How can Big Data Analytics Support People-Centred and Integrated Health Services: A Scoping Review

People-centred, integrated health services (PCIHS)

Designing health services in accordance with the determinants of health spanning biophysical, lifestyle-related, social, health system-related, and environmental factors challenges traditional disease-centred, fragmented models of health service delivery [17, 18]. In response to the challenges in healthcare, different concepts of integrated care emerged, centred on the needs of patients, their families, and their communities [19]. The concepts vary in size and scope and are designed around the idea to put people in the centre of service delivery to improve value-creation [3, 20]. Several of these approaches including the rainbow model of care were considered when researches designed the framework for people-centred and integrated health services (PCIHS) for the World Health Organization (WHO) [18]. In the WHO’s global vision, not only does it outline achieving a seamless patient experience but also focusing on health promotion and disease prevention for the people, which may not necessarily be patients yet [2]. Improving healthcare following this people-centred perspective must focus on all the potential interrelations of the determinants of health and uniting the diverse objectives of healthcare stakeholders [21, 22, 23, 24, 25, 26] across the continuum of health promotion, disease prevention, disease detection and acute, chronic, and palliative care [24, 25, 27, 28, 29].

The PCIHS framework proposes five strategies for health services to become more integrated [30, 31, 32]:

• Empowering and engaging people and communities(e.g. personalized health plans, shared decision making, access to health records)

• Strengthening governance and accountability(e.g. acting upon user experience, decentralization, performance evaluation)

• Reorienting the model of care(e.g. strengthening primary and community care, population health, prevention)

• Coordinating services within and across sectors(e.g. care coordination, effective referral and discharge systems, coordinated systems)

• Creating an enabling environment(e.g. large scale systems change, strong leadership, financial support, cultural change)

Big data

Although a consensus about the definition does not exist, it can be agreed upon that massive data storage alone does not define big data [27, 33]. The definition referenced most often is rooting in the 3-V model focusing on the characteristics of volume, velocity, and variety [34], which was gradually enhanced to the 5-V model by adding veracity and value [14, 35, 36, 37, 38, 39, 40]. Accordingly big data is characterized by

Big data types in healthcare

The fragmentation of patient care is also reflected in the decentralization of health data [41, 42]. In general, any source contributing information to one of the factors influencing people’s health can be valuable [22], although not all data types abide by all criteria of the 5V-model. The most common types data in healthcare are billing data, clinical data, patient- or people-generated data, health-related research data and data collected externally to the health care environment including socio-economical, societal, community-based, demographical, environmental, and other health-related data (see Table 1) [27, 43, 44].

A good overview on sources, stakeholders and capabilities in the health data ecosystem is provided by Vayena et al. [45].

Big data analytics (BDA) in healthcare

For big data analytics there is also no consented definition. Compliant to other industries analytical types in healthcare [38, 46, 47] can be categorized in

• descriptive analytics (What happened or is happening?),

• predictive analytics (What is likely to happen next?),

• explorative analytics (Why is it happening? What is unknown yet?),

• prescriptive analytics (Which decision is best to reach a desired outcome?).

From a methodological perspective the terms “prediction” and “exploration” do not define different approaches, but different analytical purposes [48]. Taken together predictive and explorative analytics are also referred to as advanced analytics [49]. Performing advanced analytics on big data is one approach to define big data analytics (BDA) [14, 15]. In a broader sense all kinds of predictive or explorative models applied to big data would meet this definition, also including statistical methods [50] and most often when the aspect of high velocity is inconclusively. In a narrower sense only inductive approaches like data mining or machine learning suited for high-dimensional data sets define big data analytics [10, 27, 46, 51, 52]. For this paper the broader focus was adapted. Big data analytics (BDA) can provide complementary information to those derived from hypothesis-based experiments which have a long tradition in healthcare [46, 51, 53, 54, 55]. As there is plenty of literature on statistical methods they are not further explained (see e.g., Hohmann et al. [56]). Machine learning has the potential to enhance statistical analytics by providing models that allow for more multivariate effects and complex relationships. While supervised learning is used to train algorithms in predictions, unsupervised learning is used for exploring unknown patterns within data sets [7, 57], whereas the analytical methods are basically the same as for both tasks [48, 58].

