科研之友
微信
新浪微博
Facebook
分享链接
摘要
Objectives: Design and implement an intelligent free-text summarization system: The system's input includes large numbers of longitudinal, multivariate, numeric and symbolic clinical raw data, collected over varying periods of time, and in different complex contexts, and a suitable medical knowledge base. The system then automatically generates a textual summary of the data. We aim to prove the feasibility of implementing such a system, and to demonstrate its potential benefits for clinicians and for enhancement of quality of care. Methods: We have designed a new, domain-independent, knowledge-based system, the CliniText system, for automated summarization in free text of longitudinal medical records of any duration, in any context. The system is composed of six components: (1) A temporal abstraction module generates all possible abstractions from the patient's raw data using a temporal-abstraction knowledge base; (2) The abductive reasoning module infers abstractions or events from the data, which were not explicitly included in the database; (3) The pruning module filters out raw or abstract data based on predefined heuristics; (4) The document structuring module organizes the remaining raw or abstract data, according to the desired format; (5) The microplanning module, groups the raw or abstract data and creates referring expressions; (6) The surface realization module, generates the text, and applies the grammar rules of the chosen language. We have performed an initial technical evaluation of the system in the cardiac intensive-care and diabetes domains. We also summarize the results of a more detailed evaluation study that we have performed in the intensive-care domain that assessed the completeness, correctness, and overall quality of the system's generated text, and its potential benefits to clinical decision making. We assessed these measures for 31 letters originally composed by clinicians, and for the same letters when generated by the CliniText system. Results: We have successfully implemented all of the components of the CliniText system in software. We have also been able to create a comprehensive temporal-abstraction knowledge base to support its functionality, mostly in the intensive-care domain. The initial technical evaluation of the system in the cardiac intensive-care and diabetes domains has shown great promise, proving the feasibility of constructing and operating such systems. The detailed results of the evaluation in the intensive-care domain are out of scope of the current paper, and we refer the reader to a more detailed source. In all of the letters composed by clinicians, there were at least two important items per letter missed that were included by the CliniText system. The clinicians' letters got a significantly better grade in three out of four measured quality parameters, as judged by an expert; however, the variance in the quality was much higher in the clinicians' letters. In addition, three clinicians answered questions based on the discharge letter 40% faster, and answered four out of the five questions equally well or significantly better, when using the CliniText-generated letters, than when using the clinician-composed letters. Conclusions: Constructing a working system for automated summarization in free text of large numbers of varying periods of multivariate longitudinal clinical data is feasible. So is the construction of a large knowledge base, designed to support such a system, in a complex clinical domain, such as the intensive-care domain. The integration of the quality and functionality results suggests that the optimal discharge letter should exploit both human and machine, possibly by creating a machine-generated draft that will be polished by a human clinician.
- 出版日期2016-6
