Using Machine Learning and Predictive Artificial Intelligence to Determine Cage Change Frequency for Mice Housed in Individually Ventilated Cages and Drive Vivarium Operational Efficiency

Abstract

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.

Introduction

Cage change is an essential activity in a vivarium to prevent feces accumulation, reduce ammonia buildup in the cage, and ensure adequate health and welfare of rodents. Cage change intervals are dependent on multiple variables such as the number and size of animals in a cage, the animal strain, temperature and relative humidity in the macro- and microenvironment of the cage, number of air changes, choice of housing (static or ventilated cages), and type and volume of bedding.¹ Pelleted and refined diced cellulose beddings provide better ammonia control as compared with corncob bedding and aspen shavings.^2,3 Certain strains, such as CD-1 mice, have been found to produce urine having a higher pH, which is associated with a higher concentration of ammonia.⁴ A low volume of bedding has been shown to result in higher ammonia levels as compared with using a larger bedding volume.⁵ An increase in intracage ammonia may result in the need for more frequent bedding changes. The Guide for the Care and Use of Laboratory Animals (Guide) states that soiled bedding should be removed and replaced with fresh materials as often as necessary to keep animals clean and dry and to keep pollutants, such as ammonia, at a concentration below levels irritating to the mucous membrane.¹ However, the Guide does not recommend a minimum frequency for bedding change. Identifying the ideal cage change frequency can be difficult, as a balance must be found between the multiple variables affecting the animals, operations, and animal welfare. Mice produce pheromones that are essential for territory marking, predation defenses, alarm responses, and behaviors indicating social status.⁶ Cage change disrupts the pheromonal milieue and scenting behaviors, with resulting impacts on stress,⁷ behavior, social hierarchy, sleep pattern,^2,3 and heart rate,⁸ all of which can affect experimental results. Changing cages too frequently will increase stress and affect reproducibility of research, whereas a reduced frequency can lead to unsanitary conditions and compromise animal health and welfare. Many institutions have adopted a 14-d cage change for IVCs as a uniform standard without regard to variables such as the type and volume of bedding, cage density, age, sex, and strain of the mice. This practice assumes that older and bigger animals will ‘soil’ the cage to the same level as younger animals or that a single animal can soil the cage to the same extent as 5 animals.

Multiple studies published over the past decade have employed novel techniques and modalities to optimize cage change frequency.^9–15 A study¹⁶ using CD-1 mice housed in IVCs showed that cages with 1, 3, and 5 mice can be changed at 4, 2, and 1 wk, respectively, with no impact to the welfare of the animals. Other studies using mice housed in IVCs have shown that use of a spot-change-only technique¹² or replacement of 75% of the bedding¹⁷ can increase the duration between cage change intervals without affecting the welfare of the animals. These studies reveal that adherance to the traditional 2 wk cage change frequency may not be needed and could lead to operational inefficiencies. For example, reduced cage changes using the spot-change-only technique resulted in 18% fewer daily cage changes.¹² Despite the apparent advantages of the extended cage change frequencies proposed in these studies,^12,16 changes to standard practice is challenging to implement across facilities due to the logistical problem of keeping track of each individual cage change. Visual assessment for soiled bedding can vary between different technicians, leading to an inefficient and inconsistent practice even within the same facility. Thus, there is a need to develop a quantitative and logistically easy method independent of observer bias to implement extended cage changes in rodents.

Performance standard is defined as a standard or guideline that, while describing a desired outcome, provides flexibility in achieving that outcome by granting discretion to those responsible for managing the animal care and use program, the researcher, and the IACUC. In this study, we aimed to develop a quantitative performance standard for cage change frequency by using continuous digital monitoring technologies to assess the intracage conditions. We used Digital Ventilated Cages (DVCs) from Tecniplast, which monitors ‘wetness’ of bedding using a digital measure called the Bedding Status Index (BSI).¹⁸ We validated the accuracy of the BSI’s assessment of bedding wetness with visual inspection by trained animal care technicians. This continuous quantitative data was analyzed using artificial intelligence (AI), specifically a machine learning (ML) algorithm, to initiate tasks for cage change. Using the ML/AI algorithm, we have determined cage change frequency using a quantitative measure that is free of observer bias and takes into consideration the variables of cage density, age, and sex of the mice from commonly used strains (C57BL/6 [B6], Swiss Webster [SW], and BALB/c). Our results show that using this method can lead to substantial savings for facilities and operational efficiency without compromising animal health or welfare.

Materials and Methods

Animals and housing conditions.

A total of 892 mice were used in this study. All mice were maintained in an AAALAC-accredited facility in accordance with the Guide. All procedures for animal use were approved by the Rutgers University IACUC. Mice were sourced either locally from faculty researchers or an approved vendor (The Jackson Laboratory, Bar Harbor, ME). All mice were housed on a commercially available DVC system manufactured by Tecniplast (Buguggiate, Italy). In this system, each cage location has a digital sensor board located directly below each cage on the rack. Each board has 12 electrodes connected to an integrated circuit that continuously measures changes in electrical capacitance at a rate of 4 times/s/electrode. Water has high permittivity compared with air, and materials containing water have high capacitance. Because the body composition of a mouse is ∼75% water, the presence of a mouse leads to a measurable change in capacitance that disrupts the electromagnetic field generated between the electrodes.¹⁹ As the mouse moves along the cage bottom, this active disruption in electromagnetic fields is recorded as activity. Similar to the mouse body composition, increased fluid content due to cage flooding or urination causes a measurable drop in the capacitance of electromagnetic fields.¹⁹

The cages were bedded with 350 mL of ⅛-in. pelleted cellulose bedding (no. L0107, BioFresh performance bedding; BioFresh, Patterson, NY), and each cage received crinkle paper nesting material (Enviropak; W.F. Fisher and Son, Somerville, NJ) as enrichment. Single-housed mice received a small enclosure as a secondary form of enrichment (Mouse Igloo; Bio-Serv, Flemington, NJ). The room was set to a 12-h light/12-h dark cycle (lights on at 0600). Room air was set to 10 to 15 air changes per hour (ACH). Intracage environmental parameters were set to 75 ACH as recommended by the manufacturer, humidity was set between 40% and 70%, and temperature was set at 70 to 72 ±2 °F (21 to 22 ± 2 °C). The animals had access to water from an automated watering system (hyperchlorinated, reverse osmosis water). Mice were provided extruded 5058 diet (PicoLab mouse diet 20; LabDiet, Richmond, IN) ad libitum. All mice were housed in an SPF facility that excluded the following pathogens: mouse parvovirus, minute virus of mice, mouse hepatitis virus, Theiler murine encephalomyelitis virus, murine rotavirus, Sendai virus, pneumonia virus of mice, reovirus, Mycoplasma pulmonis, Corynebacterium bovis, lymphocytic choriomeningitis virus, adenovirus, ectromelia virus, polyoma virus, Hantaan virus, murine chapparvovirus/mouse kidney parvovirus, mites, and pinworms.

