Introduction

Inpatient falls represent one of the most prevalent and costly preventable adverse events in acute care. Contemporary U.S. and international reports continue to show substantial event burden and injury risk across inpatient settings. Falls prevention remains a national accreditation and patient-safety priority, with The Joint Commission and national safety organizations continuing to emphasize this domain as a persistent quality challenge. Global guidance similarly identifies hospital falls prevention as an ongoing systems-level priority.

A central measurement gap in the inpatient falls literature is denominator choice. Most hospital fall metrics are reported per 1,000 occupied bed-days — a denominator that merges bed time, chair time, transfers, and room movement into a single exposure bucket. When a patient falls from a chair, the event is still counted against a bed-day denominator, which can mask position-specific hazard and dilute clinically actionable signal about chair setup and supervision.

Continuous AI-based patient monitoring systems create an opportunity to replace that blended denominator with position-specific exposure time. By assigning probabilistic chair and bed positions at sub-minute resolution, these systems can estimate how many falls occur per hour of chair-seated time versus per hour of bed-bound time within the same monitored population. This denominator-first approach aligns with standard rate modeling and supports adjusted comparisons that routine incident-reporting systems cannot produce.

This retrospective cohort analysis uses continuous AI monitoring data to estimate position-specific exposure-normalized fall rates and then test whether the observed descriptive separation (17.8 vs 4.3 falls per 1,000 exposure-hours) persists after adjustment in a Poisson rate model. The aim is not to argue for less chair use or to promote bed confinement, given the recognized harms of low mobility in hospitalized older adults. Rather, the aim is to determine whether chair-positioned time concentrates fall opportunity enough to justify safer chair setup, clearer transfer workflows, and prospective denominator-aware monitoring studies that integrate mobility promotion with targeted fall prevention.

Methods

Study Design and Setting

Retrospective observational cohort study using continuous AI monitoring data from a single regional health system, analyzed at the division level. Study period: August 2024–December 2025. Conducted under an existing data use agreement between LookDeep Health and the health system as a secondary analysis of de-identified monitoring records generated during routine clinical operations.

Data Source

The AI monitoring platform continuously processes room-level video feeds and assigns probabilistic position estimates for each monitored patient. Per-hour position fractions (pct_chair, pct_bed, pct_ambulatory) were extracted from the hourly monitoring cache using a validated study-window filter and percentage-sum constraint. The extraction yielded 356,391 hourly rows with a valid position-percentage sum of 100% across all rows. Fall events were identified from AI-detected alarm records and linked to the hourly exposure structure via monitor and date-hour keys.

Cohort Definition and Eligibility

A monitoring unit was defined as a unique monitor for cohort membership. Units were classified as intervention-type if the monitor appeared in the extracted fall-event source during the run, and as control-type if it appeared in the hourly exposure source without a linked extracted fall event. Eligibility gates: at least 4 observed monitor-hours (min_observed_hours=4) and a position-coverage ratio of at least 0.95 (min_coverage_ratio=0.95). Of 5,531 units in the hourly data, 3,980 passed both eligibility criteria (intervention: 42 eligible; control: 3,938 eligible). High fall-risk context was informed by standard clinical screening including the Hendrich II model.

Exposure Measurement

Exposure expressed in fractional person-hours by position. Chair exposure per eligible hourly row = pct_chair / 100 hours; bed exposure = pct_bed / 100 hours. Total exposure across all intervention-eligible unit-hours: 320.51 chair-hours and 5,121.42 bed-hours. Analysis base: 292,914 rows after eligibility filtering.

Fall Ascertainment

Within the study window (August 2024–December 2025), 43 fall events were identified from the adjudicated consensus record; 40 linked to eligible analysis-base hours and entered the primary exposure-normalized analysis. A broader monitoring feed (2022–2026, n=91 deduplicated events) was retained for descriptive context only. Hard labels at alarm time: chair (n=5), bed (n=23), room/no_patient (n=12). Adjusted modeling used probabilistic rather than hard-label event allocation.

Auxiliary Annotation Cohorts

Four non-interchangeable cohorts support different descriptive tasks: (1) the study-window adjudicated cohort (n=43); (2) the 40-event inferential base; (3) the broader observation cohort (32 deduplicated events, 37 source rows) for mechanism coding; and (4) the departure-aware subset (30 truth rows, 31 scored sequences) for exit-concordance diagnostics. Only the 40-event inferential base contributes to the primary RR estimate.

