The Stroke Riskometer™ App: Validation of a data collection tool and stroke risk predictor

Background The greatest potential to reduce the burden of stroke is by primary prevention of first-ever stroke, which constitutes three quarters of all stroke. In addition to population-wide prevention strategies (the ‘mass’ approach), the ‘high risk’ approach aims to identify individuals at risk of stroke and to modify their risk factors, and risk, accordingly. Current methods of assessing and modifying stroke risk are difficult to access and implement by the general population, amongst whom most future strokes will arise. To help reduce the burden of stroke on individuals and the population a new app, the Stroke Riskometer™, has been developed. We aim to explore the validity of the app for predicting the risk of stroke compared with current best methods. Methods 752 stroke outcomes from a sample of 9501 individuals across three countries (New Zealand, Russia and the Netherlands) were utilized to investigate the performance of a novel stroke risk prediction tool algorithm (Stroke Riskometer™) compared with two established stroke risk score prediction algorithms (Framingham Stroke Risk Score [FSRS] and QStroke). We calculated the receiver operating characteristics (ROC) curves and area under the ROC curve (AUROC) with 95% confidence intervals, Harrels C-statistic and D-statistics for measure of discrimination, R2 statistics to indicate level of variability accounted for by each prediction algorithm, the Hosmer-Lemeshow statistic for calibration, and the sensitivity and specificity of each algorithm. Results The Stroke Riskometer™ performed well against the FSRS five-year AUROC for both males (FSRS = 75·0% (95% CI 72·3%–77·6%), Stroke Riskometer™ = 74·0(95% CI 71·3%–76·7%) and females [FSRS = 70·3% (95% CI 67·9%–72·8%, Stroke Riskometer™ = 71·5% (95% CI 69·0%–73·9%)], and better than QStroke [males – 59·7% (95% CI 57·3%–62·0%) and comparable to females = 71·1% (95% CI 69·0%–73·1%)]. Discriminative ability of all algorithms was low (C-statistic ranging from 0·51–0·56, D-statistic ranging from 0·01–0·12). Hosmer-Lemeshow illustrated that all of the predicted risk scores were not well calibrated with the observed event data (P < 0·006). Conclusions The Stroke Riskometer™ is comparable in performance for stroke prediction with FSRS and QStroke. All three algorithms performed equally poorly in predicting stroke events. The Stroke Riskometer™ will be continually developed and validated to address the need to improve the current stroke risk scoring systems to more accurately predict stroke, particularly by identifying robust ethnic/race ethnicity group and country specific risk factors.


