This paper presents a methodology to formulate natural language rules for an adaptive neuro-fuzzy system based on discovered knowledge, supported by prior knowledge and statistical modeling. Relationships between disease related variables and fluctuations in Parkinson’s disease is often complex. Experts have simplified but mostly reliable “fuzzy” rules based on experience. These rules could be improved using statistical methods and neural nets. This gives clinicians a valuable tool to explore the importance of different variables and their relations in a disease and could aid treatment selection. A prototype using the proposed methodology has been used to induce an Adaptive Neuro Fuzzy Inference Model that has been used to “discover” relationships between fluctuation, treatment and disease severity. More data is needed to confirm these findings. The project shows that artificial intelligence techniques and methods in combination with statistical methods offer medical research and applications valuable opportunities.