Skip to main content
NCBI home page
International Journal of Nursing Sciences logoLink to International Journal of Nursing Sciences
. 2025 Dec 12;13(1):3–10. doi: 10.1016/j.ijnss.2025.12.001

Image recognition-based detection system for preventing accidental dislodgement of head-and-neck medical supplies in ICU patients: A feasibility randomized controlled trial

Zhongjie Shi a,1, Taotao Shi b,1, Xin Gao a, Jian Li b, Hong Xu b, Xiaojun Li a, Zhanxiang Wang a, Sifang Chen a,⁎
PMCID: PMC12891787  PMID: 41684616

Abstract

Objectives

This study aimed to design and evaluate a detection system for the accidental dislodgement of head-and-neck medical supplies through hand position recognition and tracking in Intensive Care Unit (ICU) patients.

Methods

We conducted a single-center, prospective, parallel-group feasibility randomized controlled trial. We recruited 80 participants using convenience sampling from the ICU of a hospital in Ningbo City, Zhejiang Province, between March 2025 and June 2025, and they were randomly assigned to either the control group (routine care) or the intervention group (routine care plus image recognition-based detection system). The system continuously tracked patients’ hand positions via bedside cameras and generated real-time alarms when hands entered predefined risk zones, notifying on-duty nurses to enable early intervention. System stability was assessed by continuous system uptime; system performance and clinical feasibility were evaluated by the frequencies of risk actions and accidental dislodgement of medical supplies (ADMS).

Results

All 80 participants completed the intervention, with 40 patients in each group. The baseline characteristics and median observation time of the two groups were balanced (intervention group: 48 h/patient vs. control group: 49 h/patient). Compared with the control group, the intervention group showed fewer ADMS (2/40 vs. 9/40) and detected more risk actions per 100 h (36 vs. 25); all system-detected events had corroborating images with complete concordance on manual review, and all nurse-recorded hand-contact events were accurately captured.

Conclusions

The study demonstrated that the image recognition-based detection system can function stably in clinical settings, providing accurate and continuous surveillance while supporting the early detection of risk actions. By reducing the observation burden and offering real-time cognitive support, the system complements routine nursing care and serves as an additional safety measure in ICU practice. With further optimization and larger multicenter validation, this approach could have the potential to make a significant contribution to the development of smart ICUs and the broader digital transformation of nursing care.

Keywords: Accidental dislodgement of medical supplies, Feasibility randomized trial, Image recognition, Intensive Care Unit, Risk monitoring

What is known?

  • •

    Intensive Care Unit (ICU) patients often require various medical supplies. Because of impaired consciousness or agitation, they may inadvertently dislodge these supplies, leading to complications such as infection, bleeding, airway obstruction, or interruption of treatment.

  • •

    Current nursing strategies rely mainly on manual surveillance and the use of physical restraints. However, manual observation is labor-intensive and cannot ensure continuous 24-h monitoring, while physical restraints may cause discomfort, restrict mobility, and are not acceptable or appropriate for all patients.

What is new?

  • •

    In this study, an image recognition-based detection system was developed and evaluated for ICU patients. The system continuously tracks patients’ hand positions through bedside cameras and automatically triggers an alarm when the hand enters a predefined risk zone. This approach enables timely detection of risk actions, supporting early nursing interventions and reducing the likelihood of accidental dislodgement of medical supplies.

  • •

    In real-world ICU use, the system operated stably and detected more high-risk actions than routine nursing rounds. Image reviews showed high concordance with observed actions, and nurse-recorded events were consistently captured.

  • •

    Compared with routine care, the intervention was associated with fewer accidental dislodgements of medical supplies, indicating a clinically meaningful reduction that supports subsequent multicenter studies.

1. Introduction

Patients admitted to the Intensive Care Unit (ICU) are typically critically ill or dependent on life support and often require multiple medical supplies simultaneously to sustain respiration, circulation, nutrition, or drainage. These supplies are distributed across various regions of the body, including the head and neck, thoracoabdominal area, and extremities, and provide essential support for clinical treatment [[1], [2], [3], [4]]. However, due to discomfort, impaired consciousness, agitation, delirium, or poor compliance, patients may grasp, pull, or resist these supplies, leading to accidental dislodgement of medical supplies (ADMS) [5,6]. In this study, medical supplies are not limited to airway supplies but also involves the accidental dislodgement of nasogastric tubes, central venous catheters, thoracoabdominal drainage tubes, urinary catheters, and surgical dressings, making it a common safety concern in the ICU care. ADMS may lead not only to common complications associated with catheter displacement but also to severe adverse outcomes [7,8].

