A Machine Learning-Informed Dynamic Fuzzy Cognitive Map for Interpretable Prioritization of Multidomain Geriatric Vulnerability in Older Adults

Multidomain vulnerability in older adults is often assessed across separate domains, despite important interactions among cognitive, frailty-related, functional, nutritional, swallowing–airway, respiratory, medication-related, and socioeconomic factors. This study developed and internally evaluated a machine learning-informed dynamic Fuzzy Cognitive Map (FCM) model for prioritizing geriatric vulnerability. The analysis included data from 77 participants: 59 from a hospital outpatient setting and 18 from a nursing home. The model comprised 22 concepts: 18 inputs, three pathway outputs, and one global output. An expert-defined topology represented clinically informed relationships across domains. Dynamic simulation used an input-preserving updating rule, while ridge regression was used to calibrate selected readout weights using fuzzy targets derived from the transformed model inputs. These targets were not independent clinical outcomes. Internal analyses examined site-stratified contribution profiles, expert-only versus calibrated outputs, leave-one-out calibration, dense input-preservation sensitivity (α = 0.50–0.90), alternative missing value imputation, frailty-scale holdout comparisons, and perturbation of manually specified modeling choices. All participant simulations converged. Expert-only and calibrated review classifications agreed for 73/77 participants (94.8%), and the swallowing–airway watchlist was unchanged; leave-one-out calibration remained an internal audit because its targets were derived from model inputs. The nursing home subgroup had higher mean scores than the hospital subgroup across all four outputs. Under the common FCM, stratified contribution analysis identified larger realized contributions from functional dependence, clinical frailty, cognitive impairment, mobility impairment, and respiratory/cough-related burden in the nursing home subgroup. These findings were considered descriptive differences within this cohort and were not interpreted as effects of care setting. The global geriatric vulnerability score classified 22 participants as baseline/low, 32 as elevated, 21 as moderate-high, and 2 as high. The swallowing–airway watchlist identified 46 participants: 11 assigned to priority review, 15 to targeted review, 15 to routine follow-up, and five to monitor. The model transformed multidomain information into interpretable pathway scores, a global priority score, and a swallowing–airway review watchlist. The findings demonstrate internal computational stability and sensitivity characteristics rather than independent clinical validity; external and prospective validation is required before clinical implementation.