VOCALIZATION DATA MINING FOR ESTIMATING SWINE STRESS CONDITIONS

This study aimed to identify differences in swine vocalization pattern according to animal gender and different stress conditions. A total of 150 barrow males and 150 females (Dalland genetic strain), aged 100 days, were used in the experiment. Pigs were exposed to different stressful situations: thirst (no access to water), hunger (no access to food), and thermal stress (THI exceeding 74). For the control treatment, animals were kept under a comfort situation (animals with full access to food and water, with environmental THI lower than 70). Acoustic signals were recorded every 30 minutes, totaling six samples for each stress situation. Afterwards, the audios were analyzed by Praat 5.1.19 software, generating a sound spectrum. For determination of stress conditions, data were processed by WEKA 3.5 software, using the decision tree algorithm C4.5, known as J48 in the software environment, considering cross-validation with samples of 10% (10-fold cross-validation). According to the Decision Tree, the acoustic most important attribute for the classification of stress conditions was sound Intensity (root node). It was not possible to identify, using the tested attributes, the animal gender by vocal register. A decision tree was generated for recognition of situations of swine hunger, thirst, and heat stress from records of sound intensity, Pitch frequency, and Formant 1.


INTRODUCTION
Pork is the most consumed meat in the world, and the production systems have been searching for an automated way to estimate animal welfare in the livestock, since this is an important demand from the overseas market.Among ways to audit animal welfare, there is monitoring of their behavior and social interactions (BAPTISTA et al., 2011;SPINKA, 2012), observation of health and rearing conditions (BAPTISTA et al., 2011;CALDARA et al., 2012), and monitoring of their vocalization expressed during production phases (MARX et al., 2003;SAMPAIO et al., 2007;NÄÄS et al., 2008).
Vocalization is an objective and noninvasive tool that may evidence lack of herd welfare based on the release of distinct vocal patterns (NÄÄS et al., 2008).MARX et al. (2003) characterized different types of swine sounds and identified diverse parameters of emitted energy, frequency, and duration.Animal vocalization is, thus, an expression of its current condition, and may occur spontaneously or, yet, be the result of an event, such as hunger and pain; for this reason, it became an important tool to assess animal welfare (DÜPJAN et al., 2008).There is evidence that each animal has individual vocalization characteristics.Animals use vocalization as a means of communication among individuals of the same species (GRANDIN, 1998).Female pigs express a composite frequency of grunts as identified by SCHRADER & TODT (1998).
For sound data analysis, it is important to use intelligent techniques of data processing, which can help to identify stressful situations.These techniques are able to transform data into relevant information and knowledge, providing support for decision making.The techniques of data mining have shown promising results for knowledge increase on animal production (NÄÄS et al., 2008;CORDEIRO et al., 2012).
This study aimed to identify differences in swine vocalization pattern according to gender and exposure to diverse stress conditions.

MATERIAL AND METHODS
The experiment was conducted at a commercial swine farm in April 2012, located in the northwest of Rio Grande do Sul State, Brazil (27°55'23'' S, 53°02'21'' W, mean altitude of 385 m).According to Köeppen classification, the climate in the region is humid subtropical.
We used 150 surgically castrated males and 150 females (Dalland ® genetic strain), aged 100 days (growth phase), distributed in 10 pens.Pigs were housed in sheds of east-west solar orientation, ceiling height of 3.0 m, side walls of 1 m height with curtains at side openings, and fiber cement tiles (0.006 m).Sheds had 10 pens of 30 m² each, with 30 animals allocated in each pen.The floor was built in concrete with a running water channel of 0.05 m at the end of each pen.Nipple drinkers were also placed at the end (three per pen).The automatic feeder had the capacity to feed six pigs.At the shed outside, there was a bush line with the aim to provide shade on the shed upper part (roof).
Environmental data (temperature and relative humidity) were recorded every 30 min using a digital thermo-hygrometer.Values of wet bulb temperature (WBT, °C) were calculated using mean values of dry bulb temperature (DBT, °C) and relative humidity (RH, %), using Psicrom ® software (RORIZ, 2003).The Temperature-Humidity Index (THI) was calculated using eq.(1), as described by ROLLER & GOLDMAN (1969): Animals were exposed to different stressful managements, such as thirst (no access to water), hunger (no access to fodder), and thermal stress (TS, THI≥74); welfare conditions (WC, animals had continuous access to fodder and water, and were reared under THI<73).
Stressful situations by hunger and thirst were determined after a period of 11 hours of fasting, for both cases; data recording began after this period.
Acoustic signals were recorded every 30 min, totaling six samples for each stress exposure.Acoustic recording was performed with a unidirectional microphone Marantz ® PMD 660.The microphone was positioned at 1 m above the pigs, at the pen central area, and connected to a recorder.Recorded signals were digitalized using a frequency of up to 44,100 Hz, during 3 min.
After finishing the swine vocalization recording, sound signals were downloaded to a computer.Each pen provided the record of a sequential range of "screams and grunts" (vocalizations).Subsequently, the audios were analyzed using Praat ® 5.1.19software, and the acoustic parameters were extracted using Fourier transform, generating a sound spectrum.The attributes generated by the software revealed the signal energy (Pa²*s), maximum and minimum amplitudes (Pa), pitch frequency (Hz), sound intensity (dB), and four levels of formants (F1, F2, F3, and F4) that are also called harmonic (Hz).
To determine stress conditions, data were processed using WEKA ® (3.5) software, with J48 algorithm, and considering cross-validation with samples of 10% (10-fold cross-validation).The attributes used in data mining processing are described in Table 1, and the target attribute was the stress condition.

