Non-generative readout with few-label calibration versus LoRA fine-tuning of a large language model for ASA physical status classification: accuracy, speed, cost and label efficiency in three data sources

Background. The American Society of Anesthesiologists Physical Status (ASA-PS) class underpins perioperative risk communication, but its assignment is subjective. Generative large language models (LLMs) are slow and costly, and LoRA fine-tuning needs site-specific labels and a separately maintained adapter for every site. Non-generative decision models return option probabilities in one forward pass, but their comparison with LoRA in anesthesia, and the number of local labels they need, is unstudied. Objective. To compare, on one open base model, generative zero-shot prompting, a non-generative readout, the readout with cutpoint calibration on a few local labels, and task LoRA for four-class ASA-PS (I, II, III, IV-V) in accuracy, speed and label requirement. Methods. Retrospective study, not preregistered. MOVER (57,473 operations) was the development source, the local hospital's 2025-2026 operations (n = 2,753) the temporally external validation set, and VitalDB (n = 6,242) a second site. The base model was Qwen3.5-9B; input was preoperative structured variables (T1) or T1 plus diagnosis text (T2). The primary metric was quadratic-weighted {kappa} against each site's recorded ASA class, with patient-level cluster-bootstrap intervals; the non-inferiority margin was 0.05. Cost means tokens and GPU time. Results. In the local validation set the readout reached {kappa} = 0.460 versus 0.227 for generative zero-shot prompting with a requested rationale (difference +0.234, 95% confidence interval [CI] 0.201-0.264; 999 parsable cases of 1,000). Median latency was 111 ms versus 7,067 ms. Cutpoints fitted on 100 and 1,000 local labels gave {kappa} = 0.526 and 0.534; LoRA gave 0.423, 0.542, 0.559 and 0.561 with 100, 300, 1,000 labels and the full training set, the two curves crossing at about 270 labels. Calibrated readout (1,000 labels) minus LoRA@1,000 was -0.019 (-0.042 to +0.005) for the prespecified seed and -0.025 (-0.046 to -0.003) for the three-seed mean, both within the 0.05 margin; in stricter sensitivity analyses the lower bound sat just below it. LoRA had higher exact agreement (0.746 versus 0.666). In VitalDB the difference was 0.000 (-0.023 to +0.025). Conclusions. The non-generative readout needs no training and ran about 64 times faster than generative prompting with a rationale. A few hundred local labels bring it close to LoRA in {kappa} without training and storing a per-site adapter, although {kappa} of 0.52-0.58 is not evidence of human-level performance. LoRA keeps a higher {kappa} once several hundred or more labels are available, and higher exact agreement. Non-inferiority holds for the prespecified analyses and sits at the margin in stricter sensitivity analyses, so it should be read as conditional.