Real-world data for precision public health of noncommunicable diseases: a scoping review

We conducted our scoping review using systematic methodology (search strategy, selection criteria, data extraction and synthesis) according to Joanna Briggs Institute (JBI) guideline methodology for ‘systematic scoping reviews’ to ensure rigour and replicability [23]. Recent commentary has justified the growing role of scoping reviews when applied to the rapidly advancing and transdisciplinary nature of digital health and medicine [24]. We adhered to the Preferred Reporting Items for Systematic reviews and Meta-Analyses Extension for Scoping Reviews (PRISMA-ScR) checklist [25] (see Supplementary File 1). Our review protocol was made publicly available online with OSF Registries [26].

A search strategy was created in consultation with a research librarian and three researchers (OC, ZK, CS), two of whom are subject matter experts (OC, CS). The search strategy was developed and implemented using the PCC (Population, Concept and Context) format [25]:

The search strategy was developed in PubMed (see Table 2) and then translated to Scopus, EMBASE, IEEE Xplore, ACM Digital Library and Google Scholar for articles published in English from inception up to February 2021 (Supplementary File 2 presents the full search strategy for each database). We chose not to define NCDs in the (P) Population category due to the scoping nature of the review and as many NCDs are addressed singularly (e.g. cardiovascular disease (CVD), type 2 diabetes). Instead, articles were screened for application to NCDs according to the eligibility criteria.

After removal of duplicates via Endnote (phase 1) and Covidence (a web-based screening and data extraction tool for reviews) [27] (phase 2), title and abstract screening, full-text review and conflict resolution was conducted by two independent reviewers (OC, ZK) using the Covidence platform according to pre-defined eligibility criteria. A snowballing approach was implemented to identify potential additional articles – the reference lists of all articles that underwent full-text review were independently screened by two reviewers (OC, ZK).

Table 3 presents the inclusion and exclusion criteria for the present review. We consolidated some types of real-world data in the framework proposed by Swift et al [15]. to reflect their typical inclusion in a single data asset e.g. an EMR. We defined real-world data as routinely collected (1) Clinical, Medication and Family History (e.g. EMR) (2) Claims/Billing (3) Mobile Health (e.g. wearables/sensors/applications) (4) Environmental (e.g. lifestyle factors, retail data) (5) Social media (e.g. Twitter) (6) Molecular profiling (e.g. genomics) (7) Patient-centred i.e. EHR. Traditional data were defined as collected through specialised methods designed by the user, such as via surveys, registries, websites and administrative databases.

NCDs were defined according to aggregated definitions from the World Health Organisation (2013) [28] (excluding communicable diseases) to include:

A data extraction tool was customised in Covidence and specified according to the purpose of our scoping review. The following data were extracted from all included articles: (a) primary author and year of publication (b) country (c) aim (d) setting (e) population characteristics (f) target noncommunicable disease/s (g) types of real-world data (h) real-world data (i) traditional data (j) refresh frequency (k) aggregation of real-world data (l) analytics and (m) implementation. We performed two methods of data synthesis: (a) a narrative, descriptive synthesis and (b) mapping each included study to the three horizons roadmap for precision public health for NCDs [13] (see Fig. 2). Our discussion comments on the strengths and limitations of included studies and describes implications of these for advancing the field of precision public health for NCDs.

The search yielded 2,005 articles from the databases PubMed, Scopus, EMBASE, IEEE Xplore, ACM Digital Library and Google Scholar (see Fig. 3 for the PRISMA flow diagram). After article duplicates were removed, 1,862 remained for title/abstract screening. A total of 44 articles were identified for full-text review. Six studies met our inclusion criteria and were included in the present review. Four of these studies were identified through the screening process and two were identified at full-text review via snowballing reference lists of screened articles.

Table 4 presents the characteristics of studies included in this review. Five studies were conducted in high-income countries; three were conducted in USA [30‐32], two in Canada [33, 34] and one in China, uan upper-middle-income country [35]. Broadly, the objective of included studies was to describe the development of a population health surveillance platform. Studies aimed to (a) describe population surveillance methodology for people with CHD [30] (b) develop an EHR-based public health information exchange [31] (c) create a socially-generated health data model for disease surveillance (via Social InfoButtons) [32] (d) track the NCD epidemic in China (NCDCMS) [35] (e) describe development of a distributed model (CCDSS) for chronic disease surveillance in Canada [33] and (f) develop PopHR, a big data platform for population health surveillance in Canada [34].

