Harnessing temporal patterns in administrative patient data to predict risk of emergency hospital admission
File(s)
Published version
Author(s)
Post, Benjamin
Klapaukh, Roman
Brett, Stephen
Faisal, Aldo
Type
Journal Article
Abstract
Background. Patients often undergo unexpected hospitalisations that may be foreseeable or even
preventable. Unplanned hospital admissions are associated with worse patient outcomes and cause
strain on health systems worldwide. Primary care electronic health records (EHRs) have success fully been used to create prediction models for emergency hospitalisation, but these approaches
require a broad range of diagnostic, physiological and lab values. Some methods require in excess
of 100 clinical parameters making society-wide implementation impractical, overly costly or too
taxing for many health providers. In this study, we introduce a novel approach to capture temporal
patterns of patient activity from EHR data and evaluate their effectiveness in predicting emergency
hospital admissions compared to conventional methods.
Methods. We used the SAIL Databank to extract temporal patterns of primary care activity from
undifferentiated electronic health record timestamp data for 1.37 million patients from the country
of Wales. Using Gaussian Mixture Modelling we grouped patients into distinct temporal clusters,
performed a 3-stage validation of our approach and calculated the risk of emergency hospital
admission for each temporal cluster group. Finally, these temporal clusters were combined with
5 administrative variables and incorporated into emergency hospital prediction models (Logistic Regression, Naïve Bayes, XGBoost and Multilayer Perceptron (MLP)) and compared with a more
traditional, but data-intensive, modelling technique. The primary outcome was emergency hospital
admission as the next healthcare event.
Findings. Six distinct temporal cluster patterns of primary care EHR activity were identified,
associated with varying risks of future emergency hospital admission risk. These patterns were
visually interpretable, repeatable at a population-level and clinically plausible. The best
emergency hospital admission prediction model (MLP) achieved an area under the receiver
operating characteristic (AUROC) of 0.82 and precision of 0.94. In external validation, similar
model performance was observed (AUROC 0.82 and precision 0.92). This model also matched
the performance of a more complex model requiring 33 clinical parameters (AUROC 0.82 vs
0.83; precision 0.94 vs 0.90) for the same task on the same dataset.
Interpretation. We developed a novel machine learning pipeline that extracts interpretable tem poral patterns from simple representations of EHR data and can be incorporated into emergency
hospital admission predictors. Our approach is independent of specific data types, reducing pre processing requirements and making it applicable to heterogeneous datasets. This framework may
enable more rapid development of parsimonious clinical prediction models.
Funding. UKRI CDT in AI for Healthcare, UKRI Turing AI Fellowship, NIHR Imperial Biomedical
Research Centre (BRC) and Research Capability Funding.
preventable. Unplanned hospital admissions are associated with worse patient outcomes and cause
strain on health systems worldwide. Primary care electronic health records (EHRs) have success fully been used to create prediction models for emergency hospitalisation, but these approaches
require a broad range of diagnostic, physiological and lab values. Some methods require in excess
of 100 clinical parameters making society-wide implementation impractical, overly costly or too
taxing for many health providers. In this study, we introduce a novel approach to capture temporal
patterns of patient activity from EHR data and evaluate their effectiveness in predicting emergency
hospital admissions compared to conventional methods.
Methods. We used the SAIL Databank to extract temporal patterns of primary care activity from
undifferentiated electronic health record timestamp data for 1.37 million patients from the country
of Wales. Using Gaussian Mixture Modelling we grouped patients into distinct temporal clusters,
performed a 3-stage validation of our approach and calculated the risk of emergency hospital
admission for each temporal cluster group. Finally, these temporal clusters were combined with
5 administrative variables and incorporated into emergency hospital prediction models (Logistic Regression, Naïve Bayes, XGBoost and Multilayer Perceptron (MLP)) and compared with a more
traditional, but data-intensive, modelling technique. The primary outcome was emergency hospital
admission as the next healthcare event.
Findings. Six distinct temporal cluster patterns of primary care EHR activity were identified,
associated with varying risks of future emergency hospital admission risk. These patterns were
visually interpretable, repeatable at a population-level and clinically plausible. The best
emergency hospital admission prediction model (MLP) achieved an area under the receiver
operating characteristic (AUROC) of 0.82 and precision of 0.94. In external validation, similar
model performance was observed (AUROC 0.82 and precision 0.92). This model also matched
the performance of a more complex model requiring 33 clinical parameters (AUROC 0.82 vs
0.83; precision 0.94 vs 0.90) for the same task on the same dataset.
Interpretation. We developed a novel machine learning pipeline that extracts interpretable tem poral patterns from simple representations of EHR data and can be incorporated into emergency
hospital admission predictors. Our approach is independent of specific data types, reducing pre processing requirements and making it applicable to heterogeneous datasets. This framework may
enable more rapid development of parsimonious clinical prediction models.
Funding. UKRI CDT in AI for Healthcare, UKRI Turing AI Fellowship, NIHR Imperial Biomedical
Research Centre (BRC) and Research Capability Funding.
Date Issued
2025-02
Date Acceptance
2024-11-13
Citation
The Lancet: Digital Health, 2025, 7 (2), pp.e124-e125
ISSN
2589-7500
Publisher
Elsevier
Start Page
e124
End Page
e125
Journal / Book Title
The Lancet: Digital Health
Volume
7
Issue
2
Copyright Statement
© 2025 The Author(s). Published by Elsevier Ltd. This is an Open Access article under the CC BY 4.0 license
License URL
Publication Status
Published
Date Publish Online
2025-01-29