on knowledge from methodsof predictionto propose a lateststrategies and to survey its viability of strategies implemented [1]. The new coronavirus disease (COVID-19) was accounted for having infections in excess of 2 million individuals, in excess of 132,000 acknowledgedmortalityacross the globe. The new global COVID-19 deadly disease has nonlinearity and complexity nature in its structure [2]. Moreover, this flare-up is contrasted with another ongoing episodes, develops aquery the capacity of normalmethod to convey precise outcomes [3].
Notwithstanding various recognized and unrecognized factors engaged with the multiplicity, the intricacy of populace-extendedactivities in different international regions and contrasts in control methodologies significantly expand the model vulnerability [4]. Thus, standard epidemiological models features new difficulties to convey more solid outcomes [5]. To conquer this difficulty, numerous novel models have arisen which initiates a many suspicions with demonstration (e.g., adding social distancing as curfews, isolation, and so forth) [6].
Because of the worst circumstance, important devices dependent on artificial intelligence (AI) is considered; the process of machine learning utilizes recognizing pattern with big data, clarification, and forecast dependent on datainput [7]. Consequently, AI can possibly configure tools to battle COVID-19. In this examination, we used SSLPNN and GPR to foresee the classes to which specifically investigates a place and the quantity of COVID19 confirmed cases in particular topographical areas[8]. However, climatic and financial situationhas solid affiliationin the rate and irresistible infections spread[9].Nonetheless, this examination will assist for designingpolicies for
public health to control reasonable advancement approaches [10]. ANN is an all around reported AI method enthusedthrough the organization of organic human neurons. It is effectively put on various issues in various platform [11]. Fundamentally, it is an amazing asset for finding a connection among data of input and output. To achieve this reason, it is needed to be prepared utilizing a bunch of records including input and the respective output data [12]. The procedure of training is usually directed by the adaptable system of ANN with three layers: (1) an input layer, (2) a hidden layer, and (3) an output layer. The first and the third ones enclose neurons related with the input and output vectors, separately [13]. Then again, neurons in the hidden layer are associated with the neurons of the input and output layers, are fundamentally accountable for transforming thedata input to the relating data output
2. Related works
In this manner, numerical, energetically and geometrical methods areexploited to calculate the COVID-19 outbreaks. Few modelsproposed for this reasonincorporatesSusceptible-Exposed- Infectious-Recovered (SEIR) method [14], Logistic Growth method [15] and Adaptive Neuro-fuzzy Inference System (ANFIS) [16], e.g, Al-qanesset al. [17] altered ANFIS method utilizing Pollination and Salp Swarm Algorithm for calculation of the dispersal of COVID-19. Fu et al. [18] used a Boltzmann work-supportedmethodologyin favor ofassessment of aggregatecasesconfirmed in China .Niazkar and Niazkar [19] stated multi-genehereditary program, an AI model, China, Republic of Korea, Japan, Italy, Singapore, Iran and United States of America are the 7 countries predicted with COVID-19 developing numerical method with the remarkable capacity. They recommend the prediction methods depends on country and alsosuggested that the COVID-19 eruption in every country needs to be explored independently. Besides, Li et al. [20] recommended an outstandingcapacity for the prediction of the COVID-19 episode. They assessed the COVID-19 pandemic end in China to be after 20 March 2020, while around 52,000 to 68,000 tainted and 2400 mortality were forecasted. Moreover, Hu et al. [21] stated a strategy called modified stacked auto-
encoder, which was motivated by an artificial intelligence (AI) model, for COVID-19 real-time forecasting. Additionally, they anticipated the center of April as the COVID-19 pandemic closes.
Besides, Yang et al. [22] built up a SEIR model and AI approach, which was prepared by the 2003 SARS data for COVID-19 prediction in China. As far as , Artificial Neural Networks (ANN) application for the prediction of COVID-19 outbreak is restricted. Al-Najjar and Al-Rousan [23]
used ANN for forecastingthe recuperated and mortality by utilizingclinical distinctiveness, as it will difficult to assemble such prolonged information needed for expectation purposes.Shawni Dutta et al.,[24]stacked-GRU based Recurrent Neural Network model, abbreviated as, stk-G, stated this paper considering the interference issues from previous medical historyat the same time asfinding patients with heart diseases. This proposed model is compared with two benchmark classifiers known as Support Vector Machine (SVM) and K-Nearest Neighbour (K-NN). The comparative analysis concludes that the proposed model offers enhanced efficiency for heart disease prediction. A promising result is given by the proposed method with an accuracy of 84.37%, F1-Score of 0.84 and MSE of 0.16.Tingyan Wang et al., [25]represents a numerous infection threat prediction scheme to orderlydisease risks reviewexpectations for patients depends upon their clinical reports. Depends on this examination, clinicalanalyzewill depend upon International Classification of Diseases (ICD) are cumulated by various stages for prediction facing its demands of various stakeholders. The methodology described obtainsauthenticated by a pair of autonomousclinical datasets, including 7105 patients with 18, 893 patients and 4170 patients with 13, 124 visits, correspondingly.
