Large language model analysis of real-world phone calls reveals prodromal and progressive biomarker of parkinsonism: A two-year proof-of-concept study.

Abstract

No sufficiently sensitive biomarker exists to monitor disease progression to assess treatment efficacy in synucleinopathies such as Parkinson's disease (PD), particularly during the prodromal phase when interventions are likely to be most effective. Existing digital biomarkers often rely on active tasks or clinic-based assessments, limiting their scalability and real-world applicability. In this proof-of-concept study, we evaluated whether linguistic features derived from real-world phone call recordings using large language models can serve as a language-based progression biomarker in isolated rapid eye movement sleep behavior (iRBD). In this two-year study, we enrolled 74 participants, including 21 iRBD (20 men), 26 PD (25 men), and 27 healthy controls (26 men) age-matched to iRBD participants. Speech data collection occurred remotely in participants' natural environments through routine phone calls. Over 34,000 phone calls (<1,400 hours) were recorded over the two-year period. Compared to healthy controls, individuals with iRBD exhibited significant declines in sentence coherence (p = 0.016), semantic-syntactic diversity (p < 0.001), topic diversity (p < 0.001), and sentence probability (p < 0.001) over the two years. Prodromal changes in iRBD were detectable with an area under the curve of 0.82 from as few as 21 calls. For a two-year neuroprotective trial targeting 50% drug efficacy, the estimated sample size was 78 iRBD participants per arm based on a time-to-event analysis. These findings demonstrate that fully automated phone call analysis can detect both prodromal and progressive changes in alpha-synucleinopathy. This approach is scalable, minimizes the effort required from both patients and clinical staff, and enables remote, low-burden monitoring in screening at-risk populations and therapeutic trials.