A dynamic machine learning framework mitigates data drift for predicting responses to dietary arginine-to-lysine ratios in lactating sows
Year: 2026
Type: article
Abstract: Arginine is critical for optimizing sow lactation, yet its dose-response relationship in low-protein (LP) diets remains poorly characterized. This study developed a dynamic machine learning (ML) framework, trained on historical sow performance data ( n = 380), to predict the optimal arginine-to-lysine (A/L) ratio. Among the ensemble models, the decision tree regression model achieved the highest predictive accuracy ( R 2 = 0.956) and identified an A/L ratio of 1.03 as optimal. To validate this prediction, a 21-d feeding trial was conducted using 40 lactating sows (Landrace × Large White, 3rd parity, 220 ± 3 kg). Sows were randomly assigned to four treatment groups ( n = 10), including a normal-protein diet (CON) with A/L ratio of 1.21 and an LP diet supplemented with A/L ratios of 0.83, 0.93, or 1.03. Results indicated that while the LP diet significantly compromised piglet growth and intestinal integrity compared to the CON diet ( P < 0.05), the ML-predicted A/L ratio of 1.03 effectively restored performance. Specifically, it increased piglet average daily gain (ADG) relative to the A/L ratio of 0.83 ( P = 0.030). Regarding the intestinal barrier, A/L ratio of 1.03 enhanced the piglet jejunal, increased villus height to crypt depth ratio, and upregulated ZO-1 mRNA expression ( P < 0.05). From a metabolic perspective, A/L ratio of 1.03 optimized systemic nutrient status by elevating sow plasma glutamine, milk cysteine, and piglet serum total protein ( P < 0.05). Furthermore, A/L ratio of 1.03 mitigated LP-induced oxidative stress by increasing total superoxide dismutase (T-SOD) and glutathione peroxidase (GSH-Px) activities in sow serum and total antioxidant capacity (T-AOC) in milk, while reducing malondialdehyde (MDA) content ( P < 0.05). This ratio also bolstered immunity through higher sow serum secretory immunoglobulin A (sIgA) and piglet mucosal immunoglobulin M (IgM) levels ( P < 0.05). In addition, it beneficially modulated the intestinal microbiota by increasing the Shannon index and the abundance of Firmicutes ( P < 0.05). Finally, newly generated experimental data were subsequently integrated into the ML framework. A Kolmogorov-Smirnov test comparing data distributions triggered incremental model updating, which mitigated data drift and maintained predictive accuracy. In summary, this study establishes a novel closed-loop ML framework that not only guides targeted nutritional experimentation but also continuously self-improves. The findings confirm that an A/L ratio of 1.03 optimizes LP diets for lactating sows and provides a scalable drift-resilient tool for precision nutrition in animal production.
Source: Animal nutrition
Language: English
Subfield: Animal Science and Zoology
Domain: Life Sciences
SDG: Zero hunger
Open Access status: gold
Estimated APC paid: $1,462
OpenAlex ID: W7221088872
