Frailty Predictive Models for Length of Stay, Complications, and Discharge Status in Patients with Traumatic Spinal Cord Injury Using Supervised Machine Learning Algorithms: A TRACK-SCI Study

Introduction: Frailty-related features may be associated with spinal cord injury (SCI) outcomes. We identified individual predictors of outcomes and compared the performance of our models to established frailty indices (Ganapathi frailty index (GFI), modified frailty index-5 (mFI-5), modified frailty index-11 (mFI-11)).

Methods: The prospective Transforming-Research-and-Clinical-Knowledge-in-SCI (TRACK-SCI) registry was queried for clinical data. Random forest (RF), a supervised machine learning algorithm, was used to identify feature importance for outcome variables (ICU-length of stay (LOS), hospital-LOS, unfavorable-discharge, and in-hospital-complication). Extended ICU-LOS and hospital-LOS defined as >75 percentile of our cohort. Unfavorable-discharge defined as discharge to skilled-nursing-facility, hospice, or deceased in-hospital. mFI-5 and mFI-11 includes variables related to past medical history. GFI includes age, BMI, psoas volume, albumin levels, hemoglobin levels. We developed our own models by incorporating top-five most predictive frailty-related features based on RF for each outcome into a 5-fold cross-validated logistic regression. These models were compared to regression models including variables from mFI-5, mFI-11, and GFI.

Results: 170 patients with SCI were included (mean age=57.1 years (SD=18.5), 26.5% females). RF showed admission albumin levels was consistently within top-five most predictive frailty-related feature for all four outcomes, while age and psoas volume were in top-five for three of the four outcomes. Our model for ICU-LOS (includes albumin, age, hemoglobin, h/o CVA w/o deficit, psoas volume) did not perform better than the established models. Our model for hospital-LOS (albumin, psoas volume, Hounsfield Units, BMI, h/o hypertension) performed better than miF-5 (AUC: 0.75 vs 0.56, P=0.002) and miF-11 models (AUC: 0.75 vs 0.62, P=0.047), but not the GFI model. Our model for unfavorable-discharge (albumin, age, h/o CVA w/ deficit, h/o impaired sensorium, hemoglobin) only performed better than miF-5 model (AUC: 0.70 vs 0.57, P=0.040). Our model for in-hospital-complication (albumin, psoas volume, Hounsfield Units, h/o of CHF, and age) only performed better than miF-5 model (AUC: 0.71 vs 0.58, P=0.018).

Conclusion : Certain frailty-related features (age, albumin levels, psoas volume) are highly predictive of our examined outcomes. Our models, which included physical measures and lab values (Ex: albumin, psoas volume), generally performed better than the mFI-5, suggesting these measurements should be more readily utilized as prognostic markers of frailty.