1 Introduction

Clinical laboratories provide medical test services to a variety of customers. From a business perspective, providing high-quality services at lower costs mandates the need for efficient utilization of clinical resources (McDonald et al., 1997). Historical and current clinical resource utilization data sets are now digitally available at almost all clinical facilities (Kessel et al., 2014). The improvement of digital equipment facilitates the acquisition and storage of all kind of clinical data at a rate much faster than what can be processed using traditional processing systems (Peters and Buntrock, 2014), which give rise to the clinical Big Data era.

Clinical Big Data is the technical term used to describe massive databases with varying high volume, variety, and velocity [e.g., electronic medical record (EMR) and biometrics data]. These databases present difficulties with storage, processing, and visualization (Rajaraman and Ullman, 2012; Coulouris et al., 2005).

Business intelligence (BI) is defined as the use of data and dedicated processing to assist informed decisions in diverse administrative settings (Pine et al., 2011). An intrinsic trait of an BI is that it integrates data from a diversity of sources, resulting in an effective information framework for clinical laboratory decision makers (Negash, 2004).

A clinical laboratory facility can have many managers with different perspectives (Bennett, 2007) (e.g., clinical section manager, logistic manager, finance manager, etc.). BI can deliver assistances to clinical laboratory directors comprising effective utilization of human resources (Crist-Grundman and Mulrooney, 2010), improved process efficiency, and unnecessary cost prevention (Foshay and Kuziemsky, 2014). BI should help clinical laboratory managers with different roles in describing and diagnose current performance and predicting the future performance. To deal with these challenges, new software programming frameworks to multithread computing tasks have been developed (Coulouris et al., 2005; de Oliveira Branco, 2009; Dean and Ghemawat, 2008). However, many clinical laboratory settings have not yet applied BI systems (Foshay and Kuziemsky, 2014), and there has been very limited research on the factors that contribute to the successful implementation of BI in a health care–specific context (Foshay and Kuziemsky, 2014).

Several BI applications already exist (Anon, 2014e, 2014f, 2014d, 2014b, 2014c, Publishing, 2014), however they usually suffer from two drawbacks: firstly, they offer a rigid and less flexible framework that addresses the need of a specific group of users; secondly, they adopt the scope of scenario-based simulation rather than Big Data–driven analytics.

New challenges have emerged, especially with respect to processing of the massive data sets that are being produced daily. This drives the need for a new framework that can be used by different clinical laboratory managers to support the informed decisions.

In this chapter, we explore a combination of reusable data analytics (i.e., BI framework) that helps clinical laboratory managers with different roles to become more effective by supporting their decisions using the BI framework. Two real-life case studies (namely, clinical laboratory test usage pattern visualization and estimation of clinical test volumes) are designed to identify potential analytics to support decisions of laboratory managers. The objective is to identify the reusable components of a user-centered framework (i.e., clinical laboratory managers with different roles and perspectives) based on Big Data analytics of clinical data sets. The implementation and validation of other framework components/analytics will be explored in future research.

2 Motivation

The case studies have emerged from real-life concerns of laboratory managers. Increases in population have a stunning influence on health care, including laboratory diagnostics. Clinical laboratories in Canada have experienced substantial growth in utilization in recent years. However, in Alberta, the volumes of all types of laboratory testing are increasing at a rate much faster than population growth, with a 36% increase in chemistry test volumes between 2003 and 2009 (Di Matteo and Di Matteo, 2009). There have been repeated statements that a substantial amount of laboratory tests are tangibly redundant. A current meta-analysis assessed the proportion of redundant tests to be in the range of 30% (Zhi et al., 2013). In Calgary alone, this equates to as many as 8 million potentially unnecessary laboratory tests per year, representing a cost of at least $80 million per year in direct and indirect costs. Indications that additional test utilization does not advance clinical outcomes have been reported in a number of studies that have found no association between the volume of tests ordered and clinical outcome (Daniels and Schroeder, 1977; Ashley et al., 1972; Bell et al., 1998; Powell and Hampers, 2003). Redundant laboratory tests not only waste valuable resources, but also may lead to patient impairment as false positives. If a healthy person is subjected to 10 unrelated (unnecessary) tests, the probability of at least 1 deviant test result is 40% (Axt-Adam et al., 1993). This deviant test result may lead to unnecessary, far-reaching, expensive, and time-consuming diagnostic examinations (Lewandrowski, 2002).