Review method and content analysis

To first of all provide a comparative overview on the “data types” and “analytical methods” used most often in healthcare, rapid literature reviews were conducted in Medline/PubMed combining the search terms of the scoping review described in the following with terms specifying the data types and analytical models (see Table 4 and Table 5 in the appendix).

To answer the main research question “How can big data analytics support people-centred and integrated health services?” a scoping review following the recommendations of the Preferred Reporting Items for Systematic Reviews and Meta-Analyses – Scoping Review (PRISMA-ScR) statement [64] was conducted. To better define the search term text mining algorithms were applied [65]. The search term “big data analytics” was used as a starting point and checked for similarities and thesaurus on Medline/PubMed using the search results clustering algorithm Lingo [66, 67]. The clustering was based on the first 200 results from a search conducted on April 1^st, 2019 and revealed overlap of BDA with the terms “predictive analytics”, “advanced analytics”, “machine learning” and “big data analysis methods”.

A combination of these overlapping terms and Boolean operators was used to build the final search term. The search was conducted in Medline/PubMed as well as in the computer science database dblp (see Table 6 in the appendix). To limit the search results some inclusion and exclusion criteria were applied followed by a qualitative classification of two researchers working independently (see Table 7 in the appendix). For instance articles before 2013 were excluded as the number of articles meeting the inclusion criteria before that date were rather low and Natural Language Processing as a subfield of BDA was excluded because it yielded too much technical articles with few links to integrated care interventions as most often textual information were extracted and analysed from one single source of medical documentation.

To further extract information about strategic interventions in context of the PCIHS framework and about challenges for big data analytics in healthcare content analyses were performed during which the articles chosen for the review were classified (see Tables 8 and 9 in the appendix).

Development of a role model of a people-centred health platform (PCHP)

The role model of a people-centred health platform presented in Figure 1 is purposely meant as a roadmap for decision makers to realize data analytical capabilities in healthcare like the PCIHS framework also is an illustration of options healthcare decision makers might consider in optimizing health services dependent on and adapted to their context conditions.

Figure 1

In compliance with the concept of PCIHS all data potentially contributing relevant information about people’s health (rainbow model) were taken into account. Integrating these data in a central health platform as timely as possible (high velocity) would create a data asset of tremendous extent (high volume) and distinctness (high variety). In the data analytics layer big data analytical methods might be applied to the data with the purpose to produce results of high veracity which, interpreted and used by well-informed health decision makers, providers or even patients shall lead to decisions of high value in terms of the five strategies towards people-centred and integrated health services. Comprehensive personal health records are developed and tested by some research institutions [10, 53, 68, 69] as well as in some real-world initiatives such as the national health platforms of Finland [70], Estonia or Australia [71] or from the US Veterans Health Administration [72].

Types of big data and big data analytical methods in healthcare – Results of the rapid literature review

According to the search results of the rapid literature review biomarker (39.3%) and medical imaging data (30.9%) are currently used most often in publications (see Figure 2). Biomarker data include the whole spectrum of ‘-omics’ like genomic, proteomic, or metabolomic data [73, 74, 75]. Medical images are often part of electronic health records. The most common technologies are ultrasound, computed tomography, magnetic resonance, and x-ray imaging [38, 52].

Figure 2

Considerably high rates were also found for smart sensor data (16,0%) and data from electronic health records (5.4%). A smart sensor can be used to constantly track individuals and is often embedded in a smart phone/watch or in telemonitoring devices, sometimes with several devices communicating with each other (Internet of Things). A smart sensor can continuously measure large volumes of data in terms of health, fitness, behaviour or lifestyle regardless of location, potentially in real-time and even supplemented by self-reported data (quantified-self) [22, 27, 38, 43, 74]. A side-specific electronic medical record (EMR) or a cross-institutional electronic health record (EHR) stores data stemming from different source systems which is why technically speaking EMR and EHR are rather data platforms than data types. The volume of data in EHR is massive on the health system level while it varies on the organizational level [37, 76]. A typical EHR contains structured data (e.g. medical coding), semi-structured data (e.g. laboratory results) and unstructured data (e.g. narrative clinical notes, medical images) [43, 77].

