🤖 machine learning

Exposure-Normalized Bed and Chair Fall Rates via Continuous AI Monitoring

This retrospective cohort study utilizing continuous AI monitoring suggests that exposure-normalized fall rates are higher in chairs than in beds, with specific footrest-positioning failures identified as a key risk factor, thereby supporting the hypothesis that optimizing chair safety rather than reducing chair usage may be a more effective fall prevention strategy.

Original authors: Paolo Gabriel, Peter Rehani, Zack Drumm, Tyler Troy, Tiffany Wyatt, Narinder Singh

Published 2026-03-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Paolo Gabriel, Peter Rehani, Zack Drumm, Tyler Troy, Tiffany Wyatt, Narinder Singh

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to figure out why people are getting hurt in a busy hospital. For years, doctors and safety officers have been using a very rough map to track these accidents. They count how many people fall per day a patient spends in a bed.

Think of this like trying to figure out why cars crash by only counting how many miles they drove on the highway, while completely ignoring the miles they spent driving through a crowded, bumpy city street. If a car crashes in the city, you still blame the "highway miles." This makes the city look safer than it actually is, and it hides the real danger spots.

This new study, led by Paolo Gabriel and his team, decided to throw away that old, blurry map. Instead, they used a high-tech, AI-powered security camera system that acts like a super-observant nurse who never blinks. This AI watches patients 24/7, knowing exactly when they are in bed, when they are sitting in a chair, and when they are walking around.

Here is the simple breakdown of what they found:

1. The "Exposure" Switch

The researchers realized that to understand danger, you have to measure the time spent in a specific spot, not just the number of days a patient is in the hospital.

The Old Way: "We had 100 falls this month." (But we don't know if they happened in bed or a chair).
The New Way: "We had 100 falls, but let's look at the hours. For every 1,000 hours patients sat in a chair, 17.8 people fell. For every 1,000 hours they lay in bed, only 4.3 people fell."

The Analogy: Imagine you are testing two types of shoes.

Shoe A (The Bed): You wear it for 10 hours a day. You trip once.
Shoe B (The Chair): You wear it for 2 hours a day. You trip three times.
If you just count "trips per day," Shoe B looks dangerous. But if you count "trips per hour worn," Shoe B is actually four times more dangerous than Shoe A. That is exactly what this study found: Sitting in a chair is a much higher-risk activity than lying in bed, hour-for-hour.

2. The "Footrest" Mystery

The team didn't just count falls; they looked at how they happened. They found a very specific pattern with the chair falls.

The Metaphor: Think of a chair as a boat. If the boat is rocking, you need to hold on tight.
The Finding: In almost every case where a patient fell directly out of a chair, it was because their feet weren't supported. It's like trying to sit on a wobbly stool with your legs dangling in the air. The footrest was either missing, broken, or not pushed in.
The Lesson: It's not that patients shouldn't sit in chairs (sitting is good for them!). The problem is that the chairs aren't set up correctly. It's a "setup failure," not a "use failure."

3. The AI Detective Work

The AI didn't just say "Fall!" It acted like a detective. It tracked the exact second a patient stood up, the second they left the chair, and the second they hit the floor.

They found that when a patient leaves a chair, there is a tiny, dangerous "gap" of about 50 seconds before they fall. This is the "transition zone"—the moment they are moving from sitting to standing or walking.
The AI also noticed that sometimes patients fall in the room after leaving the chair. If you count those, the danger of the chair is even higher.

4. The "Hypothesis" Warning

The authors are very careful not to say, "We proved chairs are dangerous." They say, "We found a very strong clue."

Why? Because the study was done in just one hospital system, and the number of falls was relatively small.
The Analogy: It's like seeing a pattern in a few rainstorms and guessing that it always rains on Tuesdays. You need to watch more storms to be sure.
However, the pattern is so clear (17.8 vs 4.3) and the reason so logical (feet dangling) that it gives hospitals a very strong reason to test new safety protocols.

The Big Takeaway

The goal of this study is not to tell patients, "Don't sit in chairs!" (That would be bad, as sitting helps prevent other health issues).

The goal is to say: "If you are going to sit in a chair, make sure the footrest is locked, the legs are supported, and the patient is watched closely for the first minute after they stand up."

By using AI to look at the hours spent in a chair rather than just the days in a hospital, the researchers found a hidden danger zone. They are essentially handing hospitals a new, more accurate map that says: "The chair is a high-risk zone, but if you fix the footrests and the workflow, you can make it safe."

1. Problem Statement

Inpatient falls are a major preventable adverse event in acute care. Current fall metrics typically use occupied bed-days as the denominator. This approach creates a measurement gap because it aggregates all patient time (bed, chair, transfers, ambulation) into a single bucket. Consequently, falls occurring from chairs are counted against bed-time exposure, which masks position-specific hazards and dilutes actionable signals regarding chair safety and supervision.

The study addresses this by proposing a shift from "bed-day" denominators to position-specific exposure time (hours spent in bed vs. hours spent in a chair) to accurately quantify fall risks associated with specific patient postures.

