Machine Learning and Simulation/Optimization Approaches to Improve Surgical Services in Healthcare

Introduction

Data science and analytics studies are undertaken to achieve long-term goals that often go well beyond the choice and calibration of the prediction models. In healthcare analytics, the overarching objective is to decipher the patterns leading to problems and inefficiencies by analyzing the data. The future success and improvement of healthcare services rely on a well-organized recursive cycle of data analytic steps initiated with discovering the problem and ending through self-correction and system-level problem resolution. Benefits of this framework lie in dynamic bottleneck discovery and generation of robust and timely problem resolutions. These steps are defined in Figure 1.

Figure 1.

Cyclic data analytics

With recent increasing demand for healthcare services, many hospitals have been facing the constant challenge of meeting increasing demand with limited staff, equipment and capacity resources while maintaining and improving the quality of care. The operating room theatre, also known as the engine of the hospital, is the focus of many healthcare improvement projects due to its impact on the overall profitability. An effective way of achieving improved utilization and quality of care of surgical services is through the better use of available data for prediction and through the proactive management of limited resources in primary (e.g., Operating Rooms, Sterile Processing Services) and auxiliary services (e.g., ICU/PACU). The ultimate goals of better predictions and proactive coordination are increased utilization and quality of care, mitigated delays, and improved workflow through increased predictability.

Healthcare information systems gather huge amounts of textual and numeric information about patients’ schedule details, event timestamps, principal procedure description, symptoms, physician notes and many more. Data mining tools provide a unique opportunity to extract necessary information from both textual data and numerical values. For instance, the information derived from text mining of health records can be useful when predicting procedure codes and surgery durations. Surgery descriptions and related notes often provide additional information that might not be available through procedural codes or descriptions. Further, all this information (especially event timestamp) provides a data rich environment amenable to building advanced simulation models may also elevate our understanding of human behavior in the true-to-life environment in which interaction of related factors operate. Consequently, near-real analysis made possible through simulating methodologies can be used to predict the unplanned situations and make more effective decisions in a timely manner.

In this chapter, the authors describe four related studies utilizing text mining and machine learning approaches to solve the prediction problems related to procedure types and surgery durations as well as simulation-based system level operational prediction and its use case through RME inventory optimization. First study develops text mining methodology to improve the surgery schedule’s performance through accurate classification of surgeries. Second study builds on the first study to predict the surgery descriptive procedural codes, i.e., CPT codes, by integrating text, categorical, and continuous features. The predicted CPT codes are then mapped to predict the surgery durations. Next, again building on the first two studies, authors describe proactive management of resources using near real-time information and the development and pilot implementation of the modules of a comprehensive simulation-based toolkit. This will constitute a System of Systems (SoS) design to provide plug-n-play module additions to build a virtual, tactical, facility-specific and configurable decision support system for use by mid-level managers. This toolkit will enable them to assess current state, build options to plan a future state, and implement a solution to overcome barriers. Last study aims at operationalizing the predictive capabilities of the machine learning and simulation models of the first three studies by solving the inventory optimization problem for reusable medical equipment (RME).