Experimental design.

This study was performed as 4 separate studies over 12 mo (Figure 1). The initial studies were done to train the AI algorithm and are called the ‘learning phase.’ In this learning phase, a minimum of 3 experienced staff members provided input to train the BSI algorithm on the level of bedding wetness that triggers a cage change using previously published criteria for wet spots in bedding.¹⁵ The scoring criteria are described in the Data collection section below and in Table 1. The BSI algorithm correlates these ‘user’-defined inputs with the capacitance drops to develop BSI thresholds that meet the definition of a ‘soiled cage’ in our facility. In the validation phase, the system ‘cage change’ tasks generated by the BSI threshold were verified by the staff members (animal care technicians and researchers) for accuracy and feedback. The observers were given the option to score the cages as ‘too dirty’ (false negative), ‘too clean’ (false positive), or ‘fair’ (true positive). If the cage was considered fair or too dirty, the cage would be changed, and the score would be entered into the software to train the algorithm. Through this iterative process of constant feedback to the BSI algorithm, predictive AI cage change alerts were generated, and cage change frequency was established.

Figure 1.

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).

Table 1.

Study design highlighting the strain, age, sex, number of cages, scoring criteria, cage change criteria, and microenvironmental validation measures for each study

Study	Phase	Age/sex	No. of cages at beginning of study	No. of cages at end of study	Study duration (wk)	Scoring criteria	Cage change criteria	Microenvironmental validation measure
Mixed strains	Learning	Variable, unequal number of males and females	100	98	10	Visual scoring: score of 1, small (diameter 1–2 in.), light-brown urine spots with diffuse boundaries; score of 2, moderate-sized (diameter 2–3 in.), medium-brown urine spots with diffuse boundaries; score of 3, moderate- to large-sized (diameter 3–4 in.), dark-brown urine spots with distinct edges; score of 4, large (diameter 5–6 in.), dark-brown urine spots with distinct edges	Two of 3 observers score 3 or 4	None
Swiss Webster	Learning	Variable, unequal number of males and females	28	28	4		Two of 3 observers score 3 or 4, ammonia level >50 ppm	Ammonia (Draeger X-am 5000 and pump)
C57BL/6J	Learning	Six weeks old, equal numbers of males and females	120	109	7			Ammonia (Draeger X-am 5000 and pump, small animal ammonia sensor), CO2 sensor
C57BL/6J	Validation	Fifteen to 16 wk old, equal numbers of males and females	120	109	8	BSI measured 24/7 by digital caging	AI algorithm based on BSI values determined cage change interval	None
BALB/c	Validation	Six weeks old, equal numbers of males and females	30	28	7			Ammonia (Draeger X-am 5000 and pump), CO2 sensor

BSI.

The BSI evaluates the moisture level in the bedding. It is a capacitance-related measurement (measured in arbitrary units) at each of the 12 electrodes that are averaged for a selected time interval. Capacitance is the ability of a component to collect and store energy in the form of an electrical charge. As urine saturates the bedding, the capacitance, and thus the BSI, decreases in value over time. Once the cage is changed and fresh bedding is added, the capacitance acutely spikes to the initial value before beginning to drop again once the bedding starts to get wet with urine (Figure 2).

Figure 2.

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 Drop_board value that triggers an alert for cage change. When the cage is changed, the capacitance returns to the initial value.

The BSI algorithm was calculated using the formula T_cage = function(T_min, T_cutoff, Drop_board, Animal_number), where T_cage represents the absolute date of the next cage change; T_min represents the minimum time (days) before changing the cage; T_cutoff represents the maximum time (days) any cage could be on the rack before being changed; Drop_board represents the capacitance drop value for each electrode board (Figure 2); and Animal_number is the number of animals in the cage. According to this formula, a cage cannot be changed before the corresponding days of T_min are spent (unless there is excessive wetness due to water leaks) and cannot go longer than maximum days represented by T_cutoff. The animal facility management determined these 2 values based on data collected from the learning phases. For our facility, T_min was set as 1 wk and T_cutoff as 6 wk. In the case of flooding due to water valve leaks or animals playing with water valves, those cages were identified and removed from the study. Flooding detection is determined by BSI. The slope of the capacitance drop in electrodes underneath the water valve in the case of flooding is very steep and immediate with as little as 100 to 150 mL of water and a flooding alert is created within 15 s. In comparison, the BSI bedding change alert is a more gradual smaller and slower change in capacitance over time due to the increase in moisture level of bedding caused predominantly by urination. When a flooding alert was detected, the cages were flagged as ‘wet cages’ by the staff and removed from the study. All of the possible cage change requests in between T_min and T_cutoff from the BSI alerts were triggered by the capacitance drop metric (Drop_board) values, depending on the cage density. Each cage on the rack is continuously (24/7) assessed by the algorithm to determine whether the cage requires a change out, and the list of dirty cages to be changed (as generated tasks) is provided by the digital caging system early in the morning for the facility. The initial capacitance value is an arbitrary starting point established when the board is calibrated, meaning this number has no significance. As the bedding becomes saturated with fluid over time, this value steadily decreases until it reaches the value that triggers an alert for cage change (Figure 2).

Mixed strains study.

We used 300 male and female mice of varying ages and strains, sourced from multiple researchers across the Rutgers University campus. These mice were not actively being used in a study and were phenotypically normal; no mice with adverse phenotypes or physical/neural defects were enrolled in the study. Strains included B6 (and transgenics made on this background), FVB, KCna null, and AdkL mice. Mice were either group-housed according to sex or single-housed. The mice were divided into 20 cages each of 5, 4, 3, 2, and 1 mouse, respectively. Once all of the mice were acquired (300 mice total, 100 cages), all cages on the rack were fully changed and received fresh bedding. This was considered day 0. The amount of soiled bedding in each cage was then scored 3 times a week by 3 independent trained observers (Figure 1, Table 1). Once a cage was determined to be soiled per the assessment criteria (see the Data collection section), the cage was changed. This study was considered only as a learning phase and was concluded after 10 wk. During the study we had to remove 1 cage of single-housed mice and 1 cage containing 4 mice due to health concerns. Two cages containing 5 mice had 2 mice each removed due to fight wounds, and this group was added to the 3 mice per cage group. Thus, the final groups were 19 cages of single housed mice, 20 cages of 2 mice/cage, 22 cages of 3 mice/cage, 19 cages of 4 mice/cage, and 18 cages of 5 mice/cage (98 cages total).

SW mice study.

