Predicting the outcome of intramuscular psoas lengthening in children with cerebral palsy using preoperative gait data and the random forest algorithm

This study used the random forest algorithm to predict outcomes of intramuscular psoas lengthening as part of a single event multi-level surgery in patients with cerebral palsy. Data related to preoperative medical history, physical exam, and instrumented three-dimensional gait analysis were extracted from a historic database in a motion analysis center. Data from 800 limbs of patients with diplegic cerebral palsy were analyzed. An index quantifying the overall deviation in pelvic tilt and hip flexion was used to define outcome categories. The random forest algorithm was used to derive criteria that predicted the outcome of a limb. The criteria were applied to limbs that underwent psoas lengthening with outstanding results (accuracy. =. .78, sensitivity. =. .82, specificity. =. .73). The criteria were then validated using an extended retrospective case-control design. Case limbs met the criteria and underwent psoas lengthening. Control limbs met the criteria, but did not undergo psoas lengthening. Over-treated limbs failed the criteria and underwent psoas lengthening. Other-treated limbs failed the criteria and did not undergo psoas lengthening. The rate of good outcomes among Cases exceeded that observed among controls (82% vs. 60%, relative risk. =. 1.37), and far exceeded that observed in Over-treated limbs (27%). Other-treated limbs had good outcomes 52% of the time. Application of the criteria in the future is estimated to increase the overall rate of good pelvis-hip outcomes from 58% to 72% among children with diplegia who undergo single-event multi-level surgery (SEMLS).