Three of six studies were conducted in a partnered academia (university) and healthcare (health centres, department of health) setting [30, 31, 34]. Two studies were conducted in government settings [33, 35] (public health agencies) and one was conducted exclusively in academia (university) [32].

The target population for included studies was large and varied. Three studies leveraged their sample from large geographical areas (Ningbo City [35], all Canadian provinces and territories [33], Montreal [34]) with populations ranging from 1 to 6 million persons. Two studies drew their samples from clinical health services (via EHRs) that ranged from 73,000 to 5 million persons [30, 35]. One study drew their sample from social media and government and health websites and quantified the sample in terms of patients, posts and reviews (~ 160,000 total) [32].

Five studies described digital aggregation of real-world data and traditional data to support targeting of 28 NCDs (e.g. asthma, type 2 diabetes mellitus, arthritis, cancer, chronic kidney failure, chronic obstructive pulmonary disease, mental illness, obesity). One study focused exclusively on congenital heart disease (CHD) [30].

Table 5 presents the characteristics, aggregation and application of real-world data and traditional data to support precision public health for NCDs. Four included studies leveraged ‘Clinical, Medication & Family Hx’ as the primary real-world data type [30, 31, 34, 35]. The data source generating this real-world data were organic and included an electronic health or medical record (EHR/EMR) [30, 31, 35] or public health insurance provider [34]. Data aggregation then occurred with real-world types of claims/billing [30, 34] and environmental (retail transactions) [34], and with traditional data via the country-specific Census [31, 34] (lifestyle factors, environmental characteristics), a disease registry (CHD) [30], electronic clinical history database [35] and administrative hospital discharge abstracts [33]. The social media platform ‘Twitter’ was aggregated with traditional health data via government and health websites and bibliographic databases (Centers for Disease Control, PubMed, WebMD, MedHelp) in Social InfoButtons [32].

Four of six included studies performed a static cross-sectional extraction of real-world data and traditional data to build their population health surveillance platform [30‐33]. These studies used EHRs, claims databases, health information systems, registries, Census, social media and government and health websites and administrative hospital datasets. The NCDCMS—a public health data exchange platform to track NCDs in China—was updated using real-time (1-day refresh frequency) data extracted from an EHR and electronic clinical history database [35]. PopHR performed near real-time updates that varied in refresh frequency (2 weeks to 1 year) depending on the type of real-world data (insurance provider, retail transactions) or traditional data (Census) source [34].

Methods for digital aggregation of real-world data and traditional data were heterogenous across all studies. Static extracts of real-world data were aggregated via an external, independent vendor (Fine-Grained Records and Linkage Tool) and hosted in Microsoft Access [30], using a public health information network (PHIN) data system [31], via semantic web technology in a digital platform (Social InfoButtons) [32] or data reconciliation of pre-aggregated data (Canadian CCDSS) [33]. Real-world data were aggregated in real-time via public health data exchange with uniform data standards (NCDCMS) [35]. In PopHR, real-world data and traditional data were aggregated in near real-time using server-client architecture to execute data processing, integration and semantics [34].

The primary end-users of platforms developed by the included studies were public health professionals and government. PopHR was implemented with test users in public health and health service agencies to (a) verify software (b) generate rapid feedback and (c) conduct usability testing [34]. Social InfoButtons described government as an end-user to assist in disease surveillance and increase awareness of social health trends [32]. The CCDSS was implemented by the Public Health Agency of Canada and produces data reports, publications and interactive web-based open data resources [33]. The NCDCMS has been implemented in five cities in Zhejiang Province to improve public health surveillance, although the exact mechanism of this was unclear [35]. In Guilbert et al [31]., static summary reports were produced via their EHR-based HIE to describe demographic, clinic and community disease characteristics. One study transferred their deidentified and deduplicated surveillance dataset for CHD to the Centers for Disease Control for review [30].

