The Stroke-Data collection was performed as a prospective national multicenter study at 2 hospitals in Finland: Oulu University Hospital (OUH) and Kuopio University Hospital (KUH), during the period between October 2021 and December 2022. The data collection was performed by 2 study nurses in both locations and assisted by a research assistant in Oulu. The study was based on direct observations, and the aim was to collect information from healthy control participants and patients who have had a stroke or TIA. Before the actual data collection started, the nurses rehearsed the study protocol and use of the data collection devices. In addition, the personnel in both hospitals were informed about the study.

To be included in the main study group, the patient had to be admitted to a neurological or neurosurgical ward at the hospital, be 18 years and older, and been diagnosed with a cerebrovascular accident or TIA within the last 0-3 days according to the International Classification of Diseases (ICD-10: I60-69 and G45; ICD-9: 430-438; and ICD-8: 430-438). Healthy people were included in the control group, if they were 18 years and older and had no history of CVD or other major chronic diseases expected to affect physiological or sensor-based measurements. The use of blood pressure and cholesterol medications was permitted. People who had other long-term medication, smoked, or had other addictions were excluded from the study. This was done to establish a relatively homogeneous control group without major behavioral or vascular risk factors that could influence physiological and sensor-based measurements and act as confounding factors in multimodal signal analysis. Prisoners or forensic patients, patients with reduced capabilities, and pregnant or nursing women were excluded from all study groups.

The stroke and TIA classifications were based on the final clinical diagnosis established during the participants’ hospital stay, in accordance with the Finnish Current Care Guidelines. Stroke diagnoses were confirmed using clinical assessment and neuroimaging findings. Participants whose final diagnosis was revised to a condition other than stroke or TIA were excluded from the final cohort (n=5). Healthy controls were recruited according to the predefined eligibility criteria and screened by the study nurse using a questionnaire covering medical history, chronic diseases, medication use, smoking status, previous cerebrovascular events, and other relevant health conditions. One control participant was excluded from the final cohort because a previous CVD was identified. All analyses were performed using these final classifications.

The patients for the study were recruited at OUH and KUH by the study nurses with the assistance of health care professionals, immediately after the patients had been diagnosed with a cerebrovascular accident or TIA, upon arrival to the hospital, and after they had received the most acute treatment and were treated in the ward at the time of recruitment. The recruitment happened within the first 3 days after hospital admission.

The healthy controls in Oulu were recruited by the study nurse from the OUH staff email distribution list, whereas in Kuopio, they were recruited by the study nurses, with the help of newsletter and wall announcements, from service centers for older people, community colleges, libraries, the Alzheimer Society of Finland, and pensioners’ associations. The recruitment of healthy controls aimed to achieve a similar age and sex distribution to that of patient groups; however, this was not fully achieved due to recruitment challenges.

The study was conducted according to the principles of the Declaration of Helsinki and approved by the ethics committee of North Ostrobothnia Hospital District (22/2021; current name: the regional medical research ethics committee of the Wellbeing Services County of North Ostrobothnia). The study was registered and approved as a clinical investigation of medical devices by the Finnish Medicines Agency Fimea (FIMEA/2021/003738). In addition, the trial was approved by the administration of OUH and KUH. Written informed voluntary consent was obtained at the time of recruitment. The study protocol was explained to the participants. The participants were given a written description of their rights, the purpose and nature of the research, the methods used in it, and possible risks and harm (Act on Medical Research 488/1999). Participants had the right to withdraw from the study without a reason at any stage without affecting their care, treatment, or services in any way. Two copies of the consent forms were signed, one of which remained with the patient him- or herself, and the other was kept in a prearranged lock cabinet on the hospital premises by the study nurse. A personal study ID was used for all participants to pseudonymize the data.

The data collection protocol included data collection from the patient data register, clinical assessments and questionnaires, and data collection with devices. Follow-up telephone interviews were conducted 3 months after hospital discharge and included the assessment of the modified Rankin Scale (mRS), patient-reported outcome measures, and patient-reported experience measures. The data collection with the devices, assessments, and questionnaires was similar for all the study groups. However, for healthy controls, neither the patient register data were collected nor the follow-up telephone interviews made. The patients were recruited between October 2021 and September 2022, and the healthy controls between October 2021 and December 2022. The data collection, including the follow-up telephone interviews, was performed between October 2021 and December 2022. The duration of individual recordings varied between participants because data collection was performed according to a standardized protocol but adapted to participants’ clinical condition and the time required to complete the assessment tasks.