We acquired 28 cages of SW mice (5 cages of 5 mice/cage, 7 cages of 4 mice/cage, 7 cages of 3 mice/cage, 9 cages of 2 mice/cage) that were retired breeders from our colony. These animals were not age- or sex-matched and were experimentally naive. These cages were scored for wetness by visual observation.

B6 mice studies.

Equal numbers of 6-wk-old male and female B6 mice were divided into 10 cages each of 2, 3, and 5 mice (60 cages total). This study was performed over 13 wk and was divided into 2 phases: a learning phase (6 wk) and a validation phase (7 wk). During the learning phase, the cages were scored and changed as described above in the mixed strains study. In the validation phase, the mice from the learning phase were placed into new cage bottoms with fresh bedding. This was marked as day 0 for the second phase. The cages on the rack were changed only when the BSI-based soiled cage tasks were generated. No visual assessment was performed at this stage. During the first study, 2 cages of 3 male mice/cage had to be removed from the study due to fighting that happened within the first week of the learning phase (final total of 58 cages were used in this study; 20 cages of 2 mice/cage, 18 cages of 3 mice/cage, and 20 cages of 5 mice/cage).

The B6 study was repeated a second time with a new cohort of mice. This replicate was performed to verify reproducibility of the data and to increase the size of the dataset used to train the BSI algorithm. During the learning phase of the second study, some of the animals received from the vendor did not adapt to the automatic watering system and had acute weight loss (4 cages of 2 males/cage, 2 cages of 2 females per cage, and 2 cages of 3 females/cage). These were immediately removed from the learning phase. One cage of 2 male mice/cage was also found dead within the first week of the learning phase, and they were removed from the study as well. As a result, only a total of 51 cages were used in study 2 (13 cages of 2 mice/cage, 18 cages of 2 mice per cage, and 20 cages of 5 mice/cage).

BALB/c mice study.

This study involved 5 cages each of 6-wk-old male and female BALB/c mice housed at 2, 3, and 5 mice per cage (30 cages total). Data collected from the mice in this study were only assessed by cage density, not sex (each group contained 10 cages). This study was performed over 7 wk using BSI-determined cage change tasks; no visual examination was done. Two cages of 5 males/cage had to be removed from the study, as these animals were constantly playing with the water valve and causing flooding in the cage. Also, 1 cage of 3 males/cage were noted to be fighting right after they were housed, leading to 1 animal being euthanized due to fight wounds. This cage was kept as 2 males/cage for the rest of the study. The final groups included 8 cages of 5 mice/cage, 9 cages of 3 mice/cage, and 11 cages of 2 mice/cage (28 cages total).

Data collection.

Visual scoring for cage soiling.

Manual scoring was based on a previously validated soiled bedding scale¹⁵ (Table 1). Briefly, soiling was scored on a 1 to 4 scale assessing the latrine area, with 1 being the cleanest (light urination, small size) and 4 being the most soiled (heavy urination, large size). Scoring was performed by 3 staff members who had been trained to use the scoring system. Prior to scoring, observers were provided a printout of the previously validated soiled bedding scale¹⁵ to review the visual criteria associated with each score. Random cages on the DVC rack were removed and individually scored by the observers, which was then compared with the author’s score (which was considered our standard). The staff members were scored as proficient observers once they provided the same score as the author for 3 consecutive scores. On scoring days, each cage was removed from the rack and lifted slightly above the observer’s head to allow viewing of the underside of the cage. Scoring was generally performed in the morning between 0800 and 1000 and was done independent of other scorers. A cage was determined to have met cage change criteria when 2 out of 3 scorers provided a score of 3 or above (Figure 1). For the second B6 study the observers included 2 research staff members.

Ammonia.

Intracage ammonia levels were measured for only the learning phases of the B6 studies, the BALB/c study, and the SW mice study. Measurements were taken using an ammonia reader attached to a pump and hose (X-am 5000; Draeger, Telford, PA) (Figure 3A). Measurements from each cage were taken 3 times weekly, and the device was calibrated to room air prior to beginning measurements for a session. The cage was removed from the rack and the hose attached to the pump and the ammonia reader was fed through the water grommet in the back of the cage. Readings were taken in the center of the cage with the tip of the hose placed at the nose level of the animals for the first B6 and SW study. In the second study with the B6 mice and the BALB/c mice, we made a minor modification to obtain readings over the nest location and over the latrine area twice weekly. Along with this modification we also measured ammonia from 18 cages (3 cages of each sex and density) using a colorimetric ammonia detector (small animal ammonia sensor, Pacific Sentry) that was hung in the center of the cage from the food hopper (middle of the cage) at the beginning of the study and remained until cages were changed (Figure 3B).

Figure 3.

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.

CO₂ measurements.

CO₂ measurements were performed for 30 out of 60 cages (5 from each sex and density) in the learning phase of the B6 study and for 18 out of 30 cages (3 from each sex and density) in the BALB/c study. Measurements were taken twice weekly using a portable CO₂ detector (Lsenlty) (Figure 3C). The detector was calibrated to the room air before every use. The CO₂ detector was equipped with a clip and was latched onto the center of the food hopper in a cage (Figure 3C). The cage was then returned to the rack, and the detector was monitored until it reached a peak value, which was recorded.

Weight.

Individual mouse weights were taken upon arrival and then once weekly until the study ended for the learning phases of the B6 studies and the BALB/c study.

Statistical analysis.

Data were analyzed using GraphPad Prism software. Interrater reliability was determined by calculating a Kendall W (coefficient of concordance) for each scoring day of the learning phases. A noninferiority testing design, with a power of 0.8 and a significance level of 0.05, was used to calculate the sample sizes. The sample size per group was determined to be 7 cages. Initial studies were performed with 20 cages per group due to unequal numbers of males and females in the experiments. In latter studies, 10 cages were used per group (density and sex) to account for any potential loss of cages during the study. Cage change data were assessed for normality using Q-Q plots. Pairwise comparisons were performed with parametric t tests with a Welch correction. One-sample Wilcoxon tests were performed to assess whether the average number of days to cage change for each cage density (data from 1, 2, 3, and 5 mice per cage did not have a normal distribution) was significantly different from the traditional 2-wk cage change schedule. A one-sample t test was performed to assess whether the average number of days to cage change for cages containing 4 mice per cage (which had normally distributed data) was significantly different from the traditional 2-wk cage change schedule.

