Skip to main content

How can we discover the most valuable types of big data and artificial intelligence-based solutions? A methodology for the efficient development of the underlying analytics that improve care

Abstract

Background

Much has been invested in big data and artificial intelligence-based solutions for healthcare. However, few applications have been implemented in clinical practice. Early economic evaluations can help to improve decision-making by developers of analytics underlying these solutions aiming to increase the likelihood of successful implementation, but recommendations about their use are lacking. The aim of this study was to develop and apply a framework that positions best practice methods for economic evaluations alongside development of analytics, thereby enabling developers to identify barriers to success and to select analytics worth further investments.

Methods

The framework was developed using literature, recommendations for economic evaluations and by applying the framework to use cases (chronic lymphocytic leukaemia (CLL), intensive care, diabetes). First, the feasibility of developing clinically relevant analytics was assessed and critical barriers to successful development and implementation identified. Economic evaluations were then used to determine critical thresholds and guide investment decisions.

Results

When using the framework to assist decision-making of developers of analytics, continuing development was not always feasible or worthwhile. Developing analytics for progressive CLL and diabetes was clinically relevant but not feasible with the data available. Alternatively, developing analytics for newly diagnosed CLL patients was feasible but continuing development was not considered worthwhile because the high drug costs made it economically unattractive for potential users. Alternatively, in the intensive care unit, analytics reduced mortality and per-patient costs when used to identify infections (− 0.5%, − €886) and to improve patient-ventilator interaction (− 3%, − €264). Both analytics have the potential to save money but the potential benefits of analytics that identify infections strongly depend on infection rate; a higher rate implies greater cost-savings.

Conclusions

We present a framework that stimulates efficiency of development of analytics for big data and artificial intelligence-based solutions by selecting those applications of analytics for which development is feasible and worthwhile. For these applications, results from early economic evaluations can be used to guide investment decisions and identify critical requirements.

Peer Review reports

Background

With the increasing ability to collect healthcare data, billions of dollars have been invested in (big) data analytics and artificial intelligence (AI) by private (e.g. IBM, Google, hospitals) and public institutions worldwide (e.g. Agency for Healthcare Research and Quality, the Patient-Centered Outcomes Research Institute, European Commission) [19]. Analytics can be applied in many ways, and it has often been suggested that they can improve care for a wide variety of clinical fields [1015]. Bates et al. define big data analytics as the discovery and communication of patterns in datasets that are extremely complex due to their size (volume), rapid collection (velocity) and/or the need to combine multiple data sources (variety) [14]. The term Artificial Intelligence was first mentioned many years ago and is defined as the ability of computers to mimic or simulate the human mind [16]. However, despite many publications on the potential of big data analytics and AI, few analytics have been implemented [6, 1720] and resulted in health benefits and/or cost savings [2123].

Data availability can be an important barrier to the development of analytics that improve healthcare [4, 12, 17, 2426]. The datasets required to develop machine learning models should be large and, depending on the method used, should contain sufficient data on relevant features [11, 27]. Data-related problems mentioned in the literature include limited sample size [4, 2426, 28], a short duration of follow-up [24], validity of results with heterogeneous patient populations and selection bias [4, 13, 17, 24, 28, 29] and bias due to missing data [12, 24, 29, 30]. Moreover, successful development does not mean easy implementation; important barriers to implementation include the need for prospective validation [4, 24, 28] and the high costs of validation and implementation [4, 19, 24, 3133].

For other healthcare technologies, such as drugs, medical devices and diagnostic tests, economic evaluations are used to assess the potential impact of anticipated barriers early on during development [3438]. In economic evaluations, the health benefits and costs of novel technologies are compared to the benefits and costs of an alternative such as current care. Use of these economic evaluations alongside development is recommended to assist decision-making by developers, to analyse the impact of uncertainty in performance of the technology on outcomes, and to identify critical requirements (e.g. price) for successful market access and dissemination [36, 37]. A key aim of this approach is to increase the likelihood of successful market uptake and avoid wasting investments due to failed implementation.

Very few economic evaluations of analytics exist [13, 17, 2023, 39, 40] and the ones that do have omitted relevant costs [19, 22]. Moreover, recommendations on how and when to perform economic evaluations of analytics do not exist, even though their use could improve development efficiency by identifying analytics with the greatest potential health impact. In this paper, we present a framework that can assist developer decision-making by selecting applications of analytics that are not only worth developing but also feasible.

Methods

We present a framework that efficiently selects analytics that are relevant, feasible and capable of generating important health and economic benefits (Fig. 1). The framework was developed based on challenges of analytics development defined in the literature and best practice recommendations for economic evaluations. It was then further refined by applying it in three clinical use cases. The use cases were selected from a European Horizon 2020 funded project (AEGLE) that aimed to develop a cloud-based big data analytics platform. The three use cases focused on chronic lymphocytic leukaemia (CLL), the intensive care unit (ICU) and diabetes.

Fig. 1
figure 1

Flowchart for assessing health economic benefits of novel analytics alongside development. p = problem

Step 1: select clinically relevant problems

This first step involves selecting relevant clinical problems. Whether problems are considered clinically relevant depends on the setting for which analytics are developed and the experts involved. When analytics are developed for a local hospital (i.e. for a learning health system), local experts should be consulted to identify relevant problems. When the aim is to develop analytics for a wider audience such as clinical experts in different countries or continents, then interviews with multiple potential users are recommended alongside a review of guidelines and the literature. Needless to say, a multidisciplinary approach throughout this step is crucial [10, 41].

Step 2: assess data for development

After relevant problems are selected, it is necessary to assess whether the data available, or to be collected, is of sufficient quantity and quality to address the problem. Such an assessment may include careful scrutiny of the sample size, duration of follow-up, expected frequency of missing data, potential sources of bias and heterogeneity in care practices between sites. Moreover, the timing of data collection and the types of outcomes collected during follow-up may differ between clinical sites.

