BACKGROUND: Deep brain stimulation (DBS) for Parkinson disease provides significant improvement of motor symptoms but can also produce neurocognitive side effects. A decline in verbal fluency (VF) is among the most frequently reported side effects. Preoperative factors that could predict VF decline have yet to be identified. OBJECTIVE: To develop predictive models of DBS postoperative VF decline using a machine learning approach. METHODS: We used a prospective database of patients who underwent neuropsychological and VF assessment before both subthalamic nucleus (n = 47, bilateral = 44) and globus pallidus interna (n = 43, bilateral = 39) DBS. We used a neurobehavioral rating profile as features for modeling postoperative VF. We constructed separate models for action, semantic, and letter VF. We used a leave-one-out scheme to test the accuracy of the predictive models using median absolute error and correlation with actual postoperative scores. RESULTS: The predictive models were able to predict the 3 types of VF with high accuracy ranging from a median absolute error of 0.92 to 1.36. Across all three models, higher preoperative fluency, digit span, education, and Mini-Mental State Examination were predictive of higher postoperative fluency scores. By contrast, higher frontal system deficits, age, Questionnaire for Impulsive-Compulsive Disorders in Parkinson's disease scored by the patient, disease duration, and Behavioral Inhibition/Behavioral Activation Scale scores were predictive of lower postoperative fluency scores. CONCLUSION: Postoperative VF can be accurately predicted using preoperative neurobehavioral rating scores above and beyond preoperative VF score and relies on performance over different aspects of executive function.
BACKGROUND: Deep brain stimulation (DBS) for Parkinson disease provides significant improvement of motor symptoms but can also produce neurocognitive side effects. A decline in verbal fluency (VF) is among the most frequently reported side effects. Preoperative factors that could predict VF decline have yet to be identified. OBJECTIVE: To develop predictive models of DBS postoperative VF decline using a machine learning approach. METHODS: We used a prospective database of patients who underwent neuropsychological and VF assessment before both subthalamic nucleus (n = 47, bilateral = 44) and globus pallidus interna (n = 43, bilateral = 39) DBS. We used a neurobehavioral rating profile as features for modeling postoperative VF. We constructed separate models for action, semantic, and letter VF. We used a leave-one-out scheme to test the accuracy of the predictive models using median absolute error and correlation with actual postoperative scores. RESULTS: The predictive models were able to predict the 3 types of VF with high accuracy ranging from a median absolute error of 0.92 to 1.36. Across all three models, higher preoperative fluency, digit span, education, and Mini-Mental State Examination were predictive of higher postoperative fluency scores. By contrast, higher frontal system deficits, age, Questionnaire for Impulsive-Compulsive Disorders in Parkinson's disease scored by the patient, disease duration, and Behavioral Inhibition/Behavioral Activation Scale scores were predictive of lower postoperative fluency scores. CONCLUSION: Postoperative VF can be accurately predicted using preoperative neurobehavioral rating scores above and beyond preoperative VF score and relies on performance over different aspects of executive function.
Authors: Daniel Weintraub; Staci Hoops; Judy A Shea; Kelly E Lyons; Rajesh Pahwa; Erika D Driver-Dunckley; Charles H Adler; Marc N Potenza; Janis Miyasaki; Andrew D Siderowf; John E Duda; Howard I Hurtig; Amy Colcher; Stacy S Horn; Matthew B Stern; Valerie Voon Journal: Mov Disord Date: 2009-07-30 Impact factor: 10.338