Results Summary
What was the research about?
Electronic health records store a lot of data about a patient. These data often include race, age, health problems, current medicines, and lab results. Looking at these data may help doctors treating patients after a trauma predict how likely it is that they will respond well to a treatment and survive. This information can help doctors make better treatment decisions. But first, researchers need to figure out how to combine and analyze data to make accurate predictions.
In this study, the research team created new statistical methods to combine data from patient records. They used these methods to predict patient health outcomes. Then the team used health record data collected from patients in hospital trauma centers to test their predictions.
What were the results?
The research team made three new computer programs. The first program combined data from patient records to automatically predict patient outcomes. Using test data, the program predicted how likely trauma patients were to stop bleeding and to survive.
The second program looked at whether the new method could be used to show what information was likely to be important for predicting outcomes. For example, in a test case, head injury and blood clotting problems were linked with lower chances of survival among trauma patients.
The third program combined data from patients’ records to automatically suggest a treatment choice. In a test case, the team looked at what would happen if patients received the treatments that the program chose. Using the treatments recommended by the program didn’t lead to better patient outcomes.
What did the research team do?
The research team used math equations to combine data from trauma patients. The data included gender, age, weight, blood pressure, medicines, trauma severity, trauma site, blood clotting, and lab readings. Then the team used the equations to create the computer programs.
Next, the team tested the programs using data created to mimic real data. Then they tested the programs with data from medical records of patients with severe trauma.
What were the limits of the study?
The data used to create the computer programs weren’t always complete. Missing information may have made the predictions less accurate. Also, the data didn’t include information taken over time, such as hourly blood pressure readings. If the programs had included these data, they might have predicted different results.
Future research could look at adding other types of data to the programs. Studies could also test the computer programs with new patients who have had severe trauma or with patients with other health problems.
How can people use the results?
Researchers could continue to test and improve the computer programs. The programs may eventually help doctors make treatment decisions based on patients’ traits and other health problems.
Professional Abstract
Objective
To develop methods, using machine-learning platforms, for incorporating clinical data on many clinical variables to accurately and efficiently predict the effects of interventions on trauma patients
Study Design
Design Elements | Description |
---|---|
Design | Theoretical development, simulation studies, empirical analyses |
Data Sources and Data Sets | Simulated data, 4 data sets from acute trauma studies for empirical analysis |
Analytic Approach |
|
Outcomes | Estimation bias and variability, confidence interval coverage probabilities associated with each method |
This study developed and evaluated methods to improve automated assessments of prognosis and predictions of treatment effects for acute trauma care patients to support individualized recommendations for patient care.
Researchers used a machine-learning framework to apply multiple statistical methods for predicting the effects of interventions. The researchers used
- SuperLearner, an analytic approach that generates optimal predictions of clinical outcomes and treatment effects using a library of prediction algorithms
- Causal inference models to identify optimal predictor variables
- Targeted maximum likelihood estimation quantifying variable importance related to clinical outcomes and treatment effects
Researchers developed and tested machine-learning methods using both simulated data and data from four patient-outcome studies in trauma units from major hospitals across the United States. These studies included
- Activation of Coagulation and Inflammation in Trauma (ACIT) study
- Prospective, Observational, Multicenter, Major Trauma Transfusion (PROMMTT) study
- Transforming Research and Clinical Knowledge in Traumatic Brain Injury (TRACK-TBI) study
- Pragmatic Randomized Optimal Platelet and Plasma Ratios (PROPPR) trial
The data sets included information on vital signs, lab results, demographics, and trauma severity. To test applications of the methods, researchers constructed both baseline and dynamic prediction algorithms for patient outcomes, including hemostasis, or blood clotting, and death.
Results
Following refinement of the methods, researchers created three software programs to implement them:
- Origami combines SuperLearner with computationally efficient cross-validation to assess prognosis.
- Varimpact processes messy data and, using causal inference methods, provides an estimate of the relative importance of specific clinical variables to the outcome.
- Opttx learns optimal treatment rules and estimates the population mean of the outcome under these treatment rules. The program also evaluates the importance of each clinical variable by estimating change in the performance of the treatment rule if a clinical variable is removed.
In test applications of the methods, researchers found that
- SuperLearner performed well in predicting hemostasis (area under the curve [AUC] ≈ 0.7-0.8) and death (AUC ≈ 0.9) using the PROPPR trial data.
- Varimpact was able to identify the relative importance of 50 potential clinical variables for predicting death using only unprocessed data from the ACIT study.
- Opttx did not identify any treatment rules using the PROPPR trial data that increased benefits to patients. However, this approach may be useful as a screening tool to assess whether using more individualized treatment rules performs better than using a standard treatment strategy for all patients.
Limitations
The four data sets did not include clinical time-series data, such as streaming vital signs, which are often necessary for obtaining the most accurate outcome predictions. In addition, researchers needed to use statistical methods to handle missing values from these data sets, which may have affected the development and performance of the algorithms. Further, the study did not test the proposed algorithms on new patients to validate performance in the context of actual clinical practice.
Conclusions and Relevance
This study elucidated the theoretical and applied properties of new methods for estimating the importance of variables relevant to trauma patient outcomes, including the conditions under which these methods lead to more, or less, accurate findings. After further validation, the three software programs may enable researchers and clinicians to identify more individualized treatment approaches in high-intensity, fast-paced medical settings.
Future Research Needs
Future research could incorporate automatically recorded time-series data into the predictive methods to improve accuracy. Additional studies could test the prediction algorithms on new patients and across different clinical settings and medical conditions.
Final Research Report
View this project's final research report.
Journal Citations
Related Journal Citations
Peer-Review Summary
Peer review of PCORI-funded research helps make sure the report presents complete, balanced, and useful information about the research. It also assesses how the project addressed PCORI’s Methodology Standards. During peer review, experts read a draft report of the research and provide comments about the report. These experts may include a scientist focused on the research topic, a specialist in research methods, a patient or caregiver, and a healthcare professional. These reviewers cannot have conflicts of interest with the study.
The peer reviewers point out where the draft report may need revision. For example, they may suggest ways to improve descriptions of the conduct of the study or to clarify the connection between results and conclusions. Sometimes, awardees revise their draft reports twice or more to address all of the reviewers’ comments.
Peer reviewers commented, and the researchers made changes or provided responses. Those comments and responses included the following:
- Reviewers asked for additional information about when simulation studies were necessary in developing new methods for this study. The researchers explained that they found simulations were unnecessary for some of the algorithms because other methods like cross validation had been used or previous studies had used simulations, so the outcomes were well known. In addition, the researchers indicated that they did not have the time and resources to perform additional simulations given the extensive work on algorithm development.
- Reviewers asked the researchers to revise the report for this methods development project to make it easier to understand for a general-scientist audience. The researchers moved the more- technical sections into the appendix.