AI and Falls Detection: Great Promise but Gaps in Evidence

Introduction

Falls by hospitalized patients are a significant cause of morbidity and mortality within the US health care system. The Agency for Healthcare Research and Quality (AHRQ) estimates that 3 to 5 patients out of every 1000 succumb to a fall. These costs are measured in both patient harm and in wasted dollars for health systems, estimated to approach $50 billion per year. ¹ AHRQ has created a tool kit that is used by many hospitals and has prevented falls. ^2,3 And yet, recent research has shown standard fall prevention measures with technology assistance have yielded mixed results. ^{4 –6} And so, falls remain a significant patient safety issue.

Recently, artificial intelligence (AI) has demonstrated numerous benefits in identifying patients at risk of falls. ⁷ There are typically 3 large categories of use: (i) wearable sensors (gyroscopes, belt clips), (ii) environmental sensors (pressure, acoustic, inactivity devices), and more recently popular (iii) camera systems. ⁸ In recent years, health systems across the world have spent $380 million annually on technologies to reduce falls. ⁸ The use of camera systems and AI, or “computer vision” (CV), is a popular approach. CV offers the promise to reduce the costs associated with the use of staff or “sitters” to monitor patients for falls. There are several non-peer-reviewed, vendor-sponsored case studies, or “white papers,” that testify to this. ⁷

Because the Centers for Medicare and Medicaid Services does not reimburse hospitals for additional costs associated with falls, hospitals have a financial incentive to reduce falls. These costs can be significant for hospitals, and reducing even 1 fall can result in saving thousands of dollars. ¹ Thus, an AI system that could identify patients at risk of a fall and then alert team members would be welcomed.

Yet to reduce harm and costs, AI must accurately differentiate between “true” and “false” alerts. When AI incorrectly identifies patients at risk for falls, caregivers waste time responding, detracting care from other patients, and increasing the risk of harm. This is especially important given recent concern that high quality research is lacking for many use cases of AI and hospital-based CV. ⁷

We believe it is imperative to validate recent AI technologies by assessing their efficacy in clinical scenarios within hospital settings, supplemented by evidence from simulations. Evaluating the current literature is a first step to determine if there is evidence of the efficacy and feasibility of CV technology to identify fall threats in real time. The purpose of this article is to critique and highlight the peer-reviewed evidence (or lack thereof) regarding the effectiveness and impact of CV in monitoring and reducing falls.

Methodology

We performed a scoping review that adhered to the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines. A comprehensive search was conducted across multiple databases, including PubMed, EMBASE, and Scopus. The search was complemented by a manual search to ensure comprehensive coverage. The search strategy integrated terms related to AI, fall prevention/detection, video recording/image processing, and care settings. Detailed search terms for each database are provided in the Supplementary Data.

Study selection

The search yielded a total of 110 records: 56 from EMBASE, 14 from Scopus, and 42 from PubMed. After removing duplicates, 105 unique records remained for screening. In addition, 79 papers were identified through manual searches, bringing the total number of records screened to 184. Two independent reviewers assessed the titles and abstracts of these records using predefined inclusion and exclusion criteria. Discrepancies were resolved through discussion or, if necessary, consultation with a third reviewer.

Studies were eligible for inclusion if they utilized AI, machine learning, or deep learning for fall prevention or detection in real-time or hospital settings, specifically focusing on video recording or computer-assisted image processing as part of the fall detection technology. Studies needed to involve live patient monitoring systems within hospital environments. Studies were excluded if they were theoretical (e.g., simulations without real-world implementation), focused on wearable devices or other sensor-based approaches not involving video/camera systems, applied AI models without physical measurements relevant to fall detection, or explored other unrelated technologies such as IoT, audio sensors, or robotics.

After the initial screening, 144 records were excluded, leaving 40 video-based articles for full-text review to assess eligibility. Following a detailed evaluation, all full-text articles were ultimately excluded, with reasons for exclusion documented. A flow diagram showing the exclusion and inclusion of studies is included (Fig. 1).

OPEN IN VIEWER

FIG. 1.

Flow Diagram.Source: Page MJ, et al. BMJ 2021;372:n71. doi: 10.1136/bmj.n71. This work is licensed under CC BY 4.0. To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/.

Results

Overall, existing studies were limited in scope, often lacking robust methodologies or large sample sizes. There were no randomized controlled trials (RCTs) in real-life settings. In addition, nearly all of the studies reviewed nonhospital-based systems that were focused on elder populations, dementia populations, or were exclusively in a “test,” nonreal environment. In addition, no studies evaluated the effectiveness of AI in detecting falls based on a CV technology in a hospital setting. Studies that did purport to show evidence of CV-based fall reduction were either in test cases, exclusively ambulatory settings, did not provide control groups, or did not provide data analysis, or comparison between groups. The absence of comprehensive, real-world, peer-reviewed evidence in one of the costliest settings for falls (i.e., hospital) hampers the establishment of AI use best practices. This underscores the urgent need for rigorous research to substantiate the claims surrounding benefits in improving patient safety in a hospital setting.

Discussion

Although health systems are spending millions of dollars on implementing AI technologies to prevent falls, the extant literature lacks evidence for efficacy or impact in hospitalized patients. No study included a control group. Statements of impact lacked methodological rigor and discipline to support their inferences. High-quality studies are essential to assess the impact of these innovations on patient outcomes comprehensively. A robust research framework should include RCTs that measure specific outcomes, such as patient safety incidents, change in fall rates, and overall patient and provider satisfaction, including usability scoring. This is something lacking either partially or in whole in current literature that aims to evaluate the effectiveness of CV in identifying and intervening on fall risks.

There are multiple frameworks regarding how to evaluate AI. ^9,10 In a high-reliability organization, a quality study would involve multidisciplinary collaboration, ensuring that all relevant stakeholders, including clinicians, data scientists, and patient representatives, contribute to the study design. In addition, such a study would require rigorous methodologies, including clear definitions of metrics, control groups, and long-term follow-up to assess sustainability. The results would ideally be subject to peer review to enhance credibility and facilitate knowledge dissemination within the broader medical community. We hope more health systems robustly evaluate AI programs so that we can learn whether, how, and why these technologies work before more resources are committed.

Conclusion

The integration of AI and current fall prevention strategies presents exciting opportunities for enhancing patient care. Health care systems have expended significant resources on AI technologies. Yet the extant literature lacks evidence for their effectiveness and impact, limiting the ability to make inferences about whether these technologies are effective and to make informed decisions to invest in these technologies. Further research is needed to address existing barriers and validate the effectiveness of these technologies in real-life hospital clinical areas. Future studies should focus on comprehensive evaluations of patient outcomes in different inpatient settings.