📄 other

Population-Scale Functional Fall-Risk Indicator Assessment in Medicare Annual Wellness Visits Using Conversational AI

This study demonstrates that integrating conversational AI into Medicare Annual Wellness Visits enables scalable, multi-domain functional fall-risk screening that effectively identifies a substantial "hidden burden" of high fall risk among older adults who do not self-report a history of falls.

Original authors: Martha L. Buckley, Meenesh Bhimani, Avik Ray, Herprit Mahal, Michelle Voisard, Jonathan D. Agnew

Published 2026-07-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Martha L. Buckley, Meenesh Bhimani, Avik Ray, Herprit Mahal, Michelle Voisard, Jonathan D. Agnew

Original paper licensed under CC BY 4.0 (https://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 walking through a crowded, bustling city. Most people are walking fine, but some are stumbling, some are leaning on walls for support, and a few have already taken a nasty spill. In the world of medicine, there is a big concern about older adults falling. Falls are like a silent thief; they can steal a person's independence, cause serious injuries, and even lead to death. For a long time, doctors have tried to catch these "thieves" by asking a simple question: "Have you fallen in the last year?" It's like a security guard checking a list of people who have already tripped over a curb.

But here is the problem: people are terrible at remembering or admitting when they stumble. They might think, "Oh, I just lost my balance for a second, that doesn't count," or they might be scared that admitting they fell means they'll lose their driver's license or have to move to a nursing home. So, the security guard's list is often empty, even though people are still wobbling. This is where a new kind of detective comes in: Conversational AI. Think of this as a friendly, super-smart robot on the phone that doesn't ask, "Did you fall?" Instead, it asks, "Do you need to push off the armrest to stand up?" or "Do you hold onto the wall when you walk?" It's looking for the signs of a wobble, not just the memory of a crash. This paper explores whether this robot detective can find the hidden wobbles that the old list missed.

The Detective's New Tool

In this study, researchers teamed up with a large health system in Ohio to test out this robot detective. They called 511 older adults living in their own homes as part of their regular annual health check-up. Instead of a human nurse asking questions, a conversational AI (a voice-based computer program) chatted with them for about 16 minutes. The robot asked seven specific questions designed to spot "functional fall-risk indicators." These aren't about falling; they are about how the body is working. For example, it asked if they needed to use their hands to stand up from a chair, if they felt unsteady walking, or if they worried about falling.

The researchers were looking for a specific group of people: those who said, "No, I haven't fallen in the last year," but who the robot's questions revealed were actually struggling with their balance or strength. They called this the "hidden risk" population. If the robot found that someone had two or more of these trouble signs, they were flagged as having a "high fall-risk indicator burden," even if they claimed they never fell.

What the Robot Found

The results were like finding a secret layer of the city map that nobody knew existed. Out of the 378 people who said, "I haven't fallen," the robot discovered that nearly one-third of them (32.3%) actually had two or more signs of high fall risk. These are the people who might be holding onto furniture to walk or feeling shaky on their feet, but who never thought to tell a doctor they were in trouble because they hadn't actually hit the ground yet.

On the flip side, among the 133 people who admitted, "Yes, I fell," the robot confirmed that 81.2% of them had these high-risk signs. This was a good sign for the robot; it proved that when people do fall, the robot's questions correctly identify that they are in the danger zone. The study found that people who reported falling were about nine times more likely to have these multiple risk signs than those who didn't report falling. Even after adjusting for age and gender, the link remained strong.

The "Pre-Fall" Mystery

The most exciting part of the paper is what it suggests about the "hidden" group. The authors propose that these people might be a "pre-fall" population. Imagine a bridge that is starting to crack. The bridge hasn't collapsed (fallen) yet, but the cracks (the risk signs) are there. The robot is spotting the cracks before the bridge falls. The study suggests that many older adults have these functional deficits—like weak legs or unsteady balance—long before they actually take a tumble. Because they haven't fallen yet, they don't report it, and traditional screening misses them entirely.

How Sure Are We?

The paper is very clear about what it knows and what it doesn't. It measured that the robot can successfully identify these hidden risk signs in a large group of people and that these signs match up with people who have already fallen. The math shows a very strong connection between the robot's findings and the reality of falling.

However, the paper also admits it hasn't proved that these people will definitely fall in the future. This was a snapshot in time (a cross-sectional study), not a movie played forward. The authors suggest that these people are likely at higher risk, but they need to do more research, following these people over time, to see if the robot's predictions come true. They also note that the robot relies on people answering honestly, just like a human would, and that the study was done in one specific area, so it might look different in other places.

The Takeaway

In simple terms, this paper shows that asking "Did you fall?" isn't enough because people forget or hide the truth. But if you ask "How does your body feel when you move?" using a friendly, automated phone call, you can find a huge group of older adults who are wobbling on the edge of a fall but haven't fallen yet. The robot didn't solve the problem of falling, but it found a new, scalable way to spot the people who need help before they hit the ground. It's like giving the city a pair of night-vision goggles to see the cracks in the bridge before the collapse happens.

