Neural Network Predictions of Outcome from Posteroventral Pallidotomy

Abstract

Neural networks have recently been utilized to make predictions in multivariate systems where simple univariate correlations often miss key predictive relationships. Toward this end, we have implemented an application of neural network analysis to the problem of patient selection and outcome prediction for posteroventral pallidotomy (PVP) in patients with Parkinson’s Disease (PD). Forty six (46) cases were used to develop a variety of neural network structures that were able to learn and predict 6 month outcomes with a high degree of accuracy. Twelve variables were used as inputs to the nets. Networks could predict 93% of Unified Parkinson’s Disease Rating Scale (UPDRS) outcomes (choosing between less than or greater than 50% improvement). Networks were further validated by using only random sets of input data and testing their predictions on known cases. Additionally, we used real-data created nets and tested them on random-data test ‘cases’. Nets trained on random data were unable to learn at all (understandably). The networks created with real data, moreover, were no better than chance (50%) at predicting random outcomes. These results strongly suggest that the networks created were specific to the pallidotomy data.