Problem
Research questions and friction points this paper is trying to address.
clinical time-series
multimodal LLMs
risk prediction
temporal question answering
ViSTA: A Simple Bridge Extends Visual Alignment to Clinical Time-Series Understanding in Multimodal LLMs
🤖 AI Summary
Multimodal large models struggle to accurately process irregular clinical time series for risk prediction. This work proposes the ViSTA adapter, which transforms structured numerical measurements into chart representations compatible with vision-language models, enabling efficient temporal understanding by learning only visual token corrections. The approach freezes pretrained parameters and incorporates visual alignment techniques during fine-tuning on the MIMIC-IV dataset. Experimental results demonstrate that the proposed model achieves an AUC of 0.7376 in acute kidney injury and mortality prediction, significantly outperforming existing methods. Furthermore, it reduces trainable parameters by over 90%, successfully unifying a lightweight architecture with high predictive accuracy.
📝 Abstract
Clinical prediction models estimate risk from patient measurements, while large language models support medical text understanding and question answering. Yet their language capabilities do not ensure accurate prediction from structured, high-dimensional clinical time series. Improving this ability would connect risk estimation with flexible questions about a patient's evolving condition. We introduce ViSTA, a compact adapter that incorporates irregular numerical measurements into a pretrained vision-language model's chart representations. It learns corrections to visual tokens while leaving all pretrained parameters unchanged. On MIMIC-IV, ViSTA has the highest mean scores among the compared adaptations on all four metrics for acute kidney injury and mortality prediction across models with 2-9 billion parameters. With 0.516 million trainable parameters, the 2-billion-parameter model reaches an area under the ROC curve of 0.7376 for acute kidney injury, compared with GPT-5.6 Sol's 0.7380 with text input and high reasoning effort. Training for temporal question answering yields 69.27% accuracy at 4 billion parameters with over 90% fewer trainable parameters than low-rank adaptation using charts or numerical text, at a 2.82-4.88 percentage-point accuracy gap. ViSTA extends pretrained language models to numerical prediction and temporal questions.
Innovation
Methods, ideas, or system contributions that make the work stand out.
Clinical Time-Series
Multimodal LLMs
Compact Adapter
Parameter Efficiency
Visual Alignment
🔎 Similar Papers
No similar papers found.
