Medical Knowledge Representation and Predictive Data Entry

Abstract

Most existing knowledge representation systems have been created to support one of two general tasks — expert systems or natural language processing. This paper reports on a third fundamentally different task for knowledge representations — predictive data entry. In predictive data entry the task of the knowledge representation is to generate all and only statements which are ‘medically sensible’ in a given situation, where being ‘sensible’ implies that the statements are a) grammatically correct, b) semantically sound, and c) self-consistent. This work arises out of the PEN&PAD project which is developing patient care workstations for direct use by clinicians. The knowledge representation - Structured Meta Knowledge (SMK) — provides a unified view of medical records and medical terminology which is used to support the PEN&PAD intelligent user interface. SMK is a semantic network language with subsumption and multiple inheritance which allows primitive concepts such as ‘fracture’ and ‘humerus’ and ‘spiral’ to be combined into complex descriptions such as ‘spiral fracture of the humerus’ and placed unambiguously within the network. The representation has been used as part of a successful prototype workstation which has been well received in evaluations by practicing doctors. The use of the knowledge representation in predictive data entry raises a number of issues which previous systems have avoided and promises to be less application dependent than other criteria for judging knowledge representations.