Earlyexaminationexposes anincreaseddifferenceof uniqueness of patients 3. Proposed methodology for COVID prediction
In this section, we propose the TRAP-LSTM for learning long-term and multi-scale dependencies in medical data. It is a trapezoid-like structure and generally has multiple layers. In order to describe the multilayer structure of a TRAP-LSTM clearly, we will start from the first layer
to show the generation process of sub-trapezoids in a single layer, and then introducing the constructionof multi-layer structure as shown in figure-1.
Figure1:Proposed system architecture for disease prediction
3.1 Preprocessing of data
The observation time window was set for various covid cases. This time window was well thought-outadequatelyextendedfor observing the development of diverse diagnostic event patterns.
This methodology was chosen to produce an adequate count of input signals, and furthermore to envelopseveraldissimilaramalgamation of advancing diagnostic event patterns. One of the results of this methodology is that the time spansoverlieso the data patterns willillustrateresemblances. It is accepted that the last happened signal that addresses anoccasionhavingthe greatest effect on the patient condition. In this way, the time since the last happened occasion was utilized as input for task of classification . The information sources were standardized to be in the span [0.1, 0.9]. This methodology empowers greater adaptability in the application stage when patterns of data are introduced to the algorithm that surpass the data range enclosed by the training data.The enrollment of the three classes in the task of classification was characterized by fuzzy membership functions.
The fuzzy classes were characterized by data patternsfor the time until the following failure occurance. Three classes were characterized: data patterns prompting an interruption occasion temporarily, in the medium-term, or in the long haul. This classification is adequately exact for the administrators to expect the failure. The determination of three classes was persuaded by adjusting
the helpfulness of the performed forecast in reasonable applications, and decreasing the intricacy of the classification task and subsequently the stream chart is demonstrated in figure-2.
Figure-2 Flow diagram of proposed TRAP-LSTM method
3.2 Multilayer structure and learning process
We use H as the initial hidden state sequence of dimension dh and length T. The bottom of a sub-trapezoid is composed of a hidden state subsequence of length L. The hidden state hl at the bottom of a sub- trapezoid is driven by the input xl of the current time step with the hidden state hl−1 of the previous time step (consistent with conventional RNNs) or a hierarchical aggregated state hˆji (unlike conventional RNNs). If the hierarchical aggregation operation is performed every g time steps the update equation of hidden state hl can be expressed as follows:
Collect data from diagnostic devices
Time series of discrete event diagnostic data
Input data preprocessing normalization and balancing
Fuzzy membership assignment
Trapezoidal Long Short Term Recurrent Neural Network Restricted boltzman
machines
Echo state network-1
Echo state network-2
Echo state network-3 Restricted boltzman
machines
Linear regression
Hl=f(Uxl + Whl−1 + b), if (l − 1) ≡ 0 (mod g) (1) where xl and hl are the l-th input (time step) of an input subsequence and the l-th hidden state of a hidden state subsequence respectively. f(·) denotes the activation function (usually tanh(·)), and U ∈Rdh×dx , W ∈Rdh×dh and b ∈Rdh are trainable input weights, recurrent weights and the bias respectively. hˆji is the i-th hierarchical aggregated state at the j-th level of a sub-trapezoid, where j is the number of trailing-zeros of the G-ary number of the value (l − 1), and i = (l − 1)/gj .
3.3 Aggregation function
Given M dh-dimensional state vectors (e.g., hl, hˆji or ok Nk−1) to be aggregated, we can concatenate them into a state matrix by column as follows:
E = [e1, ··· ,em, ··· , eM] (2) where em∈Rdh is the m-th state vector and E ∈Rdh×M is the state matrix. Based on E, we use a one-hidden layer perceptron Fmlp(·) to obtain attention weights with fine-grained effects
S = Fmlp(E;W1,W2) = f2(W2f1(W1ET))T (3) where f1(·) and f2(·) are nonlinear activation functions (ReLU(·) and Sigmoid(·) respectively), and W1 ∈ RD×M and W2 ∈ RM×D are trainable weights, and S ∈Rdh×M is the self- attention weight matrix. We then obtain the weighted state matrix:
E˜ = S E (4) where denotes the element-wise multiplication and E˜ ∈Rdh×M is the weighted state matrix.