Data types used rather seldom were internet usage or social media data (2.6%), claims data (2.1%, most often health care data, rarely social care data), data from clinical trials (1.6%) and registry data (1.2%). Data generated by using internet technologies include access log data or click streams from websites, search engines, or forums or posts and network relationships from social media platforms or messaging services [22, 38, 78]. The most common claims data types are medical, pharmaceutical, and ancillary claims while payers hold additional administrative information [26, 79]. Claims data are rather homogenous due to specific coding schemes, but at least the data provides a rather full picture of services utilization regardless of the point of care [80], whereas an all-payer database would be ideal for BDA supporting PCIHS so that analytics are not limited to the population covered by a single payer [81].

Other sources like patient surveys, drug surveillance, aged or community care data or other health-related systems together only accounted for less than 1% of current research articles on BDA in healthcare. For example, aged or community care data were presumably underrepresented because most of the provider organization are lacking the financial opportunities to build up and work with large, standardized databases although there would be additional value in using high level information technology and analytics in these contexts [82, 83, 84]. For PCIHS the integration of as many data sources as possible seems most beneficial.

Figure 3 displays the most often used BDA models in healthcare based on the rapid literature review. Support vector machines (27.3%), neural networks (20.4%) and random forests (19.5%) were used most often. Further models used occasionally were decision trees (6.7%), k-nearest neighbour models (6.1%), k-means clustering (1.9%) and Bayesian networks (1.4%). Traditional prediction models in healthcare are primarily parametric regression models based on assumptions regarding the data distribution and a predefined set of input variables [85]. Several studies retrieved in this review labelled their analytics as BDA by applying statistical models to data sources meeting more or less the definition of big data. Therefore considerably high rates were also found for statistical models like logistic regression (12.0%) and linear regression (3.7%) while other methods like multiple regression or proportional hazard models were used rather seldom (~1.0%). The results point to the fact that non-parametric models rather meet the general understanding of BDA in healthcare than traditional statistics.

Figure 3

How can big data analytics support people-centred health services

A people-centred coordination of preventative, health, and social services (including aged and disability care) is likely impossible without an equally comprehensive integration of the underlying health information technology infrastructure [9, 86]. In the scoping literature review articles were screened for analytical applications with the potential to support the five strategies for health services to become more integrated and people-centred. Based on a matrix table (see Table 8 in the appendix) all articles retrieved in the scoping review were categorized with respect to the five strategies of the PCIHS framework or rather to the respective policy options and strategical interventions. The results are summarized in Table 2.

BDA as supporting tool to empower and engage people

At least one of the strategical interventions summarized under empowering and engaging people was named in 36 (51%) of the screened publications. Not only in this domain, but in general the ability of BDA to support the development of personalized care plans was mentioned most often (43%). This could for example be by accurately and timely predicting individual health risks (lifestyle, socio-economics, environment, genetic predisposition, etc.) [26, 28, 87], by predicting risk scores for disease conversion or progression [8, 24], by deciding about the best intervention type based on patient similarity analyses or by predicting the probability for side effects or adverse events [59, 88]. Examples found during the review are predictions for chronic diseases, heart failure, type 2 diabetes and severity stages for lung cancer or potential vaccination benefits and risks (see Table 8 for all references). Besides genome-wide association studies uncovering individual genetic predispositions for disease development [93], the full potential of BDA stems from the integration of data on all factors influencing health including also population-based, socio-economic, community-based or environmental factors. By providing information about the likelihood of an individual to benefit from different therapy options more targeted decision aids and medications could be developed and greater satisfaction on the patients’ side be achieved [9, 62]. Also, self-diagnostics and self-management activities (7%) could be supported as people could regularly and timely be updated about their situation, their status and their current treatment options [22], e.g. based on sensor or patient-reported data (quantified self) [89]. By sending targeted information accessible via the personal health record or PCHP (3%) the support of peoples health education based on their individual risk factors might be improved (4%), as well as the process of shared decision making (6%) as patients can better define their individual care plans and therefore better adhere to their personal health goals. The PCHP could allow patients not only to access but also to administer and share their health-related data and to use the platform as a tool to communicate e.g. with providers.