The driving forces for the efforts illustrated in this chapter are derived from the literature cited above, and the current BI frameworks are based on text and flowchart process simulation for best scenario practices and optimization rather than Big Data analytics (Anon, 2014b, 2014c, Publishing, 2014). Text- and flowchart-based approaches are subjective (i.e., users use different synonymous and acronyms, and different flowcharts are created for the same process). On the other hand, Big Data are more representative of the variations in a given process. Moreover, representing a complex process using text and flowcharts may be inaccurate, while complex process traits are encapsulated in the Big Data representation form. Furthermore, as more and more data are collected for the same measurement, the data model becomes more and more accurate.

Another driving motivation is that there is no efficient statistical method to highlight the overutilization or underutilization of clinical laboratory Big Data (Naugler, 2013, 2014; MacMillan, 2014). Moreover, physician test ordering patterns have no analyzing/feedback mechanism (Plebani, 1999; Plebani et al., 2014; Kiechle et al., 2014), and thus, a significant amount of clinical tests and resources (e.g., medical equipment utilization, more technician workload, and cost, etc.) are misused. Furthermore, there is no efficient method to measure or visualize a human performance index (Wennberg, 2004; Ashton et al., 1999; Monsen et al., 2008)—analysis of variance. In addition, most managerial decisions are based on descriptive analytics that tell what happened, and little are based on predictive analytics that tell what is going to happen (Davenport, 2013).

4 Use-cases

In this section, we present the details of the two use-cases. The first reflects a typical scenario for the use of analytics for visualization of the clinical laboratory test usage problem, and the second employs Big Data analytics for prediction. Many of the typical clinical decision support use-cases fall into either of these categories.

Precisely stated, the purpose of these use-cases is to estimate the clinical test volumes of a given test per unit time—a time series analysis (i.e., Alberta provincial test volume estimation). Moreover, the clinical laboratory test utilization pattern service has a novel method to visualize the individual physician usage pattern for a given test or a test panel (i.e., a group of standard tests). In the following discussion, the clinical laboratory test utilization pattern visualization and provincial laboratory test volume estimation are illustrated and detailed, along with the associated design limitations.

5 Case Study 1: Clinical laboratory test usage patterns visualization

Excessive usage of diagnostic tests is a key challenge in health-care systems. Data sets of clinical test volumes for individual physicians are available from hospitals’ laboratory information systems (LISs). The unjustified usage of clinical tests drives the need for a tool to underline the usage pattern (utilization) of a given clinical test or test panel of a physician among his or her peers. The data sets must be normalized for different physician characteristics (e.g., number of working hours, years of experience, etc.,), patient status (e.g., condition of the patient, number of visits, etc.), and department workload (e.g., emergency, outpatient, hospitalized patient, etc.).

6 Data source and methodology

For illustration, we use a simple example representing a set of five different clinical tests with varying test volumes are acquired for 35 physicians over a 3-month time span at one medical facility at the city of Calgary. The data were anonymized to comply with the Personal Information Protection and Electronic Documents Act of Canada. (Anon, 2014j).

In this section, a novel graphical tool based on the z-score to identify extremes of practice variance from laboratory test ordering data is used to analyze and visualize the usage pattern of the recorded five clinical tests. This methodology can be used to assess the efficacy of a test usage control criterion (Murphy and Henry, 1978; Plebani et al., 2014; Bates et al., 1999). Evaluation and visualization of physician usage patterns of clinical tests could yield a beneficial tool to visualize physician utilization of clinical tests. In the following discussion, we describe a novel method for comparing the utilization patterns of different physicians using laboratory data. In order to use this method, the number of physicians in a given data set had to be greater than or equal to 30; otherwise, we used the T-test (Rosner, 2010).