Step 3: identify critical barriers to realising successful development and implementation

The scope of the problem should be narrowed down and used to identify critical barriers prior to estimating costs and benefits. Narrowing down the scope is a critical step in any economic evaluation [37] and one way to achieve this is through the Population (or Patient), Intervention, Comparator, and Outcomes (PICO) method [37]. First, the target population (P) is defined, which can include a description of the setting and the population size. The intervention (I) should include a description of the care pathways involved, including the analytics to be developed, the additional software and hardware needed to use the analytics, and the actions that follow from use of the analytics. The description of the comparator (C) entails a discussion on treatments available and relevant software and hardware elements used in current care. The final component of outcomes(O), refers not just to clinical outcomes but all outcomes considered relevant by users and purchasers, including mortality, life years gained, quality-adjusted life years gained(QALYs) and economic benefits. Ideally, they should go beyond diagnostic performance metrics like Area Under the Curve (AUC) [4, 17, 42, 43] and include outcomes related to health benefits, satisfaction and costs.

The detailed description of the scope, formulated using the PICO method, can then be used to identify potential barriers to successful development and implementation of the analytics. An example of a critical barrier is whether the health information system currently used in a health centre is sufficient to support the analytics or whether major upgrades are needed. If the examination of possible barriers does not reveal any insurmountable barriers, the health and economic benefits can be estimated. When continuing development seems risky, for instance because of the limited availability of required software and hardware elements in current practice, a developer can decide to select a new problem or cease development altogether.

Step 4: economic evaluation

The next step is to perform an economic evaluation of the analytics that are considered feasible to develop. An evaluation starts by developing a conceptual model and collecting input data. A conceptual model can be developed in different ways, including the estimation of the number needed to treat [44], decision curve analysis [42, 43], decision trees, and Markov models. Depending on the stage of development, the models may vary from very simple to very complex. The validity of the model should be assessed according to best practice guidelines [37, 45]. Information on relevant input parameters required to populate the model can be collected alongside model development from sources such as patient-level data and the literature, but are sometimes limited to expert opinion or assumptions, particularly in the early stages of development. Uncertainty surrounding parameter estimates generally decreases as development progresses and more information becomes available [36, 38].

Base case estimates of potential benefits can then be determined using the most likely parameter values. Results can be presented using the incremental cost-effectiveness ratio (ICER) but more importantly; results should be presented such that they are understandable to the target audience (investors, future users and purchasers). The uncertainty in these point estimates should always be analysed using uncertainty analyses. Uncertainty analyses can include scenario analyses and sensitivity analyses, but also analyses to determine critical thresholds of relevant parameters, such as accuracy and pricing thresholds needed to realise health and economic benefits. The headroom can also be estimated according to the following formula:

$${\text{Headroom}} = {\text{N}} + \uplambda *{\text{Q}}$$

Here N refers to the potential savings where the costs of the technology are set to zero, λ is the willingness to pay threshold and Q are the health effects gained [46]. Moreover, probabilistic sensitivity analyses can be used to estimate the impact of uncertainty in all parameters simultaneously. For each parameter, random estimates are drawn many times (e.g., n = 1000) from their underlying distribution. For these estimates, the costs and effects are calculated and presented using a cost-effectiveness plane and a cost-effectiveness acceptability curve. In a cost-effectiveness acceptability curve, the probability that an intervention is cost-effective is plotted against a range of willingness to pay thresholds.

Iterative approach

When a developer decides to continue development, the different steps (assess data for development, critical barriers to realising success, and the economic evaluation) should be revisited as needed throughout development, represented by the dotted line in Fig. 1.

Clinical use cases

Chronic lymphocytic leukaemia

The first clinical use case, focused on developing cloud-based analytics using next generation sequencing (NGS) data of CLL patients from three clinical sites across Europe (Sweden, Italy & Greece). CLL is characterised by considerable heterogeneity in disease progression [47, 48] and after diagnosis, the majority of CLL patients are followed according to a ‘watch and wait’ (W&W) strategy. Roughly 60% of these patients progress to having active disease requiring treatment [47]. The treatment they receive depends on their molecular profile and general fitness as well as on treatment approval and availability [47].

Intensive care

In the second use case, the aim was to develop analytics for ICU care using routinely collected data. Data from electronic health records (EHRs) and mechanical ventilators of patients from a Greek ICU was available for development. There are many ways in which analytics can improve ICU care and a variety of applications have been suggested in the literature [10, 11]; these include analytics to determine readmission risk, predict length of stay, diagnose sepsis, and improve the interaction between patients and mechanical ventilators [11].

Diabetes mellitus (diabetes type 2)

Many diabetes treatments are available, and these can often be combined to improve effectiveness. However, evaluating efficacy for all combinations, types of patients and treatment lines in randomised controlled trials would not be feasible, and using EHRs to evaluate effectiveness of treatment combinations has previously been suggested [30]. In this third use case, the aim was to develop analytics using EHRs in the United Kingdom to personalise diabetes treatment for patients.

Results

The framework was applied to three clinical use cases (e.g. CLL, intensive care and diabetes) (Table 1). The results for each case are described one by one.

Table 1 The methodology applied to address problems in care for chronic lymphocytic leukaemia, the intensive care and diabetes

Case 1: CLL

Because of the heterogeneous nature of CLL progression and treatment response, stratifying patients according to their expected prognosis could improve care [47]. In discussions with clinical experts, problems were selected based on the three decision points suggested by Baliakas et al. The first is upon diagnosis, when clinicians want to determine which patients are likely to progress to active disease. The second decision point is the moment when patients have active disease, and a first-line treatment needs to be selected. The third is the decision point when first-line treatment has failed, and a decision needs to be made about which second-line treatment is best for a patient [47]. CLL experts stated that decision points two and three were the most clinically relevant.

