Two Achievements of Shenzhen MSU-BIT University Artificial Intelligence Research Institute Accepted on the International Artificial Intelligence Conference IJCAI2025

Recently, two papers of Shenzhen MSU-BIT University Artificial Intelligence Research Institute were accepted on the international artificial intelligence conference - the 34^th International Joint Conference on Artificial Intelligence (“IJCAI” for short). This conference was highly competitive, with just 19.3% of submitted papers accepted for presentation. Under this competitive background, Shenzhen MSU-BIT University emerged as the first author and the first corresponding institution of two accepted papers. The two researches introduced a deep ECG-report interaction (DERI) framework and a semantic tokenizer (ECG2TOK) respectively. These achievements significantly enhanced the clinical semantic expression capability of ECG self-supervised learning, and had excellent performance in many ECG downstream tasks.

Achievement 1: DERI: Cross-Modal ECG Representation Learning with Deep ECG-Report Interaction

The first author of the paper was Chen Jian, a 2022 Class master student, and its corresponding authors were Professor Hu Xiping and Professor Wang Wei from Artificial Intelligence Research Institute. It proposed a cross-modal self-supervised learning framework DERI integrating ECG with clinical reports. Through multi-alignment and feature reconstruction mechanisms, ECG representation was integrated with more clinical semantic information. Additionally, a masked modeling-based RME module was developed and extended to report generation tasks, demonstrating effective representation quality assessment and clinical application potential.

Achievement 2: ECG2TOK: ECG Pre-Training with Self-Distillation Semantic Tokenizers

The first author of the paper was Yuan Xiaoyan, a 2023 Class doctorate student, and its corresponding author was Professor Wang Wei. It proposed a semantic tokenizer-based ECG self-supervised pre-training framework ECG2TOK to generate discrete labels with rich semantics. This semantic tokenizer was capable of extracting semantic embedding through self distillation and online clustering to generate training objectives. In combination with the masking strategy and the label prediction task, ECG2TOK performance was improved significantly in 6 downstream tasks, especially in the low resource scenarios.

It is known that IJCAI is one of the top-level academic conferences in the artificial intelligence field, with the longest history and the greatest influence, representing the advanced achievements in the global AI research. In the ranking of international academic conferences prepared by the China Computer Federation (CCF), IJCAI is classified as Class A.

During this conference, the two achievements of Shenzhen MSU-BIT University were accepted at the same time, symbolizing that the Artificial Intelligence Research Institute’s researches in the intelligent healthcare field have reached international advanced level. In future, the Artificial Intelligence Research Institute will continue the technological innovation in this field, to promote the application of more research achievements.