The steps of the proposed method are:

1. Acquire test volumes for a group of physicians for a given test or test panel. The test volumes must be normalized per physician for a given range of physician characteristics, patient status, and department workload, as described earlier in this chapter.

2. Calculate the mean and standard deviation of the test volumes per test as shown by Eqs. (32.1) and (32.2):

${\mu }_T=\frac{1}{N}{\sum }_{p=N}^{p=1}{X}_p\text{,}$

where μ_T is the mean of a given test (T); N is the number of physicians using test (T); and X_p is the volume of test (T) used by physician (p).

${\sigma }_T=\sqrt{\frac{{\sum }_{p=N}^{p=1}{\left(X_p-{\mu }_T\right)}^2}{N}}\text{,}$

where σ_T is the standard deviation of a given test (T).

3. Calculate the z-score for each test per physician, as shown by Eq. (32.3):

${\mathit{zscore}}_{p\left|T\right.}=\frac{X_p-{\mu }_T}{{\sigma }_T}\text{,}$

where zscore_p|T is the z-score of test (T), which has test volume X_p that is used by the physician (p).
The adaptation of test volumes into z-scores permits assessment of clinical laboratory tests with different average volumes.

4. Calculate the mean z-score for all tests and the standard deviation of their z-scores per physician, as shown in Eqs. (32.4) and (32.5):

${\mu }_{{\mathit{zscore}}_p}=\frac{1}{T}{\sum }_{i=\tau }^{i=1}{\mathit{zscore}}_{p\left|T_i\right.}$

${\sigma }_{{\mathit{zscore}}_p}=\sqrt{\frac{{\sum }_{i=\tau }^{i=1}{\left({\mathit{zscore}}_{p\left|T_i\right.}-{\mu }_{{\mathit{zscore}}_p}\right)}^2}{T}}$

5. For each physician, plot the mean z-scores against the standard deviation of z-scores for that physician.

6. Define performance marker lines at the mean z-score and the mean standard deviation of z-scores to divide the z-score space into four regions. The explanation of these regions is described in Table 32.1.

7 Results and discussion

The data set for 35 physicians utilizing five different tests (i.e., chemical seven test panel CH7; D-Dimer, DDEL; throat swab for beta hemolytic streptococcus, M BETA; urine culture, M URINE; and Troponin I, TNIV) were processed via the proposed method using the R statistical package (Anon, 2014m), the z-scores for each test were calculated, and the average z-score for all five tests was also calculated. Figure 32.3 illustrates the usage pattern of the CH7 test by the 35 physicians, illustrating that most of the physicians belonged to group D, a small number of physicians were in groups A and B, and only a few physicians belonged to group C. This reveals an optimal degree of usage of this test.

The vertical, solid black line dividing the golden rectangle characterizes the mean usage pattern of the five tests (x-axis). A physician who equally uses the five tests is characterized as a red dot on this line. A physician who uses all five tests at a slower rate than average will be characterized by a red dot on the left side of this line, and vice versa.

The y-axis characterizes the usage probability of this test (i.e., CH7), represented by its computed z-score. It would not be possible to convert the z-scores into probabilities unless it is assumed that the test volumes are drawn from a normal distribution. A adaptation table to transform z-scores into probabilities can be found in Held and Bove (2014) and Rosner (2010).

