What does this research mean for the field?

Data-driven clustering of multiple long-term condition trajectories produces overlapping subgroups that explain little variance in healthcare outcomes and perform worse at predicting hospitalisation and emergency department attendance than simple disease counts. Novelty: ClaimNovelty.INCREMENTAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

To generate data-driven clusters of patients based on their disease trajectories and evaluate their clinical interpretability and association with subsequent health outcomes.

September 23, 2026Communications MedicineOpen Access

Challenges of clustering disease trajectories in people with multiple long-term conditions using data from 7.2 million patients

Q: What is the clinical evidence from this study?

Study design: Cohort. Population: Multiple Long-Term Conditions (MLTC) (n=7285210). Intervention: Cluster membership based on disease trajectories vs. Patients with no long-term conditions. Primary outcome: ED attendance, hospitalisation and mortality.

View Full Paper

Ask AI

Bookmark

Key result

Data-driven disease trajectory clusters predict ED attendance and hospitalization worse than simple condition counts.

n=7,285,210

Why the study?

Generating clusters of people with similar patterns of multiple long-term conditions could help target healthcare services, but clinical interpretability and outcome associations remain unclear.

Do data-driven clusters of disease trajectories accurately predict 1-year health outcomes compared to simple counts of long-term conditions in adults?

Comparison

Data-driven clustering of disease trajectories vs no long-term conditions and prediction using count of conditions, individual diseases, or embeddings

Design

Cohort study using transformer model embeddings and k-means clustering

Follow-up

1 year

Authors

TBThomas Beaney JCJonathan Clarke TWThomas Woodcock

Discussion

Member takes

Overview

Clusters add little predictive value beyond condition counts; challenges data-driven trajectory clustering for short-term multimorbidity outcomes.

Key Points

To generate data-driven clusters of patients based on their disease trajectories and evaluate their clinical interpretability and association with subsequent health outcomes.
Population-based cohort study of 7,285,210 adults (5,981,091 with multiple long-term conditions and 1,304,119 without chronic conditions) using CPRD Aurum linked to hospital records as of January 1, 2015.
Disease trajectories were transformed into vector embeddings using a transformer model and grouped into trajectory clusters using k-means clustering.
Participants were followed for 1 year to evaluate cluster associations with emergency department attendance, hospital admissions, and mortality compared with individual diseases, raw embeddings, and condition counts.
The model identified eight optimal clusters of patients with multiple long-term conditions, with the largest representing 21.3% of the population dominated by cardio-kidney-metabolic conditions, though dominant conditions heavily overlapped between clusters.
Trajectory clusters explained only 1.2% to 3.1% of the total variance in 1-year outcomes, and initial differences in odds of emergency attendance, hospitalization, and death substantially attenuated after adjusting for age, sex, ethnicity, and deprivation.
Clusters performed substantially worse at predicting outcomes than individual diseases or raw embeddings, and performed worse than a simple count of long-term conditions for predicting emergency department visits and hospital admissions.

Study Design

Type

Cohort (n=7,285,210)

Structured PICO

Do data-driven clusters of disease trajectories accurately predict 1-year health outcomes compared to simple counts of long-term conditions in adults?

Population

7,285,210 adults, including 5,981,091 with multiple long-term conditions and 1,304,119 without chronic conditions, followed for 1 year.

Exposure

Data-driven clustering of disease trajectories using a transformer model followed by k-means clustering.

Comparator

Patients with no long-term conditions, or prediction using simple count of long-term conditions, individual diseases, or embeddings alone.

Outcome

1-year emergency department (ED) attendance, hospitalisation, and mortality.hard clinical

Data-driven clustering of multiple long-term condition trajectories explains little variation in 1-year health outcomes and performs worse than simple condition counts for predicting healthcare utilization.

