Extracting clinical terms from radiology reports with deep learning.

Extracting clinical terms from free-text format radiology reports is a first important step toward their secondary use. However, there is no general consensus on the kind of terms to be extracted. In this paper, we propose an information model comprising three types of clinical entities: observations, clinical findings, and modifiers. Furthermore, to determine its applicability for in-house radiology reports, we extracted clinical terms with state-of-the-art deep learning models and compared the results. We trained and evaluated models using 540 in-house chest computed tomography (CT) reports annotated by multiple medical experts. Two deep learning models were compared, and the effect of pre-training was explored. To investigate the generalizability of the model, we evaluated the use of other institutional chest CT reports. The micro F1-score of our best performance model using in-house and external datasets were 95.36% and 94.62%, respectively. Our results indicated that entities defined in our information model were suitable for extracting clinical terms from radiology reports, and the model was sufficiently generalizable to be used with dataset from other institutions.

End-to-End Approach for Structuring Radiology Reports.

Radiology reports include various types of clinical information that are used for patient care. Reports are also expected to have secondary uses (e.g., clinical research and the development of decision support systems). For secondary use, it is necessary to extract information from the report and organize it in a structured format. Our goal is to build an application to transform radiology reports written in a free-text form into a structured format. To this end, we propose an end-to-end method that consists of three elements. First, we built a neural network model to extract clinical information from the reports. We experimented on a dataset of chest X-ray reports. Second, we transformed the extracted information into a structured format. Finally, we built a tool that enabled the transformation of terms in reports to standard forms. Through our end-to-end method, we could obtain a structured radiology dataset that was easy to access for secondary use.

Extracting Symptom Names and Disease-Symptom Relationships from Web Texts Using a Multi-Column Convolutional Neural Network.

We propose a method to create large-scale Japanese medical dictionaries that include symptom names and information about the relationship between a disease and its symptoms using a large web archive that includes large amounts of text written by non-medical experts. Our goal is to develop a diagnosis support system that makes a diagnosis according to the natural language (NL) inputs provided by patients. To achieve this, two medical dictionaries need to be constructed: one that includes a wide variety of symptom names expressed in NL and another that includes information about the relationship between a disease and its symptoms. Dictionaries will then be used to predict the patient's disease via two developed methods that extract symptom names and disease-symptom relationships. Both methods retrieve sentences using WISDOM X and then apply neural classifiers to them. Our experimental results show that our methods achieved 93.8% and 88.3% in the F1-score, respectively.

Case Report Form Reporter: A Key Component for the Integration of Electronic Medical Records and the Electronic Data Capture System.

To improve the efficiency of clinical research, we developed a system to integrate electronic medical records (EMRs) and the electronic data capture system (EDC). EDC is divided into case report form (CRF) reporter and CDMS with CRF receiver with data communication using the operational data model (ODM). The CRF reporter is incorporated into the EMR to share data witth the EMR. In the data transcription type, doctors enter data using a progress note template, which are transmitted to the reporter template. It then generates the ODM. In the direct record type, reporter templates open from the progress note and generate narrative text to make record in the progress note. The configuration files for a study are delivered from the contents server to minimize the setup. This system has been used for 15 clinical studies including 3 clinical trials. This system can save labor and financial costs in clinical research.

A Strategy for Reusing the Data of Electronic Medical Record Systems for Clinical Research.

There is a great need to reuse data stored in electronic medical records (EMR) databases for clinical research. We previously reported the development of a system in which progress notes and case report forms (CRFs) were simultaneously recorded using a template in the EMR in order to exclude redundant data entry. To make the data collection process more efficient, we are developing a system in which the data originally stored in the EMR database can be populated within a frame in a template. We developed interface plugin modules that retrieve data from the databases of other EMR applications. A universal keyword written in a template master is converted to a local code using a data conversion table, then the objective data is retrieved from the corresponding database. The template element data, which are entered by a template, are stored in the template element database. To retrieve the data entered by other templates, the objective data is designated by the template element code with the template code, or by the concept code if it is written for the element. When the application systems in the EMR generate documents, they also generate a PDF file and a corresponding document profile XML, which includes important data, and send them to the document archive server and the data sharing saver, respectively. In the data sharing server, the data are represented by an item with an item code with a document class code and its value. By linking a concept code to an item identifier, an objective data can be retrieved by designating a concept code. We employed a flexible strategy in which a unique identifier for a hospital is initially attached to all of the data that the hospital generates. The identifier is secondarily linked with concept codes. The data that are not linked with a concept code can also be retrieved using the unique identifier of the hospital. This strategy makes it possible to reuse any of a hospital's data.

Estimating the Ratio of Patients with a Certain Disease Between Hospitals for the Allocation of Patients to Clinical Trials Using Health Insurance Claims Data in Japan.