The allowed disparity within the test volumes for a given test is represented by the width of the golden rectangle, and the height represents the range of the usage probabilities. The horizontal line crossing the zero of the y-axis represents the usage probability of the mean of the underlined test. The golden rectangle space is divided into four groups by the intersection of the vertical line representing the mean usage pattern of the test and the zero z-score line. A physician with perfect average usage will be represented by a dot at the intersection of these two lines. The distribution of the points outside the rectangle represents the outliers. There are many factors that may drive the physician to utilize more tests to have a better assessment of the patient conditions, and thus, Eq. (32.3) must be normalized by these factors, if quantifiable, and the modified equation for computing the z-score becomes

${\mathit{zscoreMod}}_{p\left|T\right.}=\frac{X_p-{\mu }_T}{{\sigma }_T*NFs}$

where zscoreMod is the modified z-score and NFs are the product of all the normalization factors.

Figure 32.4 shows the average usage pattern for all five tests, the average usage pattern illustrates the most common group is Group D, with a wider probability of usage for all five tests.

8 Limitations

The visualization model assumes that the data is drawn from normal distributions to translate the z-score to the probability of usage. This visualization framework is designed to view the relative usage pattern of physicians utilizing a group of clinical tests. This framework is used to visualize practice variance, not to provide conclusive proof of redundant usage of clinical tests. The framework has some restrictions, as it does not account for prospective influencing parameters such as limited ranges of practice, patient conditions, and number of hospital visits per patient that may affect test usage characteristics, as these parameters are not quantified.

9 Case Study 2: Provincial laboratory clinical test volume estimation

One main characteristic of any clinical laboratory BI platform is the ability to estimate clinical laboratory test volumes (El-Gayar and Timsina, 2014; Ferranti et al., 2010; Ashrafi et al., 2014). Accordingly, test consumables and workload can be optimized by short-term reliable forecasting. Large amounts of clinical test volume data are accessible from a variety of available data ware house in hospitals and clinical settings (Anon, 2014g). In this section, a Holt-Winters method (De Gooijer and Hyndman, 2006; Chatfield and Yar, 1988) is used to estimate the province of Alberta laboratory test volume. The results are compared to an estimation using an Auto Regressive Integrated Moving Average (ARIMA) model (Rosner, 2010).

10 Data source and methodology

A huge volume of data has been queried to consolidate a data set of clinical laboratory test volume for all ordered tests over a 40-month span from all Alberta medical facilities. This data is provided by provincial laboratory utilization office, CLS, and at the University of Calgary (Anon, 2014g).

10.3 Stationary testing

The ADF test is based on the null hypothesis test. The more negative the test is, the stronger the rejections of the null hypothesis, which announces the stationary nature of the time series. Table 32.2 shows the result of the ADF test and the associated ρ values.

Table 32.2

ADF Test Output

Time Series	ADF Test
	Dickey-Fuller statistic	ρ (significance <= 0.01)
Provincial test volumes	− 5.6335	0.01

Auto-correlation function (ACF) is a measure of the degree of similarity between the original time series and the time lag version of the time series. ACF can evaluate the stationary nature of a time series, high correlation resulting from high similarity, and less stationary representing a slowly decaying ACF. Figure 32.5 shows the ACF of the provincial test volumes time series, which reflects the stationary nature of the recorded data when correlated with itself for different time lags.

11 Results and discussion

Figures 32.6, 32.7, and 32.8 show the performance of the Holt-Winters multiplicative, Holt-Winters additive, and ARIMA (1,0,1) models to fit the provincial volume time series and estimate future values with a 95% confidence interval. The observed data are represented by the black curve; the fitted/estimated data are shown in red, and the 95% confidence intervals are shown in blue. The 95% confidence intervals signify the disparities (i.e., precision) of the model.

The estimation of the smoothing parameters α, β, and γ are based on minimizing one step ahead of the MSE. If these parameters are found to be near zero, it means that the model adds more weight to the previous samples, resulting in more smoothing. Nevertheless, if these parameters are found to be near 1, it means the model tends to add less weight to the previous samples and thus feature less smoothing.

Minimum AIC and BIC information criteria is used as model selection criteria whenever the performance parameters (i.e., MSE and R²) are close for the model ensemble. Table 32.6 shows the AIC and BIC values for the Holt-Winters multiplicative and additive models, along with the selected model. When R² < = 0.5, then it may be better to choose the mean model to fit and estimate the time series (De Gooijer and Hyndman, 2006), which is the case for the ARIMA (1,0,1) model with R² = 0.49541.

