Use of a thermal sensor microchip and machine learning in the detection of changes in body temperature in dairy calves using anaplasmosis as a disease model

Abstract

Calf mortality and morbidity represent a significant cost in rearing and an important welfare issue. The most commonly used health assessment tool is the measure of body temperature using a rectal thermometer, which can be labor intensive, invasive and stressful for the animal. Automating temperature measurement can be useful during disease occurrence for early identification and treatment of animals. The objectives of this study were to evaluate: i) subcutaneous temperature data collected by the Bio-Thermo microchip compared to rectal temperature (RT) in calves exposed to anaplasmosis; and ii) the predictive ability of Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM) in the early identification of anaplasmosis. Additionally, we aimed to investigate: ii.a) the effect of time series length prior to disease diagnosis (5, 7, or 10 consecutive days) on the predictive performance of RNN and LSTM; and ii.b) how early anaplasmosis disease can be detected in dairy calves (3 days in advance or just on the day of clinical diagnosis). Twenty-four Holstein calves with 132.4 ± 13.9 (mean ± SD) days of age and 146 ± 23.3 kg of body weight were challenged with 2 × 107 erythrocytes infected with the UFMG1 strain (GenBank no. EU676176) isolated from Anaplasma marginale. After inoculation, animals were monitored daily by assessing packed cell volume (PCV). The lowest PCV value (14.8 ± 2.6%) and the identification of rickettsia on blood smears were used as criteria to classify an animal as sick (d0). Temperature data were collected daily using passive radiofrequency identification (RFID) and clinical thermometer. Two time series were built including last sequence of -5, -7 or -10 d preceding d0 or comprising a sequence of 5, 7 or 10 d randomly selected in a window from −50 to −15 d before d0 to ensure a sequence of days in which PCV was considered normal (32 ± 2.3%). Leave-One-Animal-Out Cross-Validation (LOAOCV) was used to assess prediction quality. The RT was increased in response to the disease, while the temperature measured by the microchip showed only small variations and the correlation between these two temperatures was low (r = 0.35, P<0.001). On d0, for both temperature datasets (microchip and rectal), the precision (ACC), sensitivity (SEN) and specificity (SPE) of RNN was lower than LSTM. The ACC, SEN and SPE of LSTM in detecting anaplasmosis on d0 using microchip data ranged from 71 to 77%, 62 to 67% and 75 to 88%, respectively. For RT data on d0, ACC, SEN and SPE of LSTM ranged from 96 to 98%, 96 to 100% and 92 to 96%, respectively. The predictive performance of the models did not improve when using longer time series. The ACC and SEN in the prediction of anaplasmosis up to 3 days before clinical diagnosis were greater than 80% only for RT using RNN models, confirming that RT allows the early identification of anaplasmosis-related changes in body temperature. The models generated with the temperatures obtained via microchip showed lower predictive quality.