Under-Mattress Temporal Sensing for Next-Day Agitation Risk Scoring in Dementia Wards
Authors: Zhen Liu, Marta Bono, Robbe Decloedt, Ajda Flisar, Maarten Van Den Bossche, Maarten De Vos
Organizations: 1STADIUS Center, Department of Electrical Engineering, KU Leuven, Leuven, Belgium · 2Geriatric Psychiatry, University Psychiatric Center KU Leuven, Leuven, Belgium · 3Neuropsychiatry, Department of Neurosciences, Leuven Brain Institute, KU Leuven, Leuven, Belgium · Department of Psychiatry, Maastricht University Medical Centre+ (MUMC+), Maastricht, The Netherlands · 4Woman and Child, Department of Development and Regeneration, KU Leuven, Leuven, Belgium
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.
Figures & tables
Fig. 1: Four-paradigm modeling pipeline for retrospective next-day agitation risk scoring. The design contrasts handcrafted features with minute-level signals and whole-night aggregation with within-night representations. B0 uses nightly handcrafted summaries. B1 uses handcrafted features from the evening period (17:00–22:00), core overnight period (22:00–06:00), and morning period (06:00–10:00). D0 encodes the full ordered minute-level sequence. M0 encodes overlapping minute-level windows followed by attention pooling. All paradigms use the same patient-grouped folds and are compared using pooled OOF predictions.
Characteristic
Value
Cohort and outcome
Patients
65
Patient-nights
423
Agitation-positive nights
306 (72.3%)
Demographic and cognitive characteristics
Male
41 (63.1%)
TABLE I: Cohort characteristics and sensing configuration for the pooled EMFIT and WSA in-hospital analysis.
Model
AUROC
AUPRC
Sensitivity
Specificity
Brier
BalAcc
Always positive
0.500 [0.500, 0.500]
0.723 [0.648, 0.796]
1.000 [1.000, 1.000]
0.000 [0.000, 0.000]
0.277 [0.204, 0.352]
0.500 [0.500, 0.500]
Train prevalence
0.441 [0.333, 0.554]
0.693 [0.594, 0.795]
1.000 [1.000, 1.000]
0.000 [0.000, 0.000]
0.201 [0.169, 0.235]
0.500 [0.500, 0.500]
Stratified random
0.516 [0.463, 0.572]
0.730 [0.652, 0.806]
0.732 [0.681, 0.787]
0.299 [0.211, 0.400]
0.388 [0.334, 0.447]
0.516 [0.463, 0.572]
B0-RF
0.565 [0.472, 0.657]
0.785 [0.690, 0.866]
0.948 [0.907, 0.980]
0.077 [0.031, 0.126]
0.206 [0.174, 0.240]
0.512 [0.482, 0.543]
B1-LR
0.622 [0.547, 0.703]
0.816 [0.737, 0.881]
0.644 [0.569, 0.727]
0.538 [0.414, 0.658]
0.246 [0.212, 0.278]
0.591 [0.525, 0.662]
D0-InceptionTime
0.692 [0.622, 0.759]
0.849 [0.778, 0.907]
0.641 [0.541, 0.745]
0.675 [0.562, 0.777]
0.218 [0.193, 0.241]
0.658 [0.598, 0.718]
TABLE II: Main pooled OOF results for the EMFIT and WSA in-hospital cohort (423 nights, prevalence 0.723). Metrics are reported as pooled OOF point estimates with 95% CIs from patient-level bootstrap resampling. Boldface indicates the best performance among B0/B1/D0/M0.
Comparison
Δ AUROC
Bootstrap 95% CI
DeLong p
B1–B0
+0.057
[-0.019, +0.133]
0.077
D0–B0
+0.127
[+0.037, +0.219]
0.0005
M0–B0
+0.114
[+0.021, +0.210]
0.0015
D0–B1
+0.070
[-0.022, +0.163]
0.057
M0–B1
+0.057
[-0.037, +0.144]
0.113
M0–D0
-0.013
[-0.103, +0.077]
0.714
TABLE III: Paired AUROC comparisons on shared pooled OOF predictions. Patient-level bootstrap intervals are the primary uncertainty estimates.
Fig. 2: Discrimination and calibration on shared pooled OOF predictions. D0 has the highest AUROC, AUPRC, and balanced accuracy, whereas B0 has the lowest learned-model Brier score despite weak specificity. Calibration remains limited relative to the train-prevalence probability baseline in Table II .