Results

In this study, we present a novel ML/AI algorithm-based quantitative performance standard for establishing cage change frequency in our facility. We used BSI, a digital measure that quantifies wetness of bedding, to determine the time of cage change. We trained the AI algorithm by performing visual observations using a previously published scoring method to correlate BSI at cage change with cage wetness. This unique iterative method helps teach the algorithm the cage wetness levels that would trigger a cage change at our institution. Similar to previous observations, mice were seen to nest at one part of the cage and urinate at the other end of the cage (latrine spot).³ There were no clear differences in the distribution of the ‘latrine spots’ between the front or the back of the cage, but for the most part they tended to be away from the nest area. As mice urinated and the bedding got wet, we could see a gradual drop in the capacitance readings at the latrine spot. A representative sample of capacitance readings from a cage with 5 mice at the time of cage change is shown in Figure 4. The data show that at the time of cage change the mice had dry bedding in the nesting area leading up to cage change, as the 6 electrodes at the back of the cage did not show any change in capacitance.

Figure 4.

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.

Mixed strain study.

In our first study (Figure 5A) we used a mixture of mouse strains, ages, and sexes to simulate a typical vivarium rack containing mice from multiple laboratories. We used 100 cages of mice in total, spread across both sides of a 160-cage rack. Visual observations were done by 3 independent observers, which included animal care technicians and veterinary staff. The cellulose pelleted bedding used in the study was very efficient at absorbing the urine, and large wet spots similar to those reported in previous studies were often not seen.¹⁵ Visual observations from the top of the cage did not show a clear picture of bedding ‘dirtiness’ (feces or wetness) (Figure 6). The cellulose pellets became swollen and dried out, probably due to the high air changes within the cage (75 ACH). As the mice moved around the cage, the dried-out pellets become powdery, sticking to the bottom of the cage. This was especially visible on the bottom of the last few cages with 5 animals/cage over 4 to 5 wk on the rack (Figure 6). In these cases, bedding adherence to the bottom of the cage was scored as a soiled cage by the observers (score of 3 or 4 depending on the size of the sticky spot) and was changed immediately. Interrater reliability was assessed for each scoring day to ensure that there was consistency among the observers. Interrater reliability was determined by calculating a Kendall W value (coefficient of concordance). In the mixed strains study, the Kendall W ranged from 0.65 to 0.86. The Kendall W was >0.8 for 53% of the scoring days, >0.7 for 35% of scoring days, and >0.6 for 12% of the scoring days.

Figure 5.

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.

Figure 6.

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.

Cage change frequency and the comparison between the different groups are shown in Figure 5A and Table 2. Cage change intervals were different for various densities of mice with 5 animals per cage needing cage change around 3 to 4 wk (32 ± 13.8 d) (Figure 5A, Table 2) and cages with 1 and 2 mice per cage needing change at 6 to 7 wk (49 ± 11.1 and 45 ± 9.8 d, respectively) before the study was terminated (Figure 5A, Table 2). Cages with 3 and 4 mice per cage needed to be changed on an average of 5 to 6 wk (39 ± 14.7 and 37 ± 12.4 d, respectively) (Figure 5A, Table 2). The corresponding average drops in capacitance at the time of cage change were 9.7 ± 7.1, 16.8 ± 11.8, 22.6 ± 12.3, 32.4 ± 13.0, and 33.7 ± 19.5 arbitrary units for cages containing 1, 2, 3, 4, and 5 mice, respectively (Table 2). Some of the cages from groups with 2 to 5 animals had to be changed as early as 16 d, while some from those same groups were changed 4 to 6 wk later, indicating that cage wetness can vary based on the number of animals within that individual cage and that a set time for cage change for all cages can lead to poor sanitary conditions for animals. The first cage change in the group with 1 animal per cage was at day 33 (Table 2). The average number of days to cage change was not significantly different between single-housed cages (49 ± 11.1 d) and cages containing 2 mice (45 ± 9.8 d, P = 0.2401) (Table 1, Figure 5A). Single-housed cages had a significantly different number of average days to cage change as compared with cages containing 3, 4, and 5 mice (P = 0.0186, 0.0060, and 0.0004, respectively). These data show that single-housed cages could remain on the rack for an extensive period without reaching the visual cage change criteria. Because there was no significant difference between single-housed mice and cages containing 2 mice, the following phases did not employ the use of single-housed cages. Similarly, cages containing 4 mice were found to have no significant differences in the average number of days to cage change and/or average drop in capacitance as compared with cages of 5 mice (P = 0.2271 and P = 0.8131, respectively). Thus, cages containing 4 mice were not included in further studies. The level of variability in the cage change interval for this study is most likely due to the various strains, ages, weight, and unequal number of males and females used.

Table 2.

Average days to cage change and average drop in capacitance values over cage change interval by cage density for experiment 1 (Mixed Strains study)

Cage density	Average (d)	Minimum (d)	Maximum (d)	Average drop in capacitance
1	49 ± 11.1	33	68	9.7 ± 7.1
2	45 ± 9.8	16	50	16.8 ± 11.8
3	39 ± 14.7	15	68	22.6 ± 12.3
4	37 ± 12.4	16	50	32.4 ± 13.0
5	32 ± 13.8	15	50	33.7 ± 19.5

SW study.

We included adult retired SW breeders to determine the effect of size of the animals on the cage change interval. These animals were at least 16 wk of age at the time of the study and ∼35 to 40 g in weight. We included 28 cages of SW mice (5 cages of 5 mice/cage, 7 cages of 4 mice/cage, 7 cages of 3 mice/cage, 9 cages of 2 mice/cage) in the study. We did not have single-housed mice in this cohort of mice. Subanalysis of this group showed that cage change intervals for SW mice were 15 ± 6.6, 15 ± 7.2, 14 ± 7.6, and 10 ± 3.0 d for 2, 3, 4, and 5 mice/cage respectively (Figure 5E). The maximum days that any cage in this study reached was 24 d, with some being changed as early as 8 d due to high scores of 4 by the observers. Average cage change intervals were not significantly different between the different animal density groups.

B6 study (learning phase).

We used age- and sex-matched B6 mice at densities of 2, 3, and 5 animals per cage for this study. Cage scoring for wetness was done by visual observation. We used 10 cages of each sex and density. The study was done as 2 independent experiments to ensure reproducibility and increase ‘n’ number for the learning phase dataset. Interrater reliability using a Kendall W was assessed for each scoring day of the B6 learning phases. During this study, the Kendall W ranged from 0.79 to 0.97. The Kendall W was >0.9 for 57% of the scoring days, >0.8 for 36% of scoring days, and >0.7 for 7% of the scoring days. Due to the difficulty with identifying clear wet spots around 4 wk in the latrine area from the mixed strain study, we also measured ammonia levels in the cage as a secondary measure for cage change. Most of the cages had latrine spots in the front of the cage, and there were no differences seen between male and female animals with respect to the location of the latrine spots (data not shown). Similar to the mixed strain study, capacitance drop was seen at the latrine spots and not at the nesting areas (data not shown). No difference in cage change intervals was seen between the 2 sexes at any of the densities tested (data not shown). As a result, we pooled the data for the 2 sexes for the cage change interval analysis. The average numbers of days to cage change were 42 ± 3.0, 40 ± 4.0, and 36 ± 6.0 d for cages containing 2, 3, and 5 mice, respectively (Figure 5B, Table 3). The average numbers of days to cage change were not significantly different between cages containing 2 mice and cages containing 3 mice (P = 0.1141). The average number of days to cage change was significantly different between cages containing 2 mice and cages containing 5 mice (P = 0.0005), as well as cages containing 3 mice and cages containing 5 mice (P = 0.0189) (Figure 5B).

