Poster Abstracts 1-7

Sam Andres and Imane Guellil 

This poster aims to highlight the pipeline, methodology and techniques used in the ACRC work package 3 and AIM-CISC objective 4 to automatically detect geriatric syndromes, adverse events and social context in clinical free-text (such as discharge summaries, radiology reports and referral letters). In order to automate this process, we rely on natural language processing which is a sub-field of artificial intelligence. This requires the creation of a manually annotated resource which will be used to train our language model to detect GS/AEs and SC automatically. Geriatric syndrome is a term used to capture complex and multifactorial clinical conditions in older adults that do not fit into discrete disease categories. Geriatric syndromes usually lack well-defined aetiological or pathogenetic factors but share multiple synergistic risk factors such as older age and functional, cognitive, and sensory impairments. They often involve multiple organs, become major obstacles in the management of chronic conditions and increase an older person’s vulnerability to situational challenges [1]. In our work we have chosen to focus on 12 geriatric syndromes: falls, frailty, malnutrition, weight loss, delirium, dementia, unspecified cognitive impairment, urinary incontinence, faecal incontinence, pressure injury, visual impairment, and hearing impairment. Adverse events are harmful events or undesired harmful effects resulting from medication or other methods of treatment[2]. They are frequently precipitated by particular trigger events such as infections or drug interactions and these precipitants are often the focus of reactive healthcare. Furthermore, the occurrence and severity of adverse events are significantly determined by the presence of certain predisposing factors, which are a complex interaction between an individual’s morbidities, medical treatments and wider social support. In our work, we have elected to detect 11 adverse events: major osteoporotic fractures (hip, wrist, proximal humeral fractures and vertebral compression fractures), intracranial haemorrhage, upper gastrointestinal haemorrhage, lower gastrointestinal haemorrhage, lower respiratory tract infection, urinary tract infection, constipation and seizures

Stella Arakelyan, Atul Anand, Professor Bruce Guthrie

As the global population ages, the burden of multiple long-term conditions (MLTCs) is also on the rise. Over 60% of UK older adults (aged > 65) are affected by MLTCs which are associated with poor health outcomes such as lower quality of life, functional decline, adverse drug events, and unscheduled hospital admissions and re-admissions. The Chief Medical Officer's Annual Report 2023 identifies an urgent need to develop, evaluate, and implement effective and sustainable interventions to improve outcomes in people with MLTCs. In this ongoing work, we integrate AI and data science, social science, and health service research methods to co-design two complex healthcare interventions (targeted at hospital front door and community settings) underpinned by risk prediction tools to improve pathways of care and health outcomes for people with multiple-long-term conditions.

Arlene Casey

Healthcare systems are mostly organised to provide care for individual diseases. However, older people often have multiple conditions or problems like poor mobility which are caused by many different factors. Most research relies on ‘coded’ data, which is a highly standardised way of recording that someone, for example, has had a stroke. Complex problems are often only written about in free-text notes made by doctors and nurses. To really understand the needs of older people, we should therefore use free-text notes. However, very limited amounts of free-text (or unstructured) data are used in research, due to the privacy-risk challenges associated with data access (i.e. details that potentially reveal a person's identity) and very manual processes to review and assess these risks.

Leading a collaboration across the Scottish Safe Haven Network we are exploring these privacy risks and building a pathway to unstructured health data without compromising confidentiality. This work has the potential to improve our understanding of later life and later-life care across many future research projects by facilitating access to large sets of health free-text notes.

We categorise privacy risks as direct identifiers (e.g., names, addresses) and indirect (e.g., unique events, specific locations that may reveal a patient’s identity). Direct risks are well defined but indirect risks are not. This lack of definition puts limitations on being able to agree on how to risk assess data.

Using ~89,000 Discharge Summaries of adults we explore indirect privacy risks. Indirect risks are relatively rare and hard to find. We employ text mining techniques to explore the data including topic modelling, with BERTopic as a clustering task at sentence level, to understand latent topics in the clinical reports. Topic cluster labelling is generated using term and inverse document frequency. These labelled clusters reveal privacy-risk categories, and we use these clusters to guide further qualitative reviewing of reports and in building a map of indirect privacy risk categories.

