This research is a component of the EFEX-CARE (Effectiveness of Exercise Cardiac Rehabilitation) study, which has been officially registered with the Identifier Record NCT01916525 on ClinicalTrials.gov. The participants in the EFEX-CARE study were drawn from a consecutive series of ACS patients in the Division of Cardiology at the Oulu University Hospital. All the measurements were performed at the Oulu University Hospital. Each participant underwent coronary angiography to confirm the presence of CAD. A comprehensive description of the EFEX-CARE study population has been previously provided (8); however, to summarize, individuals with NYHA class ≥III, those scheduled for or undergoing emergency by-pass surgery, individuals with unstable angina pectoris, severe peripheral atherosclerosis, diabetic retinopathy or neuropathy, or those unable to engage in independent daily physical activities due to musculoskeletal issues were excluded from the study.

In the present study, we will show the health care costs for a 12-month follow-up and baseline risk marker data measured approximately two to 3 weeks after their hospital discharge for patients receiving usual care and participating to international guideline prescribed ECR. Originally the EFEX-CARE study was a randomized controlled trial, where the ACS patients were randomized to the usual care or ECR groups. In the original EFEX-CARE study 109 participants from the exercise training group were involved in the study. Finally, 78 individuals completed the study as planned. Drop-outs from the study were due to a lack of motivation, loss of interest, logistic problems, loss of time mainly because of work duties or health-related problems. Since data from all measured risk markers were needed for the present analysis, data for a total of 71 ECR patients was available. The research adhered to the CONSORT guidelines and was conducted in accordance with the Declaration of Helsinki. The local committee of research ethics for the Northern Ostrobothnia Hospital District approved the study protocol, and all participants provided written informed consent.

We gathered a range of health-related data through different methods. Briefly, body weight and height were measured to assess body composition. Blood pressure was measured in a supine position after a 10-min resting period, following current guidelines. Self-rated depression was assessed using the Depression Scale (DEPS) questionnaire (9). Data on smoking status, alcohol use disorders identification (AUDIT-C) (10), medication, history of acute myocardial infarction, and revascularization were obtained from hospital records and standard questionnaires. Left ventricular systolic function was assessed using 2-D echocardiography (Vivid 7, GE Health care, Wauwatosa, WI, United States). Blood samples were collected after a 12-h overnight fast to analyze plasma glucose, glycated hemoglobin (HbA1c), blood lipids, insulin, and high-sensitivity C-reactive protein using consistent methods at the Oulu University Hospital, Finland. An incremental symptom-limited maximal exercise test was conducted at the Oulu University Hospital using a bicycle ergometer (Monark Ergomedic 839 E, Monark Exercise AB, Vansbro, Sweden) to assess maximal physical exercise capacity (measured in metabolic equivalents: METs). Health-related quality of life was assessed using the 15D questionnaire (11) and completed by patients at the hospital before discharge.

For estimating health care costs, we considered specialized and primary health care services, as well as occupational health care services. Social security ID numbers were used to determine ambulatory care visits, treatment days, and external service usage, with costs calculated based on the Diagnosis Related Groups (DRG) classification. Data for primary health care services, including doctor visits, examinations, and in-ward treatment days, were obtained from electronic health registries using unique social security ID numbers. Information on home care and institutional care (e.g., assisted care homes) was gathered from registries. Additionally, we utilized the report of the Social Insurance Institute of Finland (KELA) (12) to estimate occupational health care service costs. All costs were considered in 2015 values, and no discounting was applied due to the one-year time horizon of the analysis.

The ECR program started as soon as possible after hospital discharge, as suggested earlier (13). The 12-month ECR was planned according to the guidelines (14, 15) consisting of aerobic (30–40 min) and strength exercises (30–40 min) 3 to 5 times per week. The ECR group were invited to the Verve Rehabilitation Center in Oulu, Finland to begin a 12-month ECR program. During the initial 6 months, they attended the Cardiac Rehab gym equipped with aerobic and strength exercise devices (Smart Card system, Ab HUR Oy, Kokkola, Finland), once a week, where they received individual guidance from a physical therapist on both gym and the other exercises performed home-based. After 6 months, home-based ECR continued and only checkpoint visits to monitor the progression of exercise training were scheduled at 9 and 12 months. The intensity of the exercises was stated using the perceived ratings of exertion (RPE) scale from 6 to 20 (16) and targeted to the level between 12 and 15 RPE. Realized training load was calculated from the diaries (RPE x duration of each exercise session) (17). We refer interested readers to Hautala et al. (8) for a detailed intervention description.

In predictive modeling, feature importance scores indicate the significance of input features in predicting the target variable (18). These scores guide feature selection, helping reduce computational costs and potentially enhance model performance. They offer valuable insights for improving predictive models through feature selection (19, 20) and dimensionality reduction (21, 22). Various methods exist, such as those based on statistical correlations and variances, but the choice of method should align with the specific variables and data types, necessitating the evaluation of multiple techniques for suitability. Based on our previous study (7), we performed the feature importance analysis including Cross Decomposition (23), Partial Least Squares Canonical Analysis (PLSC), Partial Least Squares based on Singular Value Decomposition (PLSSVD), Partial Least Square Regression (PLSRegression), Canonical Correlation Analysis (CCA), and Analysis of Variance (ANOVA) test relying on p-values for ranking.