Currently, ADMS recognition primarily relies on physical restraints and bedside nursing rounds [9,10]. Bedside observation is the most common method; however, its frequency is limited by nursing staff resources, making continuous, real-time recognition difficult. This limitation is particularly evident during night shifts, when only a few nurses are on duty, and the risk is higher. Physical restraints may reduce the incidence of accidental dislodgement to some extent. However, their use increases discomfort and anxiety, and may lead to skin injury, restricted mobility, and ethical issues such as impaired communication between patients and caregivers [[11], [12], [13]]. With the rising number of ICU patients and the growing workload of nurses, this manual recognition model is becoming increasingly inadequate. Thus, there is an urgent need for an objective, continuous, and intelligent recognition approach to reduce ADMS, enhance patient safety, and improve the quality of care in the ICU.

In recent years, image recognition technology has advanced rapidly, demonstrating broad applications across multiple fields. In transportation, it enables license plate recognition and traffic monitoring. In security, it has been widely applied to facial recognition and action monitoring [[14], [15], [16]]. These practices demonstrate the robust capability of image recognition in object recognition and analysis under complex conditions, providing a strong foundation for its application in clinical care environments. Recent studies [[17], [18], [19]] have introduced image recognition into nursing practice, such as detecting patient falls, abnormal movements in bed, restraint behaviors, and nursing operations, thereby reducing adverse nursing events. These explorations suggest new opportunities for risk identification.

Given these advances, applying image recognition to identify patient actions that precede supplies dislodgement has become a feasible direction. In clinical practice, ADMS commonly follow a recognizable pattern—patients’ hands move toward the supply area. Because hands are typically exposed and easily observed, hand tracking provides a practical entry point for detecting risky actions. In contrast, supplies in the thoracoabdominal region or lower extremities are often covered by clothing or bedding, making visual recognition difficult [20,21]. Head-and-neck supplies (e.g., endotracheal tubes, nasogastric tubes, or cervical drains) are more exposed and anatomically stable, and reaching them usually requires larger upper-limb movements. These features make the head-and-neck region more suitable for early visual monitoring.

Based on these rationale and prior research, we designed an enhanced system that tracked the spatial relationship between patients’ hands and the areas of their supplies to identify risky actions in real-time. Meanwhile, we also evaluated its clinical feasibility in postoperative patients admitted to the ICU, aiming to enhance patient safety and provide a basis for extending the approach to other supplies.

2. Methods

2.1. Study design and participants

We conducted a single-center, prospective, parallel-group feasibility randomized controlled trial (RCT) (ChiCTR2500111520). The study was reported in accordance with the Consolidated Standards of Reporting Trials (CONSORT) statement and its extension for feasibility and pilot trials [22]. Patients were recruited using convenience sampling from the ICU of a hospital in Ningbo City, Zhejiang Province, from March 2025 to June 2025. It is important to note that patients with prolonged ICU stays or those primarily admitted for critical care typically receive extended sedation, which may introduce significant inter-individual variability and affect comparability. In contrast, our study specifically focused on patients with short-term ICU stays following surgery, ensuring greater consistency and comparability in our findings. Therefore, we established the following inclusion and exclusion criteria.

Eligibility criteria were: 1) age greater than 18 years; 2) temporary ICU admission after surgery for intensive monitoring, with the presence of at least one head or neck supplies (e.g., cranial drain, nasogastric tube, or endotracheal tube), regardless of the underlying disease; 3) considered at high risk of ADMS, defined as a Glasgow Coma Scale (GCS) score ≥9 combined with a Sedation–Agitation Scale (SAS) score ≥5 [23,24]. Exclusion criteria included: 1) inability to recognize or detect hand movements due to trauma or other conditions resulting in the hand being covered by dressings, or due to the absence of the hand; 2) severe cardiac, hepatic, or renal disease.

2.2. Sample size

This study was designed as a feasibility study, therefore, the sample size was primarily determined by process feasibility and resource availability, rather than by a power calculation based on effect size [25]. We aimed to recruit approximately 40 patients per group. Such a sample size is standard in pilot trials [26].

