Auxiliary Data as Surrogates for Systematic Coronary Risk Evaluation Model Version 2 Risk Calculator Inputs: Design and Methodology of the Frauenherzen Study in Apparently Healthy Women Aged 40–69 Years

Abstract

Objectives

The aim of the study was to investigate whether data collected from wearable devices, such as heart rate, step count, and sleep-related parameters, can reduce uncertainty in cardiovascular disease (CVD) risk prediction among apparently healthy women aged 40–69 years when blood pressure and blood lipid values required for the SCORE2 risk calculator are unavailable. A secondary objective was to identify wearable-derived features that contribute most to reducing uncertainty in CVD risk prediction.

Design

This was a monocentric, exploratory, cross-sectional observational study.

Participants/Materials

Women aged 40–69 years with and without cardiovascular risk factors were invited to participate. Eligible participants were biologically female, fluent in German, owned a smartphone, and provided written informed consent. Women with diagnosed CVD, diabetes mellitus, chronic kidney disease, or familial hypercholesterolemia were excluded. The study uses clinical assessments, blood sampling, validated questionnaires, and 7-day continuous monitoring with a Garmin® Vivosmart 5 wearable device.

Setting

The study was conducted at the Department of Obstetrics and Gynecology, University Hospital of Bern, Inselspital, Switzerland.

Methods

Participants attended two study visits scheduled 7 to 10 days apart. At baseline, anthropometric measures, blood pressure, pulse, waist circumference, blood samples for lipid profile and HbA1c, and questionnaire data on medical history, nutrition, menopausal symptoms, psychosocial health, anxiety, depression, stress, sleep, and quality of life were collected. Participants wore a Garmin® wearable device for 7 days, generating data on heart rate, physical activity, sleep, and related digital biomarkers. Statistical analyses will explore probabilistic approaches to impute systematically missing SCORE2 input parameters, particularly blood pressure and blood lipids, using auxiliary variables such as age, BMI, and wearable-derived features. Model performance will be evaluated by comparing probabilistic CVD risk predictions with SCORE2 estimates based on complete clinical input data.

Results

As this manuscript describes the study design and methodology, no outcome results are reported. The planned sample size is 250 participants, accounting for an estimated dropout rate of approximately 10%.

Limitations

The exploratory design and monocentric setting may limit generalizability. The sample size may be limited for selecting predictors from high-dimensional wearable data, and identified surrogate features will require external validation. In addition, adherence to wearing the device and potential recording or transmission errors may affect the completeness and quality of wearable data.

Conclusions

The Frauenherzen study provides a methodological framework for evaluating whether wearable-derived digital biomarkers can support CVD risk prediction in middle-aged women when conventional clinical parameters are missing. By combining clinical, questionnaire-based, and wearable data, the study aimed to inform future sex-specific cardiovascular risk stratification and the development of accessible preventive digital health tools.

Introduction

Background and Rationale

Menopause typically occurs at an average age of 51 and is characterized by a gradual decline in ovarian reproductive hormone production [1]. Menopause increases the risk for chronic non-communicable diseases (NCDs) in women, including cardiovascular disease (CVD), type 2 diabetes mellitus (DM), cancer, osteoporosis, and dementia [2, 3]. There is considerable uncertainty regarding whether an apparently healthy woman (without CVD, DM, chronic kidney disease [CKD], or familial hypercholesterolemia [FH]) may still have an elevated risk for CVD [4].

Furthermore, metabolic syndrome leads to an increased risk of CVD and DM [5]. It is diagnosed when three or more factors are simultaneously dysregulated, including insulin resistance, obesity, elevated blood pressure (BP), dyslipidemia, and dysglycemia [6]. Postmenopausal women with metabolic syndrome are at an especially elevated CVD risk [6, 7]. However, the causal relationship between menopause and the development of metabolic syndrome is not fully understood [7].

