Search

There are various of skin pigmented lesions with high risk. Melanoma is one of the most dangerous forms of skin cancer. It is one of the important research directions of medical artificial intelligence to carry out classification research of skin pigmented lesions based on deep learning. It can assist doctors to make clinical diagnosis and make patients receive treatment as soon as possible to improve survival rate. Aiming at the similar and imbalanced dermoscopic image data of pigmented lesions, this paper proposes a deep residual network improved by Squeeze-and-Excitation module, and dynamic update class-weight, in batches, with model ensemble adjustment strategies to change the attention of imbalanced data. The results show that the above method can increase the average precision by 9.1%, the average recall by 15.3%, and the average F1-score by 12.2%, compared with the multi-class classification using the deep residual network. Thus, the above method is a better classification model and weight adjustment strategy.

Depression is a common mental disorder characterized by high incidence, high risk for disability and suicide, which brings great harm to individuals, families and society.Moreover, it is relatively difficult to diagnose, treat, monitor and manage for its subtle onset and long course as well as the sufferers' poor help-seeking behaviors.The recently emerged artificial intelligence technologies have become a new way for the management of depression and depression-related suicide.Based on the challenging situations, achievements that have been made in the diagnosis and management of depression, and monitoring and management of depression-related suicide using artificial intelligence technologies were discussed, which could be the reference for further studies about depression.

Background Suicide is a global public health and mental health problem. With the rapid development of internet technology, more and more people tend to express their suicidal tendencies and suicidal intentions online. The difference of emotional characteristics between high and low risk of suicide messages should be analyzed to help identify suicide risk and provide early intervention. Methods The "tree hole"intelligent robot captures message data, then randomly selects the same number of high and low suicide risk messages manually, and the high frequency keywords of high and low suicide risk messages are obtained by word segmentation and a TF-IDF algorithm. The keywords are analyzed by Gephi software, and the emotion dictionary provided by Boson is used to judge the emotional tendency of high and low suicide risk users. Results The emotional score of high suicide risk messages was -3.511 ∼ 2.514, averaging (-0.225±0.405), while the total score of low suicide risk messages was -4.547 ∼ 3.403, averaging (-0.121±0.628). Low suicide risk messages mainly focused on negative emotions, interpersonal relationships and social support, while high suicide risk messages mainly centered on invited suicide, means, locations and time of suicide. Conclusion There are differences in emotional characteristics between high and low suicide risk messages. The higher the suicide risk, the more obvious the negative tendency of the users' emotions. More attention is needed to the greater potential for suicide among this group of users and psychological support and interventions should be included.

The relationship of biomedical entity is the cornerstone of acquiring biomedical knowledge. It is of great significance to the construction of related databases in the biomedical field and the management of medical literature. How to quickly and accurately extract the required relationships of biomedical entity from massive unstructured literature is an important research. In order to improve accuracy, we use support vector machine (SVM) which is a machine learning algorithm based on feature vectors to extract relationships of entities. We extract the five main relationships in medical literature, including ISA, PART_OF, CAUSES, TREATS and DIAGNOSES. First of all, related topics are used to search medical literature from PubMed database, such as disease-drug, cause-disease. These documents are used as experimental data and then processed to form a corpus. In selection of features, the method of information gain is used to select the influential entities’ own features and entities’ context features. On this basis, semantic predicates are added as a feature to improve accuracy. The experimental results show that the accuracy of extraction is increased by 5%–10%. In the end, Resource Description Framework (RDF) is used to store extracted relationships from the corresponding documents, and it provides support for the subsequent retrieval of related documents.

As an important entity in medical texts, time information plays an important role in structuring medical information and supporting clinical decision-making. In this paper, time expressions in Chinese electronic medical record text of depression are studied. The method combines regular expressions with Conditional random fields (CRFs) to recognize time expressions in Chinese electronic medical records. The test data are realistic electronic medical records of depression provided by a hospital in Beijing. The proposed method uses regular expressions to initially recognize the explicit time expression in the text, and adds a dictionary of common drugs and symptoms of depression to the word segmentation, which increases the accuracy of word segmentation. External dictionary features are optimized, and dictionaries are divided into time modifier dictionary, time representation dictionary and event dictionary, which effectively improve the accuracy and recall rate of conditional random field recognition results. Experiments show that the accuracy and recall rate of this method are 96.75% and 93.33% respectively.

Several authors have developed relation extraction methods for automatically learning or refining taxonomies from large text corpora such as the Web. However, without appropriate post-processing, such taxonomies are often inconsistent (e.g. they contain cycles). A standard approach to repairing such inconsistencies is to identify a minimally consistent subset of the extracted facts. For example, we could aim to minimize the sum of the confidence weights of the facts that are removed for restoring consistency. In this paper, we present MAP inference as a base method for this approach, and analyze how it can be improved by taking into account dependencies between the extracted facts. These dependencies correspond to rules of thumb such as “if a given fact is wrong then all facts that have been extracted from the same sentence are also likely to be wrong", which we encode in Markov logic. We present experimental results to demonstrate the potential of this idea. ispartof: pages:31-36 ispartof: Proceedings of Web-KR’14 Workshop vol:2014-November issue:November pages:31-36 ispartof: CIKM location:Shanghai, China date:3 Nov - 3 Nov 2014 status: published