An arXiv preprint, “Toward System One Clinical Intelligence: Parameter-Efficient Fine-Tuning and Counterfactual Explainability of Decision Foundation Models,” addresses a central trade-off in automated disease classification from patient-reported symptoms. The work contrasts lightweight discriminative models, which are fast but limited in contextual understanding, with autoregressive large language models, which are more context-aware but carry latency costs.
The proposed direction is to apply parameter-efficient fine-tuning to decision foundation models, adapting them for clinical use without the expense of full retraining. To address interpretability, the paper also incorporates counterfactual explainability, so predictions can be understood in terms of what would need to change to alter a classification.
Because only the abstract is currently available, specific methods and evaluation results are not yet detailed. As described, the contribution is a framework for combining speed, context, and explanation in clinical decision support.