| Abstract |
Predictive modeling can be of great value in the domain of health. It can, among other things, drive more pro-active and personalized treatments. In the past, a lot of work has been done to use structured or coded data to create such predictive models with conventional machine learning techniques. There is however a wealth of information available in unstructured data as well. Think of doctor's notes, or writing of the patient. During my stay at MIT last Summer I worked together with the Clinical Decision Making group who have ample experience in extracting useful information from unstructured medical data. We focused on how the unstructured data can complement the structured data to increase predictive performance. I will present the results of two explorations, one is to see how predictive performance for colorectal cancer can be improved by using the notes of the GP, and the second which aims at predicting therapeutic outcome for treatments of anxiety disorders based on email exchanges between patient and therapist. |