After receiving the most acute treatment, patients with stroke and TIA were enrolled in the hospital ward after discussing with health care professionals, and data collection was started either on the same or the following day. The data collection took altogether 1-2 hours per participant, but if needed, the data collection was split into 2 or 3 separate sessions during the same day or 2 consecutive days. The health condition of patients was considered while collecting the data, and if the health condition worsened, the data collection was either cancelled or postponed. All assessments were conducted within 3 days of hospital admission. Exact symptom-onset times, ambulance-call times, and treatment-to-assessment intervals were not systematically available for the entire cohort. Therefore, detailed intervals between symptom onset, hospital admission, acute treatment, and individual measurements were not analyzed. The data collection protocol is illustrated in Figure 1.

The data collection of the healthy control participants was performed during 1 session at an agreed time slot. The health condition of healthy control participants was considered while collecting the data, and if their health condition declined, the data collection was either cancelled or postponed. The data collection protocol is illustrated in Figure 2.

The collected data described in Table 1 were transferred in both hospitals to network drives accessible only by dedicated study nurses. After the data had been collected, the data were transferred in a pseudonymized format to a private cloud server at CSC ePouta [38], which is a service for researchers to analyze sensitive data. The University of Oulu IT department deployed a virtual machine to the ePouta service, and all analysis of the collected data was performed with this ePouta virtual machine.

The data access inside the virtual machine was managed with Linux user groups following the minimum access principle so that researchers had access only to that part of the data which they needed to for their specific research. Each data type had its own user group, and users were assigned to the groups according to the research permit of the study. The machine had 12 central processing unit cores, 240 GB RAM memory, and 1.5 TB storage space. Additionally, the virtual machine also had an Nvidia Tesla A100 40 GB graphics processing unit for image and video data analysis and machine learning studies.

There are approximately 500-600 patients with CVD yearly in KUH, and their average stay at the hospital is 4 days. The yearly number of patients with CVD in OUH is 600-700, and their average stay at the hospital is 3.5 days. The patients who received active treatments like thrombolysis or thrombectomy come to a control visit at the ward after 3 months in both hospitals. The number of these patients is approximately 150-200 yearly per hospital. The sample included all patients matching the inclusion criteria during the study period so that the study nurses had enough time to recruit and plan the data collection for 1 patient per day. This meant that during a 6-month period it was possible to collect data from 50 to 100 patients per hospital. In addition, the aim was to collect data from 100 to 200 healthy controls. A formal statistical power calculation was not performed, as the primary aim of the study was to establish a comprehensive multimodal dataset, assess the feasibility of multimodal data collection in an acute clinical setting, and provide a foundation for subsequent methodological development, classification, and predictive modeling studies rather than to test a specific hypothesis. The sample size was therefore based on feasibility, available resources, and the expected number of eligible patients during the study period.

We assessed normality of the continuous variables with the Shapiro-Wilk test. As age and number of diseases were not normally distributed, we used the Kruskal-Wallis test to evaluate differences between the controls, patients with TIA, and patients with stroke. To determine which groups differed statistically significantly, we conducted the Dunn post hoc test with Benjamini-Hochberg correction. To assess group differences in the categorical variables, we used the Fisher exact test. In all analyses, we set the significance level to .05. We performed the analyses using R (version 4.4.1; R Foundation for Statistical Computing), Python (version 3.8.17; Python Software Foundation), and Python packages numpy (version 1.24.2) [40], pandas (version 1.5.3) [41], pingouin (version 0.5.5) [42], scipy (version 1.10.1) [43], and scikit-posthocs (version 0.8.1) [44].

Given the multimodal nature of the dataset, data quality considerations differ across modalities and intended analyses. Therefore, quality control procedures, exclusion criteria, preprocessing methods, and handling of missing values will be specified separately for each analytical study. Measurements with substantial artifacts, technical failures, or incomplete recordings may be excluded from the corresponding analyses. The dataset is intentionally maintained as close to the original recordings as possible to allow different research questions and analytical approaches, including statistical and machine learning methods, to apply modality-specific data curation strategies as appropriate.