Introduction
Despite a steady decrease in stroke mortality over the last two decades (1), the global burden of stroke is increasing. Almost 17 million people are affected by stroke every year (68% increase from 1990) and there were 33 million stroke survivors in the world in 2010 (84% increase from 1990), many with disability (2). Unlike 30-40 years ago when most strokes occurred in people aged ≥75 years, now most (>60%) strokes affect people younger than 75 years (2). This, together with the global epidemic of major stroke risk factors (3,4), including diabetes (5) and overweight (6), suggests that the burden of stroke is likely to increase in the future, unless more effective prevention strategies are implemented.
As most (>70%) strokes are first-ever strokes, the prevention of first-ever stroke is a major priority. The two main approaches to the prevention of first-ever stroke are the population-wide 'mass' approach (reducing the level of exposure to stroke risk factors in all people in the region regardless of the individual's level of risk factors), and the individual-based 'high risk' approach. The 'high-risk' aims to identify individuals at risk of stroke (e.g. people with elevated blood pressure, dyslipidaemia, atrial fibrillation and carotid artery stenosis), and to modify their risk factors, and risk, accordingly [current methods of assessing stroke risk include two established stroke risk score prediction algorithms -the Framingham Stroke Risk Score (FSRS) (7) and QStroke (8)]. Although those with high-risk stroke benefit most from prevention strategies, the highest number of strokes and cardiovascular disease occur in people with only a mildly increased risk (9)(10)(11), mainly because there are greater numbers of people in this category of risk [according to Dalton et al. (12), about 90% of UK people aged 40-74 have low 10-year risk of stroke (<20%) as determined by QRisk2]. However, the general population, amongst whom most future strokes will arise, do not readily access and utilize these prediction models; the vast majority of people do not know their risk of having a stroke, do not know their risk factors, and do not know what to do about it (13)(14)(15).
Recent advances in mobile (smartphone) technologies and their worldwide use (about 1·4 billion users) offer unique opportunities to utilize these technologies for improving health and reducing burden from these disorders. Importantly, easily accessible and cost-effective risk-estimation systems are well suited to the developing world and other regions where access to medical facilities is limited (16), including elderly populations where smartphones are being increasingly used (17)(18)(19).
In recognition of the importance of e-research into noncommunicable disease (NCD) initiatives, the United Nations (UN) Economic and Social Council, the International Telecommunication Union (ITU) and the World Health Organization (WHO) have recently (June 2013) launched a new mHealth initiative for improving NCD prevention, treatment and policy enforcement (20). In order to inform and support these UN/ITU/WHO efforts, and to increase general awareness about stroke and its risk factors as well as to improve stroke and NCD prevention on an individual level, The National Institute for Stroke and Applied Neurosciences, AUT University recently developed an app called the Stroke Riskometer TM . This app utilizes recent advances in risk presentation/communication (21,22), international guidelines on stroke and CVD prevention (23)(24)(25)(26)(27)(28) and has the potential to significantly improve stroke and NCD prevention (29). The Stroke Riskometer TM algorithm was derived from the Framingham Stroke Risk Score (FSRS) prediction algorithm (7) and enhanced to improve accessibility and to include several additional major risk factors shown to be important for stroke, largely based on the INTERSTROKE study (4).
Endorsed by the World Stroke Organization, World Federation of Neurology and International Association on Neurology and Epidemiology, the app provides estimates of the absolute risk of stroke within the next 5 and 10 years for individuals aged ≥20 years. Importantly, the Stroke Riskometer TM provides not only their absolute risk of stroke development but also a baseline risk for comparison, thus allowing users to compare their risk of stroke with someone of the same age and gender who has no risk factors. The former represents a new paradigm for high-risk stroke prevention strategy (29), and enables a refined presentation of the traditional threshold-based approach in which people are categorized into low, moderate, and high-risk groups. This procedure enables not only those at high levels of risk, but also those at low-to moderate absolute risk, to reduce their risk of stroke. The app therefore allows a combination of both high-risk and population strategies, an approach shown to be the most effective for cardiovascular disease prevention (11).
The aim of this study was to compare the performance of the Stroke Riskometer TM prediction algorithm with two other commonly used stroke prediction algorithms -Framingham Heart Study Stroke Risk Score (FSRS) prediction algorithm (7) and QStroke (8).

Study design and data sources
Three study populations (80 308 person-years of observation in total) were used to validate the Stroke Riskometer TM algorithm: the Auckland Regional Community Stroke (ARCOS IV) 2011-2012 study (30), the Rotterdam Study (1990 -ongoing) (3,31), and Russian Cohort studies (1992 -ongoing; Dr M Kravchenko, unpublished data).
The ARCOS study is a population-based stroke register where all new stroke events (both hospitalized and nonhospitalized, fatal and nonfatal) in almost 1·2 million Auckland adult residents were prospectively ascertained using multiple overlapping sources of the information, including hospital admissions/referrals, community general practices and death certificates etc. (details of the study methodology have been described elsewhere) (30). For the purpose of the validation of the Stroke Riskometer TM we used a sub-set of ARCOS IV data on strokes in people aged 21-95 years (n = 410).
The Rotterdam Study has been described previously (3). It is an ongoing prospective population-based cohort study that focuses on the causes and consequences of chronic and disabling diseases in the elderly (31). The cohort started enrolment in 1990 and included 7983 participants aged ≥55 years living in Ommoord, a district of the city of Rotterdam in the Netherlands (participation rate 78%). Follow-up was complete until January 1, 2012, for 97·1% of potential person-years (32). The Rotterdam study contributed data from n = 7713 individuals who ranged in age from 55-90 years.
Russian cohort studies were conducted in Moscow (n = 412), Ulyanovsk (n = 512), Nal'chik (n = 177) and Minsk (n = 277) over various time periods starting from 1992. Study participants (men and women; age range 39-66 years) were followed up from 12 years (Moscow) to four-years (Ulyanovsk, Nal'chik and Minsk). The World Health Organization stroke diagnostic criteria (33) were used and a diagnosis of stroke was confirmed by a study neurologist across all these studies (over 90% of stroke patients had brain neuroimaging to establish a pathological type of stroke). All these studies have been approved by the local Ethics Committees.