BDA as supporting tool to strengthen governance and accountability

In 23 (32%) of the screened publications BDA was mentioned as a tool to strengthen governance and accountability. BDA could facilitate a deeper understanding of underlying factors for variation across providers, interventions, or regions (appropriate versus avoidable variation) to improve risk adjustment systems or performance evaluations (21%) supporting a transparent competition for outcome improvements [51, 80, 90], e.g. in performance-based contracts (11%). Also, results could be made publicly available, e.g. in league tables. Geocoded analyses could uncover community-based, regional, or environmental risk factors as well as supplier-induced problems and local disease hot spots [91, 92] and be used to establish more decentralized systems (11%) with enhanced scope for local governments or community-care to implement regional health programs enhanced with patient-reported outcomes (1%). This would offer new opportunities for people in local communities to participate in the decision making process via the PCHP as a communication tool and become co-producers of population health.

BDA as supporting tool to reorient the model of care

The biggest share of articles in the scoping review described potential applications of BDA belonging to the strategy domain of reorienting the model of care (79%). Most often mentioned in this area was incorporating BDA in clinical decision support systems (32%), informing the provider about risks for disease uptake, progression, conversion, decompensation or the development of comorbidities [58, 93]. A key factor of the PCIHS strategy of reorienting the model of care is strengthening primary and community care, whereas BDA could support more accurate diagnostics at these points of care [51, 94, 95, 96]. Clinical judgements in these sectors might e.g. benefit from proactive alerts which inform about individual risks for preventable events like (re-)admissions to hospital, for intensified resource use, for (post-surgical) complications or disease progression [93, 94, 95, 96], in the best case based on intersectoral health data from the PCHP also allowing for interdisciplinary communication. According to a survey in the USA, 15% of the healthcare providers already have access to some kind of predictive analytics and the conditions most often targeted were hospital readmissions (27%), the development of a sepsis (27%), patient deterioration (18%) and general health (10%) [97]. Using intersectoral data to stratify individuals into (chronic) care groups and identify comparable or manageable populations could support additional population health management activities (26%) in which the role of nurses and community health workers could be enhanced [22, 24, 26, 35, 98]. Also surveillance and control systems (18%) could benefit from BDA based on real world health data assets, e.g. the surveillance of adverse drug and vaccination effects or the monitoring of disease transmission patterns or outspread speed of epidemics or pandemics [91, 92] enabling for example faster reaction and better targeted campaigns [88]. Using real-world data would additionally allow for rather small risk groups or (geographically) isolated communities already suffering from under-coordination to also be taken into consideration in healthcare decision making [44, 99]. Furthermore, activities like health promotion and disease prevention (13%) might be better tailored to individuals if certain risk factors are specifically addressed. By using sensing devices as well as mobile technologies (14%) or devices within the patients’ ambient (6%) therapy results might be better tracked by patients as well as by providers.

BDA as supporting tool to coordinate services within and across sectors

In the scoping review 20 publications (28%) described BDA as a tool to support service coordination. Most articles mentioned the development and evaluation of intersectoral care pathways (11%) by exploring comparable patterns and then setting up multidisciplinary task forces of medical and non-medical providers for such multi-layered problems structured around an individual’s social experiences and comorbidities. Also, BDA respective the PCHP as enabler would simplify the exchange of medical records (8%), especially in the transition between hospital and home. Four publications (6%) described BDA as an enabling tool for intersectoral partnerships across the health sector (e.g. with social security, housing, education) to provide holistic care and one publication described a model in which BDA is used for district-based healthcare delivery [100].