Limitations

Interpretation of the clusters was challenging due to overlap of the dominant conditions across clusters
Clusters explained little of the variation in 1-year health outcomes

Cite This Study

Beaney et al. (2026) conducted a cohort in Multiple Long-Term Conditions (MLTC) (n=7,285,210). Cluster membership based on disease trajectories vs. Patients with no long-term conditions was evaluated on ED attendance, hospitalisation and mortality. Data-driven clusters of disease trajectories explained only 1.2-3.1% of variance in 1-year outcomes and performed worse than simple condition counts at predicting ED attendance and hospitalisation.

synapsesocial.com/papers/6ab39db54f6cc12e38b51328 https://doi.org/10.1038/s43856-026-01917-8

View Full Paper

Ask AI

Bookmark

September 23, 2026Communications MedicineOpen Access

Challenges of clustering disease trajectories in people with multiple long-term conditions using data from 7.2 million patients

View Full Paper

Ask AI

Bookmark

Key result

Data-driven disease trajectory clusters predict ED attendance and hospitalization worse than simple condition counts.

n=7,285,210

Why the study?

Generating clusters of people with similar patterns of multiple long-term conditions could help target healthcare services, but clinical interpretability and outcome associations remain unclear.

Do data-driven clusters of disease trajectories accurately predict 1-year health outcomes compared to simple counts of long-term conditions in adults?

Comparison

Data-driven clustering of disease trajectories vs no long-term conditions and prediction using count of conditions, individual diseases, or embeddings

Design

Cohort study using transformer model embeddings and k-means clustering

Follow-up

1 year

Authors

TBThomas Beaney JCJonathan Clarke TWThomas Woodcock

Discussion

Member takes

Overview

Clusters add little predictive value beyond condition counts; challenges data-driven trajectory clustering for short-term multimorbidity outcomes.

Key Points

To generate data-driven clusters of patients based on their disease trajectories and evaluate their clinical interpretability and association with subsequent health outcomes.
Population-based cohort study of 7,285,210 adults (5,981,091 with multiple long-term conditions and 1,304,119 without chronic conditions) using CPRD Aurum linked to hospital records as of January 1, 2015.
Disease trajectories were transformed into vector embeddings using a transformer model and grouped into trajectory clusters using k-means clustering.
Participants were followed for 1 year to evaluate cluster associations with emergency department attendance, hospital admissions, and mortality compared with individual diseases, raw embeddings, and condition counts.
The model identified eight optimal clusters of patients with multiple long-term conditions, with the largest representing 21.3% of the population dominated by cardio-kidney-metabolic conditions, though dominant conditions heavily overlapped between clusters.
Trajectory clusters explained only 1.2% to 3.1% of the total variance in 1-year outcomes, and initial differences in odds of emergency attendance, hospitalization, and death substantially attenuated after adjusting for age, sex, ethnicity, and deprivation.
Clusters performed substantially worse at predicting outcomes than individual diseases or raw embeddings, and performed worse than a simple count of long-term conditions for predicting emergency department visits and hospital admissions.

Study Design

Type

Cohort (n=7,285,210)

Structured PICO

Do data-driven clusters of disease trajectories accurately predict 1-year health outcomes compared to simple counts of long-term conditions in adults?

Population

7,285,210 adults, including 5,981,091 with multiple long-term conditions and 1,304,119 without chronic conditions, followed for 1 year.

Exposure

Data-driven clustering of disease trajectories using a transformer model followed by k-means clustering.

Comparator

Patients with no long-term conditions, or prediction using simple count of long-term conditions, individual diseases, or embeddings alone.

Outcome

1-year emergency department (ED) attendance, hospitalisation, and mortality.hard clinical

Limitations

Interpretation of the clusters was challenging due to overlap of the dominant conditions across clusters
Clusters explained little of the variation in 1-year health outcomes

Cite This Study

synapsesocial.com/papers/6ab39db54f6cc12e38b51328 https://doi.org/10.1038/s43856-026-01917-8

Also Consider

Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context:

View Full Paper

Ask AI

Bookmark