ESPECTROSCOPIA ÓPTICA E APRENDIZAGEM DE MÁQUINA COMO FERRAMENTA NA PREDIÇÃO DA CONCEPÇÃO E DIAGNÓSTICO DE GESTAÇÃO DE FÊMEAS BOVINAS

Abstract

The combination of optical spectroscopy and machine learning algorithms has attracted attention due to its potential for accurate diagnosis of various diseases through the analysis of spectra from biological samples. This study aimed to develop a rapid, accurate, and cost-effective methodology for detecting the probability of pregnancy during the pre-reproductive season and determining gestational status 30 days after insemination using blood serum from female bovines. Eighty Nellore females, including heifers and primiparous cows, aged between 14 and 36 months, were used. Blood samples were collected at three time points: at the beginning of the protocol (D-10), at the time of artificial insemination (D0), and on the day of pregnancy diagnosis (D30). After centrifugation at 600 g for 10 minutes, the serum was aspirated, divided into duplicate microtube aliquots (1000 μL each), and stored at -80 ºC until further processing. Fourier-transform infrared spectroscopy (FTIR) was employed in serum analysis, with a resolution of 4 cm-1 and a range of 1800 to 900 cm-1, using a Perkin-Elmer Spectrum 100 instrument with an ATR accessory. The prominent double vibrational bands of amide I around 1570 cm-1 were identified. Principal Component Analysis (PCA) assessed clustering trends, while the Support Vector Machine (SVM) discriminative test classified the samples. The score plot derived from the first two principal components explained 69.9% of the data variation and clearly exhibited distinct clustering formations for the pregnant group. Discriminative tests achieved pregnancy prediction accuracies with 93.8% accuracy at D0 and 62.5% accuracy at D30, successfully distinguishing between pregnant and non-pregnant animals. The serum metabolome showed discernible differences between animals with high and low fertility and those that conceived or not, enabling the development of distinct management strategies for different reproductive seasons. In summary, the methodology demonstrated a rapid and cost-effective approach, allowing fertility prediction during fixed-timed artificial insemination (FTAI) protocols and serving as a non-invasive tool for pregnancy diagnosis at 30 days. This enables the stratification of animals with different fertility levels for different reproductive goals before the breeding season. Although machine learning showed higher accuracy in evaluating samples at D0 than at D30, considering a machine learning algorithm, running more samples over time improves accuracy in this diagnosis, potentially replacing invasive methods such as rectal ultrasonography in the future.