Big Data/Machine Learning/AI
Predicting Ischemic Heart Disease Risk Using Family Health Histories from Electronic Healthcare Records Amani F. Hamad* Amani Hamad Joykrishna Sarkar Linda Ejlskov Oleguer Plana-Ripoll Lisa M. Lix
Objectives: Family health histories may improve the prediction of diseases, including ischemic heart disease (IHD). We tested whether parental disease histories from population-based administrative healthcare databases improved IHD risk prediction in two regions: province of Manitoba (MB) in Canada, and Denmark (DNK).
Methods: A retrospective cohort study of individuals 40 years old with linkages to ≥1 parent between 1974 and 2022 in MB and 1978 and 2023 in DNK. Incident IHD diagnoses and predictors were identified from inpatient and outpatient records. Predictors were selected using LASSO logistic regression models with 10-fold cross-validation. The base model included demographics and parental IHD history. Subsequent models included individuals’ and their parents’ disease histories (130 chronic conditions). Models were evaluated using area under the receiver operating characteristic curve (AUC) and prediction error.
Results: The cohort comprised 118,868 individuals in MB and 774,845 individuals in DNK. 2.5% and 1.5% had an IHD diagnosis during 10-year follow-up in MB and DNK cohorts, respectively. In MB, the base model had an AUC of 63.8%, 95% confidence interval (CI) 60.8-66.9. AUC improved after including individuals’ diseases (69.7%, 95% CI 66.7-72.6); but not after including parental diseases (67.2%, 95% CI 63.0-71.4). Similar trends were observed in the DNK cohort. All models had minimal prediction error, with Brier scores between 0.01 and 0.03. In both regions, individuals’ sex, and diabetes mellitus and hypertension diagnoses were the most important predictors; diagnoses of migraines, substance use disorders, gastritis and ulcers were also important. The most important parental predictors in MB and DNK were diabetes mellitus and IHD diagnoses, respectively.
Conclusions: Disease histories of individuals, but not of their parents, improved IHD risk prediction. Several important predictors were identified that could further improve IHD risk prediction models.