Outside the healthy bounds: reference intervals for population and individual-level medical speech analysis

Reference intervals (RIs) are a well-established clinical tool for defining healthy bounds of laboratory measures. Recently, they have been applied to acoustic features extracted from speech to provide a transparent and disorder-agnostic perspective on symptomatic speakers. We evaluate the potential and limitations of RIs across three datasets: clean, laboratory recordings from people with multiple sclerosis (pwMS) and two noisy datasets targeting respiratory symptoms and work-related stress, collected under everyday-life conditions. At the population level, we compare the proportion of features outside the RI bounds between control and symptomatic groups, test noise augmentation of the RI-estimation corpus for improved robustness, and identify the most indicative acoustic features for each health condition. At the individual level, we derive a per-speaker deviation score that quantifies how far each speaker's acoustic profile deviates from the healthy reference, and assess its discriminative ability using the area under the receiver operating characteristic curve (AUC). Noise augmentation of the healthy reference corpus yields similar population-level separation; the continuous ${\mathrm{DS}}_{Q123}$ deviation score yielded 12 features that remained significant after global Benjamini-Hochberg false discovery rate (FDR) correction. The individual-level deviation score achieves promising discrimination for male pwMS ( $\mathrm{AUC}=0.97$ for read speech), even in a clinically mild cohort, while results for audRespire and the work-related stress dataset remain limited at both levels. RIs show promise as interpretable, disorder-agnostic markers for speech-based health assessment; the individual-level deviation score complements the population-level analysis by providing a per-speaker health indicator that can surface deviations masked in group-level comparisons.