Regarding decision point one, developing analytics to improve stratification for these patients was considered feasible with the data available (Table 1). In contrast, the feasibility regarding decision point two was limited because of large variations between countries in the treatments prescribed. For decision point three, development of analytics to improve decision-making would not be feasible because it was expected that few patients in the data set received second-line treatment, which therefore meant a small sample size. Consequently, the first decision point was considered the best choice for analytics development.

When defining critical barriers, the scope included newly diagnosed Swedish CLL patients. In current care, these patients are not treated, but are regularly seen by the haematologist and undergo a blood test. When developing the analytics in 2015–2016, no treatment was available for patients with a high risk of progression. The only possible changes in care available at the time was the ability to personalise the intensity of follow-up and the ability to inform patients about their risk. These very limited options of ‘treatment’ can be considered a critical barrier for success since it is likely that costs of NGS and analytics are high while health benefits could only be expected through the reduction in a patient’s uncertainty (and anxiety) regarding prognosis. Therefore, at the time, analytics development did not continue beyond research purposes. However, a recent publication has suggested that early treatment of intermediate- and high-risk patients with ibrutinib could delay time to next treatment. Given these new findings, we updated results for this application, including the possibility of treatment with ibrutinib as part of the intervention.

After the PICO question was formulated, input parameters (probabilities, utilities, unit costs and resource use) were derived from the literature, Swedish guidelines, and expert opinion (Additional file 1: Table S1). A four state Markov model (Additional file 1: Fig. S1) was used to estimate costs, life years and quality-adjusted life years adopting a lifetime time horizon and a healthcare payer perspective. Long-term survival was estimated by combining results on time to next treatment from Condoluci et al. [49] with the hazard ratio reported in preliminary results from a randomised controlled trial comparing early ibrutinib treatment with current care [50]. More details on the model structure and input parameters used to estimate the health and economic benefits can be found in the Additional file. Even if an effective treatment is available, it is unlikely that analytics to improve stratification of newly diagnosed watch and wait CLL patients would be considered cost-effective: use of analytics would lead to a substantial cost increase (€89,985) but only a modest gain in health (0.13 QALYs) (Table 2). We demonstrated the relevance of univariate uncertainty analyses to assess the impact of parameter uncertainty (Additional file 1: Fig. S2). In univariate uncertainty analyses, the impact of an individual parameter is assessed by varying its estimate while keeping all other parameters constant. Here, the high costs of the treatment in the intervention arm are decisive in the incremental costs. The relevance of scenario analyses is demonstrated in Table 2 where even in the best-case scenario, analytics are unlikely to be cost-effective, since the incremental cost-effectiveness ratio exceeds thresholds used in Sweden. When varying all parameters simultaneously in the probabilistic sensitivity analyses, most of the estimates are in the upper right and left quadrant (Fig. 2). This means that most estimates reflect higher costs and either higher or lower QALYs. When these results are shown on a cost-effectiveness acceptability curve, we can see that better stratification of watch and wait patients and subsequent treatment with ibrutinib has an extremely low chance of being cost-effective (Additional file 1: Fig. S3).

Table 2 Results from the base case and best case scenario for analytics to improve stratification of watch and wait patients in chronic lymphocytic leukaemia compared to current care
Fig. 2
figure 2

Cost-effectiveness plane reporting the quality-adjusted life years and costs (€) from the probabilistic sensitivity analysis

Case 2: the intensive care unit

For the intensive care, relevant problems were identified through discussions with an intensivist at the Greek hospital that was involved in development.

Catheter related bloodstream infection

The first ICU-related problem selected, was that infections caused by central venous catheters were often diagnosed only after they are severe. Catheter related bloodstream infections (CRBSIs) are considered an important issue in the ICU since infected patients have an increased mortality and prolonged length of stay compared to other ICU patients [51]. The aim was to use analytics to diagnose CRBSI in an early stage to reduce disease severity, risk of death and costs.

EHR and biosignal data were available to develop the analytics (N = 2000) and additional records were to be collected prospectively. The required follow-up was short, and the relevant parameters needed to develop the analytics and evaluate outcomes (e.g. mortality, length of stay) were routinely collected. Missing data was expected to be present but manageable.

No insurmountable barriers were identified when narrowing down the scope in the early stages of development. An example of a potential barrier for the CRBSI analytics is the uncertainty in the probability of CRBSI. The frequency of CRBSI varies tremendously across countries and sites. In Western European countries, the reported incidence of CRBSI is low [52]. However, for the Greek hospital for which analytics were developed 7.5% of patients developed CRBSI during their ICU stay [53] and in other Greek hospitals reported even higher percentages (22.4%) [54]. If the target market for the analytics would have been limited to the US and western European countries, obtaining better estimates of the frequency of CRBSI would have been recommended prior to continuing with an economic evaluation. Another barrier might have been the need for EHRs to enable the analytics. However, since most Greek and European hospitals have adopted EHRs this was not expected to be an issue. Additional validation when adopting results in other hospitals would probably be required and feasible but would need to be taken into account in the economic evaluation. Based on these barriers, continuing with the economic evaluation was recommended.

A detailed description of the model and input parameters used to estimate health and economic benefits can be found in Additional file 1: Fig. S4 and Additional file 1: Table S2. A decision tree was combined with a four state Markov model (Additional file 1: Fig. S4), adopting a lifetime time horizon and including only direct medical costs. Input parameters were derived from the literature, hospital reports, and expert opinion. The effect of earlier intervention on ICU mortality and ICU length of stay were derived from a study reporting the effect of earlier prescription of antibiotics [55]. Initial estimates demonstrated that continuing development was worthwhile since analytics could reduce mortality (0.5%), improve QALYs (0.06) and lead to cost-savings (€886) per patient. All input parameters were varied extensively in uncertainty analyses but the probability of CRBSI had substantial influence on the results. When the price of the technology was below €19,216 per bed, the analytics could reduce costs compared to current care. This meant that the headroom to achieve cost-neutrality with the intervention was €19,216 per bed, which meant there was sufficient room for costs of analytics, validation, and implementation. Given the large potential for the analytics to generate savings it was considered relevant to continue with development. However, the key factor that influenced benefits was the prevalence of CRBSI (Fig. 3). In this case, it was worthwhile to closely monitor site-specific prevalence throughout development and carefully consider the appropriate target market given the large variation in prevalence across sites.