Table 3.

Average days to cage change and average drop in capacitance values over cage change interval by cage density for the learning phases of experiments 2 and 3 (C57BL/6J study)

Cage density	Average (d)	Minimum (d)	Maximum (d)	Average drop in capacitance
2	42 ± 3.0	37	46	24.9 ± 17.1
3	40 ± 4.0	30	45	23.3 ± 7.7
5	36 ± 6.0	12	42	27.1 ± 8.7

B6 validation study.

Mice from the learning phases 1 and 2 were used in the validation study. As a result, the mice were older (average of 15 to 16 wk of age) and heavier for this part of the study. The scoring for cage wetness was done using the BSI algorithm–generated tasks and was verified for accuracy by trained observers who provided feedback to the algorithm. A cage was only changed if it was a true positive (fair) or false negative (too dirty). Cages that were deemed too clean (false positives) were placed back on the rack and continued to be monitored. This study was also repeated 2 times with 10 cages of each sex and density per experiment. Results shown are combined from the 2 studies (Figure 5C). The cage change interval for cages with 2 animals per cage was 38 ± 8.1 d, 3 animals per cage was 29 ± 6.2 d, and 5 animals per cage was 29 ± 6.0 d (Table 4). The average number of days to cage change was not significantly different between cages containing 3 mice and cages containing 5 mice (P = 0.7775). The average number of days to cage change was significantly different between cages containing 2 mice and cages containing 3 mice (P = <0.0001), as well as cages containing 2 mice and cages containing 5 mice (P = <0.0001). No difference in cage change intervals was seen between the 2 sexes (data not shown). The average drops in capacitance for the validation phases of experiments 2 and 3 were 21.6 ± 16.0, 25.3 ± 20.3, and 45.7 ± 28.5 arbitrary units for cages containing 2, 3, and 5 mice, respectively (Table 4). Between the 2 validations studies we could see a significant increase in the accuracy of the algorithm, which improved from 64% true positives in study 1 to 81% true positives in study 2, along with a corresponding decrease in the number of cages that were considered as too dirty (false negatives) (Figure 7). These data show that the AI algorithm can be trained to improve its accuracy by augmenting it with active human feedback. As seen above in the learning phases, interrater reliability increased from the first learning phase study (mixed strains) to the second learning phase study (B6 mice). Stronger agreement among observers in the later studies leads to greater user agreement with the AI algorithm.

Table 4.

Average days to cage change and average drop in capacitance values over cage change interval by cage density for the validation phases of experiments 2 and 3 (C57BL/6J study)

Cage density	Average (d)	Minimum (d)	Maximum (d)	Average drop in capacitance
2	38 ± 8.1	13	55	21.6 ± 16.0
3	29 ± 6.2	15	41	25.3 ± 20.3
5	29 ± 6.0	19	41	45.7 ± 28.5

Figure 7.

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.

BALB/c study.

Age (6-wk-old)– and sex-matched BALB/c mice were used in this study. The scoring for cage wetness was done using the BSI algorithm alerts and was verified for accuracy by trained observers. The average cage change intervals were 40 ± 6.9, 31 ± 2.1, and 32 ± 3.5 d for 2, 3 and 5 mice per cage, respectively (Figure 5D, Table 5), and the corresponding drops in capacitance were 18.9 ± 8.4, 26.3 ± 8.8, and 52.5 ± 22.3 arbitrary units (Table 5). The average number of days to cage change was not significantly different between cages containing 3 mice and cages containing 5 mice (P = 0.9692). The average number of days to cage change was significantly different between cages containing 2 mice and cages containing 3 mice (P = 0.0021), as well as cages containing 2 mice and cages containing 5 mice (P = 0.0031).

Table 5.

Average days to cage change and average drop in capacitance values over cage change interval by cage density for experiment 4 (BALB/c study)

Cage density	Average (d)	Minimum (d)	Maximum (d)	Average drop in capacitance
2	40 ± 6.9	30	50	18.9 ± 8.4
3	31 ± 2.1	30	35	26.3 ± 8.8
5	32 ± 3.5	25	35	52.5 ± 22.3

Ammonia measurements.

Ammonia measurements were done for the 2 learning phases of the B6 mice study, as well as the BALB/c study. In the learning phase study 1, ammonia levels were measured in the center of the cage at the level of the nostrils of the mice using a Drager ammonia reader and pump. Ammonia levels detected were negligible except for one cage that had a water leak from the grommet (ammonia ∼122 ppm), and a second one that reached 62 ppm at 6 wk (data not shown). Both of those cages were changed out when the ammonia spike was detected.

In B6 learning study 2, we measured ammonia levels directly over the latrine and the nesting areas, and the data are shown in Figure 8A. Ammonia readings over the nest were generally negligible (∼1 to 2 ppm). Ammonia values recorded over the latrine were comparatively higher, but almost all measurements were below the threshold of 25 ppm considered as noxious to humans (Figure 8B). Two cages of 2 mice/cage (1 cage of males and 1 cage of females) had ammonia levels over the latrine that surpassed 50 ppm (53 ppm on day 39 and 102 ppm on day 46, respectively), and the cages were changed immediately.

Figure 8.

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).

Ammonia was also measured in B6 learning study 2 using a colorimetric ammonia reader hung from the food hopper. These readers were only placed in a subset of cages (20 cages total, 3 to 4 cages per sex and density) and were intended to collect the cumulative accumulation of ammonia in the cage. The colorimetric reader changes color gradually as the ammonia level builds up within the cage. Readers are initially yellow, indicating the presence of 0 to 1 ppm of ammonia. A change to a light green color indicates an ammonia level between 1 and 25 ppm, and a dark green color is associated with 25 to 50 ppm. Once the reader becomes blue, ammonia levels are assumed to be >50 ppm. Only 3 of 20 cages that had colorimetric readers presented with high scores during the study. The cages with elevated ammonia levels were 2 cages containing 2 females and 1 cage containing 2 males. One cage of 2 males was determined to be soiled according to the soiled bedding scale on the same day the colorimetric ammonia reader changed to a light green color (46 d). The reader in one cage of 2 females turned light green only at the end of the experiment (6 wk). Another cage of 2 females had a colorimetric reader that changed to a dark green color within 18 d of being placed on the DVC rack. Verification of the ammonia levels using the Drager ammonia detector and pump produced a measurement of 0 ppm, indicating that the colorimetric reader was potentially defective. We continued to monitor this cage using the Drager ammonia reader and pump, and the ammonia levels reached a maximum of 15 ppm. The cage was not determined to be soiled by the soiled bedding scale until 46 d after the last cage change.