We build a map of privacy risk categories found in Discharge Summaries across age bands, sex, and location. Our analysis reveals categories of privacy risk that go beyond the current published personal information identifiers (PII) frameworks, and we discover how these risks accumulate over reports and differ in older adults compared to younger adults.

Securing public perspectives has been embedded throughout our work with three workshops (39 participants) and one online survey (+1000 participants). An introductory workshop on health data research enabled participants to take part in deliberative workshops on privacy risk and how to potentially address risks through semi-automated processes. The learning has informed our approach to finding risks and developing our privacy-risk map.

This understanding of privacy-risk is of benefit to the entire health community, not just later-life related research. It helps to direct efforts on semi-automating privacy-risk mitigation and underpins the ability of the health data community to improve data access for research.

This work described here has been supported by funding from DARE UK and Research Data Scotland.

Longfei Chen

With ageing and chronic health conditions significantly affecting the daily lives of older adults, the timely detection of developing health conditions and critically health events are crucial. Our study focuses on monitoring behaviours related to physical weakness, inactivity, and prolonged room occupancy for older adults at their homes using a non-intrusive camera sensor. The system captures statistics related to body motion, inactivity, and environmental context related to these health conditions in real time, while prioritizing privacy.

Caroline Pearce, Ki Tong, Professor Catharine Ward Thompson

Supportive outdoor environments can enable people to remain healthy and active in older age, improving quality of life. Yet older adults continue to face difficulties accessing physical environments. This poster presents findings from a longitudinal study of older people (aged 50+) living in Scotland, UK and explores how physical environments support people to undertake the ‘personal projects’ (Little, 1983) that enable them to flourish. Personal projects are the self-generated and purpose-oriented activities an individual is doing or planning to do, and range from day-to-day but important everyday routines to ambitious, long-term endeavours. Drawing on data from 45 participants, we describe the types of personal projects participants reported as important and what aspects of the environment were associated with project enjoyment and positive outcome. Projects that involved social engagement with other people and their local community were valued highly. Projects that involved engagement with nature were by comparison reported as more difficult, and participants reported feeling less in control and less supported to spend time in nature. Concerns about safety and fears of falling or an accident were a significant deterrent to spending time in natural environments. Overall, the easier the local outdoor environment made it for people to carry out their personal projects, the higher their quality of life; this is particularly important in a post-COVID world. We examine the significance of these findings in the context of policies and frameworks concerning ‘age-friendly communities’ to address the role of the physical environment in not only supporting people to achieve the necessary tasks of daily living adequately, but also in supporting people and communities to flourish in later life.

Little, B. 1983. Personal Projects: a rationale and method for investigation. Environment and Behavior, 15(3): 273-309.

Tricia Tooman

Huayu Zhang, Arlene Casey, Imane Guellil, Clare MacRae, Charis Marwick, Honghan Wu, Bruce Guthrie, Beatrice Alex

Where people live in households or neighbourhoods is closely associated with a range of socio-economic and other factors. Many such factors are key to answering questions and providing solutions in the advanced care setting. n principle, GP registration address recorded in the Community Health Index (CHI) would allow us to accurately identify where someone lives (beyond existing uses of address postcode which is less fine grained). However, by themselves, addresses in CHI are often inaccurate, non-standardized and/or ambiguous. There is the need for a unique identifier for the purpose of standardizing operations around address information, and the most suitable is the Unique Property Reference Number (UPRN) created by the Ordnance Survey (OS). Using the UPRN identifiers, we could better geolocate where individuals live, including identifying households and mapping individuals and households to smaller areas than is currently routine.

We have developed the prototype of an open-source tool (FLAP: Framework for linking address to UPRN database) that links free-text GP registered addresses to an entry in the standard address database. We adopted a generalisable solution based on a machine learning-based matching classifier coupled feature construction with linear assignments.

FLAP achieves adjusted accuracy of 0.979 in addresses in Lothian and 0.989 in Tayside/Fife. A open-source Python library implementing the tool is created on Pypi index (https://pypi.org/project/flap-lite/). We have implemented the tool in Public Health Scotland and Dataloch environment to support services and research activities.

In conclusion, we have developed a framework, FLAP, for linking free-text UK addresses to UPRN DB with good real-world performance and usability. We enabled better geo-coding of patient data in the healthcare system.