Five of six included studies applied population health analytics to their aggregated real-world data and traditional data [31‐35]. Geospatial analytics were implemented to geocode populations by disease characteristics [31, 32, 34, 35] and conduct spatial regression to geographically contextualise disease characteristics [31]. Descriptive analytics were applied across four platforms to describe disease prevalence according to multi-levelled stratification conditions (via bar charts, data tables and scatter plots) [31‐34]. Temporal analytics were applied via time series to indicate population or geographical changes in prevalence over time [34], or to compare treatment efficacy in populations over time [32]. Inferential analytics were performed in the HIE by Guilbert et al [31]. to consider multivariate queries and mine relevant data characteristics. One study did not report use of analytics [30]. No studies incorporated predictive or prescriptive analytics.

Three horizons (see Fig. 4) for digital health transformation towards precision public health for NCDs have been proposed as a strategic roadmap to guide digital public health investment, decision-making and policy [13]. One study aligned with Horizon 1 ‘Digital public health workflows’; Glidewell et al [30]. linked real-world data and traditional data and hosted the aggregate data on Microsoft Access to provide the digital foundation for improved health surveillance of people with CHD.

Five studies aligned with Horizon 2 ‘Population health data and analytics’. Shaban-Nejad et al [34]. integrated geospatial, descriptive, temporal and comparative analytics to chronic disease surveillance and tested PopHR in a real-world public health and health service agency setting. The social analytic tool Social InfoButtons provided the government with enhanced disease surveillance capability through geocoding and descriptive analytics of social and disease topic discussion online [32].

No studies aligned with Horizon 3 ‘Precision public health’.

Our systematic scoping review identified six global examples of how real-world data and traditional data is being aggregated to support precision public health for NCDs. We discovered three key findings relevant to our research question:

Traditional decision models in public health rely upon data collected specifically for decision-making, intervention and policy for NCDs. Implementation of health information technology platforms such as EHRs is the first step to revolutionising public health by creating clinical infrastructure capacity for near real-time population health surveillance. An advanced global example is the Electronic Medical Record Support for Public Health (ESPNet) surveillance platform [17] that is the digital infrastructure for RiskScape – a data aggregation and visualisation platform that updates monthly to heat map disease prevalence and perform descriptive and time series statistics for various communicable (e.g. chlamydia, influenza) and noncommunicable (e.g., type 1 and type 2 diabetes, obesity, hypertension, asthma) diseases across approximately 20% of the Massachusetts (USA) population [12]. The real-world application of this platform in the Massachusetts Department of Public Health has identified NCD hot spots and risk factor targeting to inform prevention program design [12].

EHRs formed the foundation of population health surveillance systems in three of our included studies [30, 31, 35]. Variable time interval data extracts were aggregated with other real-world data (e.g. insurance providers, retail transactions) and traditional data (e.g. Census, disease registry). Recent systematic reviews have focused on the challenges (data quality, unstructured data) [19] and barriers (missing data, limited interoperability standards) [20], and solutions (natural language processing, technical solutions for data retrieval, clinical and public health collaboration) [19] and facilitators (data quality and management, preventive care) [20] to meaningful use of EHRs for surveillance. Barriers such as missing data, lack of interoperability and bias towards populations receiving care [17] may be overcome by leveraging additional real-world data to enrich the host surveillance platform.

Leveraging EHR data is an important transformative step to progressing precision public health for NCDs. EHR-based surveillance will require augmentation with other types of real-world data and traditional data from other sectors to treat the inherent challenges of use at a population scale. Aggregation of real-world data (e.g. insurance, retail transactions) and traditional data (e.g. Census) was performed by PopHR [34] and Guilbert et al [31]. to augment surveillance of asthma and type 2 diabetes in both examples. A “bridging” of electronic health data between the acute sector and other sectors (e.g. education, infrastructure, agriculture, transport, retail) is required to enable precision public health as NCD determinants are multifactorial, systemic and largely socially-driven [36].

Social, environmental and behavioural health determinants explain ~ 70% of health variance [22]; yet, NCD surveillance has prioritised mapping acute care and biomedical indicators. A conceptual framework for integrating the social and behavioural determinants of health into population health analytics via (a) health systems (b) non-medical organisations and (c) citizen science has been proposed in the USA to deliver comprehensive surveillance [37]. This capability will unlock precision public health and create new ethical challenges; knowledge arising from contemporaneous community data has the potential to widen health inequities if misused by certain industries. Stakeholders will retain core responsibility for ensuring ethical use of data for the public health good [38].