2.3. Randomization and blinding

A computer assigned each enrolled patient a unique, randomly generated number. Based on these numbers, patients were allocated to either the routine care group (control) or the computer-based recognition group (intervention). The randomization process was fully automated to ensure fairness and eliminate human interference. Allocation concealment was maintained using sealed opaque envelopes. Because the intervention device was visible, blinding of the nursing staff was not possible. However, outcome assessors remained unaware of group assignments throughout the study. Therefore, the trial was conducted under a single-blind design.

2.4. Interventions

2.4.1. The intervention group

2.4.1.1. System design and deployment
  • (1)

    Overall architecture: We designed and implemented a lightweight, web-based real-time monitoring application with a responsive graphical user interface (GUI). Unlike traditional compiled software, this system was developed using standard web technologies and operates directly within any modern web browser. The system continuously captures and processes live video streams entirely on the client side. The processing pipeline comprised: 1) real-time video acquisition via the browser’s media devices ‌Application Programming Interface (API); 2) efficient hand detection and bounding box extraction using a hand recognition framework; and 3) geometric logic for Region-of-Interest (ROI) intersection detection and audio-visual event triggering. All computations were performed locally on the client device, utilizing WebGL acceleration to ensure high-performance inference on standard consumer hardware (Fig. 1).

  • (2)

    Video acquisition and pre-processing: A live video stream from the bedside camera was captured. Each frame then underwent per-frame processing. At program startup, the system automatically initialized a default rectangular ROI target. The operator could translate and resize this rectangle on the live preview to define the target area. Next, a public hand-detection library was applied to each frame to locate hands in the scene. For every detected hand, the system drew a bounding rectangle (ROI-hand). Unlike the simultaneous detection of both faces and hand joints, the system only detects the hand positions. As a result, its impact on system performance during operation is relatively low.

  • (3)

    ROI intersection and event logic: During runtime, the system continuously checked whether the ROI target intersected with any ROI hand. If an intersection was observed in the current frame, indicating that a hand had entered the target area, the system saved the current image and triggered an audible alert. If no intersection was detected, processing continued with the next frame. Because the ROI target is operator-adjustable in both size and position, users can indirectly tune detection sensitivity by refining the ROI’s placement and dimensions.

  • (4)

    System deployment: Successful clinical deployment of the system required attention to several practical considerations. First, because the system relies on optical tracking, the camera must maintain an unobstructed line of sight to the patient’s head and neck [27,28] (Fig. 2). Second, whenever feasible, the camera should be positioned on the same side as the target area to enhance detection accuracy. Third, the size of the target area directly influenced alarm sensitivity; therefore, it could be adjusted dynamically during use to achieve a balance between sensitivity and specificity.

Fig. 1.

Fig. 1

Open in a new tab

Operating principle of the monitoring and recognition system. (A) Initialization: the system automatically creates a red rectangular Region-of-Interest (ROI), and simultaneously tracks detected hands in real-time (green bounding box). (B) The ROI is manually repositioned and scaled to cover the head-and-neck region. (C) An alarm is triggered when the hand bounding box touches or intersects the ROI.

Fig. 2.

Fig. 2

Open in a new tab

Camera placement relative to patient positioning. (A) Camera location with respect to the bed. (B–D) Example camera positions tailored to individual patient circumstances/positioning.

2.4.1.2. Implementation of the intervention

Patients in the intervention group were monitored with an image recognition–based detection system. A bedside high-definition camera continuously tracked hand movements in real-time and automatically identified risk actions involving approaches to, contact with ROI target. When such action was detected, the system immediately issued an alarm to the nursing station. The responsible nurse then verified the situation at the bedside and determined whether intervention was required. The entire process did not alter routine nursing procedures or clinical treatment. In addition to this system, all patients received routine postoperative care identical to that in the control group.

2.4.2. The control group

Patients in the control group received routine postoperative care. 1) Maintaining catheter patency, regularly observing and recording the characteristics, color, and volume of the drainage fluid, and ensuring a closed drainage system to prevent retrograde infection; 2) scheduled nursing rounds to assess consciousness, agitation, and any risk actions related to catheter pulling; and 3) applying physical restraints in accordance with clinical guidelines when nurses observed agitation or actions indicating a risk of catheter dislodgement—such as repeated reaching toward the catheter site, struggling against medical supplies, or verbal/nonverbal expressions of resistance, after verbal persuasion proved ineffective (e.g., soft wrist restraints).