Stroke risk factors and algorithm development
Risk scores from three stroke predictors were generated. Each scoring algorithm utilized a series of known or hypothesized stroke risk factors (Table 1), some of which are in addition to those used in the FSRS and are the central targets in the new WHO Global Action Plan for the NCD 2013-2020 (34). Distribution of each risk factor for each data set is listed in Table 2. The Stroke Riskometer TM algorithm was derived from the Framingham Stroke Risk Score (FSRS) prediction algorithm (7) but enhanced to include several additional major risk factors shown to be important for both ischaemic and haemorrhagic strokes, largely based on the INTERSTROKE study (4). The additional variables are listed in Table 1. Questions were based on recall such as 'Have you ever been told by a doctor that you have atrial fibrillation (irregular heartbeats)?' and 'Have you ever been told by a doctor that you have left ventricular hypertension?' such that no immediate medical test (e.g. an ECG is required) in order for users to provide an answer. These questions have been used and validated in cross-sectional studies (4). Beta-coefficients for each additional variable were derived from current literature and discussed amongst by a panel of stroke and health experts of the Stroke Riskometer TM Collaboration. Based on these discussions and available evidence, the following risk scores were added to the FSRS (7) risk score: 0·20 for being non-Caucasian (23,35), 0·20 for poor diet (i.e., consuming less than six servings of fruits and vegetables per day) (4), 0·10 for high alcohol consumption (i.e., consuming two or more standard drinks per day) (4,36,37), 0·10 for low physical activity (i.e., less than 2·5 hours per week) (15,23), 0·05 for family history of stroke or heart attack (23,38-41), 10 (for 5-year risk) and 15 (for 10-year risk) for previous stroke or transient ischaemic attack (TIA) (42), 1·80 for any cognitive problems and 1·40 for memory problems but no cognitive issues (43), 1·20 for previous traumatic brain injury (44), 0·20 plus 0·10 for any unit (0·01) increase in waist-to-hip ratio above 0·96 for males and 0·80 for females (45). In the absence of waist-to-hip ratio data we used BMI and scored 1·02 plus 0·10 for every unit (1 kg/m 2 ) above 24 kg/m 2 for Chinese, or above 23 kg/m 2 for South Asians or above 25 kg/m 2 for all other ethnicities (46) [different cut-off criteria for Chinese people were based on recommendations from the Chinese National Centre for Cardiovascular Disease (W. Wang, personal communication)]. In the absence of both waist-to-hip ratio and BMI data, waist circumference measures can be used adding 1·02 per unit (1 cm) above 103 cm for males and 89 cm for females (45). As each of the additional risk factors was added to the algorithm separately without taking into account interactions between the risk factors, we applied conservative beta-estimates to reduce the chance of overestimating the stroke risk (47,48). Algorithm testing prior to the app launch used a number of different methods. A very large number of hypothetical cases (many hundreds of different combination of risk factors) were entered into the tool to identify problems requiring resolution before clinical use. The tool then underwent clinical evaluation by stroke experts and general practitioners to compare the estimated 5-year and 10-year risk.

Alcohol consumption
More than 2 standard drinks per day.

Stress
Significant stress as determined by the patient. Diagnosis of anxiety or depression.

Low physical activity
Less than 2·5 hours per week.

Poor diet
Less than six servings of fruit and vegetable per day = 1, More than or equal to six servings of fruit and vegetables per day = 0 Cognitive problems or dementia    Research P. Parmar et al.

Algorithm validation
The performance of the Stroke Riskometer TM was tested across three data sets (ARCOS, Russian and Rotterdam) as greater precision is gained when assessing risk prediction models using multiple epidemiologic studies compared to single-studies (49). We also compared performance of the Stroke Riskometer TM with the FSRS (7) and QStroke (8) risk score equations. The five-year estimated risk of stroke for Russian and Rotterdam cohorts was calculated across the three different prediction algorithms. Estimates for 10-year stroke risk score were generated only for the Rotterdam study where data were available over a span of 10 years. Follow-up data for the ARCOS were limited to one-year and for Russian data sets -4 to 12 years. We calculated Harrels C-statistic and Somer's D-statistic to measure discrimination (the ability of the algorithms to discriminate between stroke and nonstroke events). C-statistic values of 0·50 represent chance and 1 denotes the ability of the risk score to discriminate perfectly. D-statistics over 0·10 indicate that the risk score has a good ability to differentiate between an event and nonevent. Receiver operating characteristics (ROC) curve, Area Under the ROC Curve (AUROC) with 95% confidence intervals within each data set, sensitivity and specificity of each algorithm were also analyzed. R 2 statistic was calculated to indicate the level of variability accounted for by each prediction algorithm. Calibration was assessed using the H-L test (for goodness-of-fit statistics to examine differences between the observed and predicted risks from each algorithm) All analyses were performed in R (version 3·0·2) (50).

