Under-Mattress Temporal Sensing for Next-Day Agitation Risk Scoring in Dementia Wards

Zhen Liu; Marta Bono; Robbe Decloedt; Ajda Flisar; Maarten Van Den Bossche; Maarten De Vos

Source-linked AI summary

Under-Mattress Temporal Sensing for Next-Day Agitation Risk Scoring in Dementia Wards

Zhen Liu, Marta Bono, Robbe Decloedt, Ajda Flisar, Maarten Van Den Bossche, Maarten De Vos

arXiv:2608.28152v1Aug 28, 2026cs.AIcs.LG

TL;DR

Agitation fluctuates over short horizons in dementia, but continuous information for estimating next-day risk is limited. This study benchmarked four representations of preceding-night under-mattress signals using patient-grouped evaluation and found stronger discrimination from minute-level modeling than conventional nightly summaries, while calibration and external validation remained necessary.

Problem
Continuous information for estimating next-day agitation risk is limited despite agitation’s short-horizon fluctuations and clinical burden in dementia.
Method
The study compared four representations of preceding-night under-mattress signals using patient-grouped out-of-fold evaluation of discrimination, calibration, fixed-threshold metrics, and period–signal attribution.
Results
Full-night sequence modeling achieved the highest AUROC (0.692) and balanced accuracy (0.658), while minute-level pipelines had higher AUROC than nightly summaries and calibration remained limited.
Takeaways & Limitations
Preceding-night under-mattress signals support modest next-day risk discrimination, making minute-level temporal modeling a promising engineering direction for agitation-risk research.
Takeaways & Limitations
Generalizability is limited by 65 participants from one specialized hospital unit, with unassessed inter-rater reliability, a broad endpoint, and exploratory model selection on the same folds.

Abstract

from arXiv · show

Agitation fluctuates over short time horizons in people living with dementia, yet continuous physiological information for anticipating next-day risk is limited. We assessed whether contactless under-mattress signals from the preceding night inform next-day agitation risk and whether preserving minute-level temporal structure improves performance over conventional nightly summaries. We analyzed 423 patient-nights from 65 subjects in a specialized hospital dementia unit using two under-mattress sensing systems. A unified four-paradigm benchmark compared nightly handcrafted summaries, three-period handcrafted features, full-night sequence modeling, and sliding-window multiple-instance learning. Source-specific preprocessing and five-fold patient-grouped cross-validation were used, with performance estimated from pooled out-of-fold predictions. Evaluation included discrimination, calibration, fixed-threshold metrics, and a comparison of period-signal attribution patterns across two temporal models. Full-night sequence modeling achieved the highest discrimination (AUROC, 0.692; AUPRC, 0.849) and balanced accuracy (0.658). Both minute-level pipelines had higher AUROC than nightly summaries, but differences from three-period handcrafted features were uncertain. Cross-model attribution prioritized activity, heart rate, and respiratory rate during the core overnight period. Calibration remained limited. The preceding night's signals supported modest next-day risk discrimination, with minute-level temporal modeling outperforming nightly summaries. Prospective calibration and external validation are needed before use in individual care decisions. This patient-grouped benchmark identifies contactless overnight sensing as a promising biomedical engineering direction for agitation-risk research in hospitalized dementia cohorts.

I. INTRODUCTION

Agitation is common and burdensome in dementia, but routine observations provide limited continuous information for next-day risk estimation. This study evaluates contactless under-mattress sensing and compares temporal representations for cross-patient prediction in a hospital dementia unit.

Agitation affects approximately 40% of care-home residents with dementia and includes motor, verbal, aggressive, and distress-related behaviors.
Earlier identification could support proactive non-pharmacological care because pharmacological management has limited efficacy and substantial safety concerns.
Night-time physiology is a plausible source of short-horizon agitation information, but its predictive value for next-day behavior remains under-specified.
Contactless under-mattress sensors capture movement, respiratory, and cardiac measures through pressure-sensitive sensing and BCG signals without adding burden to patients or staff.
The study addresses whether preceding-night signals support cross-patient next-day risk estimation by comparing four representations and evaluating discrimination, calibration, thresholds, and attribution.
The analysis paired overnight recordings from two passive sensing systems with next-day agitation assessments under a hospital monitoring protocol.

B. Four-Paradigm Modeling Pipeline

The benchmark maps each overnight record to a next-day agitation probability using four representations: whole-night summaries, period-resolved features, full ordered sequences, and sliding-window multiple-instance learning.

