[Background] To alleviate the huge burden on the medical staff of inputting electronic health records per patient visit, automatic speech recognition has drawn much attention. Although its recent implementation has known difficulties in accurate recognition and medicaldomain adaptation, phrases of the core clinical concepts would still be safely recognized because of their repetitive mentions in patient communications. [Objective] We propose an Entity-to-Text approach that automatically generates the "chief complaints" section of nursing records from a list of clinical named entities. [Material & Methods] Crowdsourced 589 sentences of pseudo "chief complaints" were annotated with patient-status entities. From these entities, a text generation model was trained to replicate the original "chief complaints". [Results & Discussion] The evaluation results showed the basic feasibility of the proposed approach. It also suggested that complete sentences could be generated even from incomplete entities.
With the progress of digitization in medical care, various medical texts, such as electronic medical record texts, case report texts, discharge summaries, and radiology reading reports, are being accumulated electronically. It is expected that medical knowledge can be extracted from these electronic medical texts and utilized for secondary purposes such as diagnostic support. In fact, in Japan, there is a case (J-CaseMap) where a case report text from a regional meeting of the Japanese Society of Internal Medicine was used as a diagnostic support system to search for diseases and pathological conditions that can be used as a reference for differential diagnosis. On the other hand, it is said that half of the medical knowledge becomes outdated in about five years, and new knowledge emerges and changes rapidly every day. Therefore, it is necessary to update the knowledge database constructed from electronic medical texts as needed. However, the construction and updating of a structured database requires a lot of expert manpower, making efficiency and automation a major issue. In fact, in J-CaseMap, the construction of diseases, pathological conditions, symptom findings, and causal chains among them are all done manually. The process of automatically constructing causal chain knowledge from case report texts involves three major steps: (1) extracting important terms and phrases (diseases, pathological conditions, and symptom findings that are important for differential diagnosis) from case report texts, (2) normalizing the extracted expressions to standard terminological expressions, and (3) estimating causal relationships among the normalized terminological expressions. In this study, we focused on the 2nd Step, i.e., normalization of important expressions extracted from case report sentences into standardized terminological expressions, utilizing an Encoder-Decoder model. The study also compared each model and analyzed the quality of the output.
In this study, we attempted to estimate primary diseases from symptoms and findings using open domain question answering model that considers symptoms and findings as questions and primary diseases inferred from symptoms and findings as answers, and we validated the effectiveness of the model for medical care. The inference results from our model resulted in a maximum Exact Match of 49.53 for the test data. In addition, we also investigated the applicability of the model to answer the National Examination for doctors.
As a representative method of pre-learning in machine learning, it is known that using a prelearning model using ImageNet is useful for improving performance classification. In this research, we used data recorded by different shooting methods for different model and verified how the performance classification of machine learning was improved by incorporating the data into prelearning. Data from TMS, a corneal tomography measurement device and CASIA, an anterior segment OCT measurement device, both of which are marketed by Tomey, were used. These two Device are capable of photographing a map of the refractive index of the cornea, although their imaging methods are different. We found that pre-training using data of different shooting methods taken by different devices improved the classification performance more than machine learning models without pre-training. Even if the imaging method is different, if the data obtained by photographing the same parts(Cornea) and obtaining similar output is used for pre-learning, it can be said that the classification performance of the machine learning model will be improved.