Construction of an Early Warning Mathematical Model for Penaeus vannamei AHPND Based on the Deep Forest Algorithm

Abstract

Acute hepatopancreatic necrosis disease (AHPND) is widely prevalent, has a rapid onset, and has high mortality in shrimp culture, making it a key limiting factor affecting shrimp aquaculture development in recent years, resulting in massive economic losses to the industry worldwide. Systematic studies that investigate which factors significantly correlate with the occurrence of AHPND, and further establishment of a prediction model for the occurrence of shrimp AHPND, are important for preventing and controlling the disease. In this study, Penaeus vannamei in pond culture were preliminarily analyzed for the coupling relationship between the occurrence and prevalence of AHPND in shrimps and pathogens, and for environmental and host autoimmune factors by assessing the environmental factors, pathogen abundance, and host health indicators under AHPND incidence. Then, a mathematical early warning model of AHPND occurrence in pond-cultured P.vannamei was constructed using Deep Forest algorithm. The occurrence of AHPND and its environment, pathogen, and shrimp immunity factors in pond-cultured P.vannamei were continuously monitored to explore the relationship between the occurrences of AHPND in relation to these factors. A total of 18 parameters were assessed, including the air and water temperature, salinity, pH, dissolved oxygen (DO), ammonia nitrogen (NH4+-N) and nitrite (NO2-N) concentrations, culturable bacteria and Vibrio in water, culturable bacteria and Vibrio in the shrimp hepatopancreas, the proportion of Vibrio in water and the shrimp hepatopancreas, and the activities of acid phosphatase (ACP), alkaline phosphatase (AKP), superoxide dismutase (SOD), lysozyme (LZM), and phenol oxidase (PO) in shrimp muscles. The parameter simulation prediction data based on the P.vannamei AHPND occurrence-related factor sequence (environmental factor, microbial factor, and shrimp health indicator) were constructed for the first time. The one-dimensional sequence was mapped into the three-dimensional space, different kernel functions were selected in combination with the actual classification problem to compare the model fitting accuracy, and the test algorithm optimized the parameters in the model. A total of 140 relevant data groups were collected under the same mode, and the groups of additional exogenous inputs during the breeding process were eliminated. After deleting invalid data, there were 100 groups of classified monitoring data, including 25 groups of morbidity data and 75 groups of health data. Moreover, the model was affected due to the dimensional and quantitative differences among different factors. In order to improve the speed of subsequent experimental training and prediction accuracy, the 100 groups of training test data processed by the mapminmax function were normalized for data processing. The relationship between 18 parameters and the occurrence of AHPND in P.vannamei was analyzed using Pearson′s correlation, and the main influencing factors were further screened using pairwise analysis between the factors. Pearson′s correlation analysis indicated that the incidence of AHPND positively correlated (P 0.05). Furthermore, parameters were removed in the model construction process according to the correlation between parameters and factors. The occurrence of AHPND in P.vannamei directly and significantly correlated with seven parameters, including the total number of shrimp bacteria, the total number of shrimp Vibrio, LZM, the proportion of shrimp Vibrio, the total number of water bacteria, salinity, and the total number of water Vibrio. The prediction performance of three popular integrated learning method algorithms based on decision tree, Deep Forest, LightGBM, and XGBoost was evaluated using Python language programming, and, finally, a four-dimensional vector early warning prediction model based on the Deep Forest algorithm for the total number of shrimp bacteria, the proportion of Vibrio shrimp, the total number of water bacteria, and salinity was established (accuracy: 89.00%). Although the prediction performance of the Deep Forest model decreased somewhat compared with that of the support vector machine model established in this study, the algorithm was gradually screened out based on the correlation between factors, including the effects of all factors. It was proven that the Deep Forest model established in this study was the ideal prediction model for predicting the occurrence of AHPND in P.vannamei among the 10 dimension parameters tried, and the superiority of the Deep Forest algorithm was also further verified. The results provide basic data and technical support for shrimp AHPND disease prediction, prevention and control, and lay a theoretical foundation for further establishment of aquaculture animal disease early warning theory.

Keywords

• shrimp
• ahpnd
• early warning mathematical model
• deep forest algorithm
• python programming language