The four paradigms differ in temporal aggregation and representation while mapping one overnight sensor record to a next-day agitation probability.
B0 – Nightly summary: B0 summarizes heart rate, respiratory rate, sleep architecture, activity, and quality across the full night using fold-wise SHAP feature selection and Random Forest prediction.
B1 – Period-resolved summary: B1 extracts handcrafted features separately for evening, core overnight, and morning periods, adds cross-period descriptors, and uses Logistic Regression after SHAP-based selection.
D0 – Whole-night raw-sequence modeling: D0 directly models the complete ordered minute-level sequence with InceptionTime.
M0 – Sliding-window MIL: M0 partitions the sequence into overlapping 90-minute windows with a 45-minute stride, then applies a Transformer encoder and gated-attention MIL pooling.
Final performance uses representative models selected through held-out validation on predefined patient-grouped folds and summarized from pooled out-of-fold predictions.

C. Preprocessing and Cross-Validation

The preprocessing and validation pipeline harmonized heterogeneous sensor streams into minute-level inputs and evaluated all nights with patient-grouped cross-validation to prevent participant leakage.

Each overnight record provided heart rate, respiratory rate, bed activity, binary sleep–wake, and a device-derived quality channel.
Sleep-stage streams encoded Deep, Light, and REM as sleep and Wake/Awake as wake, while missing or unrecognized stages were encoded as wake.
All four pipelines used the same predefined five-fold patient-grouped assignment, balancing sensor source and patient-level agitation rates across folds.
All nights from each patient remained in one fold, and every patient-night contributed exactly one held-out out-of-fold prediction.
Class-imbalance weights were estimated only from each training fold, with no imbalance parameter tuned on held-out predictions.

D. Performance Evaluation

The study evaluated four overnight-signal representations using patient-grouped out-of-fold metrics designed to separate discrimination, probability error, and threshold behavior. Attribution comparison was descriptive and hypothesis-generating rather than causal.

Evaluation reported AUROC, AUPRC, Brier score, sensitivity, specificity, and balanced accuracy at a fixed probability threshold of 0.5.
Performance was estimated with pooled out-of-fold predictions, patient-grouped bootstrap uncertainty, and paired pipeline comparisons.
Four pipelines represented each overnight record using whole-night summaries, period-resolved features, full-night sequences, or sliding windows with MIL aggregation.
Attribution rankings across 15 period–signal combinations were compared descriptively between B1 and M0, without establishing physiological mechanisms or causal relationships.
The main results table covered 423 nights with prevalence 0.723 and reported pooled out-of-fold estimates with 95% patient-level bootstrap confidence intervals.

G. Subgroup and Sensitivity Analyses

Additional analyses examined sensor-source variation and a higher-burden agitation definition within the common patient-grouped evaluation framework. The strongest overall pipelines retained higher discrimination and more balanced threshold behavior, while paired gains were clearest against B0.

Source-stratified analysis: Source-stratified metrics were calculated separately for EMFIT and WSA using pooled patient-grouped out-of-fold predictions.
Sensitivity analysis: The higher-burden sensitivity endpoint required a day-level PAS total score greater than 4, representing agitation in at least two domains with at least one score above 1.
Overall performance: D0 achieved the highest AUROC, specificity, and balanced accuracy, while M0 achieved similar balanced accuracy; B1 exceeded B0 in AUROC and specificity point estimates.
Paired comparisons: +0.127 and +0.114 were the AUROC differences for D0 and M0 versus B0, with both patient-level bootstrap confidence intervals excluding zero.

C. Calibration and Prevalence-based Baselines

High event prevalence shaped probability and threshold metrics, while calibration remained limited across learned models. Attribution rankings showed moderate agreement and emphasized activity and cardiorespiratory signals during the core overnight period.

Prevalence and calibration: 72.3% of patient-nights were agitation-positive, influencing baseline probability and fixed-threshold metrics.
Prevalence-based baselines: The train-prevalence baseline achieved the lowest Brier score overall despite having no useful discrimination, while B0 combined high sensitivity with very low specificity.
Cross-model attribution: Activity, heart rate, and respiratory rate during the core overnight period ranked relatively highly in both attribution models.
Prevalence and calibration: Calibration remained limited across learned models relative to the train-prevalence probability baseline.
Cross-model attribution: The B1 and M0 attribution rankings had descriptive Spearman correlation ρ = 0.57 across 15 period–signal combinations.
Source-stratified performance: Source-stratified performance differed between EMFIT and WSA, with D0 comparatively stable in discrimination but lower WSA balanced accuracy driven mainly by specificity.