Clinical prediction tools for assessing individual CVD risk are available but their use is limited by the need for inputs that require a physician visit and invasive assessments. For instance, the European Society for Cardiology (ESC) provided guidelines and the online risk calculator Systematic Coronary Risk Evaluation model version 2 (SCORE2) in 2021 [8]. To estimate a woman’s CVD risk, this specific risk calculator requires mandatory input parameters, including age, blood lipid levels (HDL and total cholesterol), systolic BP, and both personal and family history of CVD. The result is a percentage representing the 10-year CVD risk, which can be classified into low to moderate (<5%), high (5 to ≤10%), or very high (>10%).

Wearable device data could help predict CVD risk in cases where traditional, physician-recorded parameters (BP and blood lipids) are unavailable or incomplete. Wearable data, together with other easily obtained information such as BMI, could potentially be used to surrogate missing physician-recorded parameters. For example, metrics such as the average daily step were found to be associated with various CVD risk factors including physician-recorded parameters such as blood lipids [9, 10]. With this study, we aimed to investigate the potential of wearable data, such as the heart rate or the daily step count, to reduce uncertainty regarding underlying CVD risk in apparently healthy women in the absence of BP and two blood lipid measurements, three required input parameters of the SCORE2 risk calculator, which typically require an in-person visit to the physician.

Objectives

The primary objective of this study was to assess whether the data collected from wearable devices, such as heart rate, step count, or sleep patterns, can reduce uncertainty in predicting the CVD risk in women aged 40–69 years in a situation where information about the BP and blood lipids is unavailable. The second objective was to identify specific wearable signal-derived features that best contribute to reducing uncertainty in predicting CVD risk in the absence of information on BP and blood lipids.

Study Design

This paper describes the design and methodology of a monocentric, exploratory, cross-sectional, observational study.

Methods and Analysis

Study Setting

This study was conducted at the Department of Obstetrics and Gynecology at the University Hospital of Bern (Inselspital) in Switzerland. The study design and methodology was developed according to the STROBE Checklist (online suppl. File 1; for all online suppl. material, see https://doi.org/10.1159/000552881).

Eligibility Criteria

Inclusion criteria to participate included female as biological sex, aged 40–69 years at the time of signing the informed consent, fluency in German, ownership of a smartphone and signing the informed consent. Exclusion criteria were a diagnosis of CVD, DM, CKD, or FH; inability to provide informed consent; employment in the Gynecological Endocrinology and Reproductive Medicine section of the Department for Obstetrics and Gynecology at the Inselspital, Bern, or any other professional relation to one of the study project leaders/investigators, participating in a clinical intervention study at the same time; inability to comply with the study procedures for any reason including language comprehension, psychiatric illness, general inability to get to the study site, smartphone incompatibility with Fitrockr, and technical inability of the participant’s smartphone to connect to the smartwatch.

Informed Consent

Written informed consent or assent was obtained by the participant’s physician or trained medical professionals as authorized surrogates from all participants prior to enrollment.

Study Procedures

Interested women were contacted by the study team via telephone to verify eligibility according to the inclusion and exclusion criteria. The participants attended two visits, scheduled 7 to 10 days apart, during which they wore a wearable device (Garmin^® Vivosmart 5). The Garmin^® Vivosmart 5 is a fitness tracker designed to monitor a range of health and activity parameters. Key features include continuous heart rate monitoring, sleep tracking, and physical activity tracking, offering comprehensive data to support the analysis conducted in the study. Its long battery life ensures reliable and continuous data collection for extended periods, making it well suited for research and health monitoring.

At the baseline visit, participants provided written informed consent prior to any study-related procedures. Subsequently, they completed standardized questionnaires in Research Electronic Data Capture (REDCap^®) regarding their medical history, nutrition, menopausal symptoms, life experiences, anxiety, and depression. BP, pulse, waist circumference, body height, and weight were measured, and blood samples were drawn for the analysis of lipid profiles and HbA1c levels. Participants also installed the Fitrockr application on their smartphones, received the wearable devices, and connected them to their smartphones with guidance from the study team.