Fig. 3: Qualitative comparison of B1 segment-feature and M0 final-model attribution rankings across 15 segment–signal combinations. Each axis shows the relative attribution rank within one model, with higher values indicating higher rank. Combinations near the dashed diagonal have similar ranks, and those in the upper-right region rank highly in both models. Marker shape denotes the within-night segment and color denotes the signal. The descriptive Spearman rank correlation was ρ=0.57 .
Source
Model
AUROC
AUPRC
Sens.
Spec.
BalAcc
EMFIT
B0
0.506
0.679
0.938
0.057
0.497
B1
0.606
0.749
0.611
0.557
0.584
D0
0.688
0.802
0.625
0.743
0.684
M0
0.648
0.758
0.597
0.643
0.620
WSA
B0
0.631
0.866
0.957
0.106
0.532
B1
0.630
0.868
0.673
0.511
0.592
TABLE IV: Sensor-source-stratified performance from pooled OOF predictions. EMFIT included 214 nights with a prevalence of 0.673, and WSA included 209 nights with a prevalence of 0.775.
Paradigm
Configuration / model
AUROC
AUPRC
F1
Accuracy
Sensitivity
Specificity
B0
SHAP-FS Random Forest
0.578 ± 0.103
0.798 ± 0.085
0.823 ± 0.061
0.709 ± 0.083
0.947 ± 0.058
0.080 ± 0.037
B0
Tuned Random Forest
0.589 ± 0.080
0.811 ± 0.062
0.827 ± 0.049
0.714 ± 0.072
0.953 ± 0.027
0.087 ± 0.078
B0
Default Random Forest
0.575 ± 0.083
0.806 ± 0.061
0.829 ± 0.048
0.714 ± 0.068
0.969 ± 0.026
0.040 ± 0.068
B1
SHAP-FS Logistic Regression
0.612 ± 0.056
0.819 ± 0.057
0.701 ± 0.091
0.617 ± 0.080
0.638 ± 0.118
0.535 ± 0.168
B1
Default XGBoost
0.601 ± 0.083
0.816 ± 0.070
0.777 ± 0.049
0.656 ± 0.059
0.832 ± 0.035
0.191 ± 0.051
B1
Tuned XGBoost
0.568 ± 0.058
0.798 ± 0.046
0.751 ± 0.066
0.625 ± 0.079
0.793 ± 0.066
0.174 ± 0.095
TABLE S1: Exploratory configuration comparisons on the same patient-grouped folds. Values are mean ± sample SD of the five fold-specific held-out metrics and are not pooled OOF estimates. Only the selected SHAP-FS classifier is shown for B0 and B1. Tuned and default rows are non-SHAP-FS sensitivity references. D0 and M0 rows compare neural encoder families under the selected five-channel input set.
Window / stride
AUROC
AUPRC
F1
Accuracy
Sensitivity
Specificity
30 / 15 min
0.622
0.805
0.718
0.624
0.668
0.518
60 / 30 min
0.645
0.826
0.719
0.630
0.668
0.541
90 / 45 min
0.698
0.851
0.745
0.666
0.681
0.623
180 / 90 min
0.677
0.837
0.703
0.624
0.635
0.621
TABLE S2: M0 Transformer window/stride sensitivity under five-channel train-fold normalization. Values are simple five-fold means for sensitivity screening.
Model
AUROC
AUPRC
Brier
Sens.
Spec.
BalAcc
B0
0.566
0.255
0.178
0.178
0.900
0.539
B1
0.636
0.280
0.164
0.260
0.911
0.586
D0
0.654
0.286
0.239
0.575
0.700
0.638
M0
0.635
0.267
0.179
0.274
0.877
0.576
TABLE S3: Higher-burden agitation endpoint sensitivity using a day-level PAS total score >4 as the positive class. Metrics are pooled OOF point estimates on 423 patient-nights with 17.3% prevalence.
Fig. S1: Within-patient AUROC by paradigm. AUROC is computed only for patients with at least one agitated and one non-agitated night. Of the 65 patients, 37 were evaluable and 28 with single-class outcomes were excluded. Lines within boxes denote medians, green triangles denote means, and the dashed line marks chance-level AUROC=0.50.
Fig. S2: Selected single-subject prediction trajectories from pooled OOF predictions. Patients with at least seven observed nights were eligible. The outcome-varying panels were selected by the largest numbers of non-agitated and agitated nights, respectively. The single-class panels were selected by longest follow-up, with ties resolved by greater cross-model prediction variability. Red and green date labels indicate agitated and non-agitated next-day outcomes, respectively. Each point is predicted from the preceding overnight sensor record.