Fig. 3
figure 3

Impact of the prevalence of catheter related bloodstream infection in the intensive care unit on incremental savings

Ineffective effort events

The second ICU-related problem to be addressed with analytics, was suboptimal interaction between patients and their mechanical ventilator. One form of suboptimal interaction relates to ineffective efforts where a patient tries, but fails, to trigger the mechanical ventilator into providing a breath. Several studies have found that ineffective efforts could be associated with worse outcomes [56, 57]. Here the aim was to enable clinicians to intervene in those patients with ineffective efforts, who are therefore at risk of having worse outcomes.

EHR records were available for all patients and once again relevant parameters were routinely collected and missing data was expected to be manageable. Furthermore, recordings of > 24 h for more than 100 patients were available from a prototype monitor detecting patient-ventilator interaction.

When assessing feasibility, no barriers were considered insurmountable (Table 1). An important barrier was the need to have a monitor capable of measuring ineffective efforts in addition to analytics that could identify patients with ineffective efforts at risk of having worse outcomes. The prototype monitor available in the Greek ICU would need to be purchased in order to use the analytics. Furthermore, costs of site-specific validation would need to be included in the economic evaluation.

The model and input parameters used to estimate the health and economic benefits have been previously reported [58]. The potential impact of analytics that identify patients with ineffective efforts at risk of having worse outcomes also suggests that continuing further development is worthwhile [58] since it can reduce mortality by 3%, increase QALYs by 0.21 and reduce costs (€264) [58]. Furthermore, it was demonstrated that even if the effectiveness of intervening was varied extensively, benefits could still be achieved [58]. The headroom for the analytics to generate savings (€7307) was considered sufficient to cover relevant hardware costs and additional costs of site-specific validation. Thus, further development was considered both relevant and feasible and the potential impact of the analytics was considered substantial.

Case 3: diabetes mellitus

For diabetes, clinicians indicated that a highly relevant problem was to determine predictors of response to treatment with sodium glucose transporter-2 inhibitors combined with glucagon-like peptide-1 agonists. EHR data was available from diabetes patients treated in secondary care in the United Kingdom. However, a small sample size and substantial missing follow-up data raised questions about the feasibility of development, which resulted in the decision not to assess critical barriers and conduct an economic evaluation.

Discussion

In this paper, we present a framework that aims to promote the efficient development of high potential analytics by rapidly assessing whether it is feasible and worthwhile to continue development. The use cases demonstrate the value of first assessing the feasibility of development and identifying relevant barriers before estimating the potential health and economic benefits of analytics. Examples were presented for CLL and diabetes where development was not feasible given the data available. Furthermore, the essence of critically narrowing down the scope is demonstrated for CLL and the ICU where the absence of actionable output is an important barrier to realising success and disease prevalence strongly influences benefits.

Early economic evaluations of analytics can assist decision-making of developers and stimulates them to develop those analytics with the greatest potential benefits. These evaluations allow developers to assess the influence of certain requirements of analytics (e.g. the costs of the technology, validation and implementation) on their potential health and economic impact. In our use cases, we see risks that could strongly influence widespread adoption, such as the prevalence of CRBSI and the high drug costs for CLL. Early economic evaluations can also be used to strengthen the business case of developers seeking funding for prospective validation and evaluation. This is especially relevant since the high costs of validation and implementation are important barriers to successful use of analytics in clinical practice [4, 19, 24, 3133]. During implementation, data and tools used to perform early economic evaluations alongside development can be reused to perform a ‘late’ economic evaluation to convince payers that the analytics are worth purchasing. Elements covered in this framework align with key economic information sought by payers such as the UK’s National Institute for Health and Clinical Excellence [59].

However, for efficient development, economic evaluations should only be initiated for those applications deemed feasible and after ensuring that there are no critical barriers to success. Often multiple analytics can be developed for a single setting, disease or using a single dataset [27, 60]. For instance, for the ICU [11] and diabetes care [61] many more types of EHR-based analytics have been suggested than the ones presented here. This is an important difference compared to when early economic evaluations are used to assist decision-making during development of a technology with one or few applications (e.g. diagnostics). Since it is often unrealistic to evaluate—all potential applications of a particular type of analytics, our framework stimulates developers to select which applications are worthy of additional resources. Where feasibility is clearly a problem for the diabetes use case, the lack of an actionable output is the shortcoming for CLL; an issue often reported in the literature [10, 15, 24, 25]. The initial analyses performed in the early economic evaluation can be very simple at first but can become more complex as development progresses; this corresponds with recommendations that analytics development and validation should also be iterative [4, 62]. However, as with analytics for CLL and CRBSI, it is sometimes worthwhile to invest more time in adding additional details at an early stage, since it is better to fail fast when limited investments have been made. Using early economic evaluations in an iterative manner and providing a detailed definition of the scope aligns with best practices for early economic evaluations of other healthcare technologies such as diagnostic tests [3438]. The recommendations provided by others such as Drummond et al. [63], or Buisman et al. [37] regarding the selection of a model structure (e.g. decision tree, Markov model), estimation of input parameters, and calculating outcomes (such as the ICER) are likely to be applicable when estimating benefits. We demonstrate in the CLL and diabetes use cases how the framework may assist developers in selecting those applications that are likely to succeed, before investing additional resources in performing an economic evaluation. Similar to other papers [e.g. 4, 12, 17, 2426], we found the data available for development to be a barrier to success in the CLL and diabetes case studies. Analytics for artificial intelligence are ‘data hungry’ and therefore require large datasets [11, 27]. Furthermore, the quality of the data is an important issue when developing and using AI. Roberts et al. have emphasised in their review of AI for the diagnosis and prognostication of secondary pneumonia, that many AI analytics were hampered by poor quality data [64]. Our framework aligns with recommendations by Vollmer et al. who include critical questions regarding the data used as part of their framework to inform design and evaluate AI analytics [65]. Reviewing the data quality ensures developers select those applications of analytics for which development is most likely to succeed. For instance, rapid checks of potential sample sizes have been previously suggested [66]. For analytics with adequate data quality, additional resources can then be invested to perform an economic evaluation.