Ammonia measurements for the SW study were performed similarly as in the B6 learning study 1. SW mice were found to have high levels of ammonia within a short period of time (>50 ppm) and required cage change due to elevated ammonia levels. Ammonia measurements were also performed for the BALB/c study, which were measured directly over the nest (Figure 8C) and the latrine area (Figure 8D), similar to the B6 learning study 2. During these 4 wk of the study, ammonia readings over the nest and latrine never surpassed 50 ppm. Only one cage of 5 male mice produced readings >25 ppm in the latrine spot (29 ppm at 14 d). Ammonia readings over the nest were generally negligible, except for the same cage of 5 male mice, which produced a reading of 38 ppm at 25 d.

CO₂ measurements.

Intracage CO₂ levels were measured from a subset of cages during the B6 learning phase of study 2. CO₂ levels never reached or surpassed a reading of 5,000 ppm. The average CO₂ levels by cage density are shown in Figure 9A. CO₂ was also measured for 18 cages of BALB/c mice (3 cages per sex and density). Average CO₂ levels peaked during week 1 and stayed consistently elevated for the rest of the study (Figure 9B). Eight of 18 cages had CO₂ measurements ≥5,000 ppm, which came down at later time points. The highest level of CO₂ was detected in cages of 5 mice that peaked around 18 d. The average CO₂ levels by cage density are shown in Figure 9B.

Figure 9.

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.

Weekly growth curve.

Weekly weights were performed for individual mice upon arrival (starting at 6 wk of age) until the end of the experiment for the B6 learning studies 1 and 2 (Figure 10A, B), and for the BALB/c study (Figure 10C, D). Weekly growth rates of male and female B6 mice generally paralleled the normal weekly growth rates published on The Jackson Laboratory strain catalog. Male B6 growth rates were slightly higher than the vendor values at most time points. Female B6 growth rates did not follow the same trend as for the vendor, but their growth rates were generally greater than those published by the vendor. Male and female BALB/c growth rates followed a similar pattern as the B6 mice. These data indicate that extended cage change did not influence health or welfare of the B6 or BALB/c animals.

Figure 10.

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.

Operational efficiency.

To evaluate the savings due to reduced cage changes, we tracked the cage changes on the digital rack for a period of 10 mo. The animals on the digital racks included experimental animals, animals used in the validation phases, and training animals of various densities. In the 10 mo of the evaluation, there was an ∼65% decrease of the cages changed on the rack (1,789 cages). If these cages had been changed using the standard 2 wk change out, we would have changed 5,041 cages (Table 6). The data are a little skewed toward single-housed animals due to the transgenic mice phenotyping studies undertaken by the researchers. However, comparative efficiencies for each cage density ranged between 60% and 70% on average. The cohort with 4 animals per cage was very few, and they were housed in the facility for only 5 wk before termination of a study. Therefore, the relative efficiency is much higher than for other densities. The average cage change intervals for these cages were compared against the traditional 2-wk cage change schedule at each cage density using a one-sample Wilcoxon test (except for cages containing 4 mice, which were assessed using a one-sample t test). At each cage density, the average cage change interval was significantly higher than the facility standard of 14 d (P < 0.001). Following the validation studies, we used the BSI alerts as the primary method for cage change for >300 cages of experimental and stock animals housed on the DVC racks. Evaluation of all cage changes done over 10 mo with these cages showed that the algorithm’s accuracy increased to 84%, with accuracies for individual cage densities ranging anywhere from 76% to 90% (Table 7). While the accuracy crossed 90% for the 5 animals per cage, 24% of the single-housed animals were still being scored as too clean by staff members. The percentage of cages being scored too dirty was reduced to 2%, indicating that the system has attained a high degree of accuracy with soiled cage detection. These results indicate that the algorithm achieved an average sensitivity of 96%, a specificity of 97.4%, a false negative rate of 4%, and a false positive rate of 2.6% when assessing cages over the 10-mo period.

Table 6.

Comparison of algorithm-based cage changes and traditional cage change frequency with resulting operational efficiency savings over a 10-mo period at our facility

Cage density	No. of cages	Traditional 2-wk cage change over 10 mo	Cage change based on algorithm recommendations over 10 mo	Percent decrease
1	150	3,000	903‡	70
2	38	768	279‡	64
3	34	687	333‡	52
4	4	81	15‡	81
5	25	505	259‡	49
Total	251	5,041	1,789‡	65

A one-sample Wilcoxon test was used to compare the 14 d change out compared with BSI alert-based cage change. A one-sample t test was used for 4 animals per cage comparison. ‡, P < 0.001.

Table 7.

Percentage of accuracy (fair) and inaccuracy (too clean and too dirty) of tasks generated by the AI algorithm over a 10-mo period at our facility

Cage density	Too clean (%)	Too dirty (%)	Fair (%)
1	24	0	76
2	16	0	84
3	11	1	88
4	0	20	80
5	7	2	90
Average	12	5	84

Discussion

Cage change is an essential husbandry practice that, despite multiple studies on refinement^9–16 and methodologies, continues to rely on visual observation of the bedding condition by animal care staff or on a set cage change interval as primary methods for determining the need for changing. These practices may lead to operational inefficiencies without yielding any improvement in animal welfare. Using digital cages that monitor bedding wetness 24/7 and ML algorithms we have validated BSI as a quantitative digital measure for accurate detection of soiled cages. This biomarker relies on urine saturation of bedding and, agnostic of bedding type, has the potential to be applied across facilities and husbandry conditions.

Similar to a previous study,¹⁶ we could not determine a single optimal time point when a cage needed changing, as the wetness of cages varied based on stereotypy (mice playing with water valves resulting in wet bedding and leading to an earlier change), density (a higher cage density was associated with an earlier change), and age of the animals (older, heavier animals produce more urine and require an earlier change). Although most of the bedding change events in our mixed strains study were related to the number of animals in the cage, even the cages containing the highest density (5 animals/cage) had an average cage change interval longer than the standard 2-wk cage change in our facility. These results are consistent with previous studies that indicated cages with higher densities can have extended cage change intervals with no resultant impact on animal welfare.^{3,11,16,20,21} Cages containing higher animal densities resulted in shorter average cage change intervals, ranging from 3 wk (5 mice/cage) to 4 wk (3 mice/cage), either of which was still longer than the traditional 2-wk interval for IVC caging. Most of the cages with either 2 mice or 1 mouse per cage were generally not scored as soiled by the observers, even at times closer to the end of the study due to the absence of a clear wet spot that could be seen. The capacitance drops were also very small, indicating that these cages can have lengthy cage change intervals depending on the threshold set by the animal facility. We set the upper limit at 6 wk because of the accumulation of dried feces, which was scored as soiled by the observers, including research staff.