2.5. Study initiation and termination

At the time of postoperative ICU admission, nurses assessed each patient according to the predefined criteria. If the patient was under sedation at admission, reassessment was performed after discontinuation of sedative medications as per medical orders. Patients who met the eligibility criteria were immediately enrolled and began the intervention. The observation period for both groups started immediately after enrollment. During the observation period, if a patient exhibited actions suggestive of risk action, nurses first attempted verbal persuasion; if this was ineffective, physical restraints were applied according to clinical protocols. If physical restraint failed to ensure safety (e.g., persistent attempts to break free, severe agitation, or actions posing immediate threats to patient safety), the attending physician evaluated the need for sedative medication. Observation continued until the occurrence of any of the following events, whichever came first: dislodgement of the supplies (including both accidental dislodgement and dislodgement according to clinical orders), transfer out of the ICU, clinical deterioration requiring secondary surgery, death in the ICU, or administration of sedative drugs according to medical orders.

2.6. Measures

2.6.1. Basic demographics and clinical data

Baseline demographic and clinical characteristics were collected, including sex, age, and types of medical supplies in place. For patients with multiple head-and-neck supplies, the supply closest to the patient’s hand was designated as the target for monitoring. For example, if both a central venous catheter and a cranial drainage tube were present, the central venous catheter was selected, as the patient’s hand would reach this supply first during upward movement, warranting earlier nursing attention. The observation period for each patient was also recorded from study initiation to completion.

2.6.2. Clinical outcomes

During the study period, we prospectively recorded, for each study group, the number of ADMS, the number of risk actions identified by nursing staff, and the continuous system uptime (h).

2.7. Data collection and quality control

Before enrollment, the study purpose, procedures, data use, and privacy safeguards were explained to patients and families. Demographic variables (age, sex) and clinical information were obtained from the Electronic Medical Record (EMR), bedside assessments, and nursing documentation. ADMS were identified from EMR and nursing incident notes. When an event occurred, the on-duty nurse informed the study recorder, the event time was logged, and follow-up for that patient ceased. Risk actions were counted per patient during observation. In the intervention group, detections were automatically time-stamped in the system log. In the control group, actions observed by nurses were reported on the same day and cross-checked with bedside notes. Daily aggregation yielded the total count of risk actions per patient. Observation time was defined as the period from randomization to ICU discharge, death, transfer, or withdrawal from the study. In the intervention group, system uptime, as recorded from software logs, was documented at the study end and matched the observation time. For patients with multiple eligible supplies, any accidental dislodgement of one supply was considered ADMS, and risk actions were recorded whenever the hand entered the target area. All data were collected by an independent researcher and verified through cross-checks between Case Report Forms (CRFs), EMR, nursing notes, and system logs.

2.8. Data analysis

Data analysis was conducted by a separate researcher who was not involved in clinical implementation and was blinded to group allocation to minimize potential bias. All analyses were performed using Python 3.12. Descriptive statistics were primarily used. Normality of continuous variables was assessed using the Shapiro–Wilk test; variables that followed a normal distribution were presented as mean and standard deviation (SD); otherwise, they were presented as median and interquartile range (IQR). Categorical variables were summarized as frequencies and percentages. Between-group comparisons were conducted for baseline characteristics and clinical outcomes. Normally distributed data were analyzed using the independent-samples t-test, and non-normally distributed data were analyzed using nonparametric tests. All statistical tests were two-sided, and P < 0.05 was considered statistically significant.

2.9. Ethical consideration

The Local Institutional Ethics Committee approved this study (No. 2025171). Informed consent was obtained from all participants or their legal guardians before enrollment.

3. Results

3.1. Baseline characteristics of the study participants

All 80 ICU patients completed the intervention, with 40 in each group. Diagnostic/surgical categories were as follows: cranial and brain surgery (n = 26), fracture surgery (n = 4), lung cancer surgery (n = 9), maxillofacial trauma (n = 6), esophageal surgery (n = 7), and abdominal surgery (n = 28). There were no statistically significant between-group differences in age, sex, types of medical supplies, and observation duration (all P > 0.05). Physical restraints were applied to 7 patients in the control group and 6 in the intervention group, with no significant difference by Chi-square test (P = 0.78). Overall, the two groups were comparable at baseline. The main baseline characteristics are summarized in Table 1.

Table 1.

Comparison of baseline characteristics and clinical outcomes of the study.

