Knowledge discovery in perioperative databases using bootstrapped rule induction.

摘要

Analysis of medical databases can provide knowledge which may improve patient care. The goal of this research was to develop a Knowledge Discovery in Databases (KDD) approach for discovering relationships between perioperative variables, intraoperative events, and postoperative outcomes. A method for bootstrapping brute force rule induction was developed. The method reduces the number of rules an analyst must interpret, provides summary rules which indicate the rule's stability and variability, and uses multidimensional scaling to visualize rule similarity. A brute force rule induction algorithm is bootstrapped ten times. The variability in continuous attribute values, rule accuracy, and rule coverage is recorded for rules which persist across multiple replications. The method was applied to samples of increasing size to see the impact of sample size upon the stability of the rules generated.; The choices involved in conducting knowledge discovery in perioperative medicine are described. Rules were evaluated for clinical relevance. Experiments were conducted to determine if the inclusion of intraoperative and physiological data elements improved rule performance. Knowledge discovery was applied to the following postoperative outcomes: severe pain, the absence of pain, nausea and vomiting, and extended recovery time. Rules were also induced for the following intraoperative physiological incidents: hypertension, hypotension, low pulse pressure, low oxygen saturation, and high heart rate variability.; Where clinical guidelines are not established, practice patterns may reflect informal or unstated consensus among clinicians. Anesthetist practice patterns were induced from perioperative using bootstrapped rule induction to determine the drugs, anesthetic techniques, and postoperative orders associated with specific patient populations. Patterns were discovered for two populations: orthopedic patients undergoing arthroscopic surgery and obese women over fifty years old. The composition of practice patterns and the relationship between rule accuracy and coverage were explored. Coverage for twenty arthroscopic practice patterns exceeded ninety percent, suggesting informal consensus for managing arthroscopic procedures.