Indian Farming 69(03): 60–62; March 2019 Application of Artificial for livestock disease prediction

K. P. Suresh1, Dhemadri2, Rashmi Kurli3, R. Dheeraj4 and Parimal Roy5 ICAR-National Institute of Veterinary Epidemiology and Disease Informatics, Bengaluru 560 064

Diseases have emerged as a major constraint to the sustainable growth of the national economy. Several diseases have reduced the productivity of the livestock and have slowed the growth of the sector. Many diseases are linked to environmental deterioration and stress associated with farm intensification. According to the epidemiological triad, for a disease to occur it is the environment which helps the pathogen/vector to move towards the susceptible host. Thus based on the environmental parameters of the particular area early recognition of a serious or exotic animal disease can be done which is one of the most important factors influencing the chance of controlling the disease and reducing its economic and social impact on the whole country.

Key words: , Disease outbreak prediction, Livestock

REDICTING the occurrence of chance of controlling the disease and opportunities around the animal Plivestock disease outbreaks can be reducing its economic and social healthcare sector. of considerable value to the long- impact on the whole country. Disease outbreaks data collected term sustainable development in In an attempt to look for a more from 31 AICRP centres is stored and India. Prior research on disease reliable model, we developed a maintained in NADRES v2 prediction has essentially depended disease-climate relationship models (National Animal Disease Referral upon traditional statistical models using artificial intelligence and GIS to Expert System version 2) (Fig. 1 and with varying degrees of prediction predict 13 economically important 2) database since year 1987. The risk accuracy. Furthermore, the livestock disease outbreaks in India. factors data consisting of application of these models in Often conflated with Artificial environmental variables such as sustainable development and in Intelligence, machine is a precipitation, soil moisture, potential controlling environmental subfield and one application of AI. In evaporation, wind speed, deterioration has been very limited. practical terms, machine learning is a temperature, specific humidity, The geographic and seasonal method for automating data analysis surface pressure were retrieved from distribution of many infectious by using that iteratively GES DISC diseases are associated with climate identify patterns in data and learn “GLDAS_NOAH025_M.2.1” and therefore the possibility of using from them. Machine learning dataset and remote sensing variables seasonal climate forecasts as applications are generally classified such as NDVI, LST (MODIS data predictive indicators in disease early into three broad categories: (1) product). warning system (EWS) is an interest supervised learning, (2) unsupervised The HDF files for LST (oC), of focus. Geographic Information learning and (3) reinforcement NDVI were downloaded from the system (GIS), remote sensing (RS) learning. Supervised learning uses MODIS website using the and Global Positioning system (GPS) patterns already identified in data (ie, MOD11A2 and MOD13A1 products are the three commonly used training data). Data mining is related respectively by specifying the veterinary geo-informatics to unsupervised machine learning and coordinates and time period (dates). technologies employed in this digital involves identifying patterns in large The Livestock population data was era for rapid of data datasets. derived from 19th livestock census for better management of animal 2012, Department of Animal diseases. Early recognition of a Disease outbreaks prediction husbandry, dairying and Fisheries, serious or exotic animal disease can Combining Artificial Intelligence Krishi Bhavan , New Delhi. be done which is one of the most techniques and copious amounts of The risk data generated at village important factors influencing the livestock disease data provide new level were aggregated to block level

Indian Farming 60 March 2019 Fig. 1. Preview of NADRES home page and month for suitable representation. Discriminant Analysis (FDA) and Receiving Operating Characteristic Disease outbreak data were aligned Classification Tree Analysis (CTA) (ROC) curve, Cohen’s Kappa with generated risk variables to the were employed for disease modelling. (Heildke Skill Score) and True Skill respective latitude and longitude, Different modelling methods return Statistics (TSS). These measures were which were subjected to climate- different types of ‘model object’ and all used to evaluate the quality of disease modelling. A number of these model objects could be used for predictions based on presence-absence models were fit to aligned data and the predict function to make data. Raster Stack was used to tested for accuracy in terms of predictions for any combinations of combine the results of individual discrimination power. Two regression values of independent variables. predictions by different model models, Generalized Linear Models Response plots were created to methods. All the models were assessed (GLM) and Generalized Additive explore and understand model for overfitting because it can cause Models (GAM) and six machine predictions. Anthrax outbreak on misleading of estimated co-efficient, p learning algorithms, i.e. Random livestock is forcasted for the month of values and R-Squares values. Forest (RF), Boosted Regression Tree April 2019 in Karnataka, as shown in Overfitting is suspect when the model (BRT), Artificial Neural Network Fig. 1. accuracy is high with respect to the (ANN), Multiple Adaptive The fitted models were assessed for data used in training the model but Regression Spline (MARS), Flexible their discriminating power using drops significantly with new data. In

Indian Farming March 2019 61 Fig. 2. Disease prediction view of NADRES v2 this study, the cross-validation occurrence of disease incidence. The important livestock diseases in Pan procedure was adopted to assess the Forewarning results are generated at India. Application has been optimized overfitting of models by keeping 30% block level for 4 economically by making it easy to navigate, so that data on hold while, the rest was used important diseases in Karnataka state users stay on and engage on content for training. The accuracy of held out as pilot mode were communicated to (Figs. 1 and 2). This web application data was used to compare the accuracy 2,000 veterinary officers via auto- provides several other features such as derived from data used in the training messaging (Fig. 1). The results are analysis results of disease pattern in and the significant variance in these also available in website (http:// each state, disease maps, distribution two flags overfitting. www.nivedi.res.in/Nadres_v2/) and of livestock population, impact The outcome of best fitted model/ mobile app developed for the analysis etc. Epi calculator for s were in probability of disease purpose provide the similar features. developing sampling plan was occurrence and was categorised into incorporated in the application. All the 6 risk levels-No risk (NR), Very low NADRES v2 Outlook results in the application is dynamic risk (VLR), Low risk (LR), NADRES v2 an interactive and and automated. Moderate risk (MR), High risk (HR) dynamic web applications built using and Very high risk (VHR). The HTML and PHP software as front- results forewarning were sent to all end tool and MySQL as back-end tool. 1,2Principal Scientists, 3,4Senior Research the state animal husbandry departs The applications provides forewarning Fellow, 5Director, Corresponding e mail: and DADF for taking preventive of disease occurrence in two months [email protected] measures in advance to avoid the advance for all the 700 districts for 13

Indian Farming 62 March 2019