The SAH method was developed in a uniform multi-stage process for all included animal species and categories (Figure 1). These were dairy cows, veal calves, sheep, goats, sows, fattening pigs, broilers and laying hens (the latter two hereinafter referred to as poultry).

First, a context analysis was conducted, consisting of a review of scientific literature, data availability and the potential of existing PLF technologies. For better readability, the data and PLF review will be discussed in separate sections. In order to determine the current state of research regarding existing indicators for assessing animal health and welfare, a systematic literature research was carried out for each animal species and category involved. For literature identification, search queries were made in the scientific databases PubMED, Web of Science, Scopus, CAB Direct and Science Direct, and identified publications were analyzed based on the PRISMA guidelines (17). Detailed methods of the literature review for the different animal species are described in Gebhardt-Henrich and Schlapbach (18), Lutz et al. (19), Minning et al. (20), and Zufferey et al. (21). Subsequently, a preliminary set of promising indicators was compiled and discussed with stakeholders. The number of external stakeholders per workshop varied between 16 and 28. For cattle (n_workshop1 = 22; n_ws2 = 28), pigs (n_ws1 = 18; n_ws2 = 23) and poultry (n_ws1 = 16; n_ws2 = 23) there were two rounds of stakeholder workshops each, one before the farm visits and one afterwards. With a few exceptions, most of the participants and organizations in the second workshop were the same as in the first workshop. For sheep and goats there was only one workshop (n_ws = 17). Stakeholders from the relevant sectors participated in the workshops, including people from industry associations, producers, animal transport companies, livestock marketers, processors, retailers, veterinarians, universities, Swiss Veterinary Service and Swiss Animal Welfare Protection. Based on stakeholder feedback and relevance regarding Swiss production conditions, suitable indicators were determined and assigned to four categories, based on the Swiss Animal Welfare Act (2): (i) “Animal health”, (ii) “Husbandry and feeding”, (iii) “Appropriate behavior” and (iv) “Freedom from pain, suffering, harm and anxiety”.

To assess the feasibility of the method, farm visits were carried out in a next step, and the developed SAH method was applied in the field with an elaborated specific set of indicators per animal category. Based on a convenience sample, a total of 35 dairy cattle, 31 veal calf, 27 pig and 17 poultry farms were visited and data on SAH indicators were collected. For sheep and goats, no farm visits were conducted, as the context analysis had indicated that the development of a data-driven method for these species was not feasible at this stage. During the farm visits, the WQ protocol was also carried out to assess the animal health and welfare status of a given farm, which later served as reference value for comparison with results based on SAH indicators. For laying hens, the MTool (22) was used as reference method, because this tool is more widely used for welfare assessment in laying hens in Switzerland than the WQ protocol. Following the farm visits, a second round of stakeholder workshops was held. These workshops were conducted online, and feedback was collected using Mentimeter (23), a web-based tool to anonymously capture real-time participant feedback and stimulate discussion. Results of the field studies were presented and discussed, and the method was further refined (e.g., regarding categorization or importance of individual indicators). Finally, outcomes from the SAH and WQ methods were compared and assessed for consistency using descriptive and inferential statistical methods. Within the SAH method, a dual approach was chosen in order to account for dependencies and restrictions in the scoring and integration of the individual indicators: On the one hand, a purely descriptive statistical analysis of the scores was carried out in view of benchmarking; on the other hand, individual indicators were scored by means of thresholds with target and alarm values. For the benchmarking, indicator values were standardized by means of Z-transformation and compared to the other farms serving as reference population. The 25 and 75% quartile served as cut-offs and scores were categorized into “lowest 25%”, midfield (25–75%) or “best 25%” (>75%). The assessment against target and alert values was mainly based on thresholds retrieved from the literature (24–26). If literature data were insufficient or not applicable for Swiss production systems, thresholds were adapted or defined by informal expert consultation.

In the data context analysis, existing databases and information systems were assessed within the framework of animal health and welfare and examined for the availability of data for possible indicators. Data sources containing potential indicators can be allocated into three categories: (i) public data sources, where content and access are regulated under public law, (ii) private data sources that are owned by farmers and/or private organizations, and (iii) so-called “on-farm” data, which refers to data that do not yet exist in a database and must be collected on the farm (Table 1). Consequently, “data-based indicators” were defined as indicators for which data can be retrieved from existing public or private data sources and “on-farm indicators” as indicators for which data must be obtained “on-farm”.