Characteristics Total (n = 80) Intervention group (n = 40) Control group (n = 40) t/χ2/U P
Age (years) 50.0 ± 17.8 51.0 ± 16.9 49.8 ± 19.0 0.310a 0.760
Sex
 Male 44 (55.0) 21 (52.5) 23 (57.5) 0.045b 0.830
 Female 36 (45.0) 19 (47.5) 17 (42.5)
Types of medical supplies
 Cranial drainage catheter 13 (16.2) 6 (15.0) 7 (17.5) 0.570b 0.970
 Nasogastric tube 20 (25.0) 9 (22.5) 11 (27.5)
 Central venous catheter 20 (25.0) 11 (27.5) 9 (22.5)
 Wound dressing 16 (20.0) 8 (20.0) 8 (20.0)
 Tracheostomy Tube 11 (13.8) 6 (15.0) 5 (12.5)
Observation duration (h) 49 (38, 61) 48 (38, 57) 49 (37, 62) 786.500c 0.920
Frequency of risk actions of ADMS per 100 h 30 (24, 40) 36 (30, 48) 25 (19, 35) 1,423.000c <0.001
Frequency of ADMS 11 (13.8) 2 (5.0) 9 (22.5) – 0.048d
Open in a new tab

Note: Data are n (%) or Median (P25, P75) or Mean ± SD. a Independent-samples t-test. b Chi-quare test. c Mann–Whitney U test. d Fisher’s exact test. ADMS = accidental dislodgement of medical supplies.

3.2. Continuous system uptime

In the intervention group, the total system uptime was 1,963 h, with a median of 48 h per patient (IQR: 38, 57). The system operated continuously throughout the observation period without interruption or unexpected shutdowns, demonstrating good operational stability.

3.3. Accidental dislodgement of medical supplies

Nine ADMS occurred in the control group (5 partial catheter dislodgements and 4 wound-dressing dislodgements), resulting in an overall incidence of 22.5 % (9/40). Two ADMS occurred in the intervention group, both partial catheter dislodgements, for an overall incidence of 5.0 % (2/40). Fisher’s exact test suggested a lower incidence of actual ADMS in the intervention group compared with the control group (P = 0.048), suggesting that the automated image recognition-based detection system may help reduce the occurrence of ADMS.

3.4. Risk actions of accidental dislodgement of medical supplies

During the study, 498 risk actions were documented in the control group, corresponding to 25 events per 100 h (IQR: 19, 35). In the intervention group, the system detected 706 risk actions, equivalent to 36 events per 100 h (IQR: 30, 48). The intervention group exhibited more detected risk actions (U = 1,423.000, P < 0.001), indicating enhanced sensitivity to monitoring. Additionally, for the intervention group, the system simultaneously captured 706 images corresponding to the 706 risk actions detected. Manual review confirmed full concordance between image content and the recorded actions, yielding a 100 % match rate. Nursing staff documented 426 hand-contact events during routine rounds, all of which were included in the system detections.

4. Discussion

In this study, patients in the intervention group underwent prolonged, continuous monitoring during which the detection system operated without interruption, crashes, or premature termination—attesting to stable performance in real-world practice. In comparison to standard care, the intervention resulted in fewer ADMS and more frequent identification of risk actions, underscoring its clinical feasibility.

In the ICU, monitoring requirements vary across patients and shifts, necessitating a system capable of prolonged, continuous, and stable operation [29,30]. Undetected system failures—particularly during understaffed night shifts—can substantially increase the risk of missed detections [31,32]. In this study, patients in the intervention arm remained under continuous system surveillance; the most extended single-patient observation interval was 57 h, representing the maximum continuous runtime observed under the study conditions and, in part, meeting the clinical need for sustained monitoring [33]. Two design choices likely contributed to runtime stability: constraining inference to a fixed, nurse-defined head-and-neck ROI at start-up and limiting computation to hand localization rather than repeatedly detecting the target zone. In addition, compared with the control group, the intervention group had a lower incidence of ADMS (5.0 % vs 22.5 %). The system detected more ADMS risk actions than routine nursing rounds (706 vs 426).

In fact, for patients with tubes, one of the most critical actions that requires manual observation is whether the patient’s hands are approaching the head, as such actions could lead to ADMS. Promptly detecting these movements allows nursing staff to intervene early. However, manual observation is prone to missing these events, even though the patient’s hand may approach the head without causing extubation. Such actions should still be identified immediately. The findings of this study reinforce that routine manual nursing rounds often fail to detect such actions. In contrast, the continuous monitoring system employed in this study was able to detect significantly more instances of hand movement toward the head, highlighting its superior sensitivity in identifying risk actions and its potential clinical value.

