Professional Abstract
Background
Researchers can predict a patient’s health trajectory and response to treatments, but to do so, they need improved statistical methods to leverage information in electronic health records. Better predictions of patient health outcomes can help clinicians improve clinical diagnoses and treatments. But gaps in existing statistical methods make it difficult to analyze health data to predict changes in individual patient health outcomes. This study developed Bayesian hierarchical models for analyzing diverse types of health data in electronic health records to answer research questions about patient health outcomes. Bayesian models incorporate data from previous trials or studies in the estimation of treatment effects.
Objective
To use a Bayesian analytic approach to develop and implement new methods and software for predicting individual patient health status, changes in health status over time, and response to treatment
Study Design
| Design Element |
Description |
| Design |
Theoretical development, empirical analysis, software development |
| Data Sources and Data Sets |
Data from 3 clinical studies, including (1) the PERCH Study on childhood pneumonia, (2) the Brady Urological Institute active surveillance study on prostate cancer, and (3) the Janssen schizophrenia trial and National Network of Depression Centers project |
| Analytic Approach |
Bayesian hierarchical modeling |
| Outcomes |
Model outputs, including (1) predictions of individual health status, (2) predictions of individual health trajectory, (3) estimates of treatment effects, and (4) measurement of heterogeneity of treatment
New software package called OSLER inHealth
|
Methods and Results
The research team developed a Bayesian hierarchical statistical model. The model had four components:
- Estimation of the effects of current treatment and patient characteristics such as age and clinical history on health status
- Use of multiple clinical measures to infer health status
- Analysis of the effect of health measurements at one time point on subsequent treatment decisions
- Specification of a model with two levels in which factors affecting the estimation of treatment effects could vary over time for the same patient and could also vary across many patients with similar characteristics
The research team then developed a software package called OSLER inHealth to help other researchers implement the Bayesian hierarchical model as well as existing methods to make individual predictions on health status and treatment effects.
Finally, the research team used three case studies to test and adapt the statistical model and software to predict a patient’s health status and changes in health over time. The team applied the methods to estimate
- Whether a child’s pneumonia was caused by a virus or bacteria
- The efficacy and safety of active surveillance compared with surgery for a patient with prostate cancer
- A patient’s level and trajectory of depression, anxiety, and mania symptoms in the past, which could be used to predict future symptoms
Patients, clinicians, and a health plan administrative leader helped build and refine the statistical models for each case study. The research team also worked with a steering committee of researchers and statisticians to develop the software.
Limitations
The research team implemented the methods in three case studies; the performance of the methods may differ when applied to other health conditions.
Conclusions and Relevance
Bayesian hierarchical models can combine diverse sources of prior knowledge and data about an individual patient and a reference population to predict the patient’s health status and the likely effects of different treatment options. Such predictions can help doctors and patients monitor disease progression and make better clinical decisions.
Future Research Needs
Future research could systematically evaluate the new methods in larger and more diverse patient populations. Studies could also evaluate whether using the new methods in treatment decisions can improve health outcomes.