The most relevant public Swiss data sources are the animal movement database (TVD), the agricultural policy information system (AGIS), the laboratory information system (ALIS, recently renamed to ARES), the information system for control data (ACONTROL) which contains records on animal welfare inspections, the information system on antibiotics in veterinary medicine (IS-ABV), and the meat inspection database (FLEKO). The IS-ABV was introduced in 2019 and contains the antibiotic prescriptions according to the Therapeutic Products Act (27), with information on oral group therapies, individual animal treatments as well as the dispensing of veterinary medicinal products in stock. The updated version of FLEKO was launched in 2020 and contains information on ante-mortem inspection of live animals and condemnation of partial and whole carcasses. As IS-ABV and FLEKO were only introduced shortly before the start of the SAH project, they could not be considered in the analyses because of data quality issues. Private data sources included for example data from breeding associations, abattoir data on carcass evaluations, data from animal health services, records in electronic treatment journals, performance data, data from producer organizations and integrators (especially in broiler and veal calf production), parts of the mandatory milk quality analyses, as well as other records from the farmers or their veterinarians. The “on-farm” data include specific data on animal welfare evaluation such as qualitative behavior assessment or indicators for “Freedom from pain, suffering, harm and anxiety”.

A systematic search on available PLF technologies was carried out focusing on automated technologies for real-time monitoring of animals. These systems usually consist of a sensor and smart software, which collect and analyze data of animals and/or their environment (12). The review focused specifically on PLF technologies in the animal health and welfare context that could be used to simplify and improve data collection of on-farm indicators. First, it was determined which PLF systems are commercially available, followed by reviewing which systems or single parameters have been scientifically validated (28). The overviews for the different animal species were based on the “4D4F Technology Warehouse” format (29) structured by the main purpose of the systems (e.g., feeding, activity, lameness, etc.). Subsequently, further compilations of PLF systems (commercially available or in development) were produced to capture specific SAH indicators (30). To examine the potential of PLF systems for health and welfare assessment in livestock, obstacles and challenges in the development of PLF technologies were identified via a systematic literature search based on the PRISMA guidelines (17), and a framework for the automated detection of health- and general welfare-related issues was developed as a guideline for future studies (31).

To address the aim of implementing risk-based animal welfare inspections, a purely data-driven approach was chosen based on public databases with high data availability and coverage. Different databases and national registries (i.e., TVD, AGIS, ALIS, ACONTROL) containing information on farm demographics, animal identification, traceability, diagnostic results, and animal welfare inspections were pseudonymised and interconnected. Exploratory data analysis combined with machine learning algorithms [Support vector machine (SVM), Logistic regression, Random forest (RF) and Artificial Neural Network (ANN)] were used to identify relevant proxies for animal health and welfare. Input data (so-called features of the holding) were fed to machine learning algorithms that built a model of classification. Features were derived from extracts of the national databases cited above, like population and mortality by age, participation to additional welfare programs or labels, and geographical information. Features were preprocessed to reduce dimensionality where needed, to transform them into categorical values and standardized. The results of previous on-farm welfare inspections recorded in ACONTROL were used to calibrate a binary welfare index. Farms where 50% or more of the control points were found to be deficient were considered to be “at risk” for animal welfare violations. Different methods, including logistic regression, random forests or artificial neural networks were applied and performances of the different algorithms were assessed through a k-fold method. Meta-parameters of the methods were chosen to maximize performances. Performance was assessed using the F₁-score and similar metrics between the ground truth and the predictions of the algorithms. The training and testing of the algorithms were repeated 1,000 times to be able to assess training variability. Further details on the methodology for the risk index will be described in a separate publication.

The literature review identified a large number of potential indicators for the respective animal species and categories. For dairy cows, broilers, sheep and goats, detailed results of the reviews have been published (18–21). Some of the identified indicators are already well established and widely used in practice (e.g., mortality rate, lameness score, somatic cell count), while others lack reliability or practicability and/or require further research and validation (e.g., pain assessment, eye condition). Moreover, it was found that some indicators have only limited suitability for Swiss production systems, for example due to smaller herd sizes (when based on proportion of animals) or different welfare regulations (e.g., prohibition of tail docking).

All SAH indicator sets developed for the different livestock categories consisted of data-based indicators (indicators for which data could be retrieved from existing public or private data sources) as well as on-farm indicators for which data had to be collected “on farm”. Depending on animal category, there were considerable differences in the origin and availability of the data. In addition, the number of indicators per category varied between the different animal categories. Table 2 gives an overview of assessed indicators during pig farm visits along with information on data availability, distribution and cut-off values. The validity of the SAH method was evaluated by comparing the outcomes with the WQ or MTool protocols carried out in parallel. It became apparent that certain aspects of animal health could already be assessed with a relatively high degree of validity. If public data, private data and on-farm records can be used, valid estimations can be made for all four defined categories of animal health and welfare. If, on the other hand, only information on data-based indicators are available, the validity decreases significantly and in particular the welfare categories “Appropriate behavior” and “Freedom from pain, suffering, harm and anxiety” can usually only be assessed inadequately or not at all (33). Therefore, for sufficient validity, a combination of data-based and on-farm indicators with direct assessments of the animals is needed.

