Past decade has witnessed an increasing shift from data governance to data analytics, with emphasis on knowledge discovery from a wide variety of sources that can assist in healthcare decision making. Three broad themes that emerged under data analytics umbrella include descriptive, predictive and prescriptive analytics. This article presents an overview of the essential ingredients of predictive analytics for the healthcare leadership and enthusiasts. Predictive analytics has continued to gain increasing visibility across a spectrum of areas. Business Wire reports the 2019 Global Healthcare Predictive Analytics Market to reach $7.8 billion by 2025 with a market growth of 21.17% CAGR during the forecast period. The 2019 Predictive Analytics in Healthcare Trend Forecast Survey by the Society of Actuaries also reported a marked increase in the adoption of Predictive Analytics by healthcare executives from 2018 (47%) to 2019 (60%). While some of the challenges are acknowledged, the report clearly emphasizes significant investment (~15%) in predictive analytics by payers with an expected return over the next five years.
Figure 1. An example of predicting high-risk and low-risk clinical outcomes in a given cohort from their baseline characteristics
What is Predictive Analytics?
The primary objective of predictive analytics is to predict future events or outcomes from their baseline characteristics using appropriate models. For instance, predicting high-risk and low-risk clinical outcomes from baseline characteristics using a predictive model is shown in Fig.1.
Why Predictive Analytics?
Predictive Analytics is Evidence-Based: Data-driven aspect of predictive analytics renders it context-specific and truly evidence-based. More importantly, predictive analytics has the ability to validate established benchmarks as well as reveal patterns that are unique to the data and the patient population. Therefore, it enables discovery and hypothesis generation in addition to hypothesis testing.
Predictive Analytics can assist in Prevention: Predictive analytics has the potential to provide preliminary cues into subsets of baseline characteristics that are actionable and can significantly impact outcomes, Fig. 2. Such an understanding is especially critical in developing effective preventive strategies and interventions so as to alter the trajectories ahead of time from a predicted outcome to a desired outcome, Fig. 2.
Figure 2. Altering the predicted outcome to a desired outcome using effective interventions on actionable baseline characteristics
Predicted outcomes may include clinical, operational, and financial outcomes, as well as key performance indicators (KPIs). Impact of predictive analytics on the three outcomes and KPIs is illustrated in Fig. 3. Of interest is to note that these outcomes may not necessarily be independent and their potential associations are shown by vertical bi-directional arrows in Fig. 3. Consider the case of identifying the high-risk and low-risk groups from their baseline characteristics, Fig. 1. Identifying high-risk subjects ahead of time can assist in developing targeted treatment regimens and disease management strategies that can favorably impact the clinical outcome. Improved clinical outcome in turn can minimize the overall economic burden on the patient, provider and payer, hence has a direct impact on the financial outcome. Improved clinical outcome can also facilitate optimal care delivery by minimizing resource utilization (e.g. medical equipment, beds, staffing), hence impact the operational outcome. These outcomes singularly or in combination can in turn impact KPIs (e.g. Mortality rate, Occupancy rate, Bed turnover, Avg. Length of Stay, Treatment Costs) in the care continuum.
Figure 3. Impact of Predictive Analytics on Outcomes and KPIs.
Predictive Analytics Using Machine Learning
Machine learning approaches such as supervised learning (e.g. classification) have been used widely for predictive analytics. It is important to note that there are subtle differences between machine learning approaches and traditional statistical models. In general, statistical models are hypothesis driven and often report p-values to establish the significance of the association between outcomes and baseline characteristics. They may also impose a particular relationship (e.g. linear) between the baseline characteristics and outcomes of interest. In contrast, machine learning approaches such as classification are relatively flexible in that they learn the decision boundaries separating the outcomes of interest from the given baseline characteristics. They also report performance measures (e.g. accuracy) and focus on establishing the generalization ability of the classifier on previously unseen data. Essential ingredients of a typical predictive analytics classification framework is presented in Table 1.
Five key stages in a typical predictive analytics classification framework include training, validation, testing, deployment and demonstrating value, Fig.4. Critical evaluation and refinement accompanies all stages and the difficulty increases from left to right culminating in demonstrating value, Fig. 4.
Figure 4. Five Key Stages Accompanying Predictive Analytics Implementations
Predictive Analytics and Artificial Intelligence
Operational definition of artificial intelligence often relies on the Turing test. Informally, Turing test is useful in assessing whether the behavior of machines is indistinguishable from that of humans. Within the context of healthcare, recent adoption of speech recognition and advances in natural language processing (NLP) in interpreting unstructured clinical text in conjunction with widespread adoption of machine learning approaches such as deep learning as a part of the predictive analytics implementation builds the nexus to artificial intelligence.
Predictive Analytics – a Team Sport
Successful predictive analytics implementations demand multidisciplinary teams working in concert. A brief description of common roles in a typical predictive analytics team is enclosed below.
Concluding Thoughts
Predictive analytics has the potential to assist in developing preventive strategies with a significant impact clinical, financial, operational outcomes and KPIs. In contrast to consensus-based and hypothesized benchmarks, predictive analytics machine learning frameworks are data-driven and harness the intricate patterns in the given data, hence evidence-based. Data-driven aspect of predictive analytics does not subscribe to “one-size fits all” norm and is especially appealing, since a number of characteristics (e.g. patient population, source systems, data quality) can vary across healthcare environments. For these reasons, failure of predictive analytics solutions to translate across distinct healthcare environments is not necessarily a failure of these approaches. While machine learning and artificial intelligence approaches in predictive analytics frameworks rely on several decades of research and experimentation, a majority of them are optimization approaches. Therefore, critical evaluation and refinement of the models by domain experts in a team setting should accompany all stages of predictive analytics implementation. Data science expertise is at the core of predictive analytics implementation, since a good handle on the assumptions, convergence and limitations of the models as well as hardware demands precedes successful deployments. From an ethical standpoint, it has been emphasized that incorporation of evidence from predictive analytics may need to be discretionary and not necessarily binding. On a related note, GDPR (General Data Protection Regulation) by the European Commission also provides regulations on the ethical use of healthcare data and importance of transparency of machine learning and artificial intelligence models. These regulations have also been argued to minimize the administrative burden in accessing healthcare data for analytics. Predictive analytics is a team sport and successful and seamless implementations in workflows demand concerted efforts across multiple entities including healthcare information technology. Finally, a nurturing culture with realistic expectations and timelines for demonstrating value is critical for successful implementations and sustaining predictive analytics teams as critical members of healthcare organizations.