Early detection of neurodegeneration remains a critical clinical challenge. This study investigates whether sleep EEG signal criticality, quantified via Multifractal Detrended Fluctuation Analysis (MFDFA), serves as a non-invasive biomarker for future cognitive decline. We analyzed longitudinal data from the National Sleep Research Resource (NSRR) Study of Osteoporotic Fractures (SOF) cohort, comparing baseline sleep EEG dynamics between women who remained cognitively normal and those who later progressed to dementia-related impairment (3MS<78).Our results reveal significant group-level differences in Hurst exponent H(q) distributions, particularly during non-REM stages N2 and N3. Cognitively healthy individuals exhibited signal dynamics significantly closer to an optimally critical state across all electrode locations (p⩽0.001), supporting the Brain Criticality Hypothesis. Supervised UMAP projections confirmed clear spatial separation between groups throughout the overnight sleep architecture.The dementia group demonstrated a shift in DFA exponents toward 1.0, suggesting that a reconfiguration of scale-free neural dynamics during sleep precedes clinical symptoms. These findings highlight the potential for MFDFA-derived measures to be integrated into automated, sleep-based screening tools, enabling earlier preventative interventions during the prodromal window of dementia.
Stanisław Narębski, Tomasz Komendziński, Tomasz M. Rutkowski
Academia Copernicana Nicolaus Copernicus University Toru´n, Poland · Department of Cognitive Science Nicolaus Copernicus University Toru´n, Poland · The University of Tokyo, Tokyo, Japan +2
Modern clinical practice increasingly depends on reasoning over heterogeneous, evolving, and incomplete patient data. Although recent advances in multimodal foundation models have improved performance on various clinical tasks, most existing models remain static, opaque, and poorly aligned with real-world clinical workflows. We present Cerebra, an interactive multi-agent AI team that coordinates specialized agents for EHR, clinical notes, and medical imaging analysis. These outputs are synthesized into a clinician-facing dashboard that combines visual analytics with a conversational interface, enabling clinicians to interrogate predictions and contextualize risk at the point of care. Cerebra supports privacy-preserving deployment by operating on structured representations and remains robust when modalities are incomplete. We evaluated Cerebra using a massive multi-institutional dataset spanning 3 million patients from four independent healthcare systems. Cerebra consistently outperformed both state-of-the-art single-modality models and large multimodal language model baselines. In dementia risk prediction, it achieved AUROCs up to 0.80, compared with 0.74 for the strongest single-modality model and 0.68 for language model baselines. For dementia diagnosis, it achieved an AUROC of 0.86, and for survival prediction, a C-index of 0.81. In a reader study with experienced physicians, Cerebra significantly improved expert performance, increasing accuracy by 17.5 percentage points in prospective dementia risk estimation. These results demonstrate Cerebra's potential for interpretable, robust decision support in clinical care.
Sheng Liu, Long Chen, Zeyun Zhao +16
Department of Biomedical Data Science, Stanford University · Department of Radiation Oncology, Stanford University · Center for Data Science, New York University +13
Delirium is a common and serious complication in the Intensive Care Unit (ICU), associated with increased morbidity, prolonged hospital stays, and higher healthcare costs. Despite its prevalence, early prediction and prevention remain challenging. Environmental factors such as ambient sound and light may influence the onset of delirium, yet they are often overlooked in risk assessments. In this study, we examined whether light intensity and sound pressure levels can independently predict delirium across multiple prediction horizons. We evaluated four efficient sequential neural network models on data collected from 9 ICUs across 309 patients to predict delirium for 10 prediction-window sizes. We reported feature importance and direction of influence using Shapley Additive Explanations analysis. The convolutional model achieved the strongest discrimination, with AUC = 0.80 on sound data and on combined data. Sound features were the dominant predictors overall. Integrating sound with light improved short-term (<1 week) prediction, with the combined model assigning the highest risk immediately after the sensing period. These findings suggest that passive ambient sensing, especially sound, can add a clinically meaningful, interpretable signal for delirium risk estimation and offer a practical pathway to enrich multimodal ICU prediction and prevention strategies.
Jiaqing Zhang, Sabyasachi Bandyopadhyay, Miguel Contreras +8
Department of Electrical and Computer Engineering, University of Florida, Gainesville, United States · Department of Medicine, Stanford University, Stanford, United States · Department of Biomedical Engineering, University of Florida, Gainesville, United States +1