Participants wore the wearable devices for 1 week while completing a work and sleep journal. After this period, they returned for a second visit, where they filled in another standardized questionnaire about their sleeping habits and stress over the last 4 weeks, and returned the wearable device. Upon completion of these tasks, participants received a travel allowance of 20 Swiss francs.

Participant Timeline

After inclusion in the study, participants attended two visits spaced 7 to 10 days apart, during which they continuously wore a wearable device. The participant timeline is outlined in Table 1. Participant recruitment for the study began on 24 August 2023.

Table 1.

Schedule of activities

Visit name	SCR+BL (1st)	Last (2nd)
IP	IP	IP
Visit days	0	7a (to max. 10)
Screening
Inclusion and exclusion criteria	X	​
Subject information and informed consent	X	​
Medical information
Demographics, personal/family medical history	X	​
Validated psychometrics	X	X
Vital sign: height, weight, calculated BMI, waist circumference, systolic/diastolic BP, pulse measurement. BP measurements to be performed up to four times (in the same visit) by alternating the arms, in case of high variation in the initial reading and a recommendation to verify this medical aspect with the treating general practitioner will be made to each such case/participant	X	​
Blood sample draw, for blood lipids (HDL, total cholesterol, and triglycerides) and HbA1c laboratory analysis: to be performed only once for each study participant	X	X, if not performed at the 1st visit)
Dispensation of the wearable, the participant’s Garmin®-Fitrockr troubleshooting guide, of the participant’s word and sleep journal, study’s flyer	X	​
Training in use of Garmin®, Fitrockr, and REDCap®	X	​
PROM and COA {12}
Wearable raw signal and derived data (Table 2)	→ continuously	​
Safety
Concomitant medications	X	X
Study exit
Collection of wearables, participant’s reimbursement, and documenting the eventual participant’s study-related feedback	​	X
Signing of the anamnesis sheet, to confirm the above	X	X

BL, baseline; COA, clinical outcome assessment; PROM, participant-related outcome measure; (S)AE, (serious) adverse event; SCR, screening; V, visit; IP, in-person.

^aWearable will be continuously worn for 7 days, but V1 can take place up to 10 days after BL.

Modifications

No modifications were anticipated, as participants were only required to wear a watch, and this procedure is noninvasive. However, if a participant exhibited an allergic reaction to the material of the watch, they would need to be excluded from the study.

Adherence

During the visits, a member of the study team will always be present to address any questions participants may have while completing the standardized questionnaires. This approach ensures the completeness and accuracy of the participants’ responses. To maintain consistent data transmission, participants will be reminded via the Fitrockr Hub application if they forget to synchronize their wearable devices daily.

In the event of technical issues with the wearable device or other inquiries, participants will be able to contact the study team via email. Issues will be addressed and resolved through email correspondence or by telephone consultation.

Sample Size

Preliminary analyses were conducted using a synthetic dataset based on data collected at the gynecological department of the Inselspital in Bern, Switzerland [11]. The rationale for this approach was to simulate a scenario in which a patient cannot visit a physician. In this simulation, physician-recorded data, such as BP and blood lipids, were omitted, and CVD risk was estimated based on the remaining variables and easily accessible auxiliary information, including age and BMI.

These preliminary analyses were repeated for differently sized subsets of the original coded dataset and assessed the loss of predictive performance as the sample size was decreased. Among all repetitions, a sample size of 200 represents the minimum acceptable sample size to maintain adequate predictive performance.

To account for a modest reduction in predictive performance due to the inclusion of wearable data and an estimated dropout rate of about 10%, we aim for a sample size of 250 participants. We are aware that the high-dimensional nature of wearable data and the need to identify appropriate surrogates of risk calculator inputs from a vast set of wearable-derived features place higher demands on sample size than the mentioned analysis using only small set of auxiliary variables. However, this is an exploratory study and identified surrogates will require external validation.

Recruitment

The women were mainly recruited at the Inselspital, Bern, through the Intranet, study portal of the Inselspital, digital flyers presented on the Inselspital’s TV screens, and printed flyers distributed throughout the hospital. At the first visit, participants received a flyer to share with interested acquaintances. Interested women were able to register via a QR code, through which they provided their personal details for contact purposes and completed the inclusion and exclusion criteria to pre-assess their eligibility.

