How do different machine learning methods compare to one another when used to predict inpatient rehabilitation outcomes for patients with TBI?

The Study:  Machine learning is a type of artificial intelligence (AI) used to look for trends in complex datasets.  The purpose of the study was to determine which machine learning can help predict patient outcomes better than traditional statistical analyses. We used the same dataset collected from 2008-2011 for the TBI Practice Based Evidence study, which included 1946 participants who were 14 years or older, and diagnosed with a TBI of sufficient severity to require admission to one of nine inpatient rehabilitation facilities in the US. The data collected included variables about participants, their injury and treatment they received during rehabilitation, and six outcome variables. The outcomes were: inpatient rehabilitation length of stay; discharge to home; discharge cognitive function; discharge motor function; cognitive function at 9 months post discharge; and motor function at 9 months post discharge.  We divided the dataset in two. Both datasets contained the same patient and injury variables; however, the two datasets were different for the types of therapies used in the analyses. One dataset used all collected therapy from occupational, physical, and speech-language therapies (OT, PT, ST), therapeutic recreation, psychology, and social work/case management. The second dataset used only the therapy collected from the primary inpatient therapies (OT, PT, ST). We compared different types of machine learning to see which predicted outcomes best.

Results showed one machine learning method (Gradient Boosting Tree Model; GBM), performed the most consistently at predicting outcomes across both datasets. We found specific treatment factors associated with the outcomes, including level of patient effort, how quickly rehabilitation started following injury, how old a participant was when they started rehabilitation, and specific types of physical activities performed during inpatient rehabilitation (e.g., advanced gait training, walking out in the community, and stairs). We also found the top ranked predictor was different for each outcome. For example, the number one predictor for discharge motor function was level of motor function at admission, but the number one predictor for motor function 9 months later was the number of days from injury to start of inpatient rehabilitation. The ability to pair outcomes with specific predictors is one of the advantages in using machine learning methods over traditional analyses.

Who may be affected by these findings? Providers of inpatient rehabilitation and researchers studying how to improve outcomes from inpatient TBI rehabilitation.

Caveats: The dataset used in this study is over ten years old, which means some of the collected data may be outdated and no longer relevant since treatment changes over time as new research becomes available. Also, the dataset may have missed some relevant patient, injury and treatment features that could impact outcomes.

Bottom Line: By figuring out which patient, injury, and treatment factors can predict better outcomes, we can help make inpatient rehabilitation more efficient and help clinical practice. Using advanced computer methods like machine learning may help researchers identify these factors more precisely.