TL;DR
This paper outlines the DynAIRx project, which aims to use AI to improve medication management for patients with multiple chronic conditions (multimorbidity). By integrating data from various healthcare sources and applying machine learning, DynAIRx will create tools to help clinicians identify patients most in need of medication reviews and optimize their treatment plans, ultimately aiming to reduce harm and improve patient outcomes.
What Problem Does This Paper Address?
Managing patients with multimorbidity, defined as having two or more long-term health conditions, is a significant challenge in healthcare. These patients often take multiple medications (polypharmacy), increasing the risk of adverse drug reactions, drug-drug interactions, and suboptimal treatment. The UK's NHS Long Term Plan and the 2021 National Overprescribing Review highlight the need for better medicines optimization, particularly through structured medication reviews (SMRs). However, current SMRs are hampered by fragmented health records across different providers, making it difficult to gather comprehensive patient information. There's also a lack of guidance on how to identify which patients would benefit most from an SMR, leading to inefficient resource allocation and potentially missed opportunities for intervention. DynAIRx seeks to bridge this gap by leveraging AI to create a more efficient and effective system for identifying and supporting these patients.
Methods
DynAIRx employs a multi-faceted approach combining qualitative and quantitative methods. The core of the project involves developing AI-driven tools that integrate structured clinical data from general practice, hospital, and social care records, along with unstructured clinical text processed via Natural Language Processing (NLP). Machine learning algorithms will be used to:
- Identify patterns: Detect combinations of conditions, medications, tests, and clinical contacts that precede adverse events.
- Predict risks: Develop models to estimate the risk of adverse outcomes and predict multimorbidity trajectories.
- Visualize data: Create interactive dashboards that present longitudinal patient journeys, risk trajectories, and key AI-learned patterns alongside clinical guidelines.
A key methodological innovation is the emphasis on co-design and iterative refinement. The project involves extensive engagement with end-users, including general practitioners, pharmacists, patients, and carers, through interviews, focus groups, and "think-aloud" studies to ensure the developed tools are clinically useful, user-friendly, and explainable. The project also explicitly incorporates causal inference methods to move beyond simple correlations and understand the causal relationships between interventions and outcomes, aiming for explainable AI.
Datasets
The project will utilize large-scale, integrated health data from various NHS Integrated Care Systems across England, covering approximately 11 million people. This includes:
- Structured clinical data: From general practice, hospital (Secondary Uses Service data), and community/mental health records.
- Unstructured clinical text: Specifically, psychiatric narratives from Merseyside will be processed using NLP.
The paper mentions the use of existing healthcare NLP applications and annotated datasets like WEB-RADR and AVERT for extracting information related to adverse drug reactions and mental health narratives. The data will be curated within Trusted Research Environments (TREs) to ensure data security and privacy.
Key Insights and Ideas
- The power of integrated data: The project hinges on the increasing availability of integrated health records, which are crucial for a holistic view of patients with multimorbidity.
- AI for proactive identification: Instead of reactive approaches, AI can proactively identify patients at high risk of adverse outcomes, enabling timely interventions.
- Explainability is paramount: Clinicians need to understand why an AI tool is making a recommendation, especially when dealing with complex patient cases. DynAIRx prioritizes explainable AI through causal inference and clear visualizations.
- Co-design drives utility: The iterative development process, involving end-users at every stage, is critical for ensuring the tools are practical and adopted in real-world clinical settings.
- Beyond simple correlations: The focus on causal inference aims to provide more robust insights into treatment effects and potential harms, moving beyond just identifying associations.
Key Results
As this paper is a project protocol, specific quantitative results are not yet available. However, the expected key outputs are:
- Evaluation of challenges and opportunities in medicine optimization within general practices.
- A pipeline for processing structured and unstructured care data for multimorbidity research.
- An AI framework for identifying high-risk patients and predicting disease trajectories that warrant SMRs.
- Novel visualizations of patient journeys to enhance medication reviews.
- The development of a clinically useful learning healthcare system that integrates these outputs, co-developed with end-users and accessible to patients and carers.
Limitations and Future Work
- Data availability and quality: While aiming for integrated data, the paper acknowledges that data may not always be complete or perfectly structured across all sources.
- Generalizability: The study will initially focus on specific multimorbidity groups and geographical areas, with future work potentially expanding to broader populations.
- Implementation challenges: The paper acknowledges that translating research findings into daily clinical practice is a major challenge and that the effectiveness of feedback systems depends on their content and delivery.
- Future work: The project is designed as a multi-phase initiative. DynAIRx 2 is envisioned to build upon the developed visualisations and outcome predictions to create multi-morbidity decision support tools for direct use in consultations. The project also aims to explore the applicability of social-norm and practice/prescriber-level feedback to medicines optimization.
Summary updated