Routine milk records reveal novel two-way interactions shaping ketosis risk in German dairy cows

Ketosis is one of the most prevalent metabolic disorders in dairy herds, with significant economic and welfare implications. However, existing literature often examines risk predictors in isolation, and their interactions and dependencies remain poorly understood. This study investigated how production, health, and management factors - and their interactions - are associated with ketosis risk (KR) at the animal level in German dairy cows, with a focus on German Holstein (GH) and German Simmental (SIM) breeds. Data from 76,809 cows across 717 German dairy farms collected between 2015 and 2019 were analyzed using multivariable generalized linear mixed-effects logistic regression with two-way interactions and random forest algorithms. KR was predicted using a non-invasive, widely applicable approach based on milk composition thresholds derived from monthly test-day records. Key risk factors included breed, parity, energy-corrected milk (ECM) yield, body condition, lameness, housing, and lactation stage. Twelve significant two-way interactions were identified, highlighting the multifactorial and interdependent nature of KR, most frequently involving lactation stage (six interactions), followed by ECM yield (five interactions) and breed (four interactions). The overall predicted probability of KR was 7.1%. Although high-yielding multiparous cows showed the highest absolute KR (prob. = 13.6% [95% CI 12.5-14.8]), high-yielding primiparous cows exhibited an unexpectedly elevated KR (prob. = 11.4% [95% CI 9.9-13.1]). Elevated KR was also observed in underconditioned cows during mid-early lactation (days in milk [DIM] 30-99, prob. = 21.0% [95% CI 18.6-23.5]), and SIM cows producing high ECM yields (>35 kg/day; prob. = 19.5% [95% CI 16.2-23.4]). The highest predicted KR occurred in underconditioned, first-parity cows housed in tie stalls, lame, and producing high ECM yields during mid-early lactation, indicating that specific combinations of risk factors amplify KR beyond the sum of their individual contributions. These findings highlight the importance of modeling interactions among animal-level and environmental factors to enable context-specific risk assessment. Evaluating breed, lactation stage, body condition, production level, housing, and health status jointly may support more targeted prevention strategies, particularly during the first 99 DIM, and improve identification of cows at elevated ketosis risk.