Statistical Analysis

Unadjusted rates calculated as events per 1,000 exposure-hours. Adjusted analyses used a Poisson GLM with log-link and offset log(exposure_hours). Covariates: patient position (chair vs. bed; reference = bed), seven local-time windows (00:00–05:59 through 21:00–23:59), day of week, calendar quarter, division identifier. Adjusted RR = exp(β_chair); 95% CIs via Wald intervals on the log-RR scale. Overdispersion assessed via deviance/df and Pearson χ²/df. Pre-specified negative-binomial fallback at threshold >1.5 (not triggered; deviance/df = 0.13). Robustness: HC3 heteroskedasticity-consistent SEs and division-level cluster-robust covariance. All analyses in Python (statsmodels).

Sensitivity Analyses

Pre-specified sensitivity analyses: seven local-time window restrictions; weekday-only and weekend-only restrictions; position-certainty hours only (chair or bed fraction ≥0.50); five misclassification scenarios (symmetric 10%/20%/30% chair–bed swaps; one-sided 20% swaps); eligibility thresholds at 4, 12, 24, and 48 hours; furniture-origin reclassification of chair-origin room falls.

Furniture-Exit Detection Concordance

Two AI departure signals were evaluated against consensus-annotated furniture_departure_time using a ±5-minute search window: (1) nudge boundary crossing (first green→yellow/red transition in the operational nudge_state indicator); (2) location-label transition (first change in dominant_location_label away from the patient's known last-furniture position). Concordance quantified as bias (AI minus consensus), mean absolute error (MAE), and percentile-based distributions.

Label Evaluation

Automated label quality evaluated against the departure-aware benchmark subset (30 truth rows, 31 scored sequences). Metrics: macro F1, ECE (10-bin), detection F1, detection precision, detection recall, latency MAE.

Ancillary Multimodal LLM Benchmark

Three multimodal LLMs (Gemini 2.5 Flash, Gemini 2.5 Pro, Gemini 3.1 Pro Preview) independently processed de-identified adjudicated event clips (81 clips; v2 consensus package). Primary comparison: 3-way overlap subset (n=42 visible primary rows). Metrics: fall sensitivity, non-fall specificity, pre-fall location accuracy, fall-time MAE. Reported as ancillary validation only.

Ethics and Data Governance

Data collected from patients admitted to one of eleven hospital partners across three U.S. states. The study followed CHAI standards under a Business Associate Agreement. Patients provided written informed consent for monitoring as part of standard inpatient care (HIPAA-compliant). All video data blurred prior to storage; no identifiable information included. Face-blurred frames used only for training purposes. The outcomes of this analysis did not influence patient care or clinical outcomes.

Results

Using position-specific exposure hours rather than occupied bed-days as the denominator, chair time showed a higher descriptive fall rate than bed time (17.8 vs. 4.3 falls per 1,000 exposure-hours). Within the study window, 43 adjudicated events matched the monitoring pipeline and 40 linked to eligible analysis-base hours for adjusted modeling, yielding a primary chair-vs-bed RR of 2.35 (95% CI 0.87–6.33; p=0.0907). In a separate broader observation cohort (n=32), 6 of 7 direct chair falls carried a footrest/positioning tag.

Descriptive Fall Rates by Position (Unadjusted)

Among intervention-eligible units, the probability-weighted descriptive chair fall rate was 17.8 per 1,000 exposure-hours (5.69 expected falls / 320.51 chair-hours) and the bed rate was 4.3 per 1,000 exposure-hours (22.05 expected falls / 5,121.42 bed-hours).

Expected rates are probability-weighted; hard-label counts reflect AI-assigned position at event time. Analysis restricted to intervention-eligible units (n=42).

Primary Adjusted Relative Risk

The adjusted Poisson GLM executed on the 40-event inferential base. The primary adjusted chair-vs-bed RR was 2.35 (95% CI 0.87–6.33; p=0.0907). HC3-robust: RR 2.35 (95% CI 0.93–5.94; p=0.0709). Division-clustered (exploratory): RR 2.35 (95% CI 1.89–2.92; p<0.0001) — 9 intervention divisions only. The model showed underdispersion (deviance/df=0.13; Pearson χ²/df=0.49), so the pre-specified negative-binomial fallback was not triggered. Position-certainty restriction: non-estimable.

NE = non-estimable. Misclassification scenarios (estimable): RR 4.49–13.17 across symmetric 10/20/30% and one-sided swaps. Effect magnitude highly sensitive to label error assumptions. The clustered row is exploratory.

Mechanism Taxonomy (Observation Cohort)

Among 32 deduplicated fall events in the observation cohort: 7 direct chair falls, 12 direct bed falls, 13 room falls. Of the 7 direct chair falls, 6 (86%) carried a footrest/positioning tag. Transfer failure was the dominant mechanism for direct bed falls (6/12, 50%).

Convenience sample for mechanism characterization only; not a population prevalence estimate.