RESULTS AND DISCUSSION
The Decision Tree (Figure 1), obtained by using the J48 algorithm, had 77.66% accuracy and Kappa of 0.62.According to the Decision Tree result, the most important acoustic attribute for classification of the studied stress conditions was sound intensity (root node).
For noise intensity lower than or equal to 73.87 dB, there was an indication that animals were under a welfare situation; on the other hand, for sound intensity ranging from more than 73.87 to 80.18 dB, we understood that animals were thirsty.For noise intensity higher than 80.18 dB, it was necessary to check the pitch frequency.Thus, for pitch frequency higher than 276.71Hz, there was an indication that animals were under heat stress.For pitch frequency ranging from 276.71 to 212.87 Hz, we estimated that animals were hungry.From pitch frequency higher than 212.87 Hz, Formant 1 needed to be verified.In this case, when Formant 1 was greater than 1,066.4Hz, there was an indication that animals were thirsty, while for Formant 1 lower than or equal to 1,066.4Hz, animals may have been hungry.
In a study on piglets during farrowing, RISI (2010) compared acoustic characteristics of piglet vocalizations between arthritis and health.Healthy piglets emitted vocalizations with greater intensity (79.76 dB) than arthritic animals (78.15 dB).This feature is interpreted by sound intensity (dB); the explanation for its occurrence is also related to the animal performance in stimulating air expulsion through the respiratory tract, promoting the vibration of vocal cords (RISI, 2010).The use of sound intensity as a determinant factor for welfare estimate was efficient in a study carried out by NÄÄS et al. (2008).FIGURE 1. Classification of the studied stress situations using the J48 algorithm.
Piglets respond with high sound intensity cries during pain (MARX et al., 2003;LEIDIG et al., 2009), what is easily identified by humans, facilitating the sound identification for only one attribute.Responses to hunger and cold sensations during extreme exposures that may cause pain are more difficult to identify, so more attributes are needed to be used in the identification process, as indicated by SCHRADER & TODT (1998).SAMPAIO et al. (2007) found that the noise emitted by pigs under thermal discomfort tends to increase with raised heat stress.This agrees with our findings and, also, with the pitch frequency, which determines the height of sound waves, enabling identification of heat stress.
The identification of hunger and thirst by Formant 1 was also found by SCHRADER & TODT (1998) and DÜPJAN et al. (2008).The authors indicated that these vocal expressions are different from those related to other types of social stress, fear, and frustration.However, several authors highlight the difficulty of classifying these specific conditions, since these occur concomitantly and need more than one attribute, or combination of them, to more precisely identify the stress type (SCHRADER & TODT, 1998;MARX et al., 2003;CORDEIRO et al., 2012).
It was not possible to identify gender by sound intensity, or to use the other studied attributes.This is different from studies on other animal species, which show that animal gender may be identified by sound signals (BALSBY & ADAMS, 2011;MATROSOVA et al., 2011;GOGOLEVA et al., 2012).

CONCLUSIONS
A decision tree was generated to identify situations of swine hunger, thirst, and heat stress, using records of sound intensity, pitch frequency, and formant 1.
Considering the studied attributes, it was not possible to identify animal gender by their vocalization.

TABLE 1 .
Attributes used for data mining.