Background and Objectives: Diabetes mellitus (Type 2 diabetes) is a significant and rapidly growing global healthcare challenge and a leading cause of severe complications, including cardiovascular disease (CVD), chronic kidney disease and other conditions. In Latvia, more than 103 000 individuals have been registered in the Latvian Diabetes Patient Register, of whom over 90% have type 2 diabetes. Since early diagnosis can reduce complication risk, considerable effort has been devoted to the development of numerous CVD risk prediction tools using traditional statistics as well as data science-based methods. This article augments these efforts with predictive models built specifically for Latvia’s diabetes patient population. Materials and Methods: Models used anonymized diabetes patient data stored at Latvia’s Center for Disease Prevention and Control, containing routinely collected information on patients’ utilization of publicly funded healthcare services. No clinical data were included. In keeping with best practices in model development, 96,302 qualifying available patient records were randomly partitioned into a 70% training data portion and 30% for out-of-sample validation. We tested three modeling methodologies—logistic regression, Random Forests, and XGBoost. Results: The binary dependent variable was defined as the occurrence of an ischemic heart disease hospitalization during 2023 or 2024. The predictor variables were limited to data from the prior two-year period, 2021–2022. The three model types all produced similar fit to the data—area under the Receiver Operating Curve (ROC) of 0.69–0.71 in the validation data, which is comparable to results obtained in other studies, including SCORE2. Prediction reliability, as measured by the Brier Score, was 0.11. Conclusions: We find ischemic heart disease hospitalization risk prediction models based on routinely collected Latvian health care utilization data to be feasible. Such a model could be applied to identify diabetes patients with an elevated (30–40%) risk of experiencing an ischemic heart disease-related hospitalization over the next two years. In a retrospective validation cohort, the highest risk decile had an observed two-year hospitalization rate above 30%. An illustrative projection of a similarly sized decile is about 8000–10,000 patients, but this estimate is exploratory. The model is not ready for clinical implementation; prospective studies are required to test calibration, clinical utility, and benefit. We suggest future research to evaluate whether integrating the model into primary care can lead to earlier interventions and a reduction in preventable ischemic heart disease hospitalizations.
Background and Objectives: Diabetes mellitus (Type 2 diabetes) is a significant and rapidly growing global healthcare challenge and a leading cause of severe complications, including cardiovascular disease (CVD), chronic kidney disease and other conditions. In Latvia more than 103 000 individuals have been registered in the Latvian Diabetes Patient Register, of whom over 90% had type 2 diabetes. Since early diagnosis can reduce complication risk, considerable effort has been devoted to the development of numerous CVD risk prediction tools using traditional statistics as well as data science-based methods. This article augments these efforts with predictive models built specifically for Latvia’s diabetes patient population. Materials and Methods: Models used anonymized diabetes patient data stored at Latvia’s Center for Disease Prevention and Control containing routinely collected information on patients’ utilization of publicly funded healthcare services. No clinical data were included. In keeping with best practices in model development, 96 302 qualifying available patient records were randomly partitioned into a 70% training data portion and 30% for out-of-sample validation. We tested three modeling methodologies – logistic regression, Random Forests, and XGBoost. Results: The binary dependent variable was defined as the occurrence of an ischemic heart disease hospitalization during 2023 or 2024. The predictor variables were limited to data from the prior two year period, 2021-2022. The three model types all produced similar fit to the data – area under the Receiver Operating Curve (ROC) of 0.70-0.71 in the validation data, which is comparable to results obtained in other studies, including SCORE2. Prediction reliability, as measured by the Brier Score, was 0.11. Conclusions: We find ischemic heart disease hospitalization risk prediction models based on routinely collected Latvian health care utilization data to be feasible. Such a model could be applied to identify diabetes patients with an elevated (30%-40%) risk of experiencing an ischemic heart disease-related hospitalization over the next two years. While the actual size of Latvia’s diabetes patient population in the near future is not precisely known, we estimate that this highest risk decile may contain a group of 8 000 – 10 000 diabetes patients. We suggest future prospective studies to evaluate whether integrating the model into primary care can lead to earlier interventions and a reduction in preventable ischemic heart disease hospitalizations.
Various predictive modeling methodologies utilizing machine learning models are making their way into the practice of public health and medicine. In this article a numerical forecasting method called the XGBoost algorithm is tested on a dataset of anonymized health care data records for the purpose of assessing cardiovascular disease (CVD) mortality risk in a future 12-month period, using only data from the previous year. The data originates from Latvia’s Centre for Disease Prevention and Control and contains over 116 000 anonymized observations from persons who were living as of January 1, 2018 and may have utilized government-subsidized medical services in 2018. CVD case fatality in 2019 occurred among 13% of the sample. The machine learning mortality risk model was based on six input variables (age, sex, the number of out-patient visits and in-patient hospital stays, prescriptions filled, and ambulance calls) and produced an out-of-sample Concordance (AUC) measure of 0.90, which is comparable with the accuracy of cohort-based SCORE2 models recommended for use in Europe today. Potential model applications could be in public health risk quantification, future capacity and load balancing estimates for medical facilities, as well as patient stratification for individual physician practices. An earlier English language version of this article is available upon request from the author: usprudzs@chicagobooth.edu.