The use of machine learning algorithms for predicting the functional recovery following traumatic spinal cord injury
1Facchinello Y, 2Beauséjour M, 1Richard-Denis A, 1Thompson C, 1Mac-Thiong J
1Research Center, Hôpital du Sacré-Coeur de Montréal, Montréal, QC, Canada; 2Department of Surgery, Faculty of Medicine, University of Montreal, Montreal, QC, Canada
Traumatic spinal cord injury (TSCI) occurs at an annual incidence of 10 to 60 cases per million of inhabitants depending on the country . The functional outcome is typically considered as the most useful primary outcome for patients with TSCI as patients are mostly concerned with their ability to engage in activities of daily living . An accurate prediction of the functional recovery is important to set realistic goals and to plan an optimized rehabilitation. Existing predictive model are mainly in the form of linear or logistic regression based on clinical and demographical data collected during the acute care hospitalisation. These models are generally complex and not used clinically as they do not offer satisfactory performances. Recently, machine learning algorithms have been used in various medical fields to build predictive models . Among machine learning methods, classification and regression trees (CART) are promising as CART have been found to be consistently as good as, or better than linear and logistic regression models, particularly for datasets with high skew and kurtosis. The objective of this work is to evaluate the potential of CART for predicting the functional outcome following TSCI, using predictors collected during the acute care hospitalization. It also evaluates the influence of varying the number of variables considered for predicting the functional outcome.
This prospective study was performed on 172 patients treated in a single Level 1 trauma center specialized in TSCI. Patients were included if they sustained a TSCI between C1 and L2 levels, had complete data and were followed for a minimum of 6 months after the trauma. Functional outcome was quantified using the Spinal Cord Independence Measure (SCIM) collected within the first-year post injury. Age, delay prior to surgery and Injury Severity Score (the burden of associated traumatic injuries) were considered as continuous predictors while energy of injury, trauma mechanism, neurological level of injury, AIS grade, development of early spasticity, urinary tract infection, pressure ulcer and pneumonia were coded as categorical inputs. A simplified model built using AIS grade, NLI, energy and age as predictors was compared to a more complex model considering all predictors mentioned above. Both models were subjected to cross-validation.
The models built using 4 and 11 predictors were found to explain 51.4% and 62.3% of the variance of the SCIM total score after cross-validation, respectively. The severity of the neurological deficit (AIS grade) at admission was found to be the most important predictor. Other significant predictors were the Injury Severity Score, age, neurological level of injury and delay prior to surgery.
Compared to multivariate regressions, CART offer an easily interpretable and clear mapping of the links between the predictor and the outcome variable combined with good predictive performances. The simplified model is of particular interest in a clinical setting, as fewer predictors would lead to an easier data collection and a better use of resources.
This work was supported by the U.S. Army Medical Research and Material Command and by the Rick Hansen Institute.
You Must Be Logged In
This Reminders tab will display for your scheduled lectures 15 minutes before start time and for new messages waiting.