Abstract
<title>Abstract</title> <p> <underline>Background</underline> : Functional decline in older adults is a critical predictor of falls, hospitalization, and mortality, yet early-stage impairment frequently goes undetected in routine clinical care. Self-rated health (SRH), widely used in population health screening, demonstrates poor sensitivity for functional limitations because older adults often normalize mobility changes or employ compensatory strategies that mask emerging deficits. Medicare AWVs remain heavily dependent on self-report and inconsistently identify early decline. <underline>Objective</underline> : To evaluate whether a conversational voice AI health assessment system can detect meaningful functional decline among older adults who rate their health positively, thereby identifying a "hidden risk" population that conventional SRH-based screening would overlook. <underline>Methods</underline> : This cross-sectional observational study analyzed data from a voice AI agent leveraged telephone-based health risk assessment deployed within OhioHealth (Columbus, Ohio). Community-dwelling older adults (N = 461) with valid responses on SRH and six functional indicators (meal preparation, personal care needs, difficulty standing from a chair, furniture walking, unsteady gait, and curb navigation) were included. SRH was dichotomized as high (good/very good/excellent) versus low (poor/fair). Functional decline was defined as ≥2 impairments. Chi-square tests, unadjusted and adjusted logistic regression, and sensitivity analyses with alternative thresholds were conducted. <underline>Results</underline> : Participants had a mean age of 75.6 years (SD = 6.9); 59.7% were female. Most participants (83.1%) reported high SRH. Among those with high SRH, nearly half (49.3%) demonstrated high functional decline when assessed by the AI system. In contrast, 87.2% of participants with low SRH exhibited high functional decline, confirming clinical coherence. The association between SRH and functional decline was statistically significant (OR = 6.99, 95% CI: 3.48–13.89; p < .001) and remained robust after adjustment for age and gender (adjusted OR = 6.41, 95% CI: 3.17–12.99; p < .001). Sensitivity analysis using a ≥3 impairment threshold found that 24.3% of the high-SRH group still demonstrated substantial multi-domain functional limitations. <underline>Conclusions</underline> : Conversational AI can identify substantial functional decline among older adults who perceive themselves as healthy—a hidden-risk population that traditional screening could miss. These findings support the potential of voice AI-based telephone assessments to enable scalable, proactive detection and early intervention for at-risk older adults before functional impairment progresses to disability. </p>