3.4 n-gram masking
The n-gram context is having each word w h = wk−n+1 . . . wk−1 is entrenchedby E(w)mapping; the vectors obtained as result are progressed to create a d · (n − 1) directional vector which is applied in first to input layer and later tooutput layer pursued by a tanh non-linearity.
Obatained output in “hidden” layer which is laterfeed into a output layer and then followed by an output layer O that result is with similar dimension likeexpressions. An exponential “soft-max” layer exchanges the establishments formedin the basispreviousoutput layer into in excess ofdatasetprobabilities.
To summarize:
X = concat(E(wk−n+1), . . . , E(wk−1)) D(X) = dropout(X; Pkeep)
Y = tanh(H · D(X) + Hbias) D(Y ) = dropout(Y ; Pkeep)
P(·|wk−n+1 . . . wk−1) = exp(O · D(Y ) + Obias)
This method is based on constraints of such as embedding matrix E ∈ R d×V , the maintainits probability for layers which are dropped out Pkeep, the input layer is constrained by H ∈ R s×(n−1)·d , Hbias∈ R s and the output layer is constrained by O ∈ R V ×s , Obias∈ R V . The hyper-parameters scheming the preparation are: numerous training epochs, order ofn-gram, parameterdimensionality d, s, keep probability value, gradient norm clipping value, initialize SD and the Adagrad learning rate and initialvalue of the accumulator.
3.5 Algorithm
D=[d1,d2,d3….dn]
Fixthe interval F F=[0.1, 0.9]
Membership function=M
M[short,medium,long]
Arrangement of sequence hl−1 If hl−1>F
Then
Calculate HI Form the matrix E
E = [e1, ··· ,em, ··· , eM]
Obtain the weight from matrix S n-gram context h = wk−n+1 . . . wk−1 summarization process ‟S‟ for prediction 4. Performanceanalysis
The proposed method is evaluated on two well-known publicly available covid datasets from database . Contradicting of our algorithm is done with the existing algorithms in terms of various parametric metrics like accuracy, precision, hamming score, recall are chosen. To confirm suggested Trapezoidal Long Short Term Recurrent Neural Networkis efficient than the existing methods the graphs are given here. For the purpose of simulation here we have chosen python.
If precisionrefers to the detectionability to accurately detect covid in dataset, the sensitivity calculation doesn‟ttakeindeterminate test results into account asa test cannot be repeated, and indeterminate samples should all be excluded from analysis.
precision= 𝑇𝑃
𝑇𝑃+𝐹𝑁
The table 1 presents the precision Analysiswith existing GRU and LSTM and proposed TRAP- LSTM
Table 1: Precision Analysis
Number of datasets GRU (%) LSTM(%) TRAP-LSTM (%)
50 56 64 82
75 59 69 86
100 62 72 90
125 65 75 95
150 69 79 98
Figure 3: Comparison of precision
Figure 3 shows the sensitivity comparison of existing and proposed algorithm. The X axis and Y axis shows that number of datasets and precision in percentage respectively. Precision of suggested TRAP-LSTM achieves better performance than the existing methods.
Recallrefers to the detection ability to correctly reject non-covid patients in dataset.
Mathematically, this can also be given as follows,
recall= 𝑇𝑁
𝑇𝑁+𝐹𝑃
The table 2 presents recallanalysis with existing GRU and LSTM andproposed TRAP-LSTM
Table 2: Recall Analysis
Number of datasets GRU (%) LSTM(%) TRAP-LSTM (%)
50 45 66 74
75 52 69 79
100 58 67 82
125 62 75 85
150 66 79 89
Figure 4: Comparison of recall
Figure 4 shows the recall comparison of existing and proposed algorithm. The X axis and Y axis shows that number of datasets and recall in percentage respectively. recall of suggested TRAP-LSTM achieves better performance than the existing methods.
