< Back to Industry Collaborators
Funder: UCB Pharma
Overview: The study aimed to predict the development of ankylosing spondylitis (AS), a form of inflammatory arthritis, using machine learning techniques. AS often takes years to diagnose, leading to delays in treatment. By leveraging the Secure Anonymised Information Linkage (SAIL) databank, the research utilized routinely collected primary care data to create predictive models. These models focused on identifying patterns, such as diagnosis codes, medications, and laboratory tests, associated with AS before diagnosis, offering a potential pathway for earlier identification and intervention.
Why SAIL was valuable:
- SAIL offered access to anonymized, longitudinal datasets from general practice, hospital admissions, and prescribing records.
- SAIL enabled tracking of individual patient pathways over extended periods.
- SAIL combined primary and secondary care datasets, facilitating a holistic view of patient histories.
- SAIL ensured data security and compliance with ethical standards.
Results:
- The study identified key patterns in the use of diabetes treatments and their impacts on patient outcomes.
- Findings demonstrated the effectiveness of specific treatment regimens, helping to inform clinical practice and healthcare policy.
- Cost-effectiveness analyses provided insights into the financial implications of treatment options, supporting resource allocation decisions in healthcare systems.
- For test datasets, the models achieved a positive predictive value (PPV) of 70-80%.
- However, due to the low prevalence of AS in the general population, the PPV dropped to 0.15%-0.25% in population-level validation.
- Women had more complex diagnostic pathways compared to men, often requiring more tests and specialist referrals.
- Men were typically diagnosed earlier, with symptoms presenting during teenage years.
- The models highlighted recurring interactions with healthcare providers and specific diagnostic codes as key indicators.
- Suggested further research into integrating patient-reported data (e.g., family history, symptom severity) to enhance prediction accuracy.