In clinical trials, investigating the ratio of patients with each disease who are treated in a hospital is important for determining the number of patients who are allocated to hospitals. The Japanese health insurance claims data includes standardized disease and medicine data. However, the disease data has some problems in terms of reliability, because the healed diseases are sometimes not deleted or because a disease that a patient does not actually have is registered to claim the cost of the examination. On the other hand, therapeutic medicines are administered to target particular diseases. In this study, we developed a system for estimating the number of patients with each disease using the disease data and the therapeutic medicine data. We converted the ICD-10 code to a 4-grade classification code so that we could predict the diseases in the shallow layer (e.g. gastrointestinal disease) when it was difficult to predict the precise diseases in the deep layer (e.g. gastric ulcers). A table showing the disease code and the corresponding therapeutic medicine code was provided by the Japan Pharmaceutical Information Center (JAPIC). We calculated the disease probability score from the diseases and therapeutic medicines and recorded the predicted disease. For the system evaluation, we used the health insurance claims data from Osaka University Hospital for January 2015. A total of 58,526 diseases were predicted from the health insurance claims data of 18,393 patients. One hundred twenty patients were randomly extracted for use in a chart review that was performed by an expert physician. Two hundred twenty-four of 329 predicted diseases, were correctly predicted; 56 were reasonably predicted, and 49 were incorrectly predicted. The main disease was correctly predicted in 71 patients. In conclusion, we could estimate the number of patients with each disease using the health insurance claims data with a certain degree of accuracy.

A template-based computerized instruction entry system helps the comunication between doctors and nurses.

In a hospital, doctors and nurses shares roles in treating admitted patients. Communication between them is necessary and communication errors become the problem in medical safety. In Japan, verbal instruction is prohibited and doctors write their instruction on paper instruction slips. However, because it is difficult to ascertain revision history and the active instructions on instruction slips, human errors can occur. We developed template-based computerized instruction entry system to reduce ward workloads and contribute to medical safety. Templates enable us to input the instructions easily and standardize the descriptions of instructions. By standardizing and combine the instruction into one template for one instruction item, the systems could prevent instructions overlap. We created sets of templates (e.g., admission set, preoperative set), so that doctors could enter their instructions easily. Instructions entered via any of the sets can be subdivided into separate items by the system before being submitted, and can also be changed on a per-item basis. The instructions were displayed as calendar form. Calendar form represents the instruction shift and current active instructions. We prepared 382 standardized instruction templates. In our system, 66% of instructions were entered via templates, and 34% were entered as free-text comments. Our system prevents communication errors between medical staff.

Interconnection of electronic medical record with clinical data management system by CDISC ODM.

EDC system has been used in the field of clinical research. The current EDC system does not connect with electronic medical record system (EMR), thus a medical staff has to transcribe the data in EMR to EDC system manually. This redundant process causes not only inefficiency but also human error. We developed an EDC system cooperating with EMR, in which the data required for a clinical research form (CRF) is transcribed automatically from EMR to electronic CRF (eCRF) and is sent via network. We call this system as "eCRF reporter". The interface module of eCRF reporter can retrieves the data in EMR database including patient biography data, laboratory test data, prescription data and data entered by template in progress notes. The eCRF reporter also enables users to enter data directly to eCRF. The eCRF reporter generates CDISC ODM file and PDF which is a translated form of Clinical data in ODM. After storing eCRF in EMR, it is transferred via VPN to a clinical data management system (CDMS) which can receive the eCRF files and parse ODM. We started some clinical research by using this system. This system is expected to promote clinical research efficiency and strictness.

Development of a system for transferring images via a network: supporting a regional liaison.

We developed a system that transfers images via network and started using them in our hospital's PACS (Picture Archiving and Communication Systems) in 2006. We are pleased to report that the system has been re-developed and has been running so that there will be a regional liaison in the future. It has become possible to automatically transfer images simply by selecting the destination hospital that is registered in advance at the relay server. The gateway of this system can send images to a multi-center, relay management server, which receives the images and resends them. This system has the potential to be useful for image exchange, and to serve as a regional medical liaison.

Usage experience with the document archiving and communication system for the storage and retrieval of medical records.

Standard Japanese electronic medical record (EMR) systems are associated with major shortcomings. For example, they do not assure lifelong readability of records because each document requires its own viewing software program, a system that is difficult to maintain over long periods of time. It can also be difficult for users to comprehend a patient's clinical history because different classes of documents can only be accessed from their own window. To address these problems, we developed a document-based electronic medical record that aggregates all documents for a patient in a PDF or DocuWorks format. We call this system the Document Archiving and Communication System (DACS). There are two types of viewers in the DACS: the Matrix View, which provides a time line of a patient's history, and the Tree View, which stores the documents in hierarchical document classes. We placed 2,734 document classes into 11 categories. A total of 22,3972 documents were entered per month. The frequency of use of the DACS viewer was 268,644 instances per month. The DACS viewer was used to assess a patient's clinical history.

Proposal of diagnostic process model for computer based diagnosis.

We aim at making a diagnosis support system that can be put to practical use. We proposed a diagnostic process model based on simple knowledge which can be gleaned from textbooks. We defined clinical finding (CF) as a general concept for patient's symptom or findings etc., whose value is expressed by Boolean. We call the combination of several CFs a "CF pattern", and a set of CF patterns with concomitant diseases "case base". We consider diagnosis as a process of searching an instance from the case base whose CF pattern is concomitant with that of a patient. The diseases which have the same CF pattern are candidates for diagnosis. Then we select a CF which is present in part of the candidates and check whether it is present or absent in the patient in order to narrow down the candidates. Because the case base does not exist in reality, the probability of CF pattern is calculated by the product of CF occurrence rate assuming that occurrence of CF is independent. Therefore the knowledge required for diagnosis is frequency of disease under sex and age group and CF-disease relation (CF and its occurrence rate in the disease). By processing these two types of knowledge, diagnosis can be made.