Estratégias de predição de anaplasmose bovina com tecnologias de precisão

Abstract

Bovine anaplasmosis is a widely distributed disease and the severity of clinical signs can cause direct damage to production systems worldwide. The detection of anasplasmosis is performed by clinical and laboratory analysis. These traditional methodologies are cumbersome, expensive, and difficult to implement systematically in large-scale operations. The objectives of this study were to evaluate: i) rumination (RUM) and activity (ACT) data collected by Hr-Tag in calves exposed to anaplasmosis; and ii) the predictive capacity of recurrent neural networks (RNN) in the early identification of anaplasmosis. Additionally, we investigated: ii.a) the effect of the duration of the time series prior to the diagnosis of the disease (5, 7, 10 or 12 consecutive days) on the predictive performance of the RNN; and ii.b) how early anaplasmosis disease can be detected in dairy calves (5, 3 and 1 day in advance). Twenty-three heifers with a mean age of 119 days and weighing 148 kg were challenged with 2×107 erythrocytes infected with the UFMG1 strain (GenBank nº EU676176) isolated from Anaplasma marginale. After inoculation, the animals were monitored daily by assessing packed cell volume (VG). The lowest VG value (14 ± 1.8%) and the finding of rickettsiae on blood smears were used as criteria to classify an animal as sick (d0). Rumination and activity data were collected continuously and automatically at 2 h intervals, using SCR Heattime Hr collars. Two time series (TS) were constructed including the last sequence of -5, -7, -10 or -12d preceding d0 or comprising a sequence of 5, 7, 10 or 12d selected randomly in a window of -50 to -15d before of d0 to ensure a sequence of days in which VG was considered normal (32 ± 2.4%). The quality of prediction was evaluated by the Long Short-Term Memory (LSTM) method and by Leave-One-Animal-Out cross-validation. Anaplasmosis disease reduced 34% and 11% of RUM and ACT, respectively. The precision (ACC), sensitivity (SEN) and specificity (SPE) of the LSTM in detecting anaplasmosis ranged from 87 to 98%, 83 to 100% and 83 to 100%, respectively, using rumination data. For ACT, ACC, SEN and SPE data, they ranged from 70 to 98%, 61 to 100% and 74 to 100%, respectively. The combination of RUM and ACT, as well as the use of longer time series, did not improve the predictive performance of the models. The accuracy and sensitivity in predicting anaplasmosis up to 3 days before the clinical diagnosis (d0) were greater than 80%, confirming the possibility of early identification of the disease. In a second study, the predictive capacity of the RNN was evaluated based on RUM and ACT data retrieved from two devices: necklace (Heatime HR) and earring (eSense Flex®). Fourteen dairy calves (119 days old and 148 kg) were equipped with both devices and challenged in the same methodology as in the first study. ACC, SEN and SPE for predictions at 0d were similar for necklace and earring (100%) using ACT data. ACT-based predictions for -3d on advancement decrease the ACC by 21% (ear tag) and 29% (paste) over 0d models. When the RUM data were used, the ACC, SEN and SPE were lower compared to the ACT data. The specific RNN models developed to detect anaplasmosis for both devices achieved high predictive quality and were able to detect anaplasmosis 3 days in advance. Such achievements indicate the great potential of wearable sensors in the early identification of anaplasmosis diseases. This could positively impact dairy farmers' profitability and animal welfare.