In this study, the framework was applied to three clinical use cases. Therefore, validation in other use cases is recommended. Other use cases can include different clinical areas (e.g. psychiatric disorders) but also other data sources such as data from patient devices (e.g. Fitbits), imaging and social media. Additional research could also assess criteria to value the quality of unstructured data. Furthermore, the framework presented could be easily adopted alongside initiatives such as RE-AIM used to translate research into practice [67]. This framework pays particular attention to the timing of economic evaluations intended to assist development considering relevant elements in the ‘Reach’, ‘Effectiveness’, ‘Adoption’, ‘Implementation’ and ‘Maintenance’ steps.

Since many factors can influence the successful implementation and adoption of analytics, we may have adopted a somewhat narrow approach by solely focusing on the value of economic evaluations to support developer decision-making. A wider form of decision support can be achieved through a broader evaluation of analytics, for instance using health technology assessment, which includes social, and ethical elements besides the health and economic impact [68]. Moreover, elicitation of stakeholder preferences such as patients and clinicians could ensure that potential barriers to development, acceptability and implementation are addressed [69].

In recent years, there has been an increased interest in the ethical challenges that we face relating to the adoption of artificial intelligence [70]. In this paper, we discuss that factors such as the risk of bias and small sample sizes, should be assessed at an early stage of development prior to performing an economic evaluation. Trocin et al. emphasise the severity of the consequences of failing to do so. Some of the challenges relating to the data quality mentioned in this paper have also been emphasised by Trocin et al. Moreover, these authors also provide research questions that need to be answered to ensure the responsible adoption of AI related technologies [70]. Many answers to these questions could be very relevant for future improvements of the flowchart. Depending on the setting and type of analytics, for instance, the quality of the data can be assessed according to the risk of selection bias in the data [4, 13], or the absence of ethnic variation in the data which could limit generalisability of machine learning models [4, 17, 28].

Conclusions

This is the first study providing recommendations on the use of economic evaluations to support development decisions of analytics for big data and artificial intelligence-based solutions. Many types of analytics can be developed within a specific clinical setting or disease or using a particular dataset. The framework presented in this study stimulates efficiency of development by selecting those applications worth further investment after assessing the feasibility of development and identifying critical barriers. For these applications, early economic evaluations can assist decision-making of analytics developers by estimating for instance requirements of effectiveness and the headroom for pricing, validation, and implementation.

Availability of data and materials

Not applicable.

Abbreviations

AI:

Artificial intelligence

AUC:

Area under the curve

CLL:

Chronic lymphocytic leukaemia

CRBSI:

Catheter related bloodstream infection

EHR:

Electronic health record

ICU:

Intensive care unit

NGS:

Next generation sequencing

PICO:

Population, intervention, comparator, and outcomes

QALYs:

Quality-adjusted life years

W&W:

Watch and wait

References

  1. 1.

    Morello L, Guglielmi G. US science agencies set to win big in budget deal. Nature. 2018;555(7698):572–3.

    CAS  PubMed  Google Scholar 

  2. 2.

    Banks MA. Sizing up big data. Nat Med. 2020;26:5–6.

    CAS  PubMed  Google Scholar 

  3. 3.

    Kisner J. Creating Shareholder Value with AI? Not so Elementary, My Dear Watson. [Internet] Jefferies Group LLC; 2017[cited 2021, November 10]. https://javatar.bluematrix.com/pdf/fO5xcWjc

  4. 4.

    Fröhlich H, Balling R, Beerenwinkel N, et al. From hype to reality: data science enabling personalized medicine. BMC Med. 2018;16:1–15.

    Google Scholar 

  5. 5.

    Furlow, B. ASCO announces funding for CancerLinQ Clinical Data Analysis Initiative. 2012; Available at https://www.cancertherapyadvisor.com/home/cancer-topics/general-oncology/asco-announces-funding-for-cancerlinq-clinical-data-analysis-initiative/. Accessed 12 Dec 2020.

  6. 6.

    McLachlan S, Dube K, Johnson O, Buchanan D, Potts HW, Gallagher T, Fenton N. A framework for analysing learning health systems: are we removing the most impactful barriers? Learn Health Syst. 2019;3(4):e10189.

    PubMed  PubMed Central  Google Scholar 

  7. 7.

    Mandl KD, Kohane IS, McFadden D, Weber GM, Natter M, Mandel J, Schneeweiss S, Weiler S, Klann JG, Bickel J, Adams WG. Scalable collaborative infrastructure for a learning healthcare system (SCILHS): architecture. J Am Med Inform Assoc. 2014;21(4):615–20.

    PubMed  PubMed Central  Google Scholar 

  8. 8.

    Big Data to Knowledge. 2019; Available at https://commonfund.nih.gov/bd2k. Accessed 17 Dec 2020.

  9. 9.

    Agency for Healthcare Research and Quality. 2020; Available at https://www.ahrq.gov/funding/training-grants/lhs-k12.html. Accessed 17 Dec 2020.

  10. 10.

    Sanchez-Pinto LN, Luo Y, Churpek MM. Big data and data science in critical care. Chest. 2018;154(5):1239–48.

    PubMed  PubMed Central  Google Scholar 

  11. 11.