Technical Summary: Population-Scale Functional Fall-Risk Indicator Assessment in Medicare Annual Wellness Visits Using Conversational AI

Problem Statement

Falls are a leading cause of injury, disability, and mortality among adults aged 65 and older. Current screening protocols, such as the CDC's STEADI initiative and Medicare Annual Wellness Visits (AWV), rely heavily on patient self-report of fall history (i.e., "Have you fallen in the past 12 months?"). This approach suffers from significant limitations:

Underreporting: Retrospective self-report sensitivity ranges from 56% to 89%, meaning 11–44% of falls go unreported due to recall bias, normalization of instability, fear of losing independence, or cognitive impairment.
Missed "Pre-Fall" Vulnerability: Functional deficits (e.g., gait instability, balance issues, lower extremity weakness) often precede actual fall events. Relying solely on fall history fails to identify individuals who are functionally vulnerable but have not yet fallen or do not recall falling.
Scalability Constraints: Administering multi-domain functional assessments at a population scale via human staff is labor-intensive, difficult to standardize, and often infeasible within time-constrained clinical workflows.

Methodology

This cross-sectional observational study analyzed data from an AI-powered, telephone-based health risk assessment deployed within OhioHealth's Medicare AWV program.

Participants: The final analytic sample included 511 community-dwelling older adults (mean age 75.6 years; 59.5% female) who completed the assessment with valid responses for fall history and seven functional indicators.
Intervention (Conversational AI): A conversational AI agent conducted structured, natural-language telephone interviews (15–20 minutes). Unlike static questionnaires, the AI used follow-up prompts to clarify responses.
Measures:
- Independent Variable: Self-reported fall history (Fall History [FH] vs. No Fall History [NFH]).
- Dependent Variable: A composite Fall-Risk Indicator Score based on seven binary, behaviorally anchored functional indicators:
  1. Needing hands to stand from a chair (Lower extremity strength).
  2. Holding furniture/walls while walking (Balance/Mobility).
  3. Feeling unsteady when walking (Gait stability).
  4. Trouble stepping onto a curb (Functional mobility).
  5. Worry about falling (Fear of falling).
  6. Loss of feeling in feet (Peripheral neuropathy).
  7. Medication-induced lightheadedness (Medication side effects).
- Risk Classification: "High fall-risk indicator burden" was defined as endorsing ≥2 indicators.
Statistical Analysis: The study employed chi-square tests, unadjusted and adjusted logistic regression (controlling for age and gender), and sensitivity analyses using alternative thresholds (≥1 and ≥3 indicators) to assess the association between self-reported fall history and AI-detected functional burden.

Key Results

Hidden Risk in Non-Fallers: Among the 378 participants who reported no falls in the past year (NFH), 32.3% (n=122) met the criteria for high fall-risk indicator burden (≥2 indicators).
Validation in Fallers: Among the 133 participants who reported falls (FH), 81.2% (n=108) met the criteria for high indicator burden, confirming the clinical coherence of the AI-detected deficits.
Statistical Association: Self-reported fall history was strongly associated with high AI-derived indicator burden ( $\chi^2 = 95.16, p < .001$ $χ^{2} = 95.16, p < .001$ ).
- Unadjusted Odds Ratio (OR): 9.07 (95% CI: 5.58–14.73).
- Adjusted OR (for age and gender): 8.65 (95% CI: 5.26–14.21).
Sensitivity Analysis: Even under a more stringent threshold of ≥3 indicators, 18.3% of those denying falls still demonstrated substantial multi-domain burden, compared to 58.6% of those reporting falls.
Convergent Validity: The AI-derived indicator burden showed strong correlation with functional decline ( $\phi = .698$ ) and self-rated health ( $\phi = .298$ ).

Key Contributions

Identification of "Hidden" Risk: The study quantifies a substantial population of older adults who deny falling yet exhibit multiple functional risk indicators. This suggests that reliance on fall history alone misses nearly one-third of the population with elevated functional vulnerability.
Scalable Functional Screening: The research demonstrates that conversational AI can successfully administer complex, multi-domain functional assessments via telephone at a population scale, bypassing the logistical barriers of in-person staff-led screening.
Behaviorally Anchored Assessment: By focusing on specific functional behaviors (e.g., "Do you need hands to stand?") rather than fall events, the AI approach potentially circumvents the psychological and cognitive barriers that lead to fall underreporting.
Triage Utility: The study proposes that AI-generated indicator profiles can serve as a triage signal to prioritize follow-up assessments and prevention services (e.g., home safety, medication review) for high-risk individuals who would otherwise be classified as low-risk.

Significance and Claims

The paper claims that conversational AI offers a pragmatic solution to the limitations of current fall-risk screening. By integrating behaviorally anchored functional indicators into automated telephone outreach, health systems can:

Detect Vulnerability Early: Identify "pre-fall" individuals who have not yet experienced a reported fall but possess the functional deficits that predict future risk.
Standardize Population Health: Deliver consistent, validated screening items to large Medicare panels without the need for specialized equipment or extensive staff training.
Optimize Clinical Resources: Use AI-derived risk scores to triage patients, ensuring that clinician time is reserved for higher-touch interventions for those identified as high-risk by the AI system.

Limitations Noted by Authors:
The authors explicitly state that the study is cross-sectional and does not establish prospective prediction of incident falls. The composite score has not been independently validated against future fall outcomes. Additionally, the sample was drawn from a single Midwestern health system, and the functional indicators, while behaviorally anchored, remain self-reported rather than objectively measured via performance testing. The authors conclude that prospective validation is necessary to confirm the predictive utility of this approach.