Validation cohorts
A total of 752 new strokes that developed in a sample of 9501 individuals over the follow-up period (80 308 person-years of observation) across three studies (ARCOS, Russia and Rotterdam) were utilized to investigate the recently derived stroke risk prediction tool algorithm Stroke Riskometer TM against two established stroke risk score prediction algorithms [FSRS (7) and QStroke (8)]. The three data sets differed in their distribution of stroke outcomes and predictor variables required for each algorithm. The ARCOS data set was comprised of stroke only data whilst the Russian database was generated through a new cohort study, with 3·2% total strokes being observed. Of the Rotterdam study, 8·4% was comprised of strokes. None of the three studies had all variables required for the Stroke Riskometer TM algorithm. The Russian data set was the most recent of the three data sets analyzed here so the average age was lowest (50 years for males and females; Table 2). Individuals in ARCOS and the Rotterdam study were similar in age (males 69 years and females 72 years). The Russian data set had the lowest average systolic blood pressure (SBP) while ARCOS had the highest. Both the Rotterdam and ARCOS studies had similar SBP values for males and females whilst the Russian data set had higher values in males. BMI was not recorded in ARCOS but was similar in males and females of the Russian data set (average for males = 27·8 kg/m 2 and females = 27·5 kg/m 2 ) and comparable with males and females in the Rotterdam study (average for males = 25·6 kg/m 2 and females = 26·7 kg/m 2 ). As BMI was not recorded in ARCOS waist circumference was used (for the Stroke Riskometer TM algorithm); women had greater waist circumference (99 cm) than men (97 cm). Due to inclusion of only patients with stroke, the ARCOS database had the highest percentage of individuals with diabetes (20%) compared to the Russian data set (range 4·4-4·8%) and Rotterdam (range 6·1-7·1%). A much higher proportion of the ARCOS database had individuals with a history of CVD, previous stroke/TIA event and were of non-European descent, compared to the Russian and Rotterdam cohorts ( Table 2).
Validation and overall performance of the Stroke Riskometer TM As none of the three studies had all variables required for the Stroke Riskometer TM algorithm, we cannot fully validate this algorithm with the emphasis for the continuing development of the Stroke Riskometer TM algorithm. We present measures of overall performance, discrimination and calibration of the Stroke Riskometer TM algorithm based on available data. The FSRS and Stroke Riskometer TM algorithms gave comparable 5-year and 10-year risk scores for males and females within each data set (Fig. 1). Risk scores differed substantially by data set, reflecting the availability of predictors within each cohort. Each algorithm (FSRS, Stroke Riskometer TM and QStroke) explained 50% of the variation observed in the ARCOS data set (R 2 statistic, Table 3). With fewer stroke outcomes in the Russian and Rotterdam data sets, the reported R 2 was low across all cohorts for all algorithms, ranging from 0·31-5·22% (Table 3).

Discrimination
All three algorithms showed poor discriminative ability across each cohort (C-statistic range 0·50-0·53, D-statistic <0·05, Table 3). The ROC curves (Fig. 2) show that the FSRS and Stroke Riskometer TM algorithms behaved similarly for 5-year and 10-year risk scores for males and females, with area under the ROC curves ranging between 61% and 66% in the Rotterdam cohort (Fig. 2, Table 3). The QStroke algorithm outperformed the FSRS and Stroke Riskometer TM algorithms (Table 3).