    Gutierrez G. Artificial intelligence in the intensive care unit. Crit Care. 2020;24:1–9.

    Google Scholar 

  12. 12.

    Hemingway H, Asselbergs FW, Danesh J, et al. Big data from electronic health records for early and late translational cardiovascular research: challenges and potential. Eur Heart J. 2017;39(16):1481–95.

    PubMed Central  Google Scholar 

  13. 13.

    Rumsfeld JS, Joynt KE, Maddox TM. Big data analytics to improve cardiovascular care: promise and challenges. Nat Rev Cardiol. 2016;13(6):350.

    CAS  PubMed  Google Scholar 

  14. 14.

    Bates DW, Saria S, Ohno-Machado L, et al. Big data in health care: using analytics to identify and manage high-risk and high-cost patients. Health Aff (Millwood). 2014;33:1123–31.

    Google Scholar 

  15. 15.

    Phillips KA, Trosman JR, Kelley RK, et al. Genomic sequencing: assessing the health care system, policy, and big-data implications. Health Aff (Millwood). 2014;33(7):1246–53.

    Google Scholar 

  16. 16.

    El Morr C, Ali-Hassan H. Analytics in healthcare: a practical introduction. Berlin: Springer; 2019.

    Google Scholar 

  17. 17.

    Kelly CJ, Karthikesalingam A, Suleyman M, Corrado G, King D. Key challenges for delivering clinical impact with artificial intelligence. BMC Med. 2019;17(1):195.

    PubMed  PubMed Central  Google Scholar 

  18. 18.

    Magrabi F, Ammenwerth E, McNair JB, De Keizer NF, Hyppönen H, Nykänen P, Rigby M, Scott PJ, Vehko T, Wong ZS, Georgiou A. Artificial intelligence in clinical decision support: challenges for evaluating AI and practical implications: a position paper from the IMIA Technology Assessment & Quality Development in Health Informatics Working Group and the EFMI Working Group for Assessment of Health Information Systems. Yearb Med Inform. 2019;28(1):128.

    PubMed  PubMed Central  Google Scholar 

  19. 19.

    Morse KE, Bagely SC, Shah NH. Estimate the hidden deployment cost of predictive models to improve patient care. Nat Med. 2020;26:18–9.

    CAS  PubMed  Google Scholar 

  20. 20.

    Budrionis A, Bellika JG. The learning healthcare system: where are we now? A systematic review. J Biomed Inform. 2016;64:87–92.

    PubMed  Google Scholar 

  21. 21.

    Mehta N, Pandit A. Concurrence of big data analytics and healthcare: a systematic review. Int J Med Inform. 2018;114:57–65.

    PubMed  Google Scholar 

  22. 22.

    Bakker L, Aarts J, Uyl-de Groot C, Redekop W. Economic evaluations of big data analytics for clinical decision-making: a scoping review. J Am Med Inform Assoc. 2020;27(9):1466–75.

    PubMed  PubMed Central  Google Scholar 

  23. 23.

    Wolff J, Pauling J, Keck A, Baumbach J. The economic impact of artificial intelligence in health care: systematic review. J Med Internet Res. 2020;22(2):e16866.

    PubMed  PubMed Central  Google Scholar 

  24. 24.

    Shilo S, Rossman H, Segal E. Axes of a revolution: challenges and promises of big data in healthcare. Nat Med. 2020;26(1):29–38.

    CAS  PubMed  Google Scholar 

  25. 25.

    Prosperi M, Min JS, Bian J, et al. Big data hurdles in precision medicine and precision public health. BMC Med Inform Decis Mak. 2018;18(1):139.

    PubMed  PubMed Central  Google Scholar 

  26. 26.

    Beckmann JS, Lew D. Reconciling evidence-based medicine and precision medicine in the era of big data: challenges and opportunities review. Genome Med. 2016;8:134.

    PubMed  PubMed Central  Google Scholar 

  27. 27.

    Obermeyer Z, Emanuel EJ. Predicting the future—big data, machine learning, and clinical medicine. N Engl J Med. 2016;375(13):1216.

    PubMed  PubMed Central  Google Scholar 

  28. 28.

    Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med. 2019;25(1):44–56.

    CAS  PubMed  Google Scholar 

  29. 29.

    Goldstein BA, Navar AM, Pencina MJ, Ioannidis J. Opportunities and challenges in developing risk prediction models with electronic health records data: a systematic review. J Am Med Inform Assoc. 2017;24(1):198–208.

    PubMed  Google Scholar 

  30. 30.

    Farmer R, Mathur R, Bhaskaran K, Eastwood SV, Chaturvedi N, Smeeth L. Promises and pitfalls of electronic health record analysis. Diabetologia. 2018;61(6):1241–8.

    PubMed  Google Scholar 

  31. 31.

    Marsolo K, Margolis PA, Forrest CB, Colletti RB, Hutton JJ. A digital architecture for a network-based learning health system: integrating chronic care management, quality improvement, and research. eGEMs. 2015;3(1):1168.

    PubMed  PubMed Central  Google Scholar 

  32. 32.

    Bhandari RP, Feinstein AB, Huestis SE, Krane EJ, Dunn AL, Cohen LL, Kao MC, Darnall BD, Mackey SC. Pediatric-Collaborative Health Outcomes Information Registry (Peds-CHOIR): a learning health system to guide pediatric pain research and treatment. Pain. 2016;157(9):2033.

    PubMed  PubMed Central  Google Scholar 

  33. 33.

    Feeley TW, Sledge GW, Levit L, Ganz PA. Improving the quality of cancer care in America through health information technology. J Am Med Inform Assoc. 2014;21(5):772–5.

    PubMed  Google Scholar 

  34. 34.

    Sculpher M, Drummond M, Buxton M. The iterative use of economic evaluation as part of the process of health technology assessment. J Health Serv Res Policy. 1997;2(1):26–30.

    CAS  PubMed  Google Scholar 

  35. 35.