Accuracy is ascountof correct predictions divided bypredictions made totally. Mathematically it is given as follows,
Accuracy = 𝑇𝑃+𝑇𝑁
𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁 (11) The table 3 presents accuracy analysis with existing GRU and LSTM andproposed TRAP-LSTM
Table 3: Analysis of accuracy
Number of datasets GRU (%) LSTM(%) TRAP-LSTM (%)
50 31 41 62
75 35 45 66
100 39 47 70
125 45 49 75
150 50 55 78
Figure 5: Comparison of accuracy
Figure 5 shows the accuracy comparison of existing and proposed algorithm. The X axis and Y axis shows that number of datasets and recall in percentage respectively. recall of suggested TRAP- LSTM achieves better performance than the existing methods.
Rate of hamming score isas the ratio of positive samplenumber. Rather precision represents the proportion of the prediction models of cancer where cancer is actually present. The rate of precision( P) is defined as :
Hamming score = 𝑇𝑃
𝑇𝑃+𝐹𝑃 (12) The table 4 presents the analysis of the hamming scorewith existing GRU and LSTM andproposed TRAP-LSTM
Table 4: Analysis of hamming score
Number of datasets GRU (%) LSTM(%) TRAP-LSTM (%)
50 15 36 41
75 22 39 45
100 28 42 47
125 32 45 49
150 36 49 50
Figure 6: Comparison of hamming score
Figure 6 shows the hamming score comparison of existing and proposednetwork. The X axis and Y axis shows that number of datasets and hamming score in percentage respectively. Hamming score of suggested TRAP-LSTM achieves better performance than the existing methods.
Table 5 shows the overall performance of analysis for prevailingGated Recurrent Units (GTU), Long Short Term Memory (LSTM) with proposedTrapezoidal Long Short Term Recurrent Neural Network model(TRAP-LSTM).The parameters considered for analysis are recall, hamming score, accuracy and precision
Table 5: Overall Performance Analysis
Method Precision(%) Recall(%) Accuracy(%) Hamming score(%)
GRU [24] 62.2 56.6 40 26.6
LSTM [25] 71.8 71.2 47.4 42.2
TRAP-LSTM [proposed]
90.2 81.8 70.2 46.4
Figure 7: Overall comparative analysis of existing and proposed algorithm
The figure 7 compares the values achieved for the parameters. Xaxisand Y axis shows parameters considered for analysis and parameter values obtained in percentage respectively. The proposed algorithm achieves 90.2% of precision, 81.8% of recall,70.2% of accuracy and 46.4% of hamming score.
Conclusion
This paper examined the COVID risk prediction modeling, considering multi-label classification issue. This approach was authenticated by a real- lifeclinical set of data. Thisrepresents the diagnosis of the patient wascombined into various stages to congregate the demands of severalcontributors s, e.g., victims, medical technicians, hospitals, and makers of decision in polocies. Auspiciously, the consequencesestablished that TRAP-LSTM systemswillforecast furtherrisks of disease for patients havingperfect-match score of utilizingtaskhold upunit in a data model of hospital, which make possiblehealthcare decision making of professional when it is needed.Predicting COVID-19 positivecount concurrentlymonitoring thepatientshistorical data with COVID-19 is preceded further, and the majority of the external factors that affect the spread of the virus is also taken into consideration by using blob-based visualization method.
References
1. Ivanov, D. Predicting the impacts of epidemic outbreaks on global supply chains: A simulation-based analysis on the coronavirus outbreak (COVID-19/SARS-CoV-2) case.
Transp. Res. Part E 2020, 136, 101922.
2. Koolhof, I.S.; Gibney, K.B.; Bettiol, S.; Charleston, M.; Wiethoelter, A.; Arnold, A.-L.;
Campbell, P.T.; Neville, P.J.; Aung, P.; Shiga, T.; et al. The forecasting of dynamical Ross River virus outbreaks: Victoria, Australia. Epidemics 2020, 30, 100377.
3. Darwish, A.; Rahhal, Y.; Jafar, A. A comparative study on predicting influenza outbreaks using different feature spaces: Application of influenza-like illness data from Early Warning Alert and Response System in Syria. BMC Res. Notes 2020, 13, 33.
4. Rypdal, M.; Sugihara, G. Inter-outbreak stability reflects the size of the susceptible pool and forecasts magnitudes of seasonal epidemics. Nat. Commun. 2019, 10, 2374.
5. Scarpino, S.V.; Petri, G. On the predictability of infectious disease outbreaks. Nat. Commun.
2019, 10, 898.
