Large Language Model as a Universal Clinical Multi-task Decoder

TL;DR

The paper tackles the challenge of managing numerous and evolving clinical tasks by introducing ClinTS-LLM, a universal clinical multi-task decoder that fuses a Warpformer-based time-series encoder with a frozen large language model via a trainable adapter and LoRA. Predictions are produced as language prompts, enabling robust multi-task learning and strong zero-shot and few-shot transfer to new tasks. Across hundreds of tasks, ClinTS-LLM matches traditional STL/MTL performance while offering streamlined deployment and impressive transfer capabilities, especially in low-data regimes. The work highlights the practical potential of a unified, data-efficient approach to dynamic clinical task ecosystems, reducing need for task-specific model maintenance.

Abstract

The development of effective machine learning methodologies for enhancing the efficiency and accuracy of clinical systems is crucial. Despite significant research efforts, managing a plethora of diversified clinical tasks and adapting to emerging new tasks remain significant challenges. This paper presents a novel paradigm that employs a pre-trained large language model as a universal clinical multi-task decoder. This approach leverages the flexibility and diversity of language expressions to handle task topic variations and associated arguments. The introduction of a new task simply requires the addition of a new instruction template. We validate this framework across hundreds of tasks, demonstrating its robustness in facilitating multi-task predictions, performing on par with traditional multi-task learning and single-task learning approaches. Moreover, it shows exceptional adaptability to new tasks, with impressive zero-shot performance in some instances and superior data efficiency in few-shot scenarios. This novel approach offers a unified solution to manage a wide array of new and emerging tasks in clinical applications.

Figures (7)

• Figure 1: An overview of clinical data collection and task organization.
• Figure 2: The technical details for our model.
• Figure 3: The behavior of MTL, STL and LLM for inference.
• Figure 4: The zero-shot transfer effect. ClinTS-LLMs is in zero-shot setting and other models are trained with 100, 200, 400, 800, and 1600 samples, shown in the x-axis. The green lines represent the pretrain results, and the blue lines represent the scratch results. The colors transition from light to dark, indicating the process from shallow-tuning to deep-tuning. The specific intensity of the color shown in the color bars on each side of the line chart corresponds to the number of epochs tuned.
• Figure 5: The label choice transfer effect. All features in subfigure \ref{['fig:pheno_zero']} are the same as Figure \ref{['fig:all_combine']}. For subfigure \ref{['fig:pheno_few']}, we only show the best tune result (selected from all deep-tuning and shallow-tuning) for pretrain and from-scratch model.