BDA supporting the creation of an enabling environment

The strategy of creating an enabling environment is supporting the aforementioned strategies and is rather broad in scope. BDA itself is an enabler for people-centred health services, but 17 publications (24%) mentioned BDA as incorporated in other enabling factors as well. On the level of resource planning and allocation (15%) BDA might be capable of reducing financial waste by identifying common patterns of fraud and abuse or by uncovering disincentives of the renumeration system towards finding the right payment mix [79, 80, 101]. BDA could also support system research comparing the effects of different system architectures (9%). Assisting in quality assurance (4%), BDA could, e.g. by exploring care patterns, identify clinical waste and provide the opportunity to get rid of ineffective or unnecessary interventions or to reduce over- and undertreatment [37, 44]. Two publications described BDA as tool to identify those professionals benefitting the most from additional training and education, e.g. on team-based culture or open feedback (3%).

Challenges of big data analytics in healthcare discussed in the literature

As BDA has the potential to improve PCIHS it seems valuable to find solutions for the challenges stemming from big data in healthcare [102]. Currently the situation for most stakeholders is characterized by confusion or uncertainty [54]. Of the 72 articles in this review, 45 (62.5%) discussed at least one BDA challenge. Most often discussed were methodological challenges (54.2%) followed by regulatory (43.1%) and technological challenges (41.7%). Cultural challenges were less often discussed (25.0%). The five issues mentioned most often in making better use of BDA were missing modelling standards and potential bias (36.1%), a questionable evidence-base of BDA results (33.3%), poor data quality (27.8%), the lack of an appropriate framework for privacy protection (26.4%) and the lack of interoperability requirements for data linkage (26.4%). In the successive descriptions only the most relevant publications will be referenced (see Table 9 in the appendix for more details).

Regulatory challenges

From a regulatory perspective it is challenging to set up a framework to coordinate, support and financially incentivize the efforts in building a big data platform for health data [15]. Besides ensuring for targeted investments this means describing the policies of appropriate data storage [27, 36]. As the relevance of analytical results in clinical processes diminishes over time it is also a challenge to facilitate user friendly processes for data entry and timely exchange to finally enable (real-time) recommendations at the point of care [9, 36, 103]. To overcome legal or commercial barriers across domains intellectual property rights must be clearly defined, penalizing e.g. the unwillingness to share relevant (clinical) data for economic reasons or unintended uses [41]. To avoid underperforming models from mis-informing clinical decision making, a framework for transparent model development and evaluation would be needed [46, 104, 105]. Analytical modelling standards could, comparable to drug licensing, be transparently developed by quality controlled institutions which incorporate the technical and methodological expertise but also contribute domain knowledge to determine how to provide accurate, reliable and actionable information for patient care [44, 106]. Likewise, this is touching ethical issues, e.g. if a BDA model at the beginning of the learning curve provides seriously harmful recommendations for some individuals [88, 107, 108]. The most often mentioned regulatory challenge was the design of an appropriate framework finding the sweet spot between transparency and protecting privacy enabling as effective decision supporting analytics as possible without enabling a potentially manipulative misuse of the data [54, 77]. To enable as many beneficial analytics as possible, it might be an option to make deidentified data extracts from the PCHP accessible for chosen academic or even commercial purposes [9, 54, 77].

Technological challenges

Despite prices for data storage are steadily going down from a technological perspective the design of an infrastructure appropriate for storing and curating massive amounts of diverse health data is still a complex task [37, 38, 77, 108]. Also, it is challenging to deal with high-velocity data depending on considerable computational processing resources and then to use appropriate software tools for data analytics [27, 85]. Blending the extremely diverse and often unstructured health data from heterogenous sources leads to the challenge of establishing technological standards of interoperability [77, 89]. Furthermore, inaccurately calibrated measurement systems as well as hard- and software failures (e.g., wrong auto-fill-in functions) inadequate data transfer protocols or not adequately developed software pose risks for data quality. Data quality problems can possibly arise at every step during data generation while the chance for bias might be lower for recorded medical signals than for manually documented features [36, 39, 54]. Finally, all layers of a big data platform (storage, transfer, analytics, presentation) have to be technically protected against unintended uses or breaches, e.g. by data encryption, certification or access authentication [72, 77]. Big data technologies were out of the scope of this review, but at least it shall be referenced to articles discussing tools for big data storage & transformation like MongoDB or Apache HBase [9, 38, 43, 74, 108], for big data processing & analysing like Hadoop or MapReduce [38, 43, 74, 85, 109] as well as methods for (big) data security [77, 110, 111, 112].

