LLM-Guided Adaptive Cost Calibration in Active Learning for Parkinson’s Disease Biosignals

Confirmatory labels for Parkinson’s disease (PD), from expert examination or, prospectively, biological assays, are costly, and missed cases and false alarms carry unequal consequences. ClinRisk-LLM is a framework for label-efficient, cost-aware learning from gait and speech signals in which a large language model (LLM) translates a predefined safety specification into a frozen policy setting the positive-class weight of a support vector machine and the priority of potential missed cases for annotation. Cost-sensitive training, risk-aware querying and the LLM policy were ablated on gait (n=165) and speech (n=252) data and benchmarked against post hoc thresholding, a refer-all policy, held-out gait studies and an independent speech cohort. With half of the labels, cost-sensitive active learning reduced the 4:1 asymmetric cost by 34.1% (gait) and 5.7% (speech) relative to standard active learning, at the expense of specificity. ClinRisk-LLM translated the predefined safety specification into a frozen, auditable policy that adjusted risk settings during label acquisition. Component-wise analyses indicated that classifier weighting accounted for most of the observed cost reduction, while the incremental benefit of LLM-guided adaptation varied across datasets and comparison settings. A threshold-tuned symmetric model came within 1.3 cost points for gait, and gait remained roughly level with referring everyone. ClinRisk-LLM reached a gait cost of 47.60 versus 46.10 for fixed cost-sensitive learning; 40 policies generated across three LLMs from a reconstructed specification yielded gait costs of 52.1–75.8. Applications to molecular triage remain prospective. Code and data availability details are provided in the Data Availability Statement.