    Annemans L, Genesté B, Jolain B. Early modelling for assessing health and economic outcomes of drug therapy. Value Health. 2000;3(6):427–34.

    CAS  PubMed  Google Scholar 

  36. 36.

    Pietzsch JB, Paté-Cornell ME. Early technology assessment of new medical devices. Int J Technol Assess Health Care. 2008;24(1):36–44.

    PubMed  Google Scholar 

  37. 37.

    Buisman LR, Rutten-van Mölken MPMH, Postmus D, et al. The early bird catches the worm: early cost-effectiveness analysis of new medical tests. Int J Technol Assess Health Care. 2016;32(1–2):46–53.

    PubMed  Google Scholar 

  38. 38.

    Ijzerman MJ, Steuten LMG. Early assessment of medical technologies to inform product development and market access. Appl Health Econ Health Policy. 2011;9(5):331–47.

    PubMed  Google Scholar 

  39. 39.

    Shah NH, Milstein A, Bagley SC. Making machine learning models clinically useful. JAMA. 2019;322(14):1351–2.

    PubMed  Google Scholar 

  40. 40.

    Simpson L, Dr Lisa Simpson Interview, Foley T, Fairmichael F, editors. The Learning Healthcare Project: web. 2015. www.thelearninghealthcareproject.org.

  41. 41.

    Wiens J, Saria S, Sendak M, Ghassemi M, Liu VX, Doshi-Velez F, Jung K, Heller K, Kale D, Saeed M, Ossorio PN. Do no harm: a roadmap for responsible machine learning for health care. Nat Med. 2019;25(9):1337–40.

    CAS  PubMed  Google Scholar 

  42. 42.

    Vickers AJ, Van Calster B, Steyerberg EW. Net benefit approaches to the evaluation of prediction models, molecular markers, and diagnostic tests. BMJ. 2016;352:i6.

    PubMed  PubMed Central  Google Scholar 

  43. 43.

    Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models. Med Decis Mak. 2006;26(6):565–74.

    Google Scholar 

  44. 44.

    Liu VX, Bates DW, Wiens J, Shah NH. The number needed to benefit: estimating the value of predictive analytics in healthcare. J Am Med Inform Assoc. 2019;26(12):1655–9.

    PubMed  PubMed Central  Google Scholar 

  45. 45.

    Eddy DM, Hollingworth W, Caro JJ, Tsevat J, McDonald KM, Wong JB. Model transparency and validation: a report of the ISPOR-SMDM Modeling Good Research Practices Task Force–7. Med Decis Mak. 2012;32(5):733–43.

    Google Scholar 

  46. 46.

    Girling A, Lilford R, Cole A, Young T. Headroom approach to device development: current and future directions. Int J Technol Assess Health Care. 2015;31(5):331–8.

    PubMed  PubMed Central  Google Scholar 

  47. 47.

    Baliakas P, Mattsson M, Stamatopoulos K, Rosenquist R. Prognostic indices in chronic lymphocytic leukemia: where do we stand how do we proceed? J Intern Med. 2016;279(4):347–57.

    CAS  PubMed  Google Scholar 

  48. 48.

    Eichhorst B, Robak T, Montserrat E, Ghia P, Niemann CU, Kater AP, Gregor M, Cymbalista F, Buske C, Hillmen P, Hallek M. Chronic lymphocytic leukaemia: ESMO Clinical Practice Guidelines for diagnosis, treatment and follow-up. Ann Oncol. 2021;32(1):23–33.

    CAS  PubMed  Google Scholar 

  49. 49.

    Condoluci A, Terzi di Bergamo L, Langerbeins P, Hoechstetter MA, Herling CD, De Paoli L, Delgado J, Rabe KG, Gentile M, Doubek M, Mauro FR. International prognostic score for asymptomatic early-stage chronic lymphocytic leukemia. Blood J Am Soc Hematol. 2020;135(21):1859–69.

    Google Scholar 

  50. 50.

    Langerbeins P, Bahlo J, Rhein C, Gerwin H, Cramer P, Fürstenau M, Al-Sawaf O, von Tresckow J, Fink AM, Kreuzer K, Vehling-Kaiser U. Ibrutinib versus placebo in patients with asymptomatic, treatment-naïve early stage CLL: primary endpoint results of the phase 3 double-blind randomized CLL12 trial. Hematol Oncol. 2019;37:38–40.

    Google Scholar 

  51. 51.

    Blot SI, Depuydt P, Annemans L, et al. Clinical and economic outcomes in critically ill patients with nosocomial catheter-related bloodstream infections. Clin Infect Dis. 2005;41(11):1591–8.

    PubMed  Google Scholar 

  52. 52.

    Blot S, Poulakou G, Timsit JF. Catheter-associated bloodstream infection rates: how low can you go? Intensive Care Med. 2019;45(6):896–7.

    PubMed  Google Scholar 

  53. 53.

    GiViT, Gruppo Italiano per la Valutazione degli Interventi In Terapia Intensiva. Report PROSAFE project. 2014. 2014; Centre GR001.

  54. 54.

    Apostolopoulou E, Raftopoulos V, Filntisis G, et al. Surveillance of device-associated infection rates and mortality in 3 greek intensive care units. Am J Crit Care. 2013;22(3):e12–20.

    PubMed  Google Scholar 

  55. 55.

    Ferrer R, Martin-Loeches I, Phillips G, et al. Empiric antibiotic treatment reduces mortality in severe sepsis and septic shock from the first hour: results from a guideline-based performance improvement program. Crit Care Med. 2014;42(8):1749–55.

    CAS  PubMed  Google Scholar 

  56. 56.

    De Wit M, Miller KB, Green DA, Ostman HE, Gennings C, Epstein SK. Ineffective triggering predicts increased duration of mechanical ventilation. Crit Care Med. 2009;37(10):2740–5.

    PubMed  Google Scholar 

  57. 57.