Methodological challenges

From a methodological perspective it is challenging to work on a high-dimensional database likely to contain more feature variables than observable subjects [44] and to develop real-time analytical models as most documentation processes in healthcare traditionally are rather slow [36, 72]. Regarding human documentation also data entry errors like incomplete, incongruent, or missing data and a poor update status pose risks for data quality [39]. As a priori it is unclear which model is most appropriate for the targeted type of application and which model offers clinically more meaningful interpretations, the process of evaluating analytical models is quite challenging [113]. It affects the analytical results that no commonly accepted methodological standard for modelling exists offering nearly unlimited different options for the combination of variables whereas currently there is a lack of knowledge about which methods to use for which purposes and the black box design of some machine learning algorithms even exacerbates their comprehensibility. Additionally, external validity or generalizability is a challenge as it is difficult to compare the performance of different BDA models based on different data types from different regions [77, 113, 114]. It is also problematic that in some source systems data is recorded for specific reasons (e.g., medical billing) or with different coding standards potentially limiting interpretability beyond the original purpose. In a greater extent the same limitations as for observational studies also apply for BDA such that it is extremely difficult to exclude potential bias (e.g. selection bias, confounding bias, measurement bias), that due to missing randomization no causal relationships can be determined and that especially BDA has a high risk for modelling artefacts like random noise or overfitting [27, 56, 87]. Designing a methodology on how to evaluate the clinical usefulness and evidence-base of the analytical models or their effectiveness and safety in part also is a methodological issue [115]. To date, there is only minimal evidence that BDA in healthcare revealed anything surprisingly new and can effectively improve decision making or medical outcomes [93, 94, 116]. Furthermore, is has not been proven that machine learning models outperform traditional statistical models in predictive or exploratory tasks. Most often only sparse differences in the model performance are observed, maybe because they were often applied to rather small data sets limiting the ability of BDA models to optimize the inductive feature selection process [7, 8, 113]. To disseminate information about the most effective treatments to the intended providers at the point of care requires that information overload is prevented, and analytical results are timely and easily accessible, appropriately simplified, appealingly visualized and well-integrated in clinical workflows [93, 117]. A comprehensive discussions of methodological issues of BDA in healthcare is e.g. provided by Hoffman/Podgurski [54] and Van Poucke et al. [46].

Cultural challenges

An adaption of BDA models in healthcare requires appropriate education as well as a shift towards team-based analytics enhancing medical domain knowledge with skills e.g. from data science and health economics [37, 85]. Form an organizational perspective also resistances against expanding and speeding-up electronic data exchange and against redesigning clinical workflows with data-driven feedback need to be overcome by communicating potential benefits and by putting media-hyped expectations into perspective [25, 72]. A data quality culture must be developed to reduce behaviours like unreflective copy-pasting and strategical manipulation of data. From the societal perspective, a data sharing culture would be helpful to counteract personal and organizational concerns. This might be accompanied by an open science culture which ensures that peoples’ data are used as intended [22, 36, 118]. Exploratory studies point to the fact that the majority of people is willing to share health data for population-based health research, but fewer individuals are comfortable to have their data used to improve medical decision making or to adapt insurance rates [119, 120] with country-specific, cultural differences [121, 122]. As the mere existence of BDA tools does not influence value improvement a learning culture with engaged providers needs to be achieved with (clinical) usability as a precondition.

In Table 3 all challenges mentioned above were systematized by combining the domains of technological, methodological, regulatory, and cultural challenges [37, 74] with the 5-V model as each big data characteristic entails specific obstacles [36, 54, 77, 85].

Potential success factors of big data analytics or strategies to overcome the challenges can be derived as countermovement to each challenge displayed in Table 3. For example the success factor of data quality assurance would be a strategical reaction to the described data quality challenges as well as the success factor of implementing a big data governance would be a reaction to the fact that healthcare organizations are often missing a data governance. The enabling factors of the PCIHS framework [32] as well as some articles from the scoping review provide further information [105, 123].