Physician Vanderbilt University Medical Center, United States
Introduction: Surgical decompression for degenerative cervical myelopathy (DCM) is often indicated, yet it remains difficult for surgeons to predict who will have improvement after surgical treatment for DCM. The purpose of this study was to build a machine learning (ML) model that uses natural language processing (NLP) to predict clinically significant improvement in functional outcome after surgical decompression for myelopathy using the physician’s written narrative.
Methods: A single-institution retrospective cohort study of adults undergoing elective cervical decompression surgery for DCM. Age and text from preoperative notes and imaging reports were collected, along with baseline and 12-month modified Japanese Orthopedic Association (mJOA) score. Text was treated by isolating single words or word pairs as individual variables. 96 different models were trained using 80% of the data and ranked according to area under the receiver operating curve (AUC) to identify the top performing model. The final model was validated using the remaining 20% of the data.
Results: There were 565 patients, mean age 58.0, mean BMI 30.1, 55.4% male. The following surgeries were performed: 286 (50.6%) anterior cervical discectomy and fusion, 217 (38.4%) posterior decompression and fusion, 52 (9.2%) corpectomy and fusion, and 16 (2.8%) laminoplasty. The final model predicted ≥3-point improvement in mJOA with an AUC 0.76 on cross-validation and 0.74 on holdout validation. Baseline mJOA followed by imaging text, then patient age, and finally H&P text were the most important variables to the model. The words “broad-based,” “flavum,” and “prominent” were associated with < 3-point improvement in mJOA score, while the words “mild,” “edema,” and “foraminal stenosis” were associated with ≥3-point improvement in mJOA.
Conclusion : ML and NLP are underutilized in spine surgery and may be useful in identifying patients with DCM who would benefit most from surgical intervention. Here, we construct a NLP model that predicts minimum clinically significant improvement in mJOA score.