The cages in the B6 learning phase studies had a narrower range of average days to cage change as compared with those from the mixed strains study. This is probably due to the harmonization of age- and sex-matched mice used in the B6 and BALB/c studies. It is also possible that the variability of average cage change interval between the learning phases was due to different levels of agreement among the observers scoring the bedding. During the learning phase of the mixed strains study, interrater reliability calculations determined that agreement among observers was high for the first half of the study (Kendall W was between 0.7 and 0.86). However, this decreased to a moderate level of agreement toward the end of the study (Kendall W was calculated between 0.6 and 0.7 for 2 scoring days). Interrater reliability increased for the learning phases of the B6 studies, as it was calculated to be very high for 88% of the scoring days. Observer agreement was more succinct for the B6 learning phases, and so the reduced agreement among observers in the mixed strains study may have resulted in the wider range of average cage change intervals. The average days to cage change for the B6 validation studies was slightly lower than that of the learning phase. This is likely because the mice were several weeks older at the start of the validation studies (15 to 16 wk) as compared with the learning phase studies (6 wk). Increased age may be associated with increased weight gain and urine production, leading to pronounced soiling within a shorter time frame.²² It is also reasonable to speculate that the digital measure identifies the criteria of a soiled cage based on urination better and earlier than that achieved through human visualization. The average cage change intervals for the BALB/c study were very similar to those of the B6 validation studies. It has been previously shown that BALB/c mice housed in digital caging have comparable average drops in the BSI as B6 mice, whereas CD-1 mice had larger capacitance drops.¹⁸ The average days to cage change in both the B6 and BALB/c studies generally decreased with increasing cage density as in the mixed strains study, and cages of 5 mice still had average cage change intervals (3 to 4 wk) greater than those of the traditional 2-wk interval. Urine volume has been reported to be higher in B6 and BALB/c females compared with males,²³ and therefore we expected to see a difference in cage soiling between the 2 sexes. However, we did not find any difference in cage change intervals between the 2 sexes. The previous study²³ used metabolic caging for 5 d, and metabolic caging has been shown to increase stress in the animals, which could have led to higher adrenal corticosteroid production and thereby higher urine volumes. In our study, the animals were housed as stable pairs between 7 and 16 wk, which could have attenuated any stress-related effect.

On the other hand, retired SW breeders, which were the heaviest of the mice used in the study, had bedding identified as soiled earlier, probably due to the higher volume of urine produced per gram of body weight, resulting in visible wet spots, and required a cage change as early as 8 d. Almost all of the cages (including those with 2 and 3 animals per cage) were changed at or before the standard 2-wk change. We attribute the clear wet spots seen in this study, compared with the B6 or BALB/c studies, to the urine volume exceeding the absorption capacity of the cellulose bedding within a short time. Based on these results, we think that the BSI algorithm, which depends on urine volume and, thereby, bedding wetness, can adapt to any changes that might exist due to strain variation and/or urine production to adjust the cage change frequencies. We did not examine other strains of mice, such as CD-1 mice, to see if outbred mice have differences in cage change intervals. CD-1 mice were shown to have higher BSI capacitance drops,¹⁸ and therefore we would expect a shorter cage change interval than that of B6 or BALB/c mice.

Measurement of intracage ammonia levels has been used as a surrogate measure for degree of cage soiling. Urine-soaked bedding can act as a substrate for the urease-producing bacteria from the gut to break down urine, resulting in ammonia accumulation in the cage. The amount of ammonia production can depend on the type of caging (static compared with IVC),^13,20,21,24 number of air changes within the cage, type of bedding,²⁵ and the number of mice in the cage.^11,13,24 This is evident from the results of previous studies where ammonia levels have been shown to vary from 5 ppm,^5,26 65 ppm,³ and 200 to 400 ppm¹³ to 700 ppm²⁷ within a period of 7 to 17 d. While the time-weighted 8-h exposure limit for humans is 25 ppm, the minimum level of ammonia exposure is not defined for mice. The effect of ammonia exposure on nasal pathology has been disputed, with some studies showing nasal irritation and inflammation at 25 ppm¹⁶ and others documenting mild to low lesions at 50 ppm,¹³ whereas others found lesions only when ammonia levels were 3-fold the value of ammonia’s RD₅₀ (or exposure concentration producing a 50% respiratory rate decrease), a value estimated at ∼900 ppm.²⁸ Severe nasal lesions have been shown to increase with a higher number of animals in the cage, especially trio-breeders,^13,24 presumably due to higher urine production and inability of the caging type to efficiently scavenge the ammonia out of the cage. For this study, we used 50 ppm as our criterion for a cage change. Only 6 cages had ammonia levels above the human time-weighted average of 25 ppm, whereas the rest were <25 ppm. We attribute this to the high air exchanges within the IVC (75 ACH) and the ability of the cellulose bedding to efficiently absorb the urine, resulting in lack of substrate for ammonia production.

Most studies assessing extended cage change intervals have measured intracage microenvironment, such as ammonia, as a marker for cage change. One study showed¹⁶ that mean intracage ammonia levels over the 28-d period were >25 ppm for cages of 3 and 5 female mice, and >50 ppm for cages of 5 male mice. Our data contrast with these previously found results, as only a small number of cages in our study (4/116) had surpassed an intracage ammonia level of 50 ppm and even fewer cages (2/116) had ammonia measurements between 25 and 50 ppm. The likely reason for these discrepancies is the difference in husbandry conditions. We used ⅛-in. pelleted cellulose bedding and the intracage ventilation rate of 75 ACH. Vogelweid and colleagues¹⁶ used irradiated ⅛-in. corncob bedding and a ventilation rate of 68 ACH. It has been found that cellulose bedding leads to lower intracage ammonia levels compared with corncob bedding.²⁹ Also, a higher ACH would likely lead to a more rapid clearance of ammonia and evaporation of urine in the cage, resulting in a lower ammonia buildup in our study. Vogelweid and colleagues¹⁶ used CD-1 mice, while our study involved B6 and BALB/c mice. CD-1 mice fed the same diets as B6 mice have been found to have a higher urine pH, and a higher urine pH is associated with a higher concentration of ammonia.⁴ Further, the volume of bedding has been shown to influence soiling and ammonia buildup.⁵ In our study we used 350 mL of ⅛-in. cellulose pelleted bedding, as we have determined that this volume of bedding provides adequate coverage of the cage floor and facilitates efficient drying (unpublished observations). The absorption capacity of different types and amounts of bedding can vary.²⁹ Accordingly, caution must be used in translating our cage change intervals to circumstances that do not use the same type of bedding. Future studies comparing the different types of bedding and use of BSI alerts are planned. Other objective measures of cage soiling used during extended cage change intervals, such as evaluation of microbial load, were not investigated in our study, but would be beneficial in future experiments.

