Mining time-dependent patient outcomes from hospital patient records

We describe REMIND, a data mining framework that accurately infers missing clinical information by reasoning over the entire patient record. Hospitals collect computerized patient records (CPR's) in structured (database tables) and unstructured (free text) formats. Structured clinical data in the CPR's is often poorly recorded, and information may be missing about key outcomes and processes. For instance, for a population of 344 colon cancer patients, important clinical outcomes, such as disease state and its evolution, are stored only as unstructured data (doctors' dictations) in the CPR. Raw evidence (extracted directly from the CPR) is not a good predictor of disease state. Yet by combining this evidence in a principled fashion (using methods from uncertain and temporal reasoning), REMIND accurately infers disease state sequences for recurrence, a complex time-varying outcome, for these patients. These outcomes can now be added back into the CPR in structured form.