Data Collection, Management, and Analysis

Quality Measures

In order to assure quality of this study, only well-trained personnel conducted the examinations. For quality assurance, the Ethics Committee could visit the research sites. On such occasions, direct access to the source data and all project-related files and documents will be granted. All involved parties were instructed regarding measures to ensure that the participant data are kept strictly confidential. There was no patient and public involvement.

Data Management

Each participant’s personal identifying information was linked to a unique study code, which was password-protected and stored in a secure institutional storage server space. The study code was used to link wearable data in Fitrockr, record information in the electronic case report form (CRF) in the REDCap^® system, and label blood samples.

Participant-provided data were collected through CRFs, implemented electronically using a dedicated electronic data capture system (REDCap^®). All data entered in the CRFs were stored on a Linux/Windows server in the dedicated PostgreSQL database. The responsibility for implementing and hosting the electronic data capture system rested with Clinical Trials Unit (CTU), Bern.

Wearable data were collected using Fitrockr. The data were transmitted directly to Fitrockr’s cloud servers, which are hosted in a professional ISO-27001-certified hosting center that complies with the highest security and data protection standards. Upon completion of data collection, the software and all associated data will be permanently deleted from the Fitrockr servers.

Data from Fitrockr and REDCap^® will be securely transferred to the Institute of Social and Preventive Medicine (ISPM) at the University of Bern for processing and merging, using secure socket layer encryption. Participant codes will serve as unique identifiers for data merging. The merged dataset and a compressed Binary JavaScript Object Notation (BSON) database dump from Fitrockr will be transferred to Inselspital servers for long-term storage. The participant codes were hashed using the SHA-256 algorithm before the data became accessible to the data analysis team.

Statistical Methods

Our objective was to explore the potential for accurately predicting SCORE2 risks when some key clinical input parameters, namely, BP and blood lipids, are systematically missing. We propose utilizing auxiliary self-reported data, such as age and BMI, alongside selected features extracted from smartwatch-derived signals, such as heart rate, daily step count, and sleep patterns to impute probabilistically the missing inputs. We will then propagate this uncertainty through the risk calculator, resulting in probabilistic predictions of CVD risk. To evaluate the effectiveness of our methods, we will use cross-validation to compare these predictions with the values obtained from the SCORE2 risk calculator when all inputs are available. This assessment employs scoring rules such as the Continuously Ranked Probability Score (CRPS) and Brier Score.

Our first aim was to replicate the findings of Mühlemann et al. [12], demonstrating that probabilistic methods perform better than deterministic imputation techniques and that incorporating BMI as a predictor enhances performance. The second aim was to demonstrate that integrating wearable data can improve CVD risk predictions and reduce uncertainty. For this purpose, we extract specific indicators from these wearable signals that may be related to participants’ cardiovascular health. We anticipate that using the information obtained from wearable devices may help compensate for missing clinical parameters.

We consider the probabilistic prediction models used in Mühlemann et al. [12], such as Gaussian copula and multiple imputation by chained equations, as baselines. In addition, we develop methods based on other types of copulas, Gaussian copulas derived by regression models, Student copulas, Gaussian mixture copulas, or Vine copulas. Since we will consider a large set of indicators derived from the smartwatches as compared to the sample size, we will also employ dimension reduction and/or regularization techniques.

As per our secondary objective of identifying wearable-derived features that will best contribute to reduce uncertainty in CVD risk prediction without BP and blood lipid data, we will explore statistical and machine learning methods capable of handling irregularly sampled time-series biosignal data listed in Table 2. Features refer to numerical values derived from raw signals, such as mean heart rate during deep sleep.

Table 2.

