Machine learning predicts improvement of functional outcomes in spinal cord injury patients after inpatient rehabilitation.
Mohammad Rasoolinejad, Irene Say, Peter B Wu and 5 others
PMID 40927746WHAT IT FOUND
In 589 adults with spinal cord injury, machine-learning models predicted discharge independence better than using admission scores alone, but they were tested only within the same dataset from a single rehab centre.
Do not use them to guide patient care yet.
Key findings
01Tree-based models predicted discharge FIM scores best on test data, with R-squared 0.49 to 0.52 and MSE 0.82 to 1.42; using admission scores as discharge scores gave R-squared -0.57 and MSE 4.93.
02The strongest predictor for each discharge FIM item was the patient's admission score for that same item; length of stay and age were also consistently important.
03Cervical injury, wheelchair dependence at admission, and prehospital living in intermediate care or a rehab centre were linked to worse predicted discharge function, while lumbar injury and walking at admission were linked to better predicted discharge function.
STILL TO COME
How it was doneWhat they foundWhat it means for PTsWhat it means for OTs
Read the rest of this summary
You get three full summaries a month, free, and we do not ask for a card. Search, the TL;DRs and your library stay unlimited either way.
What it does not show
The study used records from a single acute rehabilitation centre from 2010 to 2015, so it may not apply to other settings, patients, or current Section GG documentation. Models were validated only by cross-validation within the same dataset, not on an external prospective cohort, so real-world accuracy is unknown. Only patients with complete admission and discharge FIM scores were included, and pediatric, pregnant, and patients who died during rehabilitation were excluded, so results do not cover those groups. Missing numerical data were filled with the mean, and missing categorical values were kept as separate categories. The best models still had only modest test performance, and discharge walking or wheelchair status and social interaction were predicted poorly. Feature associations are not causal; they do not show which therapy component caused improvement. The data used FIM before the transition to Section GG, and the authors did not directly validate the models on Section GG data.
Declared interests
The study was supported by the Jill Kort and Family Foundation Fund at UCLA. The authors declare financial support was received, and the content is solely the authors' responsibility.
The easy way to misread this
Do not use these models to predict an individual patient's outcome yet. They were tested only within the same retrospective dataset from a single centre, and the best test R-squared was 0.52, so predictions were imperfect and not externally validated.