    Georgopoulos D. Ineffective efforts during mechanical ventilation: the brain wants, the machine declines. Intensive Care Med. 2012;38:738–40.

    PubMed  Google Scholar 

  58. 58.

    Bakker L, Vaporidi K, Aarts J, Redekop W. The potential of real-time analytics to improve care for mechanically ventilated patients in the intensive care unit: an early economic evaluation. Cost Eff Resour Alloc. 2020;18(1):1.

    Google Scholar 

  59. 59.

    National Institute for Health and Care Excellence. Evidence standards framework for digital health technologies. London: NHS England; 2019.

    Google Scholar 

  60. 60.

    Zeitoun JD, Ravaud P. Artificial intelligence in health care: value for whom? Lancet Digit Health. 2020;2(7):e338–9.

    PubMed  Google Scholar 

  61. 61.

    Riddle MC, Blonde L, Gerstein HC, Gregg EW, Holman RR, Lachin JM, Nichols GA, Turchin A, Cefalu WT. Diabetes Care Editors’ Expert Forum 2018: managing big data for diabetes research and care. Diabetes Care. 2019;42(6):1136–46.

    PubMed  Google Scholar 

  62. 62.

    Newton KM, Peissig PL, Kho AN, Bielinski SJ, Berg RL, Choudhary V, Basford M, Chute CG, Kullo IJ, Li R, Pacheco JA, Rasmussen LV, Spangler L, Denny JC. Validation of electronic medical record-based phenotyping algorithms: results and lessons learned from the eMERGE network. J Am Med Inform Assoc. 2013;20(e1):e147–54.

    PubMed  PubMed Central  Google Scholar 

  63. 63.

    Drummond MF, Sculpher MJ, Claxton K, Stoddart GL, Torrance GW. Methods for the economic evaluation of health care programmes. Oxford: Oxford University Press; 2015.

    Google Scholar 

  64. 64.

    Roberts M, Driggs D, Thorpe M, Gilbey J, Yeung M, Ursprung S, Aviles-Rivero AI, Etmann C, McCague C, Beer L, Weir-McCall JR. Common pitfalls and recommendations for using machine learning to detect and prognosticate for COVID-19 using chest radiographs and CT scans. Nat Mach Intell. 2021;3(3):199–217.

    Google Scholar 

  65. 65.

    Vollmer S, Mateen BA, Bohner G, Király FJ, Ghani R, Jonsson P, Cumbers S, Jonas A, McAllister KS, Myles P, Granger D. Machine learning and artificial intelligence research for patient benefit: 20 critical questions on transparency, replicability, ethics, and effectiveness. BMJ. 2020;368:16927.

    Google Scholar 

  66. 66.

    Wallace PJ, Shah ND, Dennen T, Bleicher PA, Crown WH. Optum Labs: building a novel node in the learning health care system. Health Aff (Millwood). 2014;33(7):1187–94.

    Google Scholar 

  67. 67.

    RE-AIM. 2021; Available at https://www.re-aim.org/. Accessed 30 June 2021.

  68. 68.

    Kristensen FB, Lampe K, Chase DL, Lee-Robin SH, Wild C, Moharra M, Garrido MV, Nielsen CP, Røttingen JA, Neikter SA, Bistrup ML. Practical tools and methods for health technology assessment in Europe: structures, methodologies, and tools developed by the European network for Health Technology Assessment, EUnetHTA. Int J Technol Assess Health Care. 2009;25(S2):1–8.

    PubMed  Google Scholar 

  69. 69.

    Salloum RG, Shenkman EA, Louviere JJ, Chambers DA. Application of discrete choice experiments to enhance stakeholder engagement as a strategy for advancing implementation: a systematic review. Implement Sci. 2017;12(1):1–2.

    Google Scholar 

  70. 70.

    Trocin C, Mikalef P, Papamitsiou Z, Conboy K. Responsible AI for digital health: a synthesis and a research agenda. Inf Syst Front. 2021;26:1–9.

    Google Scholar 

Download references

Acknowledgements

We would like to thank Katerina Vaporidi for all her input for the cost-effectiveness analyses for the intensive care and all partners in the AEGLE project for the many discussions performed throughout the project.

Funding

This work was supported by European Union’s Horizon 2020 Research and Innovation Programme Grant Number 644906. The funding body had no role in the design of the study and no influence on study results.

Author information

Affiliations

Authors

Contributions

LB contributed to developing the framework, performing the cost-effectiveness analyses, drafting and revising the manuscript. JA contributed to developing the framework and drafting and revising the manuscript. CUG contributed to the cost-effectiveness analyses for chronic lymphocytic leukaemia and drafting and revising the manuscript. WR contributed to developing the framework, the cost-effectiveness analyses and drafting and revising the manuscript. All authors (LB, JA, CUG, WR) read and approved the final manuscript.

Corresponding author

Correspondence to Lytske Bakker.

Ethics declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

CUG reports unrestricted grants from Boehringer Ingelheim, Celgene, Janssen-Cilag, Genzyme, Astellas, Sanofi, Roche, Astra Zeneca, Amgen, Gilead, Merck, Bayer, outside the submitted work. The remaining authors have no competing interest to declare.

Additional information

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Additional file 1:

Description of model structure, input parameters and results for the example use cases.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and Permissions

About this article

Verify currency and authenticity via CrossMark

Cite this article

Bakker, L., Aarts, J., Uyl-de Groot, C. et al. How can we discover the most valuable types of big data and artificial intelligence-based solutions? A methodology for the efficient development of the underlying analytics that improve care. BMC Med Inform Decis Mak 21, 336 (2021). https://doi.org/10.1186/s12911-021-01682-9

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doi.org/10.1186/s12911-021-01682-9

Keywords

  • Analytics
  • Artificial intelligence
  • Big data
  • Cost–benefit analysis
  • Critical care
  • Chronic lymphocytic leukaemia