Garmin® parameters (left column) and literature providing the link between those parameters and either blood lipids, BP, or CVD risk (right column)

Parameter	Clinical link to lipids, BP, or CVD risk
Physical activity (steps), number of steps/day, and cadence (step/min)	[13–15]
Sleep (REM, non-REM, wakening), minutes per phase (awake/REM/light sleep/deep sleep)	[15]
Tracks active minutes manually or after a few automatically	[10]
Physical activity (high intensity, cardio fitness, fat burning, sedentary time)	[10]
Stress management index score from 1 to 100 (higher score means fewer physical signs of stress)	[16, 17]
Resting heart rate, bpm	[6, 18]
SpO2, %	[19]
RF in breaths per minute	[20, 21]

bpm, beats per minute; SpO₂, oxygen saturation; RF, respiration frequency.

Handling of Missing Data

Our imputation approaches primarily deal with systematically missing covariates, namely, blood lipids and BP, although other parameters may be missing for some participants. A key advantage of our probabilistic imputation approach is its ability to directly accommodate instances of sporadically missing variables. In these cases, it is sufficient to extract a restricted version of the full distribution and derive conditional distributions based only on the available data.

If a participant withdraws their informed consent form after the study is completed, their data will remain included in analyses that have already been conducted. However, their data will not be used in any future analyses. In case of missing data, the recorded participants’ data will be used as far as possible.

Monitoring

Data Monitoring

This research project was reviewed by independent monitors to ensure data integrity, participant safety, and adherence to the study design and methodology. All parties involved in the review process are obligated to maintain strict confidentiality and ensure the protection of participant privacy at all times.

Discussion

The primary aim of the study was to assess the potential of data collected from wearable devices, such as heart rate or step count, to reduce the uncertainty in predicting the CVD risk among women aged 40–69 years in the absence of information on BP and blood lipid levels. Additionally, the study aimed to identify specific wearable-derived features that contribute to reducing uncertainty in the CVD risk prediction in the absence of information on BP, blood lipids, and the presence or absence of metabolic syndrome.

Menopause, which typically occurs at an average age of 51, is associated with increased risk for chronic NCDs. However, uncertainty persists regarding CVD risk in apparently healthy women without a history of CVD, diabetes, CKD, or FH. This uncertainty highlights the importance of finding accurate and easily accessible methods that enable individuals to assess CVD risk.

Although clinical tools, such as the ESC’s SCORE2 calculator, are available to assess CVD risk, their use remains limited. A key limitation is their reliance on BP and lipid values, which may not be readily available or easily measured in nonclinical settings. Integrating wearable devices to collect relevant data could offer a promising alternative for predicting CVD risk, particularly when physician-recorded inputs, such as BP and blood lipids, are incomplete or unavailable. Many individuals already include wearable devices into their daily lives, making these tools not only accessible but also highly accepted by the general population. Data obtained from wearable devices may provide more complementary information for missing parameters than traditional reliance on BMI and age alone.

We hypothesize that, in the absence of BP and blood lipid values, variables measured by wearable devices, such as heart rate and daily step count, may improve the accessibility and usability of CVD risk prediction. The study is conducted in accordance with ethical guidelines and has received approval from an Ethics Committee, ensuring participant safety, data integrity, and adherence to the highest standards of research ethics. Informed consent is obtained from all participants, ensuring their full understanding of the study, their voluntary participation, and their right to withdraw at any time.

The target number of participants was set based on the minimum sample size needed to avoid a considerable loss in prediction performance based on previous analyses using synthetic datasets and accounts for anticipated participant dropout. However, the sample size is limited for variable selection from high-dimensional wearable data. This is an exploratory study, and any findings will require validation in an independent sample. Also, the study relies on participants consistently wearing the wearable device throughout the study period. However, lack of compliance and recording or transmission errors may reduce the usefulness of the collected wearable data for predicting CVD risk.

The use of wearable devices, which are increasingly used by individuals, offers a convenient and easily accessible way to collect various health data that could be used to assess CVD risk. By utilizing this technology, the study provides an opportunity to develop a method that could enable individuals to independently calculate their own CVD risk.