Classification, sensitivity and specificity
Mean predicted stroke risk scores were on average higher in the group of observed stroke outcomes compared to individuals with no stroke outcome ( Supplementary Fig. S1). Sensitivity and specificity was calculated for FSRS, Stroke Riskometer TM and QStroke predicted risk scores, which reached accuracy threshold of 50%, 70%, 80%, 85% and 90% ( Table 4). The predicted risk scores were then categorized into 'High' risk (based on reaching 80% accuracy and >80% specificity, Table 4) and 'Low' otherwise, which were compared against each other (Table 5). Sensitivity for QStroke in males was low (10·6% for accuracy = 80%) compared to FSRS and Stroke Riskometer TM which had a sensitivity = 53·9% (FRSR) and 52·3% (Stroke Riskometer TM ) for accuracy = 80%.
In the Russian database we observed that both the Stroke Riskometer TM and FSRS algorithms classified most participants as high risk (63·6% five-year risk in males and 78·4% five-year risk in females). As ARCOS had all stroke events we would expect these to predominately to be categorized as 'High risk' this is observed for FSRS and Stroke Riskometer TM (males = 98·1% and females = 86·4%). A very high proportion of individuals in the ARCOS data set were classified as high risk for Stroke Riskometer TM but low risk on QStroke (males = 97·2% and females = 48·1% for five-year risk). A high proportion of females in the Rotterdam study were categorized as low-risk for Stroke Riskometer TM and high-risk for QStroke (13·2% for five-year risk in females) compared to 5·5% of males classified as low-risk for Stroke Riskometer TM and high-risk for QStroke (Table 5).

Calibration
Calibration plots of the predicted risk scores against the observed event for each tenth of predicted risk, separately for males and females, are shown in Fig. 3 (all data sources) and Supplementary  Fig. S2 (Russian and Rotterdam cohorts). The Russian cohort illustrated that the QStroke algorithm was better calibrated for   Somer's D-statistic to measure discrimination (the ability of the algorithms to discriminate between stroke and nonstroke events). C-statistic values of 0·50 represent chance and 1 denotes the ability of the risk score to discriminate perfectly. D-statistics over 0·10 indicate that the risk score has a good ability to differentiate between an event and nonevent. AUROC the data set of all three algorithms, for a database with few strokes (Supplementary Fig. S2). An improved calibration for the FSRS and Stroke Riskometer TM algorithms compared to QStroke was observed for the Rotterdam data set, particularly for females. The QStroke algorithm was shown to over-estimate stroke risk in females whilst FSRS and Stroke Riskometer TM over-estimated stroke risk in males (Supplementary Fig. S2). Visual assessment of five-year risk scores from the combined data (ARCOS, Russia and Rotterdam) highlighted that the Stroke Riskometer TM algorithm was better calibrated compared to QStroke, especially for females (Fig. 3). All predicted risk scores were not well calibrated to our data sets (Table 6, H-L tests P < 0·006).