Following the ranking of health care cost risk factors, we conducted a linear regression analysis to predict costs. This involved starting with the top-ranking risk marker and gradually adding the following best markers, resulting in the creation of 13 predictive models. Descriptive statistical analyses, including means, standard deviations (SDs), and proportions where relevant, were performed. We utilized the SPSS software (version 28, SPSS Inc., Chicago, IL, United States) for these predictive data analyses. Statistical significance was established with a p-value threshold of <0.05 for all tests.

Diabetes is one of the fastest-growing global health emergencies of the 21st century. The combination of diabetes with ACS enhances the risk for cardiovascular events emphasizing an interdisciplinary approach for a personalized treatment strategy including exercise training (24) to reduce each patient’s disease burden (25). Diabetes is one of the most influential predictors of health care costs associated with a myriad of complications, including cardiovascular issues, neuropathy, and kidney disease, all of which can substantially increase health care costs. For example, the total estimated cost of diagnosed diabetes in the U.S. in 2022 was 414$ billion, including 307$ billion in direct medical costs and 106$ billion in indirect costs attributable to diabetes (26). We found diabetes as the primary contributing risk factor for the health care costs, followed by body mass index and systolic blood pressure with our ACS patients, although they obeyed carefully exercise training prescriptions and exceeded their prescribed training volume by a significant 51%. It is also important to mention that blood glucose level in terms glycated hemoglobin (HbA1c) and fasting plasma glucose were well balanced according to treatment targets in our patients. Nevertheless, diabetes itself showed to be the main contributor for the health care costs.

In our previous study we tested the applicability of ML tools to predict health care costs in a group of ACS patients treated with usual care only. The chosen methods for feature importance demonstrated their effectiveness in ranking established risk markers, determining the most critical primary targets among them. This aids in identifying key contributors to healthcare costs. We observed that depression, expressed as the higher DEPS score, was the primary contributing factor of health care costs in a 12-month follow-up (7). Interestingly, for the ECR group in the present study, the DEPS score had the lowest explanatory power in predicting health care costs without practically no contribution (0.1%). This is interesting an finding, since not only in terms of health care costs, but understanding that psychosocial risk factors, such as depression, have shown their significance in affecting cardiovascular prognosis, treatment adherence, quality of life, and even sudden cardiac death (27, 28). We also confirmed that the ACS patients involved in usual care did not differ from the ECR group in the present study at baseline in any clinical variables such as diabetes, DEPS score, or any other risk factor, or medications (Supplementary Table S1). Furthermore, it is notable that all ACS patients in the EFEX-CARE study were willing to participate and were at baseline randomized into the ECR and usual care groups.

In addition to medical treatment, exercise is a key component of comprehensive CAD management (5, 14). It is also well-documented that exercise is efficacious in treating depression and depressive symptoms. Therefore, it should be offered as an evidence-based treatment option to individuals with a diagnosis of a major depressive disorder or those with depressive symptoms (29) and similarly with cardiovascular disease patients suffering from mental health disorders (30). It is important to note that the DEPS scale we used in this study is a self-report scale that assesses the severity of depressive symptoms, and it should not be used as the sole basis for diagnosing depression (31). However, it seems that the ECR changes markedly the important risk marker profile measured at baseline compared to the group of patients treated in usual care only when predicting health-care costs in a 12-month time course (Figure 2).

In the present study, we used the selected feature importance methods to find the most preferred first-order targets of risk markers to contribute to health care costs. We also assessed if the order of the leading predictive risk marker will change if we remove one-by-one the risk markers ranked from 6 to 13. Despite excluding the risk markers from the feature importance analysis, the order of the five leading markers remained the same. Therefore, in addition to selected feature importance analysis tools, the selected risk markers included were relevant and valid to the performed analysis.

We believe that the strength of this study lies in its utilization of hospital records to determine the use of health care services, as opposed to relying on patient self-reports, thus minimizing recall bias. Furthermore, the study conducted a comprehensive assessment of baseline patient characteristics, encompassing clinical status, medication usage, thorough laboratory analysis and realized exercise training analysis from diaries. However, it is important to acknowledge a limitation in the study, namely the small and possibly selectively chosen patient sample in the EFEX-CARE study, which may restrict the generalizability to a broader population of ACS patients with significant co-morbidities. Nevertheless every participant in the EFEX-CARE study expressed a willingness to take part and was subsequently randomized into either the ECR or usual care groups.

Our findings demonstrated that when all 13 markers were combined, this model could only predict 31% of the costs. While this may initially appear low, it is worth noting that the proprietary nature of economic data and the diverse sources of health care cost components may partly account for these results. For instance, we were able to analyze direct health care costs but not indirect costs stemming from reduced work productivity due to health issues. An open question remains: can the predictive value be enhanced by incorporating additional variables? This question presents a potential avenue for future research. Additionally, although we assessed various feature importance methods to establish their stability, the relatively small sample size and multicollinearity among risk factors raise some caution in interpreting and generalizing the results. However, we believe that the proposed methodology holds promise for application in medical and health care settings where patient risk profiles and health care costs over a specific period need to be evaluated.