2.1. Experiment
An animal experiment was carried out during the winter in Spain in 2019 to predict the risk condition of ammonia levels and correlate them with the health risk of broiler chickens. The experiment was approved by the Animal Ethics Committee N2018/VSC/PEA/0067 on 16 April 2018. This study was part of a doctoral thesis.
Two identical rooms (Room 1 and Room 2) were used in this test, measuring 13.2 m × 5.95 m, totaling approximately 70 m2 for each room. An automated temperature control system was installed (DNP Climate Controller, Exafan, Spain), which controlled ventilation rates in accordance with commercial temperature recommendations. The room temperature was gradually decreased from 32 °C (day 1) to 19 °C (day 42). Temperature was controlled using the temperature control sensor and recorded along with relative humidity every 10 min using a data logger (HOBO U12, Onsetcomp, Bourne, MA, USA). Furthermore, each room was equipped with an electrochemical NH3 sensor (DOL 53, Dräger, Germany). Room 1 was programmed to maintain a maximum of 10 ppm of NH3, while Room 2 was programmed to maintain a maximum of 20 ppm. These environmental conditions (ammonia concentration) were programmed to be maintained from the fourth week onwards, that is, in the second half of the production cycle, when ammonia levels tend to be higher within broiler production systems. A propane heater was used to maintain an adequate room temperature.
The difference in ammonia concentration between rooms was obtained through ventilation. However, the different ventilation rate was not significant enough to cause large temperature differences between the rooms. For the initial four weeks, both rooms maintained a comparable temperature. Yet, starting from week 5, Room 1 elevated ventilation rates led to an average temperature reduction of 1.7 °C. It is worth noting that this 1.7 °C difference is an overall average across all weeks studied. In the crucial first two weeks, the difference did not surpass 0.3 °C, making the temperature conditions practically identical and unlikely to impact the birds’ performance. Relative humidity remained consistent between the two rooms.
The experiment was carried out during the winter period, when gas concentrations were expected to be higher due to lower ventilation rates. To ensure that the desired concentrations were achieved during a relevant part of the poultry production period, it was decided to apply a urea solution to the litter. The dosage was always 0.21 L/m2 of urea solution, with a concentration of 187.5 g/L on day 32 of the rearing cycle and 93.75 g/L on days 39, 51, and 56.
Two commercial lines of broiler chicken were used, one with fast-growing (Ross
®, slaughter age 42 days) and another with slow-growing (Hubbard
®, slaughter age 63 days). All slow-growing birds were housed at a density of 32 kg m
−2. Fast-growing birds were housed in two different housing densities: low stocking density with a final stocking density of 16 kg/m
2 and high density with a final stocking density of 32 kg m
−2. A total of 1250 birds were used in this experiment, 450 of which were fast-growing birds and 800 of which were slow-growing birds. A total of 102 birds were allocated into 6 boxes designated for the slow-growth treatment at high stock density, resulting in 17 birds per box. Similarly, another 102 birds were placed in 6 boxes for the fast-growth treatment at high stock density, also with 17 birds per box. Lastly, an additional 102 birds were distributed across 6 boxes for the fast-growth treatment, housed at high stock density, maintaining the ratio of 17 birds per box. The remaining birds were positioned in the external corridors of the boxes to simulate real-field breeding conditions. Each box had 3 nipple drinkers and a manual feeder. The dimensions of the high-density boxes were 1 × 1.3 m
2, whereas the low-density boxes, stocking the same number of animals, were more spacious, with dimensions of 2 × 1.3 m
2 (
Figure 1).
All animals were fed with Nanta brand starter commercial feed for broiler chickens from day 1 to day 18 of the trial. From day 18 until the end of the trial, each strain was fed with different feeds. The feed transition was done gradually over two days, mixing the two feeds to enhance adaptation. For feeding fast-growing breed (treatments HD and LD), two commercial feeds for fast growth (Nanta A80) and slow growth (Nanta A32) were used.