CO₂ levels were also measured in a subset of cages in both the B6 learning phase of study 2 and the BALB/c study to ensure that the presence of this gas was within an appropriate range. The current 8-h time-weighted average exposure duration for humans is 5,000 ppm,³⁰ which is considered the human occupational exposure limit for CO₂. There is no current CO₂ threshold for cage change requirements in mice, although previous studies³¹ have suggested that a measurement of 15,000 ppm should be considered as the experimental limit. CO₂ levels of 3,000³² and 5,000 ppm¹³ have been reported for CD-1 and B6 and BALB/c mice, respectively, and no effect on animal health was seen in these studies. CO₂ levels never exceeded 5,000 ppm for the B6 mice, and approximately half of the BALB/c cages tested reached a CO₂ measurement ≥5,000 ppm (the detector limit on the CO₂ reader was 5,000 ppm). CO₂ levels ≥5,000 ppm did not correlate with cage change intervals or seem to affect the growth rate for male and female B6 and BALB/c mice.

Cage change can be stressful to mice and affect the rigor and reproducibility of research.^7,2,7,3,8 Loss of pheromonal cues due to bedding change can cause mice to fight to reestablish territory, leading to loss of valuable animals, especially long-term studies involving strains that have been previously shown to exhibit aggressive behavior.^33–35 Fighting events were seen in the studies conducted only in the first week of cohousing. These were seen in both male and female mice, and some animals had to be humanely euthanized due to severe fight wounds. Surprisingly with the extended cage changes the fighting events were drastically reduced and even absent in the mice. At the time of publication, there are 2 cages of B6 male mice that are still on the racks after >12 mo with no reports of fighting or wounds. We attribute this to the animals being less stressed with the extended cage changes, which allows them to establish stable housing groups. We did not measure serum or fecal corticosteroids in the study to examine the effect of extended cage change intervals on animal stress. It would be beneficial in future studies to measure the cortisol levels of mice. The digital sensor board also detects disruptions in circadian metrics locomotion, sleep pattern changes, day and night activity, diurnal activity, response to lights on and off phases, acrophase and activity onset, and regularity disruption index.^18,36,37 In our study, there was no discernible effect on the circadian metrics (data not shown), indicating that animal welfare was not affected by the extended cage change interval.

The AI algorithm functions through an iterative process of user feedback that enhances the accuracy of the algorithm as shown by the increase in accuracy from 64% to 81% in the second study. This has been further improved to 90% for >5 mice per cage. The accuracy was found to be dependent on cage density with staff scoring too clean at a higher rate (24%) for the single-housed cages. We believe this difference to be due to the inability of visual observations to detect a clear wet spot, which is the most commonly used parameter for spot changes in many facilities. The accuracy of the bedding change tasks has reduced the incidence of too dirty cages (false negatives) to <2%, indicating that the AI algorithm accuracy for predicting a cage change is very high.

The adoption of novel technologies similar to those described in this study is low due to the costs and uncertainty surrounding return on investment. The digital caging system used in the study costs ∼30% more than a standard IVC rack and requires annual license subscriptions for the AI algorithms. Our efficiency gains for cage change and cage wash at 10 mo of implementation are shown in Table 8. The results show that we saved 34 h/wk of labor, which equates to 1 full-time employee savings (our staff work 32.5 h/wk), which equates to $64,767 (salary + 81% fringe) savings. Every time a new cage is changed, we incur a cost of $0.29 (bedding + enrichment), and changing out 65% fewer cages equates to a savings of $314.20/wk or $17,741.62/y. The average cost of one rack washer cycle is $19.64. We calculated that we need to run 10 fewer loads per week, which translates to $10,212.80 in savings on electricity, water, steam, and detergent used in the facility. Also reducing the number of cage changes leads to a reduction in 0.1 ton of solid waste per week being generated in the facility, which is a cost saving of $15 per week for waste disposal. Thus, the total cost savings was calculated as $93,501.42 per year for a facility with 3,500 cages.

Table 8.

Operational efficiency calculations

Function	Weekly quantity		Throughput (in minutes)	Weekly labor
	14-d baseline	DVC		14-d baseline (in hours)	DVC (in hours)
Clean supplies to room	21 carts	11 carts	0.1/cart	0.03	0.01
Cage change	1,750 cages	770 cages	1.5/cage	43.75	19.25
Soiled cages to hallway	21 carts	11 carts	0.1	0.03	0.013
Moved soiled cages to cage wash	21 carts	11 carts	2.3/cart	0.80	0.4
Prep cages for presentation rack	21 carts	11 carts	8.1/cart	2.84	1.5
Load rack into rack washer	21 carts	11 carts	1.3/cart	0.46	0.2
Unload rack washer	21 carts	11 carts	1.3/cart	0.46	0.2
Remove cages from presentation rack	21 carts	11 carts	2.5/cart	0.88	0.5
Bedding filling	1,750 cages	770 cages	0.2/cage	5.83	2.56
Enrichment fill	1,750 cages	770 cages	0.2/cage	5.83	2.56
Move cages into vivarium	21 carts	11 carts	1.7/cart	0.60	0.3
Total time taken (h)				61.51	27.49

A time in motion study was conducted to calculate the accurate operational savings using the digital caging system.

In conclusion, we have demonstrated that the BSI, along with the unique AI algorithm used by a digital caging system, can be used as a reliable and quantitative biomarker for cage change. We show that using this method can lead to 50% or greater savings in the facility, providing an opportunity for the animal care technicians to focus on other tasks such as animal health or research support. Novel home cage monitoring technologies such as the one discussed in this study can revolutionize animal husbandry operations and improve animal welfare and the rigor and reproducibility of research.

Conflict of Interest

Drs. Joseph Collins, Bhupinder Singh, Michael Zwick, and Jeetendra Eswaraka are employees of Rutgers, The State University of New Jersey, and do not have any conflicts of interest to declare. The technology was purchased using Rutgers institutional funds. Mara Rigamonti, Giorgio Rosati, and Cristian Uridales Garcia are employees of Tecniplast, who manufactures the technology that was used in the study.

Funding

This work was internally funded.