Related Experiment Video
Updated: Aug 5, 2026
Implementation of a Real-Time Psychosis Risk Detection and Alerting System Based on Electronic Health Records using CogStack
Published on: May 15, 2020
Dynamic prediction of cardiovascular risk from linked primary care records in New South Wales, Australia: protocol
Nicholas I-Hsien Kuo1, Heidi Welberry1, Sanja Lujic1
1Centre for Big Data Research in Health, University of New South Wales, Sydney, Australia.
Insights
This study introduces advanced transformer models for predicting 5-year cardiovascular disease (CVD) risk using electronic medical records. These dynamic models offer improved risk assessment compared to traditional methods.
Area of Science:
- Health Informatics
- Machine Learning in Medicine
- Cardiovascular Epidemiology
Background:
- Cardiovascular disease (CVD) is a major cause of preventable mortality.
- Existing CVD prediction tools often rely on static, baseline data.
- Longitudinal data from electronic medical records (EMRs) are underutilized in current CVD risk prediction.
Purpose of the Study:
- To develop and validate transformer-based sequence models for predicting 5-year CVD risk at each general practitioner (GP) encounter.
- To compare the performance of these dynamic models against traditional landmark survival models and static CVD prediction tools.
- To leverage longitudinal EMR data for more accurate and timely CVD risk assessment.
Main Methods:
- Utilizing the NSW Lumos linked health data asset, including GP EMRs, hospital admissions, and mortality data.
- Developing sex-specific dynamic landmark survival models and transformer-based sequence models for 5-year CVD risk prediction.
- Employing internal validation with patient-level and temporal separation, and evaluating geographic transportability via internal-external validation.
Main Results:
- Transformer models aim to provide updated, encounter-level CVD risk estimates.
- Performance will be benchmarked against existing static models and dynamic landmark survival models.
- The study will assess discrimination, calibration, and decision-curve analyses for model comparison.
Conclusions:
- Dynamic modelling using transformer-based sequence models holds promise for enhancing CVD risk prediction.
- This approach can improve patient monitoring and risk assessment in primary care.
- The findings could lead to more personalized and timely cardiovascular care interventions.
Introduction:
Cardiovascular disease (CVD) remains a leading cause of preventable morbidity and mortality. Most existing CVD prediction tools use a baseline survival modelling framework, in which each predictor is represented by a single value recorded at the start of follow-up. Although risk calculators can be re-run using updated values, these models do not explicitly use the longitudinal, irregularly recorded information in primary-care electronic medical records (EMRs). Dynamic modelling approaches have the potential to generate updated CVD risk estimates at each general practitioner (GP) encounter, supporting both risk assessment and monitoring of changes over time. This protocol outlines the development and validation of transformer-based sequence models for encounter-level prediction of 5-year CVD risk, compared with dynamic landmark survival models and benchmarked against static CVD prediction tools.
Methods And Analysis:
The study will use the New South Wales (NSW) Lumos linked health data asset, a large, privacy-preserving linkage of GP EMRs with hospital admissions, mortality and other administrative datasets across NSW, Australia. Eligible GP encounters will be included from 2018 onwards. The primary outcome is first fatal or non-fatal CVD event within 5 years of each eligible encounter, identified using ICD-10-AM codes in linked hospital and mortality data. Predictor variables include routinely collected demographics, chronic conditions, clinical measurements and medications, as well as healthcare utilisation patterns available at or before each encounter. Sex-specific dynamic landmark survival models (such as Cox proportional hazards models) with time-updated covariates (full and least absolute shrinkage and selection operator (LASSO)-regularised) will be developed alongside sex-specific transformer-based sequence models, fine-tuned for 5-year risk prediction at each encounter. Internal validation will use repeated resampling for landmark survival models with patient-level and temporal separation for sequence models. Geographic transportability will be evaluated using internal-external validation across primary health networks (PHNs). Performance will be evaluated using discrimination, calibration, decision-curve analysis and subgroup analyses and will be compared with existing static predictive models. Missing-data will be imputed where necessary: a range of imputation approaches will be compared to optimise computational efficiency but also support predictive generalisability.
Ethics And Dissemination:
This study is conducted under NSW Health governance arrangements with approval from the NSW Population and Health Services Research Ethics Committee (2019/ETH00660). Analyses will be performed within a secure data environment using de-identified linked data. Results will be reported in accordance with TRIPOD-AI guidance and disseminated through peer-reviewed publications, scientific conferences and policy and consumer forums.