Previously, centesimal chemical analysis was conducted. The composition obtained for the fast-growing breed was crude protein: 18%, crude fat: 1.9%, crude fiber: 3.1%, ash: 6.2%, calcium: 1.00%, phosphorus: 0.67%, sodium: 0.16%, methionine: 0.33%, lysine: 0.94%. for the slow-growing breed: crude protein: 15.7%, crude fat: 3.9%, crude fiber: 2.6%, ash: 5.8%, calcium: 1.00%, phosphorus: 0.67%, sodium: 0.16%, methionine: 0.28%, lysine: 0.82%.
All birds randomly stocked in the boxes were weighed (kilograms) weekly, and their respective feed consumption was calculated. Average daily weight gain (ADG) per bird was calculated for each week of rearing, and average daily weight gain accumulated over the entire study period (42 days for fast-growing birds and 63 for slow-growing birds). Feed conversion (FCR) was also obtained for each week and for accumulated period, dividing the amount of food consumed by each pen by the weight gain of all birds present in it
Table 1. The data in
Table 1 are from a previously published study [
10].
Room ammonia concentration data was collected by installing electrochemical sensors (DOL 53, Dräger, Germany) in each room. Assessments were carried out 24 h a day, every day of the week. In addition to the productive character determinations, during the development of the experiment, animals were sacrificed, and samples were taken at four moments: day 0, day 21, day 42, and day 63 (day 63 only for animals from the slow-growing lineage). All the sacrificed birds were previously stunned by an electric shock.
In the first sampling, on day 0 of the study, 30 animals from each lineage were randomly sacrificed before distribution into the boxes. In the second and third sampling, days 21 and 42 of the experiment, respectively, 5 animals were sampled from each pen, resulting in a total of 180 animals for each sampling day. After sacrifice, the animals were necropsied.
The indicators were assessed following the necropsy protocol, conducted by the veterinarians in our research team. Necropsy procedures were consistently deliberated upon by the same researchers throughout the study.
The fourth and final sampling was carried out on day 63 of the experiment with the same procedure as the previous two but involving only animals from the slow-growing lineage, since animals from the fast-growing lineage have a commercial production cycle of 42 days.
In the last two collections, at 42 days for fast-growing birds and at 63 days for fast-growing birds, the sampled animals were inspected for symptoms related to prolonged exposure to NH3. These exams aimed to find epidermal lesions on the legs and injuries to the eyes and respiratory tract due to this irritating gas.
2.2. Data Mining Approach
The following data analysis steps were performed: data selection, pre-processing, transformation, mining, analysis, and interpretation of results.
The data preprocessing stage covers data understanding and data preparation, which includes standardizing nomenclatures, cleaning the raw data in the spreadsheet, and dividing the database in the Weka software (version 3.8.4) [
11].
For training and testing, the model employed a “stratified remove folds” filter to bifurcate the dataset into training and testing subsets. Pre-processing also included the discretization of attributes into classes that reduces and simplifies the data, making learning faster and the results denser, according to the proposed methodologies [
11,
12].
In the processing stage, the data set was analyzed by applying predictive classification models for training (75% of the data set with 17,062 instances) and for validation of the model with the test set (25% of the data set with 5688 instances).
The classification algorithms, decision tree (J48), Sequential Minimal Optimization (SMO), Naive Bayes and Multilayer Perceptron were applied to the training and test data sets to build a rule model for predicting ammonia risk levels in broiler chickens. The cross-validation technique (test mode: 10-fold cross-validation) was used to parameterize the analysis in all models. The number of attributes used in the modeling was seven, including “housing_condition”, “age_week”, “T-hobo (temperature of hobo)”, “UR%”, “Vent (ventilation)”, “NH3_ppm” and the response attribute “Ammonia_concentration_risk”, with a total of 5688 instances.
The study developed a machine learning model to predict the risk condition of ammonia concentration in the production of chickens of slow and fast-growing breeds with low and high production densities. The study also compared the performance of all algorithms with respect to their prediction abilities and model quality. When evaluating the models, the data was divided into training and testing subsets, then the results were compared by the performance metrics of the algorithms. The flowchart used to identify the best test algorithm is shown in
Figure 2 (adapted from [
13]).
The criterion used to discretize the classes of the response attribute (target) “ammonia concentration risk condition” included ammonia concentration levels in five classes described in
Table 2 [
9,
14,
15].