Table 32.6 shows that the Holt-Winters additive model is a better model to fit and estimate the provincial test volume time series. This means that the level, trend, and the error are additive in nature. This is clearly shown in Figure 32.9, where the estimated values by the additive model (i.e., xhat), level, trend, and seasonality show additive behavior.

Figure 32.10 shows the histogram of the residual error of the Holt-Winters additive model with a normal distribution curve fitted to the residual errors. The diagram shows that the error is almost centered on 0 mean, which means that the errors are uncorrelated and normally distributed. This strengthens the choosing of the additive Holt-Winters model to fit and estimate the data.

12 Limitations

Figure 32.7 shows the estimation horizon of 6 months in the future of the additive Holt-Winters model. The estimation horizon of the 6 months in the future and the upper and lower limits (i.e., precision) are shown in Table 32.7. When the estimation horizon moves forward, the 95% confidence interval becomes wider, resulting in poor precision. This drawback results from the limitation that the additive Holt-Winters model is static (i.e., the exponential smoothing parameters are calculated only once from the historical data). This is due to the limited computational resources as the time series increases with time. A good solution is to implement an adaptive Holt-Winters model that utilizes the error between the estimated and exact value of the future points to update the smoothing parameters of the model; however, it may consume a good deal of computational resources.

13 Conclusion and future work

This chapter presents a user-centered analytics framework as a BI platform that can be used in a clinical laboratory setting to come to an informed conclusion on questions such as “What is the usage pattern of a clinical test?” and “Can we estimate the clinical test volumes for the next month?”. We selected the empirical route in order to understand and describe a BI framework for a clinical laboratory setting. The practice is based on conducting real-life case studies, identifying the commonalities between them, and extracts reusable framework elements that can be used as a BI framework to help clinical laboratory managers support their informed decisions.

This BI is composed of multiple services that can be used by clinical laboratory managers (e.g., CLS managers). The BI is based on the MapReduce framework and an open-source implementation of the Hadoop is used to process the massive clinical laboratory data sets to form the basic Big Data analytics tools. The user-centered analytics framework is designed with the clinical laboratory managers in mind, resulting in the BI platform with the necessary functionalities for different laboratory manager perspectives.

The two use-case studies present the value of BI analytics framework in managing clinical laboratory test volumes (i.e., usage pattern visualization and test volume estimation) utilizing the clinical databases available at the data warehouse. The analytics presented in the two use-cases present tools that can be reused by managers of different clinical laboratory settings to visualize performance and estimate resources (i.e., logistic manager, who estimates a specimen handling time; financial manager, who estimates employee rewards based on the facility-specific key performance indicators).

Big Data tools (Olston et al., 2008; 2014l; Chen et al., 2014; Hortonworks, 2014; 2014a; 2014h) present a development of Big Data analysis imposed by the advent of large-scale databases. The Hadoop platform and the MapReduce programming framework already have a substantial base in the bioinformatics community, especially in the field of next-generation sequencing analysis, and such use is increasing.

This study makes the following contributions to the literature:

• This is the first attempt to address a user-centered analytics framework based on Big Data analytics to process massive clinical data sets.

• The application of a novel visualization tool based on the z-score to visualize the clinical test utilization.

• The application of Holt-Winters method to estimate the provincial test volume of the province of Alberta, Canada. To the best of our knowledge, this is the first attempt to develop demand-estimation analytics for the clinical laboratory test volumes in Alberta.

Future research on the clinical Big Data analytics BI framework should accentuate the modeling of whole-business interacting subprocesses (e.g., clinical test utilization pattern, test procedures, specimen collection/handling, etc.). This can be assembled using inexpensive clusters of commodity hardware and the proper tool to construct a suitable processing framework for handling massive clinical data.