JMIR Publications

ABSTRACT

Background:

Comprehensive geriatric assessment (CGA) in long-term care is performed intermittently. Contactless mattress sensors record movement, posture, sleep-related and pressure-related signals continuously and could partly supplement CGA between assessments, but their added value beyond the previous assessment has not been established.

Objective:

This study aimed to evaluate whether mattress-derived features can supplement intermittent CGA by examining how well they reflect concurrent pressure-injury risk (Braden Scale) and activities of daily living (Barthel Index), how much they add beyond the previous assessment, and whether they predict decline before the next assessment.

Methods:

This retrospective cohort study included 192 residents aged 65 years or older at one long-term care provider in Taiwan. Because beds and their sensors are reassigned between residents, sensor data were attributed to residents from the identity recorded with each record (110,211 resident-days, March 2023 to August 2025; 666 features). Endpoints were Braden score ≤14 and Barthel score ≤5. Random forests were evaluated with resident-disjoint 5-fold cross-validation and resident-cluster bootstrap 95% CIs and compared with age, in-bed proportion, the previous score, and the previous score plus sensor features. We also examined robustness, clinical cut-points, decline prediction from 30-day checkpoints between assessments, validation schemes, and a reanalysis of the original data, in which device histories had been assigned to residents from an end-of-study configuration snapshot.

Results:

Where identity evidence existed, 52.4% of resident-days assigned by the configuration-snapshot method belonged to another person. With record-level attribution, sensor-only areas under the receiver operating characteristic curve (AUROCs) were 0.774 (95% CI 0.720-0.827) for Braden and 0.805 (0.747-0.861) for Barthel. Sensor features outperformed age and in-bed proportion, and discrimination was robust across 20 resident partitions, with one assessment per resident, at clinical cut-points (0.79-0.86) and with the original 342-feature architecture (0.794 and 0.790). The previous score alone achieved AUROCs of 0.97 to 0.98. Adding sensor features to a random-forest model of the previous score increased AUROC by 0.0067 (Holm P=.143) for Braden and 0.0096 (0.0031 to 0.0177; P=.028) for Barthel, and did not exceed a logistic model of the previous score. Sensor features did not reliably improve detection of decline before the next assessment (38 and 39 declines). Reanalysis of the original data supported the same conclusions with weaker status discrimination (AUROC 0.629 and 0.627).

Conclusions:

Mattress-sensor features attributed to residents from record-level identity carry information about concurrent pressure-injury risk and functional dependency that could partly supplement CGA between assessments. The added information beyond the previous assessment was small and comparator-dependent, and earlier detection of decline was not demonstrated; prospective studies with frequent, fixed-schedule assessments are needed. Longitudinal sensing studies in shared beds should verify which resident a device measured on each day. Clinical Trial: This retrospective analysis used routinely collected records with no change to care and no contact with residents. Ethics approval was granted by the Institutional Review Board of Kaohsiung Veterans General Hospital (Approval Number: T-VGHKS (TN)-38762), which waived the requirement for informed consent. Residents and the public were not involved in the design or conduct of this retrospective study. Analyses used pseudonymized data. Names and national identifiers were replaced by keyed pseudonyms and study codes in a restricted linkage file that was not used for analysis, and only aggregate results are reported.