Essays on predictive and prescriptive analytics for risk monitoring and population screening in healthcare management

• Main Author: HE, Yan
• Format: text
• Language: English
• Published: Institutional Knowledge at Singapore Management University 2021
• Subjects: Business Administration, Management, and Operations; Health Services Administration; Operations and Supply Chain Management
• Online Access: https://ink.library.smu.edu.sg/etd_coll/359; https://ink.library.smu.edu.sg/cgi/viewcontent.cgi?article=1357&context=etd_coll
• Institution: Singapore Management University

Due to increased aging populations and changes in lifestyles, we have witnessed an increased prevalence of various chronic and acute diseases and a drastic rise in healthcare expenditures in recent years. It is of paramount importance for public health to promote regular screening and close monitoring to detect the early onset of diseases. On the other hand, the increasing availability of healthcare data and advancement in data analytics offer a huge potential to facilitate this goal. We can analyze the vast amount of data and recommend more personalized diagnostic tests after receiving results and signals from screening tests and monitoring systems, which are critical decisions for the effective and efficient implementation of such screening programs and monitoring systems. Meanwhile, it is also necessary to consider human behavioral issues and their impact in making the recommendations. In particular, individual adherence to the recommended diagnostic tests can significantly affect the effectiveness and efficiency of the programs. This dissertation aims to integrate predictive analytics, optimization techniques, and behavioral models to improve risk monitoring and decision-making in patient monitoring systems and population screening programs. This dissertation first studies the real-time risk monitoring problem for patients in intensive care units (ICUs). We identify a critical lag in the provision of information due to the long lead time to measure some laboratory test variables (e.g., creatinine, platelets, and bilirubin) used in calculating the Sequential Organ Failure Assessment (SOFA) score, a well-established and important risk measure for patients in ICUs. We develop machine learning models to estimate such variables using easily measured bedside variables, the rate of changes in bedside variables, and time lag from the previous laboratory test, which mimics how physicians assess patient conditions in practice. Then the predicted laboratory test variables can be used to calculate an estimate of the real-time SOFA score. We further take advantage of the estimated standard deviations from these models to construct intervals of the real-time SOFA scores. We hypothesize that the estimated score intervals could capture the uncertainty in patient condition since the previous test and provide valuable information in a new dimension that complements the nominal SOFA scores. Using a dataset collected from an ICU in a tertiary hospital in Singapore, we calibrate our model and validate the hypothesis by comparing the prognostic accuracy of the proposed approach on patients’ 24-hour mortality and 30-day readmission with those from the SOFA score calculated using the conventional approaches. The proposed methodology could be applied to other risk measures to improve their prognostic accuracy and provide more reliable early warning for timely intervention. The methodologies developed in the previous chapter can help raise a warning of potential deterioration in a patient’s health condition, but the exact problem still has to be confirmed through follow-up diagnostic tests, which are typically more invasive and expensive. Medical resource overuse has become increasingly common in recent years and caused diverse problems, including unnecessary and risky diagnostic tests and overly intensive or expensive treatments. There is a growing call for more evidence-based decisions to reduce unnecessary diagnostic tests. The next part of the thesis dives into this problem to optimize the prescription of diagnostic tests during the health monitoring process, leveraging the improved risk monitoring tools developed in the previous chapter. In particular, we develop a finite-horizon, partially observable Markov decision process model to optimize the time to initiate a diagnostic test. Our model captures both measured and estimated clinical variables (including estimated intervals) in real-time to update the belief on a patient’s underlying health condition. We apply the model to monitor patients’ blood glucose levels to detect hyperglycemia, a common complication of critical illness. We calibrate the model using the same ICU dataset as in the previous chapter and demonstrate that the new approach can advance the detection time with fewer diagnostic tests. The methodology can also be applied to many other health monitoring systems, especially those powered by smart wearable health devices for chronic diseases. However, to optimally design the warning signals and recommend the diagnostic tests for such a monitoring system, one must consider the impact of human behavioral issues, especially individuals’ perception of the warning signals and adherence to the recommendations. We address this challenge in the next chapter in the optimal design of population screening programs for cancer surveillance and screening. Cancer remains one of the leading causes of human death, while early detection enables timely intervention and reduction in mortality rate. Two-stage screening programs are broadly implemented in practice among large average-risk populations to effectively and efficiently detect cancer in the early stages. Individuals receiving positive results in first-stage (initial) tests are recommended to undergo second-stage tests for further diagnosis. Notably, individuals’ adherence to the second-stage tests, which is closely associated with the initial test design (sensitivity and specificity) and personal characteristics, varies considerably across individuals and leads to different cancer detection rates and demands for second-stage tests. We adopt a Bayesian persuasion framework to model the optimal initial test design problem in the context of colorectal cancer screening. Our goal is to balance the trade-off between test effectiveness (i.e., detection rates of cancer incidences) and test efficiency (i.e., demands for second-stage tests), considering individuals’ adherence behavior. We conduct a nationwide survey in Singapore to calibrate the individual’s response to changes in the test design. With the embedded behavioral model, we next optimize the threshold selection in the initial test design (which decides the test sensitivity and specificity). We characterized the structural properties of an optimal initial test design. Using various data and information collected locally in Singapore and from the literature, we demonstrate that a well-designed initial test can detect more cancer incidences with fewer second-stage tests than the current practice. We further explore the benefits of using heterogeneous initial tests for different sub-populations and use the interpretable clustering technique to search for implementable rules to partition the population. We find that customized tests with simply an age-gender partition rule could bring significant extra benefits. To conclude, this thesis studies the optimal design of real-time patient monitoring systems and population screening programs, using a combination of techniques from machine learning, optimization, game theory and survey design. By analyzing the comprehensive datasets collected from various sources, we showcase that well-designed monitoring systems and screening programs can benefit individuals, healthcare service providers, and health systems through improved effectiveness and efficiency in healthcare service delivery.