TL;DR
This study investigates how effectively grouping people with multiple long-term conditions (MLTC) into disease clusters can predict health outcomes like mortality and hospitalizations. The findings reveal that while clustering can simplify complex health data, it ultimately loses predictive power compared to considering individual diseases, suggesting limitations for clinical applications.
What Problem Does This Paper Address?
Managing patients with multiple long-term conditions (MLTC) is a growing challenge in healthcare. These individuals often experience poorer quality of life, increased healthcare utilization, and higher mortality. To better understand and manage MLTC, researchers are exploring ways to group patients based on co-occurring diseases, creating "disease clusters" that represent distinct health phenotypes. However, a key challenge is how to assign these clusters to individual patients. Various methods exist, but their effectiveness in explaining health outcomes remains unclear. This paper aims to directly compare different strategies for assigning disease clusters to people and assess how well these assignments predict important health outcomes.
Methods
The researchers conducted a cohort study using electronic health records from England. They utilized a previously identified set of 15 disease clusters, representing 212 long-term conditions (LTCs). Seven different strategies were employed to assign these clusters to individuals:
- Binary Classification Strategies:
- "One or more": Assigns a cluster if a person has at least one disease within it.
- "Two or more": Requires at least two diseases in the cluster.
- "Three or more": Requires at least three diseases in the cluster.
- "Modal": Assigns the cluster with the most diseases for that person (ties broken randomly).
- "Majority": Assigns a cluster if over 50% of a person's diseases fall within it.
- Quantitative Strategies:
- "Count": Assigns the number of diseases a person has within each cluster.
- "Proportional": Assigns the proportion of a person's diseases within each cluster.
The performance of these strategies was evaluated by their ability to explain three health outcomes over one year: mortality, emergency department (ED) attendances, and emergency hospital admissions. This was done using logistic regression models, with age, gender, ethnicity, and socioeconomic deprivation as covariates. Crucially, they also compared these cluster-based strategies against a model that used the individual LTCs as predictors. Model performance was assessed using the Akaike Information Criterion (AIC), where lower values indicate better explanatory power.
Datasets
The study utilized data from the Clinical Practice Research Datalink (CPRD) Aurum dataset, which comprises electronic health records from General Practice (GP) in England. This was supplemented with secondary care data from Hospital Episode Statistics (HES) and death registration data from the Office for National Statistics. The dataset included over 6.2 million patients aged 18 and over with MLTC.
Key Insights and Ideas
- Loss of Information: Grouping diseases into clusters, while simplifying data, inherently leads to a loss of information. This loss significantly impacts the ability to accurately predict health outcomes.
- Individual Diseases Reign Supreme: No cluster assignment strategy outperformed using information on individual diseases for predicting mortality, ED attendances, or emergency hospitalizations. This suggests that for predictive purposes, focusing on the specific conditions a patient has is more effective than relying on broader disease categories.
- "Count" Strategy Shows Promise (Relatively): Among the cluster assignment strategies, the "Count" method (assigning the number of diseases within a cluster) performed best. This indicates that the absolute burden of disease within a cluster is more informative than just the relative proportion.
- Variability Within Clusters is High: Disease clusters, as defined, do not represent consistent relationships among the diseases they contain. The impact of individual diseases within a cluster on health outcomes can vary dramatically, making it difficult to draw generalized conclusions about the cluster as a whole.
- Clinical Relevance Questioned: The significant variability within clusters raises questions about their practical utility in clinical research and practice. If diseases within a cluster have vastly different associations with outcomes, it limits the ability to use clusters for targeted interventions or understanding disease progression.
Key Results
- A total of 6,286,233 patients with MLTC were included in the study.
- The "Count" strategy for assigning clusters performed best among the seven strategies tested for explaining mortality, ED attendances, and emergency admissions.
- However, even the best-performing cluster strategy was inferior to using individual LTCs as predictors, as indicated by higher AIC values.
- The "Majority" method performed worst among the cluster assignment strategies.
- There was substantial variation in the age at diagnosis both between and within clusters. For example, the "malignancies" cluster showed a wide range of ages at diagnosis for its constituent diseases.
- For the "malignancies" cluster, the odds ratio for mortality varied significantly, from a 1.35 times increase for prostate cancer to a 38.8 times increase for mesothelioma. This highlights the substantial within-cluster variability.
Limitations and Future Work
- Data Source Limitations: The study relied on coded information in EHRs, which may miss diagnoses recorded as free text. Coding rates can also be influenced by financial incentives, potentially introducing bias.
- Scope of Strategies: The study tested seven strategies for cluster assignment but acknowledges this is not an exhaustive list. Other methods, such as averaging disease vector embeddings to represent people, could be explored.
- Interaction Effects: The models did not account for interactions between clusters or between covariates and clusters, which could potentially improve explanatory power.
- Short Follow-up Period: The outcomes were assessed over a 1-year period. While relative differences between strategies are expected to remain consistent, longer-term evaluations could provide further insights.
The authors suggest that hybrid approaches, combining individual disease information for common conditions with clusters for rare ones, could offer a pragmatic balance. They also propose that directly clustering people, rather than diseases, might yield more meaningful segments, especially if incorporating demographics and healthcare utilization alongside disease data. Future research could explore these combined approaches and investigate the impact of interactions.
Summary updated