Aligning multimodal medical information with large language models to create a universal multimodal medical LLM assistant.
Used Skills: Neural Network Pipeline, NLP, Large Language Models, Multi-modality, Prompt Engineering, Fine-tuning
Using prompt engineering to perform self-evaluation under the zero-resource setting to test the understanding of LLMs to the instructions
Used Skills: Neural Network Pipeline, NLP, Large Language Models, Constrained Beam Search, Prompt Engineering
Developing a novel, multi-modal, and unified pretraining framework called MEDHMP for multi-modality health data pre-training.
Used Skills: Multi-modality, Pre-training, Pre-trained Language Model, Self-supervised Learning, Representation Learning, EHR, ICD Codes
Improving ICD coding performance through modeling contextualized code relations through graph network.
Used Skills: Bi-LSTM, Graph Attention Network, Synonym Fusion, ICD Coding
Incorporating the patient's EHR modality with the drug molecular level information to predict the potential adverse reaction.
Used Skills: Pre-trained Language Models, Transformers, Multi-modality, SMILES Chemical Presentation, EHR, ICD codes, Adverse Event Prediction
Deigning hierarchical matching model for trial protocols with novel group-based training loss and 2D word matching.
Used Skills: NLP, Transformers, Convolutional Network, Group Loss, Hierarchical Attention, Information Retrieval
Using information compression to reduce the clinical note noise and improve the speed of automatic ICD coding.
Used Skills: Transformers, NLP, ICD Coding
Designing a controllable medical term simplification pipeline for using external medical dictionary knowledge.
Used Skills: Neural Network Pipeline, NLP, Question Answering, Constrained Generation, External Knowledge Injection
Using two-level transformers to model the complex EHR code sequential data to predict future diseases.
Used Skills: Transformers, Time-aware Attention, EHR, ICD Codes, Disease Prediction
The project aims at using automatic rule based approaches to find the potential structures in chart designs and extracts them to generate templates for users.
Used Skills: Sequential Matching
Creating new approaches to detect chart objects. Unlike traditional common object detection, the chart data are highly homogeneous and abnormal in length-width ratio, which bring huge difficulties to the detection. Hence a new approachis needed for this new special situation
Used Skills: Computer Vision, Object Detection, Point Detection
Inspired from general text embedding approaches and is based on clustering idea.
Used Skills: Masked Pre-training, Fourier transform
An improved Pix2pix network that can assign the desired generated color based on a color mask. The original structure is divided into two parts including texture network and coloring network.
Used Skills: GAN, Computer Vision
Using net spider to collect short messages from social network and training an ID classier based on LSTM Cell to automatically classify users’ social backgrounds according to their short messages.
Used Skills: Pre-trained Language Models, Text Classfication, Web Spider
The project aims at using automatic approaches to help radiologists to analysis pulmonary nodules based on CT images.
Used Skills: Object Detection, CT images, Image Classfication
