A longitudinal geospatial multimodal dataset of post-discharge frailty, physiology, mobility, and neighborhoods

Frailty in older adults is associated with increased vulnerability to functional decline, reduced mobility, social isolation, and challenges during the transition from hospital to community living. These factors are associated with rehospitalization and may adversely influence recovery. Neighborhood environments can further shape recovery trajectories by affecting mobility opportunities, social engagement, and access to community resources. Multimodal sensing technologies combined with data-driven analytical approaches offer the potential to continuously monitor these multidimensional factors in real-world settings. This Data Descriptor presents GEOFRAIL, a longitudinal geospatial multimodal dataset collected from community-dwelling frail older adults following hospital discharge. The dataset is organized into interconnected tables capturing participant demographics, features derived from multimodal sensors, biweekly clinical assessments of frailty, physical function, and social isolation, and temporal location records linked to neighborhood amenities, crime rates, and census-based socioeconomic indicators. Data were collected over an eight-week post-discharge period using standardized pipelines with privacy-preserving spatial aggregation. Technical validation demonstrates internal consistency across geospatial, sensor-derived, and clinical measures and reports baseline performance of machine learning models for characterizing recovery trajectories.

💡 Research Summary

The paper introduces GEOFRAIL, a publicly available longitudinal dataset that integrates multimodal sensor data, bi‑weekly clinical assessments, and geospatial context for frail older adults during the eight‑week period following hospital discharge after a lower‑limb fracture. Eighteen participants (aged 60 + ) from the Greater Toronto Area were enrolled. Continuous data were captured using the MAISON platform, which combines a Google Pixel Watch 2 (1 Hz accelerometry, heart rate every 30 min, step count, GPS at 1 min when outside a 50 m home geofence), a Motorola Android smartphone for questionnaires, a Proteus M5 motion sensor in the living room, and a Withings under‑mattress sleep mat (sleep stages, total sleep time, heart rate during sleep, snoring, wake‑ups). Data streams were automatically uploaded to a secure Google Cloud instance, requiring only routine charging and wear.

Clinical assessments followed Fried’s frailty phenotype and included grip strength (CAMRY dynamometer), body weight (Withings scale), Timed Up‑and‑Go, Rapid Assessment of Physical Activity, and the Clinical Frailty Scale Health Questionnaire. Every two weeks, additional validated instruments were administered via Microsoft Teams video calls: the 6‑item Social Isolation Scale, Oxford Hip and Knee Scores, 30‑second chair‑stand test, and others. These provide reference points for physical function and social engagement.

Geospatial enrichment was performed by linking GPS coordinates to external APIs. For each unique location visited, a 1 km radius was queried via Google Places to count parks, libraries, food establishments, community centres, and places of worship. Coordinates were reverse‑geocoded to obtain Canadian postal codes, which were then mapped to Forward Sortation Areas for crime data (Toronto Police Service) and to dissemination areas for census‑derived socioeconomic variables (income, education, housing, employment, ethnicity, language, etc.). To protect privacy, raw latitude/longitude and postal codes were replaced with coded identifiers (L0001–L4096) and perturbed by 3‑5 % random noise; timestamps were also slightly shifted.

The dataset is organized into five relational tables: (1) Demographics (18 × 11 fields), (2) Daily sensor features (18 × 56 days × 46 features), (3) Clinical assessments (18 × 4 timepoints × 63 variables), (4) Temporal location records (≈ 18,843 rows of second‑level timestamps and location IDs), (5) Aggregated neighborhood attributes (2,881 unique locations with amenity counts, crime rates, and socioeconomic indicators). All tables are linked by participant ID, timestamp, and location ID.

Technical validation demonstrated internal consistency: sensor‑derived activity metrics correlated with step counts and TUG scores; higher amenity density was associated with increased daily steps and reduced social isolation scores; higher crime rates correlated with lower outdoor mobility. Machine‑learning baselines (random forest, XGBoost) predicting 8‑week recovery outcomes (e.g., step count increase, TUG improvement) achieved AUROC ≈ 0.84 when using the full multimodal + geospatial feature set, outperforming models using only sensor or only clinical data (AUROC ≈ 0.78).

Key contributions are: (1) a rare integration of continuous physiological, activity, and sleep monitoring with repeated clinical frailty assessments; (2) systematic linkage of individual movement trajectories to fine‑grained neighborhood amenities, safety, and socioeconomic context; (3) demonstration that combining these layers improves prediction of functional recovery. GEOFRAIL is hosted on Zenodo (v1) and is intended to enable reproducible research on post‑discharge frailty, to develop personalized rehabilitation interventions, and to inform community‑level policies that support aging‑in‑place.