ViSTA: A Simple Bridge Extends Visual Alignment to Clinical Time-Series Understanding in Multimodal LLMs
ViSTA adds just 0.516 million trainable parameters while keeping the underlying vision-language model frozen.
The paper says the adapter lets pretrained multimodal LLMs use irregular clinical measurements as part of chart representations. On MIMIC-IV, it led the compared adaptation methods across four metrics for acute kidney injury and mortality prediction. A 2B-parameter model reached 0.7376 AUROC for acute kidney injury, nearly matching the cited GPT-5.6 Sol result of 0.7380 using text input and high reasoning effort. For temporal question answering, the 4B model hit 69.27% accuracy with far fewer trainable parameters than low-rank adaptation, though with a 2.82–4.88 point accuracy gap. ArXiv · AI/CL/LG's note
The paper says the adapter lets pretrained multimodal LLMs use irregular clinical measurements as part of chart representations. On MIMIC-IV, it led the compared adaptation methods across four metrics for acute kidney injury and mortality prediction. A 2B-parameter model reached 0.7376 AUROC for acute kidney injury, nearly matching the cited GPT-5.6 Sol result of 0.7380 using text input and high reasoning effort. For temporal question answering, the 4B model hit 69.27% accuracy with far fewer trainable parameters than low-rank adaptation, though with a 2.82–4.88 point accuracy gap. ArXiv · AI/CL/LG's note
score 4