6. Zhan, Z.; Dong, W.; Lu, Y.; Yang, P.; Wang, Q.; Jia, P. Real-Time Forecasting of Hand- Foot-and-Mouth Disease Outbreaks using the Integrating Compartment Model and Assimilation Filtering. Sci. Rep. 2019, 9, 2661.
7. C. Garcia-Vidal, G. Sanjuan, P. Puerta-Alcalde, E. Moreno-García and A. Soriano,
“Artificial intelligence to support clinical decision-making processes,” EBioMedicine, vol.
46, pp. 27–29, 2019.
8. T. Hassan and F. Ahmad, “Transaction and identity authentication security model for e- banking: Confluence of quantum cryptography and AI, in Int. Conf. on Intelligent Technologies and Applications, pp. 338–347, 2018.
9. E. Rees, V. Ng, P. Gachon, A. Mawudeku, D. Mckenney et al., “Risk assessment strategies for early detection and prediction of infectious disease outbreaks associated with climate change,” Canada Communicable Disease Reports, vol. 45, no. 5, pp. 119–126, 2019.
10. F. R. Chowdhury, Q. S. U. Ibrahim, M. S. Bari, M. J. Alam, S. J. Dunachie et al., “The association between temperature, rainfall and humidity with common climate-sensitive infectious diseases in Bangladesh,” PLoS One, vol. 13, no. 6, pp. 1–17, 2018
11. Mohan S, Thirumalai C, Srivastava G. Effective heart disease prediction using hybrid machine learning techniques. IEEE Access. 2019;7:81542–81554.
12. Bahrami B, Hosseini Shirvani M. Prediction and diagnosis of heart disease by data mining techniques. J. Multidiscip. Eng. Sci. Technol. 2015;2(2):3159–40.
13. K. Steinhaeuser, N. V. Chawla, “A network-based approach to understanding and predicting diseases,” in Proc. Social Computing and Behavioral Modeling, Boston, USA, 2009, pp. 209- 216
14. Zhang S, Diao M, Yu W, Pei L, Lin Z, Chen D. Estimation of the reproductive number of novel coronavirus (COVID-19) and the probable outbreak size on the diamond princess cruise ship: a data-driven analysis. Int J Infect Dis. 2020;93:20
15. Rocklöv J, Sjödin H, Wilder-Smith A. COVID-19 outbreak on the diamond princess cruise ship: estimating the epidemic potential and effectiveness of public health countermeasures. J Travel Med. 2020
16. Roosa K, Lee Y, Luo R, Kirpich A, Rothenberg R, Hyman JM, et al. Short-term forecasts of the COVID-19 epidemic in Guangdong and Zhejiang, China: February 13–23, 2020. J Clin Med. 2020;9(2):596
17. Al-qaness MAA, Ewees AA, Fan H, Abd El Aziz M. Optimization Method for Forecasting Confirmed Cases of COVID-19 in China. J Clin Med. 2020;9(3): 674
18. Fu X, Ying Q, Zeng T, Long T, Wang Y. Simulating and forecasting the cumulative confirmed cases of SARS-CoV-2 in China by Boltzmann functionbased regression analyses.
J Infect
19. Niazkar M, Niazkar HR. COVID-19 Outbreak: Application of Multi-gene Genetic Programming to Country-based Prediction Models. Electron J Gen Med. 2020;17(5):em247.
2020
20. Li Q, Feng W, Quan Y-H. Trend and forecasting of the COVID-19 outbreak in China. J Infect. 2020;80(4):469–96
21. Hu Z, Ge Q, Jin L, Xiong M. Artificial Intelligence Forecasting of Covid-19 in China. arXiv preprint arXiv:200207112; 2020
22. Yang Z, Zeng Z, Wang K, Wong S-S, Liang W, Zanin M, et al. Modified SEIR and AI prediction of the epidemics trend of COVID-19 in China under public health interventions. J Thorac Dis. 2020;12(3):165
23. Al-Najjar H, Al-Rousan N. A classifier prediction model to predict the status of coronavirus COVID-19 patients in South Korea. Eur Rev Med Pharmacol Sci. 2020;24(6):3400–3.
24. Dutta, S., &Bandyopadhyay, S. K. (2020). Early Detection of Heart Disease Using Gated Recurrent Neural Network. Asian Journal of Cardiology Research, 8-15.
25. Wang, T., Tian, Y., &Qiu, R. G. (2019). Long short-term memory recurrent neural networks for multiple diseases risk prediction by leveraging longitudinal medical records. IEEE journal of biomedical and health informatics, 24(8), 2337-2346.
