Developing AI Methods for Animal Health and Welfare Monitoring

作为动物福利研究的一部分，收集了大量的自由文本数据，通常以兽医临床记录和公民科学调查的自由文本回复的形式。然而，自由文本数据通常没有像从动物福利研究中收集的定量和定性数据那样充分利用，因为它通常结构不良，语法多变且过度缩写，因此难以从中提取信号。这个博士项目的目的是从这类杂乱的自由文本中提取健康和福利信号。数据将由英国这类数据的两个最大的收集者SAVSNET（小动物兽医监测网络）和狗信托基金提供。SAVSNET收集的数据是来自400多个英国兽医实践的真实的时间数据，包括品种、性别、中性状态、处方药、描述主要主诉的症状标签和来自兽医临床记录的叙述性自由文本记录。Dogs Trust收集的数据是由Dogs Trust生成和托管的公民科学调查的免费文本数据。这些调查都是纵向的（例如，代小狗，n~8，000;收养后福利研究，n~9，000）和一次性研究项目（全国狗调查，n = 250，000;选择我的狗，n~10，000），收集的自由文本通常记录了狗主人对狗行为史的描述或狗主人对其狗的关注和意见。狗信托基金还拥有一些兽医临床笔记的形式叙述自由文本在PDF格式。我们的目的是从这些数据中检索信号，这将有利于动物福利研究。为了提取这些信号，将使用监督和无监督机器学习方法。这些方法将利用BERT等转换器和Llama和Falcon等开源大型语言模型，针对兽医和调查参与者使用的语言进行优化，以从文本中确定趋势。通过这样做，我们的目标是实现几个结果。这些包括（但不限于）：1。能够以标准化格式从PDF兽医临床记录中提取数据（例如主要表现的投诉和治疗计划）2.建立评估犬只健康和福利新趋势的渠道（如新病毒爆发）3.开发一个早期预警系统，使用狗信托所有者报告的数据，以突出需要支持或跟进的记录（例如，高风险问题行为的实例，家庭和/或动物虐待的发生）通过实现这些目标，我们希望增加对SAVSNET和狗信托持有的数据的访问和洞察力，从而扩大犬的健康和福利研究。

Large amounts of free text data are collected as part of animal welfare research, usually in the form of veterinary clinical notes and free text responses from citizen science surveys. However, free text data often isn't utilized as fully as quantitative and qualitive data collected from animal welfare studies as it is typically poorly structured, grammatically variable and over-abbreviated and therefore hard to extract signals from. The aim of this PhD project is to extract health and welfare signals from this type of messy free text.The data will be provided by two of the largest collectors of this data type in the UK, SAVSNET (Small Animal Veterinary Surveillance Network) and Dogs Trust. The data collected by SAVSNET is real time data from over 400 UK veterinary practices including breed, sex, neuter status, prescribed drugs, syndromic labels describing main presenting complaint and narrative free text records from veterinary clinical notes. The data collected by Dogs Trust is free text data from citizen science surveys generated and hosted by Dogs Trust. These surveys are both longitudinal (e.g., Generation Pup, n ~ 8,000; Post Adoption Welfare Study, n ~9,000), and one-off research projects (National Dog Survey, n = 250,000; Choosing My Dog, n ~ 10,000) with free text collected typically documenting owner accounts of canine behavioural history or concerns and opinions of owners regarding their dog. Dogs Trust also possess some veterinary clinical notes in the form of narrative free text in PDF format. The intention is that signals will be retrieved from this data which will be of benefit to animal welfare research. To extract these signals, both supervised and unsupervised machine learning methods will be used. These methods will make use of transformers such as BERT and open source large language models such as Llama and Falcon, optimized for use on language used by vets and survey participants to establish for trends from the text. By doing this, we aim to achieve several outcomes. These include (but are not limited to):1. Being able to extract data in a standardised format from PDF veterinary clinical notes (e.g. main presenting complaints and treatment plans)2. Creating pipelines for evaluating emerging health and welfare trends in dogs (e.g. new virus outbreaks)3. Developing an early warning system using Dogs Trust owner reported data to highlight records that require support or follow ups (e.g. instances of high-risk problem behaviours, occurrences of domestic and/or animal abuse) By achieving these aims, we hope to increase both access to and insight from data held by SAVSNET and Dogs Trust, allowing for the expansion of canine health and welfare studies.