Using Machine Learning and Predictive Artificial Intelligence to Determine Cage Change Frequency for Mice Housed in Individually Ventilated Cages and Drive Vivarium Operational Efficiency
A standard 2-wk cage change frequency for individually ventilated mouse cages is used in many research facilities, with negligible effects on animal health and welfare. However, these techniques rely on subjective visual evaluations and often require spot changes. In this study, we describe the use and validation of digital monitoring technology to objectively determine the necessity of a cage change for mice. We used a machine learning/artificial intelligence algorithm that was trained by annotating human observations of soiled bedding to correlate with the Bedding Status Index (BSI), a digital measure quantifying bedding ‘wetness.’ Training of the algorithm was performed using various mouse strains of different age, sex, and cage densities to account for variability of these factors. Through constant user feedback and increased datasets, we were able to identify soiled cages with an accuracy >90% for cages with higher densities (for example, 5 animals per cage), while lower densities exhibited slightly reduced accuracy levels (the lowest accuracy was attributed to single-housed mice, at 76%). Our data show that the average change intervals for most average-sized mice ranged between 3 and 6 wk depending on the number of animals in the cage, which is significantly different from the standard 2-wk change used in our facility. Retired breeders and larger mice tended to have a shorter cage change interval as determined by the algorithm. These results show that the Bedding Status Index, which measures an intracage environmental variable, namely bedding wetness, can be used as a marker for cage change. The extended cage change schedule did not affect intracage ammonia, CO2 levels, mouse growth rates, or circadian rhythm metrics. Using digital alerts to determine the need for a cage change resulted in a 65% to 70% reduction in the number of cage changes needed, indicating that this method can improve operational efficiency by reducing cage changes, cage wash time, staff labor, and resource consumption.

Pictograph representing experimental setup. The studies were done as a training phase and validation phase. In the training phase, ‘wetness’ of bedding was scored by visual observations. Cage change was done when at least 2 of 3 observers gave a score ≥3. In the validation phase, ‘dirtiness’ was determined by the BSI digital biomarker; observers could score the task as ‘fair’ (needing change), ‘too dirty’ (needing change), or ‘too clean’ (no need to change).

Determining cage dirtiness using the Bedding Status Index (BSI). The BSI is measured in units of electrical capacitance. The initial value is an arbitrary starting point established when the board is calibrated, meaning that this number has no significance. As the bedding becomes saturated with fluid over time, this value steadily decreases until it reaches the Dropboard value that triggers an alert for cage change. When the cage is changed, the capacitance returns to the initial value.

Measurement of intracage parameters. (A) Rubber tubing connected to the Drager X-am 5000 ammonia reader; the pump is fed through the water grommet in the back of the cage. Readings with the tip of the rubber hose are performed over the nest and latrine areas (or the area opposite the nest if an obvious latrine could not be identified) immediately after removal from the rack (within ∼30 s). (B) Colorimetric ammonia measurement from the center of the cage. The colorimetric card changes color with rising ammonia levels; that is, yellow, 0 to 1 ppm; light green, 1 to 25 ppm; dark green, 25 to 50 ppm; and blue, >50 ppm. (C) CO2 sensor measurement from the center of the cage.

Capacitance drop in response to bedding soiling. A representative graph of drop in capacitance across the 12 electrodes of the panel below the cages is shown from a cage with 5 female mice. The greatest drop in capacitance in this cage was seen in the front half of the cage (latrine spot) while no significant capacitance drop was seen at the back of the cage where the mice built their nest.

Average days to cage change based on cage density for all experiments. The figure presents graphs depicting the average cage change interval and SD of cages by density from the individual studies performed. (A) Mixed strain study: presents the average days to cage change and SD of cages containing 1, 2, 3, 4, and 5 mice. (B) C57BL/6J learning phase study: depicts the average days to cage change and SD based on manual scoring. (C) C57BL/6J validation phase study: depicts average days to cage change using BSI-based alerts for cages containing 2, 3, and 5 age- and sex-matched C57BL/6J mice. (D) BALB/c study: presents the average cage change interval and SD of 2, 3, and 5 age- and sex-matched mice. (E) Swiss Webster study: shows the average days to cage change and SD. *, P ≤ 0.05; †, P ≤ 0.01; ‡, P ≤ 0.001; §, P ≤ 0.0001; +, P ≤ 0.005; ×, P ≤ 0.0005.

Representative images from the visual observations from the learning phase (C57BL/6J study). This image depicts the view of the top and underside of cages containing 2, 3, and 5 mice from the learning phase of the C57BL/6J study on the day of cage change (right column). The circles and oval markings outline the ‘latrine spots’ and areas of powdered bedding sticking to the cage bottom and was scored as ‘dirty’ by observers.

Performance of the BSI algorithm. Following a BSI alert for cage change, a ‘user’ verifies if the bedding is ‘too clean’ (false positive), ‘too dirty’ (false negative), or ‘fair’ (true positive). Improvements in the accuracy of the algorithm between the first (A) and second (B) C57BL/6 validation studies are shown.

Intracage ammonia levels. (A) Comparison of average ammonia levels measured over the nest area during the learning phase (C57BL/6 study). (B) Comparison of average ammonia levels over the latrine during the learning phase (C57BL/6 study). (C) Average ammonia levels measured over the nest area (BALB/c study). (D) Average ammonia levels over the latrine (BALB/c study).

Intracage CO2 levels. (A) Average CO2 levels measured from the center of the cage during the C57BL/6 learning study 2. (B) Average intracage CO2 levels for the BALB/c study.

Average growth rates of mice housed on digital caging compared with standard growth rates provided by the vendor. This figure represents the average growth rates of male (A) and female (B) C57BL/6J mice from 6 to 12 wk of age as compared with the vendor data. The average growth rates of male (C) and female (D) BALB/c mice study as compared with the vendor data: (C) only presents data up to 10 wk, as the vendor only provided data up to this time, and (D) presents data from 6 to 13 wk of age.
Contributor Notes