2. Methodology

Study Design and Data Source

Type: Retrospective observational cohort study.
Setting: Single regional health system (11 hospital partners across 3 US states).
Period: August 2024 – December 2025.
Data Source: Continuous AI monitoring platform processing room-level video feeds. The system assigns probabilistic position estimates (bed, chair, ambulatory) at sub-minute resolution for high fall-risk patients.
Cohort: 3,980 eligible monitoring units (out of 5,531 total) provided 292,914 hourly rows of data. Eligibility required $\ge$ 4 observed hours and $\ge$ 95% position coverage.

Exposure Measurement

Denominator: Patient exposure was calculated in fractional person-hours based on position.
- Chair Exposure: Sum of pct_chair / 100 across all eligible hours.
- Bed Exposure: Sum of pct_bed / 100 across all eligible hours.
Total Exposure: 320.51 chair-hours and 5,121.42 bed-hours.

Fall Ascertainment

Event Identification: 43 adjudicated fall events were matched to the monitoring pipeline; 40 were linked to eligible exposure hours for the primary analysis.
Labeling: Events were assigned "hard labels" (chair, bed, room, no_patient) based on AI detection at the alarm time. For modeling, a probabilistic approach was used, allocating fall risk mass across position strata based on pre-fall probabilities.

Statistical Analysis

Primary Model: Poisson Generalized Linear Model (GLM) with a log link and an offset term of log(exposure_hours).
Outcome: Count of proportionally allocated fall events.
Covariates: Patient position (Chair vs. Bed), 7 time-of-day windows, day of week, calendar quarter, and division identifier.
Robustness: Used heteroskedasticity-consistent (HC3) standard errors. Overdispersion was checked; if $>1.5$ , a Negative Binomial model was considered (though not triggered here).
Sensitivity Analyses: Included time-window restrictions, position-certainty filters, misclassification scenarios (swapping chair/bed labels), and furniture-origin reclassification.

3. Key Contributions

Denominator Refinement: The study introduces a "denominator-first" approach using continuous AI to separate chair and bed exposure, moving beyond the traditional "occupied bed-day" metric.
Exposure-Normalized Rates: It provides the first probability-weighted fall rates specifically normalized by hours spent in a chair versus hours spent in a bed.
Mechanism Insight: Through a broader observation cohort ( $n=32$ ), the study identifies that footrest/positioning failures are the dominant mechanism for direct chair falls (6 out of 7 cases), suggesting a specific, modifiable safety target.
Validation of AI Pipelines: The paper details the performance of the AI monitoring system (including label quality metrics and LLM benchmarks) and explicitly addresses limitations in position classification accuracy (Macro F1 = 0.528).

4. Key Results

Descriptive Fall Rates

Chair: 17.8 falls per 1,000 chair-exposure-hours.
Bed: 4.3 falls per 1,000 bed-exposure-hours.
Observation: The descriptive rate for chairs is more than 4 times higher than for beds.

Adjusted Relative Risk (RR)

Primary Finding: The adjusted Rate Ratio (Chair vs. Bed) was 2.35 (95% CI: 0.87 to 6.33; $p=0.0907$ ).
Interpretation: While the point estimate suggests a 2.35-fold higher risk in chairs, the confidence interval crosses 1.0, meaning the result is hypothesis-generating rather than statistically confirmatory at the 95% level.
Sensitivity:
- Under a division-clustered model (exploratory), the RR remained 2.35 with a narrower CI (1.89–2.92), though this was flagged as exploratory due to underdispersion.
- Misclassification sensitivity analyses showed that the effect size is highly sensitive to label assumptions (RR ranged from 4.49 to 13.17 in various swap scenarios).

Mechanism and Latency

Mechanism: In the observation cohort, 86% of direct chair falls were attributed to footrest/positioning failures.
Post-Departure Latency: For chair-origin events, the median time between leaving the chair and falling was 50 seconds (vs. 18 seconds for bed-origin events), identifying a critical transition window for intervention.

5. Significance and Limitations

Significance

Clinical Implication: The findings do not support restricting chair use (which is vital for mobility and delirium prevention). Instead, they support safer chair setups (e.g., verifying footrests, leg support, and call-light accessibility) and targeted supervision during the transition period after a patient leaves a chair.
Methodological Advance: Demonstrates the feasibility of using continuous AI to generate granular, position-specific safety metrics that traditional incident reporting cannot capture.

Limitations

Observational Nature: Causal inference is limited due to unmeasured confounders (e.g., patient acuity, mobility status, staffing ratios). Patients selected for chairs may inherently have higher fall risks.
Label Uncertainty: The AI position classifier has a Macro F1 of 0.528 and an Expected Calibration Error (ECE) of 0.450. This calibration error introduces bias in the probabilistic allocation of falls to chair vs. bed.
Sample Size: The inferential base is modest (40 events), leading to wide confidence intervals and limited power for subgroup analyses (e.g., by time of day).
Generalizability: Results are from a single health system; transportability to other settings with different workflows or chair hardware is unknown.

Conclusion

The study successfully demonstrates that chair-seated time carries a significantly higher descriptive fall rate than bed time when normalized by exposure hours. While the adjusted relative risk (2.35) is not statistically significant at $p<0.05$ due to sample size and uncertainty, the data strongly supports the hypothesis that chair setup and transfer workflows are critical, modifiable risk factors. The authors conclude that future work should focus on prospective validation with improved AI calibration and larger datasets to confirm these findings and guide policy.