Integrating knowledge-driven and data-driven approaches for the derivation of clinical prediction rules

Clinical prediction rules are created by medical researchers and practitioners based on their knowledge and clinical experience. Such expert-generated rules are then evaluated and refined in clinical tests. Once verified, these knowledge-driven rules are used to expedite diagnosis and treatment for the serious cases and to limit unnecessary tests for low-probability cases. Alternatively, machine learning techniques can be used for automated induction of comprehensible data-driven rules from vast amount of existing clinical data. This paper investigates how the rules generated by the clinical experts compare with the data-driven rules. The paper describes three outcomes: rule confirmation, contradiction, and expansion. The study concentrates on prediction rules for the diagnosis of obstructive sleep apnea using three clinical data sets with 1,318 records. The prototype system, Hypnos, includes both a framework for rule definition, and also a mechanism for rule induction.