These findings suggest that the primary contribution of the system lies in enhancing the timeliness and comprehensiveness of risk recognition. Through continuous surveillance and real-time alarms, it transforms subtle exploratory movements into actionable alerts, enabling nurses to intervene during the critical window before catheter dislodgement occurs. At the same time, by automatically detecting and recording risk actions, the system alleviates the need for constant bedside vigilance, reduces observation fatigue, and allows nurses to devote more attention to direct patient care [34,35]. By providing traceable, image-based evidence, it also reduces subjective variation and strengthens human–machine collaboration, thus adding a barrier to the chain of ICU nursing safety [36].

Previous studies [37,38] have shown that image recognition–based fall detection systems significantly reduce fall events in elderly populations, and AI has also proven feasible in recognizing restraint actions and monitoring nursing procedures. However, most existing research has focused on general care settings or geriatric cohorts [39]. This study extends visual monitoring to the ICU—a complex environment.

Our study demonstrated that the image recognition-based detection system showed favorable signals on key nursing safety indicators. Compared with standard care, the intervention group had a lower incidence of ADMS, and the system more frequently identified ADMS-related risk hand actions, suggesting potential advantages for earlier warning and timelier intervention. Unlike wearable solutions, the system anchors the risk zone to the head-and-neck region, which nurses define once at startup as an ROI; thereafter, the algorithm tracks only hand positions and evaluates hand–ROI intersections [40,41]. This design imposes no additional burden on patients, better aligns with the comfort and humanistic needs of the critically ill, and lowers the threshold for clinical adoption. Through continuous capture and graded alerts, the system provides nurses with traceable, image-based evidence, reducing subjective variability and the burden of sustained bedside vigilance, with the potential to optimize staffing and streamline workflows. In addition, operational and interaction logs can support scheduling optimization, training evaluation, and alarm management, while providing a data foundation for integration with hospital information systems. As digital nursing technologies evolve, this monitoring paradigm could become an integral component of intelligent ICU care, advancing smart ward deployment and the broader realization of digital healthcare [[42], [43], [44]]. These observations warrant confirmation in larger, multicenter, and multiple medical supplies’ studies, alongside systematic evaluation of alert burden, cost-effectiveness, privacy compliance, and interoperability, to inform a scalable, practice-ready model for precision nursing safety monitoring.

5. Limitations and future direction

This study has several limitations. First, it was a single-center exploratory trial with a relatively small sample size, and the incidence of ADMS was low. Therefore, the current findings are insufficient to fully demonstrate the effectiveness of the recognition system in reducing ADMS. In subsequent work, broader multicenter studies across diverse hospital settings are planned to assess external validity and the feasibility of implementation. These investigations will follow standardized protocols, include a wider range of medical supplies’ types and patient populations, and incorporate pre-specified endpoints—such as ADMS incidence, restraint use, alert burden, patient-centered outcomes—and cost-effectiveness analyses. Second, the intervention group received the detection system in addition to routine nursing care, making it impossible to evaluate the system’s independent effectiveness. Future research may consider multi-arm RCTs to disentangle the relative contributions of the system and conventional nursing interventions. Third, the current optimized version of the system, while significantly improving performance by focusing on hand recognition, still relies on optical cameras as the primary sensing modality. As a result, hand movements may be obscured by bedding, bedrails, or nursing staff, leading to blind spots that compromise the continuity and completeness of monitoring. Future iterations of the system could benefit from the integration of multi-angle cameras or additional sensing devices to mitigate these blind spots and enhance robustness.

6. Conclusions

We designed and feasibility-tested a bedside camera–based image detection system that continuously tracks patients’ hands and issues real-time alarms when predefined risk zones around head- and-neck medical supplies are breached, thereby promptly identifying risk actions and notifying on-duty nurses for early intervention. The results showed that the system operated stably and had the potential to reduce the occurrence of ADMS, while also increased the frequency of identifying risk actions of ADMS. By complementing routine nursing rounds with continuous surveillance and timely alerts, this approach may enhance patient safety and advance innovation in practice. Future studies should include adequately powered multicenter trials, algorithmic refinements to handle occlusions and enable adaptive thresholds, integration with hospital information systems and alarm-management strategies to mitigate alert fatigue, extension to additional medical supplies’ types and populations, and assessments of restraint use, patient-centered outcomes, and implementation metrics.

Data availability statement

The datasets generated during the current study are included in the article.

