Estimating Optimal Inpatient Treatment for Type 2 Diabetes Open Access

Zhang, Yuchen (Spring 2021)

Permanent URL:


Background: Hyperglycemia contributes to a significant increase in morbidity, mortality, and healthcare costs in the hospital. The basal insulin regimen is recommended as the mainstay of diabetes therapy in the inpatient setting; however, it simultaneously amplifies the risk of hypoglycemia and other complications. While non-insulin agents could effectively improve glycemic control with a low risk of hypoglycemia, they may only fit for the patients diagnosed with mild and moderate hyperglycemia. It is not clear how determine the most appropriate treatment regime for Type 2 Diabetes patients with different characteristics to achieve optimal glycemic outcomes.

Methods: We explored the optimal treatment regime for targeted patients with Type 2 Diabetes by utilizing cutting-edge Dynamic Treatment Regime (DTR) methodology. We applied Q-Learning, inverse probability weighted estimator (IPWE), and augmented inverse probability weighted estimator (AIPWE), to determine the optimal treatment decision rules and estimate the expected outcomes. Model selection was conducted to decide the outcome regression models and propensity score models involved in these statistical procedures. The utility/value function for optimal treatment regime was defined either by the continuous outcome of mean blood glucose (BG) from day 2 to day 7 or the binary outcome of achieving BG target (i.e.,70-180 mg/dL) without hypoglycemia (<70 mg/dL).

Results: Using different DTR methods, we identified data driven treatment decision rules that utilized linear scores of admission BG and creatinine level to achieve optimal expected mean BG from day 2 to day 7, and treatment decision rules to achieve optimal chance of reaching BG target without hypoglycemia that utilized linear scores of admission BG and age. Based on the 10-fold cross-validation, the predicted mean BG by the optimal treatment regime derived from Q-learning, IPWE, and AIPWE are respectively 156.75 mg/dL, 155.79 mg/dL, and 161.63 mg/dL, which are all lower than the observed actual mean BG level, 163.1 mg/dL. 

Conclusions: Our treatment rules suggest treating Type 2 diabetes patients who are older, with higher admission BG, or higher creatinine concentration with basal insulin over oral agents. Our results are consistent with current clinical practice but provide more specific data-driven guidance.

Table of Contents

1. Introduction

1.1 Type 2 diabetes (T2D)

1.2 Therapy for Type 2 diabetes

1.2.1 Basal Insulin

1.2.2 Oral Antidiabetic Drugs (OADs)

2. Background

3. Methods

3.1 Data Sources

3.2 Statistical Methods

3.2.1 Q – Learning

3.2.2 Value Search Estimator Based on AIPWE

3.2.3 Value Search Estimator Based on IPWE

3.2.4 Methods for Model Selection

3.2.5 Methods for Cross-Validation Analysis

4. Results

4.1 Data Summary

4.2 Model selection

4.2.1. Continuous Outcome Regression Model

4.2.2 Binary Outcome Regression Model

4.2.3 Propensity Score Regression Model

4.3. Optimal Treatment Decision Rules

4.3.1. Continuous Outcome

4.3.2 Binary Outcome

4.4. Cross-Validation Analysis Results

4.4.1. Continuous Outcome

4.4.2 Binary Outcome

4.5. Comparison on Optimal Treatment and Actual Treatment

4.5.1. Continuous Outcome

4.5.2 Binary Outcome

5. Discussion


About this Master's Thesis

Rights statement
  • Permission granted by the author to include this thesis or dissertation in this repository. All rights reserved by the author. Please contact the author for information regarding the reproduction and use of this thesis or dissertation.
Subfield / Discipline
  • English
Research Field
Committee Chair / Thesis Advisor
Committee Members
Last modified

Primary PDF

Supplemental Files