Machine Learning Algorithms for Prediction of Ambulation and Wheelchair Transfer Ability in Spina Bifida.

Objective: To determine which statistical techniques enhance our ability to predict ambulation and transfer ability in people with spina bifida (SB).

Design: Retrospective cohort study.

Setting: Thirty-five US outpatient SB clinic sites.

Participants: Individuals (n=4589) with SB aged 5-73 years (median age=13.59y).

Intervention: Not applicable.

Main outcome measures: Ambulation ability, which consisted of the following categories: community ambulators, household ambulators, therapeutic ambulators, and nonambulators.

Secondary outcome: Wheelchair transfer ability, as defined by the ability to transfer in and out of a wheelchair unassisted.

Results: A recurrent neural network (RNN) using a multilayer perceptron discarded 76 cases during case processing, resulting in 4513 that were run through the RNN. The predictions in the resulting testing dataset were 83.22% accurate. Recall was 93.21% for community ambulators, 10.00% for household ambulators, 23.96% for therapeutic ambulators, and 76.70% for nonambulators. Precision was 85.34% for community ambulators, 16.05% for household ambulators, 16.67% for therapeutic ambulators, and 93.47% for nonambulators. Total predictions included 68.39% for community ambulators, 2.25% for household ambulators, 3.83% for therapeutic ambulators, and 25.53% for nonambulators. Correspondingly, the model accurately classified 70% of wheelchair transfers while correctly identifying 97.3% of those able to transfer unassisted.

Conclusions: RNN models hold promise for the prediction of functional outcomes such as ambulation and transfer ability in people with SB, particularly for community ambulators and nonambulators. Compared with the previous work using traditional logistic regression approaches which misclassified 16% of cases, the RNN resulted in greater prediction accuracy with fewer than 7% of cases misclassified.