Extracting Patient Case Profiles with Domain-Specific Semantic Categories

Introduction

The medical diagnosis and treatment of a patient is a complex procedure which requires relevant knowledge and clinical experience of illnesses. In order to distinguish different illnesses that show similar signs and symptoms, or to decide the best available option for treatment of a medical condition, physicians need to have adequate observations of similar patient cases, either from their own previous medical practices, or some external resources, such as the knowledge of more experienced colleagues, and the medical literature on the latest progress in the field.

Clinical case reporting therefore plays an important role in both educating young physicians in the practice of medicine, and sharing clinical experience and exceptional findings among physicians (Jenicek, M., 2001). There are two types of clinical case reports, namely routine patient case reports and clinical case studies. Routine case reports are raw patient records of the diagnostic and treatment procedure and provide information which is necessary for the continuity of patient care, such as progress notes, discharge summaries and pathology reports. Clinical case studies, on the other hand, report rare and abnormal patient cases that are considered as of significant scientific value to the field. They are reported by clinicians, usually in the form of formal journal articles in medical press.

Narrative patient records produced by nurses and physicians everyday in hospitals and clinics, provide first-hand the richest information about the progress of patients. However, the confidentiality of personal records has always been a concern which prevents the research community having access to enough data to develop useful learning systems comparable to human performance. Confidential information includes names of patients and physicians, dates, and geographic clues which are required to be anonymized before any raw patient data can be released to the public. Moreover, the anonymisation task often requires human annotators to manually check every single patient record to satisfy certain laws and ethical guidelines specified by governments. For instance, the 2007 Computational Medicine Challenge had to use human annotators to review all of the 4,055 raw patient records and to remove nearly half of them from the final gold-standard corpus to meet United States HIPAA standards.

With the emergence of publicly available on-line knowledge bases such as MEDLINE/PubMed and BMC Central, clinicians now have access to a large number of full-text journal articles of clinical case studies. Each case study records a detailed discussion of the patient’s abnormal signs and symptoms, or novel conditions which are considered as report-worthy. While all the sensitive privacy information has been carefully removed from the text, a clinical case study still contains rich information about patient case profiles, such as patient demographics, signs and symptoms, laboratory test readings and interpretations, and treatments and subsequent outcomes for patients. This patient profile information is key in answering two fundamental questions dominating the daily practice of physicians: (1) Given the case profile of a patient, what is the best explanation or diagnosis of the condition? (2) Given the specified circumstances of a patient, what is the best treatment available? By exploring clinical case studies with similar patient profiles, physicians can learn, and therefore improve their practices of medicine, from the successes or failures of their peers.