Decision tree analysis to better control treatment effects in spinal cord injury clinical research

Abstract

<jats:sec><jats:title>OBJECTIVE</jats:title><jats:p>The aim of this study was to use decision tree modeling to identify optimal stratification groups considering both the neurological impairment and spinal column injury and to investigate the change in motor score as an example of a practical application. Inherent heterogeneity in spinal cord injury (SCI) introduces variation in natural recovery, compromising the ability to identify true treatment effects in clinical research. Optimized stratification factors to create homogeneous groups of participants would improve accurate identification of true treatment effects.</jats:p></jats:sec><jats:sec><jats:title>METHODS</jats:title><jats:p>The analysis cohort consisted of patients with acute traumatic SCI registered in the Vancouver Rick Hansen Spinal Cord Injury Registry (RHSCIR) between 2004 and 2014. Severity of neurological injury (American Spinal Injury Association Impairment Scale [AIS grades A–D]), level of injury (cervical, thoracic), and total motor score (TMS) were assessed using the International Standards for Neurological Classification of Spinal Cord Injury examination; morphological injury to the spinal column assessed using the AOSpine classification (AOSC types A–C, C most severe) and age were also included. Decision trees were used to determine the most homogeneous groupings of participants based on TMS at admission and discharge from in-hospital care.</jats:p></jats:sec><jats:sec><jats:title>RESULTS</jats:title><jats:p>The analysis cohort included 806 participants; 79.3% were male, and the mean age was 46.7 ± 19.9 years. Distribution of severity of neurological injury at admission was AIS grade A in 40.0% of patients, grade B in 11.3%, grade C in 18.9%, and grade D in 29.9%. The level of injury was cervical in 68.7% of patients and thoracolumbar in 31.3%. An AOSC type A injury was found in 33.1% of patients, type B in 25.6%, and type C in 37.8%. Decision tree analysis identified 6 optimal stratification groups for assessing TMS: 1) AOSC type A or B, cervical injury, and age ≤ 32 years; 2) AOSC type A or B, cervical injury, and age &gt; 32–53 years; 3) AOSC type A or B, cervical injury, and age &gt; 53 years; 4) AOSC type A or B and thoracic injury; 5) AOSC type C and cervical injury; and 6) AOSC type C and thoracic injury.</jats:p></jats:sec><jats:sec><jats:title>CONCLUSIONS</jats:title><jats:p>Appropriate stratification factors are fundamental to accurately identify treatment effects. Inclusion of AOSC type improves stratification, and use of the 6 stratification groups could minimize confounding effects of variable neurological recovery so that effective treatments can be identified.</jats:p></jats:sec>

Development of a machine learning algorithm for predicting in-hospital and 1-year mortality after traumatic spinal cord injury

“…We also graphically examined the association between the SCIRS and ISS (see Supplementary Figure S5). Results showed a positive correlation between the SCIRS and ISS scores (Pearson's correlation = 0.30, p<0.05) which supported the validity of the SCIRS given that ISS is a well-known index that has previously been used in tSCI research [11,[35][36][37]. This figure also demonstrates that the SCIRS is more sensitive than ISS in measuring mortality in the datasets.…”

Characterizing Natural Recovery after Traumatic Spinal Cord Injury

“…57,93 There is stronger literature for age >50 being associated with poorer outcomes. 61,92,[95][96][97][98][99][100][101][102] Oleson and associates 92 (in patients with AIS B grade) and Penrod and colleagues 97 (in patients with motor incomplete injuries) found that, overall, patients age >50 years had a decreased likelihood of recovering walking ability one year post-injury, relative to patients under age 50. Burns and associates 98 reported an increased likelihood of achieving community ambulation in patients with incomplete tetraplegia and paraplegia on discharge from rehabilitation in individuals with AIS C who were <50 (91%) compared with those ‡50 years of age (42%), although for individuals presenting with an initial AIS D, age was not a factor in ambulation because all were able to ambulate 3-6 months post-injury.…”

The Role of Machine Learning in Spine Surgery: The Future Is Now

“…Finally, the branch containing AOSpine A and B injuries in the cervical spine were passed through another internal node for "age, " generating another three leaves or clusters. The final six clusters are detailed in Table 4 (32). The results of this study provide a platform for external validation studies with other patient cohorts to compare this unique classification system with current ones.…”

“…Tee et al application of decision tree learning for optimizing patient risk stratification after spinal cord injury provides a framework for understanding this modality (32). They combined different methods of assessing spinal cord function after trauma, including the American Spinal Injury Association (ASIA) Impairment Scale, total motor score (TMS) and the AOSpine classification system, to allow a decision tree to identify patient clusters that respond differently to treatment.…”

“…Six unique terminal nodes or clusters were identified. Reproduced with permission by Tee et al (32). the spine literature will be inundated with publications running multiple statistical and ML models in parallel for comparative analysis.…”