New AI Research Uses Autoregressive Sequence Modeling for Missing Healthcare Data

A new research paper addresses the challenge of missing data in multimodal machine learning models for healthcare by reframing clinical diagnosis as an autoregressive sequence modeling task. The work utilizes causal decoders from large language models to model patient trajectories across multiple data types. Researchers introduced a missingness-aware contrastive pre-training objective that integrates various modalities with incomplete data into a shared latent space. Transformer-based architectures employing autoregressive sequence modeling demonstrated superior performance compared to baseline methods on the MIMIC-IV and eICU fine-tuning benchmarks. The approach specifically tackles the inherent temporal and sparse nature of clinical datasets, where different diagnostic modalities frequently contain gaps. By leveraging techniques from large language models, the methodology aims to capture underlying predictive signals while maintaining model explainability. The research was announced on arXiv under the identifier 2604.18753v1 with a cross-announcement type.