F. Sensitivity to Agitation Endpoint Definition

The selected agitation endpoint materially affects prevalence, operating characteristics, and practical interpretation, while device and evaluation boundaries limit generalizability. Cross-model attribution nevertheless identifies recurring overnight signals relevant to next-day risk.

Endpoint definition: 17.3% prevalence followed thresholding at PAS total score greater than 4, down from 72.3%, with substantial changes in operating characteristics.D0 retained the highest AUROC, AUPRC, sensitivity, and balanced accuracy, while B1 had the lowest Brier score and highest specificity.
Generalizability: Performance values from prior predictive studies should not be compared directly because their tasks, input representations, and evaluation designs differed.The present benchmark addresses held-out-patient next-day risk scoring rather than earlier explanatory or multi-day settings.
Endpoint definition: High event prevalence made sensitivity or probability error alone potentially misleading, because prevalence baselines and B0 combined favorable sensitivity or Brier behavior with poor discrimination or specificity.Calibration remained limited across learned models.
Temporal attribution: Activity, heart rate, and respiratory rate during the core overnight period were common high-ranking period–signal combinations across B1 and M0.The convergence prioritizes these signals for targeted physiological validation beyond static nightly summaries.
Temporal attribution: B1 ranked morning activity higher, whereas M0 ranked the core-period binary sleep–wake channel higher, suggesting representation-specific emphasis on transition activity versus local sleep–wake composition.These are attribution patterns rather than mechanistic claims.
Generalizability: Patient-grouped evaluation supports cross-patient risk estimation but does not establish patient-specific day-to-day forecasting or transportability across devices and clinical settings.Source-stratified estimates differed between EMFIT and WSA, despite comparatively stable D0 rank discrimination.

D. Clinical Interpretation

The study supports cautious clinical interpretation of contactless overnight risk scores as adjunctive prompts rather than standalone treatment triggers. Their scope is constrained by endpoint breadth, limited calibration, sample size, single-site recruitment, and exploratory configuration selection.

Clinical interpretation: Any positive primary endpoint includes minor or transient PAS behaviors and does not diagnose clinically actionable agitation.A positive output represents elevated probability of any observed agitated behavior under the study definition.
Clinical interpretation: Sensor-based alerts should prompt timely review of the patient and care context alongside staff judgment, not independently indicate treatment.Agitation may reflect unmet needs, discomfort, or clinical and environmental stressors.
Clinical interpretation: Predicted probabilities cannot yet be interpreted as reliable absolute risks because calibration is limited.A predicted risk of 70% does not necessarily correspond to a 70% observed probability of next-day agitation.
Study design: 423 repeated overnight observations from two passive sensing systems enabled a common four-representation, patient-grouped evaluation under routine hospital care.Source-aware preprocessing and patient-level bootstrap resampling addressed device heterogeneity and within-patient dependence.
Limitations: Cross-patient generalizability remains limited by 65 independent participants recruited from one specialized hospital dementia unit.Inter-rater reliability was not assessed, and exploratory configuration selection on the same folds may make reported performance optimistic.
Future work: Future work requires prespecified selection, external validation across hospital and device settings, prospective source-stratified calibration, and workflow evaluation.These steps are intended to establish robustness and determine whether scores can support a defined care workflow.

V. CONCLUSION

In this hospital dementia cohort, preceding-night under-mattress signals contained modest information about next-day agitation risk. Minute-level temporal models were stronger than nightly summaries, but external validation and prospective calibration remain necessary before individual care use.

Conclusion: Under-mattress signals from the preceding night contained modest information about next-day agitation risk in this in-hospital dementia cohort.The conclusion concerns cross-patient risk estimation in the studied setting.
Conclusion: Full-sequence and sliding-window modeling achieved stronger discrimination than conventional nightly summaries, while advantages over period-specific engineered features remained uncertain.The comparison supports minute-level temporal modeling as a promising engineering direction rather than a definitive winner.
Conclusion: Activity and cardiorespiratory signals during the core overnight period were consistently prioritized across model attributions for further physiological validation.This finding concerns attribution patterns across temporal models.
Conclusion: External validation and prospective calibration are required before resulting scores can guide individual care decisions.The study included ethics approval, representative consent procedures, pseudonymization, and voluntary uncompensated participation.