The primary advantage of a precision approach to public health is to shift the dial towards a more efficient predict-prevent system, as opposed to the currently dominant but inefficient break-fix system [7]. The foundation of advancing precision public health for NCDs is to utilise this aggregated data to deliver the right health intervention – ideally preventive or early intervention – to the right population at the right time and monitor the outcomes in a virtuous learning cycle. One pragmatic conceptual example related to obesity prevention is: community public health units access a digital dashboard containing (1) retail transactional data for fast food that is geotagged with hotspots in real-time (2) weight trajectory risk data in the early years leveraged from a statewide EMR (3) census data that augments geotagging with understanding of social position e.g. education, household income and the built environment. Then by harnessing a richly-curated, digitally-connected community, public health professionals can ‘push’ digital messages and content via social media and mHealth applications to promote healthy, family-friendly alternatives that are low cost and readily available to individuals in high-risk areas.

The true gap between current public health practice and precision public health is the progression from precision surveillance (i.e., to monitor health status) to precision action (i.e., to prevent disease). Surveillance platforms in our included studies were primarily implemented or intended for use in public health organisations, government agencies and health services to improve monitoring and reporting of chronic disease and population health and inform public health and health system planning. Evidence of their real-world application to improve public health decision-making was unclear. True precision public health for NCDs will require digital disruption of the public health sector and workflows must be optimised for data-driven, responsive and accurate decisions.

One possible reason for the gap between surveillance and action is immaturity of (a) data and (b) analytics. Most platforms identified in this review integrated analytics that favoured precision surveillance e.g., descriptive, comparative, geospatial (Horizon 2). A precision public health system is likely to be responsive to risk based on predictive (e.g., machine learning) and prescriptive (i.e., decision-support) analytics (Horizon 3) [7, 9]. One example is applying a dynamic simulation model to an NCD of interest (e.g., CVD or obesity) to quantitatively model the reach, consequences and effectiveness of various population health interventions in demographically high-risk areas. Leveraging real-world data that is updated with contemporaneity enables counting of improvement in public health that drives investment.

The digital transformation of the public health sector to prevent and manage communicable diseases is further advanced than for NCDs and was accelerated by the COVID-19 pandemic [17, 39‐41]. There are lessons and opportunities from this transformation for precision public health of NCDs. Population surveillance for infectious or environmentally-driven diseases (e.g. sexually transmitted, influenza A, tuberculosis and COVID-19) is more advanced due to the requirement for fast data – transmission happens in real-time and immediate population risk is more acute, thus requiring dynamic data sources that facilitate continuous and granular population tracking. Communicable disease management requires concurrent break-fix (current state) and predict-prevent (future state) models of public healthcare. Surveillance must enable precision responsivity to outbreaks to soften their reach and impact in real-time while also enabling predictive modelling to plan resourcing, investment and policy action. The prevalence of high-burden chronic diseases (e.g., CVD, obesity, cancer) increases slowly due to compounding risk factors generated over the lifecourse. A break-fix response is necessary but problematic – once established, lifestyle-related NCDs (e.g. obesity) are difficult to reverse and attempts at reversal are often unsuccessful in the long-term [29].

The primary limitation of included studies was limited evaluation of population health surveillance solutions for usability and effect on public health decisions, policy and interventions. All studies were conducted in high-income countries; yet 77% of all deaths from NCDs occur in low and middle-income countries [1] where infrastructure is less digitally mature to enable precision public health. The refresh frequency of data in the surveillance platforms of included studies was mostly static and performed at cross-sectional timepoints. This generated a snapshot understanding of disease/s of interest. Lastly, some studies [30, 31, 35] concentrated on the digital architecture of solutions and missed explanation or visualisation of user experience and navigation.

The primary limitation of this review is its exclusion of population health surveillance platforms that only used EHR data for precision public health of NCDs. This was decided due to a significant body of existing review literature and to fill the clear research gap of understanding non-EHR sources of real-world data that contribute to precision public health of NCDs.