As two important textual modalities in electronic health records (EHR), both structured data (clinical codes) and unstructured data (clinical narratives) have recently been increasingly applied to the healthcare domain. Most existing EHR-oriented studies, however, either focus on a particular modality or integrate data from different modalities in a straightforward manner, which usually treats structured and unstructured data as two independent sources of information about patient admission and ignore the intrinsic interactions between them. In fact, the two modalities are documented during the same encounter where structured data inform the documentation of unstructured data and vice versa. In this paper, we proposed a Medical Multimodal Pre-trained Language Model, named MedM-PLM, to learn enhanced EHR representations over structured and unstructured data and explore the interaction of two modalities. In MedM-PLM, two Transformer-based neural network components are firstly adopted to learn representative characteristics from each modality. A cross-modal module is then introduced to model their interactions. We pre-trained MedM-PLM on the MIMIC-III dataset and verified the effectiveness of the model on three downstream clinical tasks, i.e., medication recommendation, 30-day readmission prediction and ICD coding. Extensive experiments demonstrate the power of MedM-PLM compared with state-of-the-art methods. Further analyses and visualizations show the robustness of our model, which could potentially provide more comprehensive interpretations for clinical decision-making.
Background: The automatic generation of radiology reports, which seeks to create a free-text description from a clinical radiograph, is emerging as a pivotal intersection between clinical medicine and artificial intelligence. Leveraging natural language processing technologies can accelerate report creation, enhancing health care quality and standardization. However, most existing studies have not yet fully tapped into the combined potential of advanced language and vision models. Objective: The purpose of this study was to explore the integration of pretrained vision-language models into radiology report generation. This would enable the vision-language model to automatically convert clinical images into high-quality textual reports. Methods: In our research, we introduced a radiology report generation model named ClinicalBLIP, building upon the foundational InstructBLIP model and refining it using clinical image-to-text data sets. A multistage fine-tuning approach via low-rank adaptation was proposed to deepen the semantic comprehension of the visual encoder and the large language model for clinical imagery. Furthermore, prior knowledge was integrated through prompt learning to enhance the precision of the reports generated. Experiments were conducted on both the IU X-RAY and MIMIC-CXR data sets, with ClinicalBLIP compared to several leading methods. Results: Experimental results revealed that ClinicalBLIP obtained superior scores of 0.570/0.365 and 0.534/0.313 on the IU X-RAY/MIMIC-CXR test sets for the Metric for Evaluation of Translation with Explicit Ordering (METEOR) and the Recall-Oriented Understudy for Gisting Evaluation (ROUGE) evaluations, respectively. This performance notably surpasses that of existing state-of-the-art methods. Further evaluations confirmed the effectiveness of the multistage fine-tuning and the integration of prior information, leading to substantial improvements. Conclusions: The proposed ClinicalBLIP model demonstrated robustness and effectiveness in enhancing clinical radiology report generation, suggesting significant promise for real-world clinical applications.
29 March 2024 in IEEE Transactions on Intelligent Vehicles
The advent of Scenarios Engineering (SE) paves the way to a new era of intelligent vehicles (IVs), driven by Artificial Intelligence (AI)-enabled strategies. It aims at shaping the IVs to be a form that is more relevant to the underlying scenario, thereby accomplishing validation, verification (V&V), and calibration, certification (C&C) of each vehicle. However, such improved capabilities relies on the accumulation and analysis of an unprecedented volume of scenarios. Recently, Sora and other video generation models have opened up new horizons for Imaginative Intelligence. As an extension of TIV-DHW (Distributed/Decentralized Hybrid Workshop) on SE, this letter discusses the potential of Sora to change the scenario generation process by reducing physical shooting, increasing extreme scenario generation, thereby enabling more comprehensive training and testing of IVs. This letter also analyzes the limitations of Sora in accurately model physics and understand cause and effect, which may affect its effectiveness in SE applications. Last, through a comprehensive outlook, this letter aims to provide a potential direction for the development of Sora-like AI technology, thereby promoting the safety, efficiency, reliability, and sustainability of IVs.
28 September 2023 in IEEE Journal of Biomedical and Health Informatics
Effectively medication recommendation with complex multimorbidity conditions is a critical task in healthcare. Most existing works predicted medications based on longitudinal records, which assumed the information transmitted patterns of learning longitudinal sequence data are stable and intra-visit medical events are seriaized. However, the following conditions may have been ignored: 1) A more compact encoder for intra-relationship in the intra-visit medical event is urgent; 2) Strategies for learning accurate representations of the variable longitudinal sequences of patients are different. In this paper, we proposed a novel Sample-adaptive Hierarchical medicAtion Prediction nEtwork, termed SHAPE, to tackle the above challenges in the medication recommendation task. Specifically, we design a compact intra-visit set encoder to encode the relationship in the medical event for obtaining visit-level representation and then develop an inter-visit longitudinal encoder to learn the patient-level longitudinal representation efficiently. To endow the model with the capability of modeling the variable visit length, we introduce a soft curriculum learning method to assign the difficulty of each sample automatically by the visit length. Extensive experiments on a benchmark dataset verify the superiority of our model compared with several state-of-the-art baselines.
Text summarization namely, automatically generating a short summary of a given document, is a difficult task in natural language processing. Nowadays, deep learning as a new technique has gradually been deployed for text summarization, but there is still a lack of large-scale high quality datasets for this technique. In this paper, we proposed a novel deep learning method to identify high quality document–summary pairs for building a large-scale pairs dataset. Concretely, a long short-term memory (LSTM)-based model was designed to measure the quality of document–summary pairs. In order to leverage information across all parts of each document, we further proposed an improved LSTM-based model by removing the forget gate in the LSTM unit. Experiments conducted on the training set and the test set built upon Sina Weibo (a Chinese microblog website similar to Twitter) showed that the LSTM-based models significantly outperformed baseline models with regard to the area under receiver operating characteristic curve (AUC) value.