S1. TEMPORAL SEQUENCE PREPROCESSING DETAILS

The preprocessing pipeline harmonizes heterogeneous EMFIT and WSA recordings into source-aware 1-minute overnight sequences, then derives temporal representations for tabular and learned models. B1 preserves coarse timing through three periods and cross-period descriptors, while fold-specific screening supports benchmark comparison without isolating representation effects from classifier choice.

Sequence preprocessing: Overnight streams were restricted to 17:00–10:00, paired with the following daytime agitation label, physiologically checked, and marked for missingness and unreliable minutes.The resulting cleaned sequences support downstream temporal modeling.
Source-specific processing: EMFIT data were range-checked, placed on a complete 4-second grid, and aggregated to 1-minute resolution using median HR/RR and mean activity.Its quality flag captures simultaneous signal absence and severe missingness.
Source-specific processing: WSA data were processed at native 1-minute resolution with median-filtered HR/RR, unfiltered movement score, and device-derived sleep-stage output.Its quality flag marks minutes where HR, RR, and movement are all missing.
Source-specific processing: EMFIT and WSA were treated as related but non-exchangeable sources because their resolutions, device processing, activity fields, sleep-stage outputs, and quality channels differ.Source-aware imputation and scaling avoid forcing source-specific streams into a common measurement assumption.
Input channels: The five temporal channels were HR, RR, ACT, binary sleep–wake, and quality, with sleep stages encoded as 1 and wake as 0.Missing or unrecognized stages were also encoded as wake, while quality used 0 for valid and 1 for invalid or unknown data.
Period-resolved features: B1 partitioned each overnight record into evening, core overnight, and morning periods to preserve coarse within-night timing for tabular classifiers.It also computed cross-period descriptors such as HR dipping, morning surge, RR dipping, ACT escalation, and variability ratios.
Benchmark design: B0 and B1 shared patient-grouped folds, source-aware preprocessing, candidate classifiers, top-30 SHAP-FS, and mean held-out-fold AUROC selection, but selected classifiers differed.Thus performance differences cannot be attributed to temporal representation alone.

S4. D0/M0 TEMPORAL MODEL CONFIGURATIONS

D0 and M0 were selected through exploratory, patient-grouped configuration screening before pooled out-of-fold evaluation. InceptionTime represented D0, Transformer represented M0, and 90-minute windows with 45-minute strides represented the selected M0 configuration.

Screening limitations: Candidate selection and performance estimation reused the same patient-grouped folds, so configuration comparisons are exploratory.This limitation applies to the screening evidence rather than the pooled out-of-fold primary analysis.
D0 configuration: InceptionTime achieved the highest mean held-out-fold AUROC among evaluated D0 candidates and maintained greater specificity than TCN.TCN showed high sensitivity but low and variable specificity.
M0 configuration: The Transformer encoder was selected as the M0 representative after comparison with alternative neural backbones.The comparison used the shared five-channel input set and patient-grouped folds.
M0 configuration: The 90-minute window with a 45-minute stride had the highest mean AUROC and AUPRC among evaluated M0 windowing settings.These values were simple five-fold means used for sensitivity-based configuration selection.

S7. HIGHER-BURDEN AGITATION ENDPOINT SENSITIVITY

A stricter higher-burden agitation definition changed model operating characteristics while retaining the leading D0 ranking for several discrimination and threshold metrics. Patient-level analyses remained limited by short follow-up, single-class outcomes, and substantial heterogeneity.

Higher-burden endpoint: D0 retained the highest AUROC, AUPRC, sensitivity, and balanced accuracy under the higher-burden endpoint.The endpoint defined positivity as a day-level PAS total score greater than 4 and had 17.3% prevalence.
Higher-burden endpoint: B1 achieved the lowest Brier score and highest specificity under the higher-burden endpoint.This indicates that model operating characteristics varied across agitation definitions.
Interpretation: The higher-burden sensitivity analysis showed partial robustness of D0 ranking while demonstrating dependence of prediction interpretation on agitation definition.The analysis used pooled out-of-fold point estimates on 423 patient-nights.
Per-patient heterogeneity: D0 had the highest median within-patient AUROC at 0.60, but only 37 of 65 patients were evaluable and follow-up ranged from 3 to 10 nights.B0 followed at 0.58, while B1 and M0 had medians of 0.50; broad interquartile ranges indicated substantial heterogeneity.
Per-patient heterogeneity: Selected patient-level trajectories were illustrative rather than inferential comparisons.Patients with at least seven observed nights were eligible, with examples chosen by predefined outcome and follow-up criteria.