Three AI4DH Papers Accepted at NeurIPS 2026

We are delighted to announce that three papers from the AI for Digital Health (AI4DH) Lab have been accepted at the Conference on Neural Information Processing Systems (NeurIPS) 2026. The accepted work spans multimodal survival analysis, model merging and data-efficient foundation models, reflecting the breadth of the group’s research in machine learning and healthcare.

Munib Mešinović’s paper, Structure-biased Graph Learning for Multimodal Survival in EHR, has been accepted to the main track. His work addresses the challenges of scalability and interpretability in survival analysis using electronic health records (EHRs). The proposed framework learns sparse graph representations tailored to predicting how patient risk changes over time, bringing together information from multiple clinical data sources. It outperformed previous survival models in intensive and emergency care settings, while achieving training speeds up to 20 times faster than the structure- and graph-learning models evaluated. The work provides a promising basis for more scalable clinical risk prediction and further clinical validation.

Chenqi Li’s paper, Standing on the Shoulders of Giants: Rethinking EEG Foundation Model Pretraining via Multi-Teacher Distillation, addresses the difficulty of developing electroencephalogram (EEG) foundation models with limited, noisy data. The team developed Multi-Teacher Distillation Pretraining (MTDP), a two-stage framework that combines knowledge from established foundation models in other modalities and transfers it into EEG models. Evaluated across two model architectures, nine downstream tasks and 12 datasets, MTDP improved on conventional self-supervised pretraining and remained competitive when using only 25% of the pretraining data. The findings highlight the potential of cross-modal knowledge transfer to make EEG foundation model development more data-efficient. For more information, please refer to the arXiv.

Zhikang Chen’s paper, From Coefficients to Directions: Rethinking Model Merging with Directional Alignment, explores how models trained for different tasks can be combined more effectively. His framework, Merging with Directional Alignment (MDA), aligns task-specific directions in both model parameters and learned features, reducing interference between tasks when models are merged. Rather than focusing only on how much each model contributes, the approach also considers how their representations fit together. Experiments across computer vision and natural language processing benchmarks demonstrated consistent improvements in multi-task performance and generalisation to unseen tasks. For more information, please refer to the arXiv.

We warmly congratulate Munib, Zhikang, Chenqi and their co-authors on these achievements, and look forward to sharing their research with the NeurIPS community.