Discussion
The Stroke Riskometer TM is comparable in performance to two widely used stroke risk scoring systems. The variation found in our study may be due to several factors. The higher R 2 values for ARCOS are explained by the high number of stroke outcome data available. Many variables required for the QStroke algorithm were not available within the ARCOS data set (such as rheumatoid arthritis, chronic kidney disease, Table 2) therefore it is likely that the QStroke risk scores we observed under-estimate stroke risk, particularly amongst males in ARCOS. A large proportion of females in ARCOS were classified as high risk in both Stroke Riskometer TM and QStroke scoring (>50% , Table 5). Conversely, the Rotterdam study had a more complete data set of variables required for the QStroke algorithm calculation (Table 2), however this appears to have led to over-estimation of the stroke risk amongst females (Fig. 1, Supplementary Fig. S2C) and an inconsistency across predicted risk scores with 13% categorized as lowrisk for Stroke Riskometer TM and high-risk for QStroke (Table 5). Sensitivity was low for the QStroke risk scores generated for males (10·6% for 5-year and 8·9% for 10-year risk scores) and females (35·9% for 5-year and 36·7% for 10-year risk scores, Table 4), when specificity was high (= 80%, Table 4) compared to the sensitivity for FSRS and Stroke Riskometer TM for males (53% for 5-year and 10-year risk scores) and females (42% for 5-year and 10-year risk scores, Table 4) for FSRS, and (45% for 5-year and 10-year risk scores, Table 4 for Stroke Riskometer TM , when specificity was high (= 80%, Table 4). The developers of QStroke have previously highlighted that their algorithm over-predicts stroke risk in females (8). It should also be noted that QStroke was developed for predicting ischaemic stroke specifically, and not for predicting any type of stroke as developed for Stroke Riskometer TM and FSRS.   Whilst the discriminative abilities of all three algorithms across all data sets appeared to be comparable, they were also very low (C-statistic ranging from 0·51-0·56, D-statistic ranging from 0·01-0·12). H-L calibration statistics illustrated that all of the predicted risk scores did not align well to observed event data, P < 0·006. This may be due to the QStroke risk score algorithms being developed from UK-based data and while the data sets being utilized here are predominately European, they were not UK-based individuals. The FSRS has been externally validated in several different European cohorts but with inconsistent result, some studies attaining appropriate levels of discrimination but over-estimation of risk of stroke (51), however other studies have shown the FSRS has poor discrimination and under-estimates stroke risk (52). QStroke was recently created and validated in a subset of the British cohort data used to develop their algorithm and showed good levels of discrimination; however the authors did acknowledge a tendency to overestimate female stroke risk (8). In a large cohort of black and white adults the FSRS overestimated observed stroke rates, particularly in certain ethnic subgroups where the FSRS suggested there should be approximately twice as many strokes occurring than was detected (53).
This indicates that there is still a need to improve current stroke risk scoring systems to more accurately predict stroke risk across different populations/countries. We have shown that there is a level of overlap in the variables considered in these algorithms, however it may be that the weights assigned to each risk factor need to be generated to be country/or ethnic-specific as some risk factors may hold more importance in some groups compared to Table 5 Comparing the scoring of the three risk score algorithms as 'High' or 'Low' risk for Framingham Stroke Risk Score (FSRS), Stroke Riskometer TM and the Qstroke algorithm across all validation cohorts (ARCOS, RUSSIA and ROTTERDAM). Thresholds for 'High' risk in each algorithm for males and females was selected for 80% accuracy and >80% specificity (Table 4  others (54). It is also likely that there are further unknown stroke risk factors that still need to be identified and included in a stroke prediction assessment tool. For one such example we refer to the recent evidence from Yusuf et al. (55) that populations from low to middle-income countries are at highest risk of cardiovascular events have the lowest risk factor burden (55), suggesting that the major 'missing piece in the equation' of the effective CVD prevention is the impaired ability of resource-limited health systems to effectively identify and modify cardiovascular risk. It is our expectation that the Stroke Riskometer TM will be further developed to account for these factors (we are currently collecting data on country) such that in future iterations of the Stroke Riskometer TM we hope to refine the algorithm to be able to provide country and ethnic specific-stroke risk prediction estimates, using both current research such as Yusuf et al. (55) and data collected from the current Stroke Riskometer TM App to improve overall predictability and applicability of our algorithm across all populations. Furthermore, an algorithm for all major noncommunicable disease, such as stroke, ischaemic heart disease (IHD), dementia and diabetes mellitus that share common risk factors, should be developed and validated in different populations. The main weakness of this validation study was that analyses were restricted due to the lack of currently available data on the variables shown to be important determinants of stroke. The Stroke Riskometer TM availability on a portable device (smartphone) with constant proximity to the user, enables individuals to assess their own risk of stroke in the privacy and comfort anytime, anywhere. Unlike web-based versions, no internet connection required to use the app or access its information. In addition, the app offers a higher level of interactivity via sending direct reminders to the smartphone that is always on hand when needed. Moreover, the availability of the app on the smartphone app stores that has global reach, and vast consumer base of various age groups allows wide range of consumers to benefit from the stroke risk assessment tool and allows the crowdsourcing of large research database. Finally, users who are at increased (even slightly increased) risk are provided with ways to reduce their risk of stroke according to their individual risk profile and recommended to seek medical attention. This could rapidly transform epidemiologic research and monitoring of health status of individuals, especially in the area of chronic NCD (17).
Current risk scores will inevitably become outdated with improvements in clinical outcomes and data recording and changes in population demographics (56). With the Lite version of the Stroke Riskometer TM being made freely available globally on both iOS and Android smartphones and users invited to partake in a large-scale study we will have the potential to amass a large database. Ethical approval for the study has been received. Anonymous data from individuals who consent to participate in the study will be collected and securely stored at study coordinating centre (AUT University, NZ). The aim of these planned epidemiological studies based on the Stroke Riskometer TM will be to generate a global, population-specific stroke and NCD risk scoring system. We will further assess the Stroke Riskometer TM in a cohort study to establish the efficiency of the algorithm and assess if the new collections of recommendations are useful for motivating users to actively reduce their risk of stroke.