Assessing Neural Network Prediction in Kidney Disease Data

Neural networks can be used as a potential way to predict continuous and binary outcomes. With their ability to model complex non-linear relationships between variables and outcomes, they may be better at prognosis than more traditional regression methods such as logistic regression. In this thesis, the prognostic abilities of neural networks will be assessed using data from the Consortium for Radiological Imaging Studies of Polycystic Kidney Disease (CRISP) using clinically significant variables such as BMI, Blood Urea Nitrogen (BUN), Height Adjusted Total Kidney Volume (htTKV), baseline estimated glomeruler filtration rate (eGFR), and type of PKD. Both a logisitic regression and variations of neural networks were modeled. The neural networks had hidden units from 2 to 10, and weight decays from 0.01 to 0.05. Each of these models was assessed by looking at Receiver Operator Characteristic (ROC) curves, specifically the area under the curve (AUC). The complexity of these models was also looked at by calculating the degrees of freedom for each model. More complex models could lead to an overfitting of the data, and were therefore examined in this study. The study showed that neural networks have the capability to predict the outcome of stage 3 kidney disease better than a more traditional logistic regression, however, the models become increasingly complex as the predictive ability increases. These findings could have a great impact on public health in the future. They could greatly impact the methods that are used for prognosis. The use of neural networks might lead to better prognosis and earlier treatment for kidney disease based on an individuals baseline measurements of the aforementioned variables, or any other new biomarkers that are discovered in the future