Several types of possible visualizations of the scores and benchmarking were developed with regard to a future implementation of the method. Figure 2 shows the benchmarking after Z-transformation, as a possible tool for authorities to monitor the health status at population level. Figure 3 shows an illustration for an individual farm, containing benchmarking information compared to other farms. In Figure 4, the scores of the SAH indicators of a single farm are displayed in relation to the corresponding target and alarm thresholds.

Most relevant data and information for the assessment of animal health and welfare are generally found in private data sources and data collected “on-farm” (Table 1). However, public data sources are the most accessible to government agencies and also have the widest coverage. In the animal movement database TVD, there are major differences in data quality among animal species. For cattle, there is a high level of detail at the individual animal level. However, there is no specific category for “veal calf”, a common production type in Switzerland, which means that animals intended for fattening cannot be distinguished from dairy calves to be raised as replacement heifers. For pigs, the information is only available at farm and animal group level, and, for poultry, data is limited to the number and age of the animals at the time of entry (stabling) and the production type. The laboratory information system ALIS contains, among other things, test results on notifiable animal diseases. However, these have often limited informative value with regard to the animal health and animal welfare status of a farm. The agricultural policy information system AGIS contains demographic data on farms and information on direct subsidies received for public and ecological services. It mostly does not contain any direct records on animal health, but does contain auxiliary variables (proxies) which can correlate with animal health and welfare indicators.

Exploratory data analyses have shown that, in general, integrating multiple data sources increases their potential compared to using each data source as a stand-alone source. Furthermore, it is important that entire data sets are available for these analyses and that possible parameters do not have to be selected a priori. Since the public databases were pseudonymised, they could be integrated with each other, but not linked to the private databases. The availability and quality of data from private data sources also varies considerably for the different animal categories. For dairy cattle and pigs, various relevant data from existing sources could already be used (with the agreement of the respective data owner). For poultry, potential health and production data are available, but these are mainly owned by integrators, kept in separate data silos by each company, and were only accessible to a very limited extent. The least data is available for veal calves, goats and sheep.

The review of PLF technologies, which can be a time-saving and objective alternative to the manual collection of indicators, has revealed two main findings: firstly, the range of PLF technologies varied greatly between different animal categories and, secondly, there is a large discrepancy between scientifically validated and commercially available PLF systems (28, 34). The largest number of PLF technologies is available for dairy cows, both in terms of availability of the number of system types and the number of suppliers within a system type. Fattening pigs and broilers follow in second place, while for sows, laying hens, veal calves, sheep and goats the choice of available PLF technologies that could be used to identify health and welfare issues is very limited (28). Furthermore, the review has shown that there are commercially available PLF technologies for many of the animal-based SAH indicators, however information on the validity of these systems was often not available (30). Based on the systematic literature search, a major weakness for the development of digital support systems was identified which needs to be addressed: It was found that many tools are aimed at identifying specific issues with non-specific indicators, such as activity or feeding. Consequently, most of the current PLF systems, whether in development or commercially available, are potentially able to detect welfare-related changes and rarely more specific issues such as defined diseases or symptoms. To improve the accuracy of PLF technologies, the aim of detection has to be specified and appropriate indicators that are strictly related to the purpose must be chosen (31). The described PLF framework suggests a categorization of the aim of detection into three levels, with each level providing a different degree of information and therefore requiring indicators with different specificity: (i) general welfare, using a large number of non-specific variables; (ii) disease and distress, with a higher degree of information about the cause of an issue; and (iii) defined disease, the most specific decision level, where the aim is to automatically detect signs related to defined diseases (e.g., mastitis).

Due to limited data availability for some species, a risk index could only be estimated for cattle and pig farms. Most of the machine-learning techniques applied reached similar classification performances (Kuntzer et al., submitted). As the explainability of the model was important both for the public acceptance of such a data-driven index as well as for the planning of on-farm welfare inspections, the random forest model was considered to be the most suitable method. Median sensitivity with random forests was 81.7% for cattle farms (Figure 5) and 81.8% for pig farms (Figure 6). Overall, the precision was increased by a factor 3–5 compared to randomly choosing farms for animal welfare inspections. The most important predictors for a lower likelihood of violating animal welfare law were participation in and compliance with federal organic (“ÖLN”) and animal welfare (“BTS” and “RAUS”) programs. Similarly, animal movement reports, notification discipline (farmers are obliged to report events such as births, deaths, leaving or arriving of animals at the farm in the animal movement database TVD within a specified time period) and structural characteristics of the farms, such as the type of husbandry or the standard labor force had high predictive weights.