CRediT authorship contribution statement

Zhongjie Shi: Software, Writing - original draft. Taotao Shi: Conceptualization, Methodology, Writing - review & editing. Xin Gao: Methodology. Jian Li and Hong Xu: Data curation and formal analysis. Xiaojun Li: Data analysis. Zhanxiang Wang: Project administration and Validation. Sifang Chen: Supervision, Resources.

Funding

None.

Declaration of competing interest

The authors declare no competing interests.

Acknowledgments

We would like to express our gratitude to all participating patients and their families, as well as the nursing staff.

Footnotes

Peer review under responsibility of Chinese Nursing Association.

Appendix A

Supplementary data to this article can be found online at https://doi.org/10.1016/j.ijnss.2025.12.001.

Appendix A. Supplementary data

The following is the Supplementary data to this article.

Multimedia component 1
mmc1.docx (12.4KB, docx)

References

  • 1.Reignier J., Rice T.W., Arabi Y.M., Casaer M. Nutritional support in the ICU. BMJ. 2025;388 doi: 10.1136/bmj-2023-077979. [DOI] [PubMed] [Google Scholar]
  • 2.Bertuccio A., Marasco S., Longhitano Y., Romenskaya T., Elia A., Mezzini G., et al. External ventricular drainage: a practical guide for neuro-anesthesiologists. Clin Pract. 2023;13(1):219–229. doi: 10.3390/clinpract13010020. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3.Cañamares-Orbís P., García-Rayado G., Alfaro-Almajano E. Nutritional support in pancreatic diseases. Nutrients. 2022;14(21):4570. doi: 10.3390/nu14214570. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4.Onsombi D., Mayaya G., Herrera V., Manyanga A., Leonald W., Byabato S., et al. The usefulness of surgical drains on short term outcomes among patients undergoing craniotomy at the Bugando Medical Centre, Mwanza Tanzania. World Neurosurg X. 2024;22 doi: 10.1016/j.wnsx.2024.100323. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Anis A., Patel R., Tanios M.A. Analytical review of unplanned extubation in intensive care units and recommendation on multidisciplinary preventive approaches. J Intensive Care Med. 2024;39(6):507–513. doi: 10.1177/08850666231199055. [DOI] [PubMed] [Google Scholar]
  • 6.Xiong B.B., Bailey D.X., Prudon P., Pascoe E.M., Gray L.C., Graham F., et al. Identification and information management of cognitive impairment of patients in acute care hospitals: an integrative review. Int J Nurs Sci. 2023;11(1):120–132. doi: 10.1016/j.ijnss.2023.11.001. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.Pandit A.S., Palasz J., Nachev P., Toma A.K. Mechanical complications of external ventricular and lumbar drains. World Neurosurg. 2022;166:e140–e154. doi: 10.1016/j.wneu.2022.06.127. [DOI] [PubMed] [Google Scholar]
  • 8.Barber J.A. Unplanned extubation in the NICU. J Obstet Gynecol Neonatal Nurs. 2013;42(2):233–238. doi: 10.1111/1552-6909.12009. [DOI] [PubMed] [Google Scholar]
  • 9.Franks Z.M., Alcock J.A., Lam T., Haines K.J., Arora N., Ramanan M. Physical restraints and post-traumatic stress disorder in survivors of critical illness. A systematic review and meta-analysis. Ann Am Thorac Soc. 2021;18(4):689–697. doi: 10.1513/AnnalsATS.202006-738OC. [DOI] [PubMed] [Google Scholar]
  • 10.Basic D., Huynh E., Gonzales R., Shanley C. Structured interdisciplinary bedside rounds, in-hospital deaths, and new nursing home placements among older inpatients. Clin Interv Aging. 2018;13:2289–2294. doi: 10.2147/CIA.S171508. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11.Ye J.R., Xiao A.X., Yu L., Wei H.M., Wang C., Luo T.Y. Physical restraints: an ethical dilemma in mental health services in China. Int J Nurs Sci. 2017;5(1):68–71. doi: 10.1016/j.ijnss.2017.12.001. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12.Salehi Z., Najafi Ghezeljeh T., Hajibabaee F., Joolaee S. Factors behind ethical dilemmas regarding physical restraint for critical care nurses. Nurs Ethics. 2020;27(2):598–608. doi: 10.1177/0969733019858711. [DOI] [PubMed] [Google Scholar]
