Natural Language Processing to Predict Deep Brain Stimulation Outcomes in Parkinson's Disease

The Unified Parkinson's Disease Rating Scale (UPDRS Part III) is often not obtained postoperatively in routine clinical practice, limiting objective assessment of treatment response following deep brain stimulation (DBS). We examine whether natural language processing (NLP) offers a potential approach to use narrative clinical documentation to predict motor outcomes and evaluate therapeutic responsiveness when formal postoperative scoring is unavailable.

Neurology notes and UPDRS Part III assessments were obtained from 25 patients with Parkinson's disease pre- and postoperatively. DBS response was defined as percent improvement in UPDRS Part III from the preoperative OFF-medication state to the postoperative DBS-ON/medication-ON assessment. We assessed whether UPDRS Part III scores could be predicted from clinic notes and whether improvement could be predicted from preoperative notes. Four NLP approaches were evaluated: TFIDF with linear models, BERT-base, ClinicalBERT, and Bio-ClinicalBERT embeddings. Models were assessed using leave-one-out cross-validation, with performance evaluated using mean absolute error, coefficient of determination (R2), Spearman correlation, and area under the ROC curve.

All approaches differentiated preoperative notes from postoperative notes. The TFIDF model demonstrated moderate accuracy at predicting UPDRS Part III total scores and percentage change. Several individual motor subscores demonstrated significant positive correlations between predicted and observed values, but only neck rigidity remained signifcant after BH-FDR correction (p=0.04).

NLP applied to routine DBS notes reliably distinguishes pre- from postoperative documentation and recovers a narrow set of consistently documented features (rigidity neck being a genuine positive signal). NLP-derived outputs may complement, but should not replace, formal UPDRS Part III assessment.