As this paper presents the study protocol and cohort description, the statistical analyses reported here are limited to characterization of the study population and comparison of participant groups. However, the multimodal dataset was specifically designed to support subsequent data-driven analyses aimed at improving CVD diagnosis, treatment response assessment, recovery prediction, and identification of risk factors and biomarkers. In addition to descriptive and group comparison analyses, the multimodal dataset may enable the development of data-driven models for classification and prediction tasks. These may include differentiating between stroke, TIA, and healthy controls, as well as predicting clinical outcomes such as recovery based on the mRS. Feature extraction from sensor-based data, including time-series, image, and video data, can be combined with clinical and questionnaire data to develop supervised machine learning models. As this is a protocol study, specific modeling approaches and performance evaluations will be reported in subsequent studies. These may include supervised machine learning models (eg, classification and regression models), including multimodal data fusion approaches integrating physiological, clinical, and behavioral features.

This study describes the design and implementation of a comprehensive multimodal data collection protocol aimed at supporting improved diagnosis, outcome prediction, and biomarker identification in CVDs.

Cerebrovascular accidents are critical neurological emergencies that require immediate diagnosis and treatment since they may have devastating consequences on the quality of life of the patients. Furthermore, accurate differential diagnostics between stroke and TIA is important so that the appropriate treatments for these different accidents can be made efficiently. Therefore, novel methods for faster and more accurate diagnosis of patients with CVD are needed.

This study describes the design and cohort characteristics of the Stroke-Data study, which aimed to address the need for prompt diagnosis by investigating the physiological markers from patients with CVD and controls comprehensively and combining information collected from different sources in a new way. More specifically, we successfully collected EEG, ECG, NIRS, gait, fundus imaging, and video recordings of neurological examination and combined them with clinical assessments, questionnaires, and EHR data.

The strength of the study is its uniqueness, since, to our knowledge, there are no other studies that have applied this large number of different modality measurements to the same participants. Studies in the literature collected only 1 modality for stroke detection, such as EEG signals [45-47], fundus imaging [48-50], and neurological examination video analysis [28,51]. Another strength of the study is that it included both patients with stroke and TIA, in addition to healthy controls.

The study has several limitations. First, the number of participants was relatively low, as collecting comprehensive multimodal data from patients undergoing a life-threatening condition requiring immediate diagnosis and care is challenging. Second, the study does not include longitudinal follow-up of healthy participants and therefore does not enable prediction of first-ever stroke or TIA events or prospective risk modeling. Third, there were differences in characteristics between healthy controls and disease groups due to the challenges in recruiting comparable participants. Although the original aim was to recruit healthy controls with a similar age and sex distribution as the patient groups, exact matching was not achieved because of recruitment challenges. Specifically, the healthy controls included a higher proportion of female participants and were more highly educated, generally younger, and more often employed than participants in the disease groups. These differences may introduce confounding effects when comparing groups and may influence observed associations in both statistical analyses and data-driven models. Consequently, the findings should be interpreted with caution, particularly when generalizing results or developing predictive models based on group comparisons. In addition, the relatively small size of certain subgroups, such as patients with TIA and hemorrhagic stroke cases, as well as incomplete data across some modalities, may limit the development and generalizability of predictive models and should be addressed in future larger-scale studies. Future studies should aim to include better-matched control groups or apply statistical approaches to adjust for these differences.

Some of the technologies used in this study, such as EEG and NIRS, are resource-intensive and require trained personnel, which may limit their direct applicability in routine clinical practice, particularly in acute stroke settings. These modalities were included to provide comprehensive physiological information and to support methodological development. Future research should evaluate which subset of modalities provides sufficient diagnostic or predictive accuracy while remaining feasible for implementation in clinical environments.

Although initial findings from the dataset have already been published, this protocol paper is the first to provide a comprehensive description of the study design and multimodal data collection framework. This documentation is important for transparency, reproducibility, and interpretation of both existing and future studies using the dataset.

To conclude, we successfully collected multimodal data from healthy controls and patients with TIA and stroke, and the data can be used for studying stroke- and TIA-related biomarkers, disease progression, recovery, and differences between patient groups and healthy controls, as well as for creating novel methods that support early and accurate diagnosis of CVDs. With these multimodal data, data-driven prediction models can be developed, which aim to identify the patients or groups of patients who are most likely to benefit from certain treatment or rehabilitation. Simultaneously, data sources and features can be identified to predict the patient’s recovery. This enables health care resources and procedures to be targeted more precisely, so that the treatment result benefits the patient, and the resulting treatment and rehabilitation costs can be optimized at the societal level.