Furniture-Origin Chain Analysis

Including the 3 chair-origin room events with the 7 direct chair falls yields 10 chair-associated events. Post-departure latency: bed-origin events median 18 s (IQR 6–69; max 136; n=10); chair-origin events median 50 s (max 150; n=3).

Label Evaluation Metrics

Assessed against v3 adjudicated benchmark subset (30 truth rows, 31 scored sequences): macro F1 = 0.528; ECE (10-bin) = 0.450; detection F1 = 0.846; detection precision = 1.000; detection recall = 0.733; latency MAE = 37.9 s (p50=22s; p90=83s). Perfect detection precision (no false positive fall alarms), but ~1-in-4 falls missed at detection level.

Ancillary Multimodal LLM Benchmark

Ancillary validation only. Different denominator structure from primary livestream pipeline; direct ranking not appropriate.

Discussion

Magnitude of Chair-Seated Risk

The main contribution of this analysis is denominator refinement rather than a definitive effect estimate. Once chair and bed time were expressed as separate exposure-hours, the descriptive rate contrast was substantial (17.8 vs 4.3 per 1,000 exposure-hours), and the adjusted RR remained above 1.0 at 2.35. However, the primary confidence interval crossed 1.0 (95% CI 0.87–6.33; p=0.0907), so the complete experiment should be interpreted as an elevated but uncertain signal rather than a confirmed multi-fold effect. The inferential event base remains modest (43 adjudicated study-window events; 40 eligible linked events).

Mechanism Insights

In the broader observation cohort, 6 of 7 direct chair falls carried a footrest/positioning tag. This concentration points to a consistent upstream problem: incomplete chair setup or unsupported lower-extremity position. Because this 32-event cohort is descriptive, broader than the study window, and not denominator-linked, the finding should be treated as signal generation for prospective testing rather than a population prevalence estimate.

Latent Bias Assessment

Position classification ambiguity. Scenario-based misclassification analyses remain the main quantitative guardrail. Across estimable scenarios, effect size changed materially (RR range 4.49–13.17), indicating sensitivity to label error assumptions even when direction is preserved.

Temporal confounding. Fall mass concentrates in waking and care-transition hours, with the highest mean chair probability in those same windows. The primary model adjusts for seven local-time windows, but sparse chair-event counts limit time-of-day clustering control. Residual temporal confounding would most likely inflate the observed chair-vs-bed RR.

Selection into position. No patient-level acuity, mobility, or staffing variables are available. Confounding by indication remains the most important unmeasured threat.

Bias audit summary. The primary interval (0.87–6.33) already allows for a substantially smaller effect than the point estimate suggests. The true causal effect could be materially attenuated or null once temporal and acuity confounders are better measured.

Post-Chair Risk Window and Furniture-Origin Analysis

Three room-classified events were preceded by chair departure, expanding chair-associated events from 7 to 10. Post-departure latency findings identify a short interval between leaving furniture and falling, supporting a clinically relevant transition window for future prospective monitoring work.

Ancillary LLM Benchmark

The Gemini models exhibited a sensitivity-specificity tradeoff consistent with the primary pipeline's label-quality findings: higher recall came at the cost of lower specificity and worse timing. No model achieved both high sensitivity and high location accuracy. The findings support the conclusion that prospective improvement in automated position classification is needed before label-derived metrics can support confirmatory inference.

Limitations

(1) Observational, single health system; no causal inference supported. (2) Label quality material constraint (macro F1=0.528; ECE=0.450; detection recall=0.733). (3) Single regional health system limits generalizability. (4) Injury outcomes not captured. (5) Observation mechanism cohort is a non-representative convenience sample. (6) Furniture-exit concordance non-estimable in this run due to instrumentation gap. (7) Small chair-fall event count (40 modeled events) produces wide CI.

Clinical Implications

Chair positioning remains clinically important for mobilization, delirium prevention, and avoidance of the functional harms of prolonged bed rest. The implication of these findings is safer chair use, not less chair use. The mechanism pattern supports targeted chair-setup confirmation before leaving chair-positioned patients unattended (footrest position, leg support, call-light accessibility) — framed as a QA hypothesis rather than a proven intervention. Stable adjusted estimates require prospective data with larger chair exposure, improved label calibration, and clinical confounders (acuity, mobility, staffing).

Conclusions

Position-specific exposure denominators identified higher descriptive fall rates during chair time than bed time (17.8 vs 4.3 probability-weighted falls per 1,000 exposure-hours), and adjusted modeling estimated a chair-vs-bed RR of 2.35 (95% CI 0.87–6.33). In a separate broader observation cohort, 6 of 7 direct chair falls involved footrest-positioning failures, pointing to a plausible modifiable prevention target. These findings remain hypothesis-generating and support safer chair use workflows, but require prospective validation before causal or policy-level conclusions.

Appendix

References