Date of Award

Spring 2021

Document Type


Degree Name

Doctor of Philosophy (PhD)



First Advisor

Pinker, Edieal


Wide implementation of electronic health record systems provides rich data for personalized medicine. One topic of great interest is to develop methods to assist physicians in prognosis for example mortality. While many studies have reported on various new prediction models and algorithms there is relatively little literature on if and how these new prediction methods translate into actual benefits. My dissertation consists of three theses that aims at filling this gap between prognostic predictions and clinical decisions in end-of-life care and intensive care settings. In the first thesis, we develop an approach to using temporal trends in physiologic data as an input into mortality prediction models. The approach uses penalized b-spline smoothing and functional PCA to summarize time series of patient data. we apply the methodology in two settings to demonstrate the value of using the "shapes" of health data time series as a predictor of patient prognosis. The first application a mortality predictor for advanced cancer patients that can help oncologists decide which patients should stop aggressive treatments and switch to palliative care such as that provided in hospice. The second one is a real-time near term mortality predictor for MICU patients that can work as an early alarm system to guide timely interventions. In the second thesis, we investigate the integration of a prediction algorithm with physician decision making, focusing on the advanced cancer patient setting. We design a retrospective study to compare prognoses made by doctors and those that would be recommended by the IMPAC algorithm developed in Chapter 1. We used the doctor's discharge decision as a proxy of what they predict the patient as dying in 90 days and show that doctor's predictions tend to very conservative. Although IMPAC on its own does not perform better than doctors in terms of precision and recall, we find that IMPAC and doctors identify significantly different group of positive cases. IMPAC and doctors are also good at identifying very different groups of patients in terms of survival time. We propose a new way to augment decisions of doctors with IMPAC. At the same recall, the augment method identifies 43\% more patients close to death than the doctors do. We also estimate potential hospitalizations and hospital length of stays avoided if the doctors use augmented procedure instead of acting on their own beliefs. In the third thesis, we look at the integration of a prediction algorithm with physician decision making, focusing on the ICU setting. We use a POMDP framework to evaluate how decision support systems based on ICU mortality predictions can help physicians allocate time to inspect the patients at highest risk of death. We assume physicians have limited time and seek to optimally allocate it to patients in order to minimize their mortality rate. Physicians can do Bayesian updates on observations of patient health state. A prediction algorithm can augment this process by sending alerts to physicians. We represent the algorithm by an arbitrary point on an ROC curve representing a particular alert threshold. We study two approaches to using the algorithm input: (1) Belief based policy (BBP) that integrates algorithm outputs using Bayesian updating; (2) Alarm triggered policy (ATP) where the physician responds only to the algorithm without updating, and compare them to benchmarks that do not rely on the algorithm at all. By running simulations, we explore how the accuracy of predictions can translate into lower mortality rates.