A future objective was to develop a digital health application that integrates data from wearable devices to enable personalized risk predictions for the development of NCDs, with the ultimate aim of supporting disease prevention in menopausal women. Furthermore, the implementation of this approach could establish a framework for addressing additional health-related research questions, thereby broadening the scope of wearable technologies in preventive medicine. Future research should focus on validating this method in larger, more diverse populations and integrating it into user-friendly digital health platforms to enable widespread applicability.

In conclusion, this study represents an important step toward integrating wearable technology into clinical risk prediction models. By utilizing easily accessible health data, the study aimed to provide a more accurate and straightforward method for predicting CVD risk, potentially contributing to the development of personalized healthcare approaches that promote prevention.

Ethics and Dissemination

Amendments

There were no substantial changes to the project design, the design, and methodology.

Confidentiality and Coding

Project data will be handled with utmost discretion and are accessible only to authorized clinical personnel who require access for study purposes. Participants are only identified by a unique participant number on the CRFs and other project-specific documents. All data transferred for the study occur under secure socket layer encryption.

Access to Data

The data analysis team does not have access to any identifying information. In addition, study IDs will be hashed before the data are disseminated to the analysis team to further safeguard the participants’ identities. The analysis data files will be stored on secure servers at ISPM and be made accessible exclusively to authorized members of the analysis team. Only individuals authorized to treat the data in the context of this project will be granted access to the corresponding folders. The personally identifiable information of the participant and the association with the participant code entered in Fitrockr and REDCap^® are accessible only to the clinical study team.

Ancillary and Posttrial Care

Since no interventions were being implemented as part of this study, no additional medical treatments or follow-up measures are planned. In the event of incidental findings of elevated BP, participants were informed and provided with appropriate guidance on subsequent steps. Should a participant be identified as having an increased CVD risk, as determined by the SCORE2, calculated after the completion of the study’s data analysis, the participant will be notified following the assessment period, unless they have indicated in the informed consent form that they do not wish to be informed.

Retention and Destruction of Study Data and Biological Material

All study data will be archived for 10 years after study termination or premature termination of the study. Any remainder of the analyzed blood samples will be destroyed after analysis. In accordance with regulatory requirements, the project leader is responsible for ensuring the retention of all trial-related documents and for maintaining all subject documents after the completion of the clinical trial, ensuring their availability for inspection by the relevant authorities.

Publication

The study results will be published in several peer-reviewed journals with different focuses (e.g., medical, mathematical, engineering). Coded data (participant code for the data collection team and hash code for the analysis team, but no study participants’ identifiers) will be shared within the study team. Master students and statistical consultants will have access to the data files with hashed IDs and only to the specific data they need to write their thesis.

Statement of Ethics

This study involving human participants was reviewed and approved by the Cantonal Ethics Committee of Bern, Switzerland (Decision No. 2022-02080; approval date: 22 February 2023). All procedures were conducted in accordance with the Declaration of Helsinki and its later amendments. Written informed consent to participate in the study was obtained from all individual participants prior to enrollment.

Conflict of Interest Statement

The authors declare that they have no conflicts of interest.

Funding Sources

This study was supported by a grant from the Center for Artificial Intelligence in Medicine (CAIM). The funding body had no role in the design of the study, data collection, analysis, interpretation of data, or in writing the manuscript.

Author Contributions

Elena Pavicic: writing – original draft and investigation. Miriam Strasser: writing – original draft. Patric Wyss: methodology. Serena Lozza-Fiacco, Manuela Moraru, and Danielle V. Bower: investigation. Rowan Iskandar: data curation. Petra Stute: conceptualization, methodology, supervision, and funding acquisition.

Funding Statement

This study was supported by a grant from the Center for Artificial Intelligence in Medicine (CAIM). The funding body had no role in the design of the study, data collection, analysis, interpretation of data, or in writing the manuscript.

Data Availability Statement

The datasets generated and analyzed during the current study are not publicly available due to ethical and legal restrictions but are available from the corresponding author on reasonable request and with permission of the responsible Ethics Committee where required.