  • 13.Jang S.G., Lee W., Ha J., Choi S. Is physical restraint unethical and illegal? : a qualitative analysis of Korean written judgments. BMC Nurs. 2024;23(1):94. doi: 10.1186/s12912-024-01781-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14.AlDahoul N., Tan M.J.T., Tera R.R., Abdul Karim H., Lim C.H., Mishra M.K., et al. Multitasking vision language models for vehicle plate recognition with VehiclePaliGemma. Sci Rep. 2025;15(1) doi: 10.1038/s41598-025-10774-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15.El F.N. Facial recognition algorithms: a systematic literature review. J Imaging. 2025;11(2):58. doi: 10.3390/jimaging11020058. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16.Izumi K., Suzuki K., Hashimoto M., Jinzaki M., Ko S., Takeuchi T., et al. Ensemble detection of hand joint ankylosis and subluxation in radiographic images using deep neural networks. Sci Rep. 2024;14(1):7696. doi: 10.1038/s41598-024-58242-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Zhang J.X., Li Z., Liu Y., Li J., Qiu H.L., Li M.H., et al. An effective deep learning framework for fall detection: model development and study design. J Med Internet Res. 2024;26 doi: 10.2196/56750. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18.Dai P.Y., Wu Y.C., Sheu R.K., Wu C.L., Liu S.F., Lin P.Y., et al. An automated ICU agitation monitoring system for video streaming using deep learning classification. BMC Med Inform Decis Mak. 2024;24(1):77. doi: 10.1186/s12911-024-02479-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19.Kim MS, Park B, Sippel GJ, Mun AH, Yang W, McCarthy KH, et al. Comparative analysis of personal protective equipment nonadherence detection: computer vision versus human observers. J Am Med Inform Assoc. 2025;32(1):163–171. doi: 10.1093/jamia/ocae262. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20.Davoudi A., Malhotra K.R., Shickel B., Siegel S., Williams S., Ruppert M., et al. Intelligent ICU for autonomous patient monitoring using pervasive sensing and deep learning. Sci Rep. 2019;9(1):8020. doi: 10.1038/s41598-019-44004-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21.Rehouma H., Noumeir R., Essouri S., Jouvet P. Advancements in methods and camera-based sensors for the quantification of respiration. Sensors (Basel) 2020;20(24):7252. doi: 10.3390/s20247252. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22.Eldridge S.M., Chan C.L., Campbell M.J., Bond C.M., Hopewell S., Thabane L., et al. CONSORT 2010 statement: extension to randomised pilot and feasibility trials. Pilot Feasibility Stud. 2016;2:64. doi: 10.1186/s40814-016-0105-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23.Mahajan C., Sengupta D., Kapoor I., Prabhakar H., Kumar V., Purohit S., et al. Evaluation of the GCS-pupils score for PrOgnosis in trauMatic brAin injury- the COMA study. Brain Inj. 2023;37(9):1041–1047. doi: 10.1080/02699052.2023.2227943. [DOI] [PubMed] [Google Scholar]
  • 24.Riker R.R., Picard J.T., Fraser G.L. Prospective evaluation of the sedation-agitation scale for adult critically ill patients. Crit Care Med. 1999;27(7):1325–1329. doi: 10.1097/00003246-199907000-00022. [DOI] [PubMed] [Google Scholar]
  • 25.Teare M.D., Dimairo M., Shephard N., Hayman A., Whitehead A., Walters S.J. Sample size requirements to estimate key design parameters from external pilot randomised controlled trials: a simulation study. Trials. 2014;15:264. doi: 10.1186/1745-6215-15-264. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 26.Kunselman A.R. A brief overview of pilot studies and their sample size justification. Fertil Steril. 2024;121(6):899–901. doi: 10.1016/j.fertnstert.2024.01.040. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27.Lindroth H., Nalaie K., Raghu R., Ayala I.N., Busch C., Bhattacharyya A., et al. Applied artificial intelligence in healthcare: a review of computer vision technology application in hospital settings. J Imaging. 2024;10(4):81. doi: 10.3390/jimaging10040081. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 28.Diao J.A., Marwaha J.S., Kvedar J.C. Video-based physiologic monitoring: promising applications for the ICU and beyond. NPJ Digit Med. 2022;5(1):26. doi: 10.1038/s41746-022-00575-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29.Lin F.F., Chaboyer W., Foster M., Hervey L., Marshall A.P. Facilitating the relocation of an intensive care unit from shared room to single room setting: staff perceptions on the effectiveness of a multicomponent intervention. Aust Crit Care. 2019;32(4):279–284. doi: 10.1016/j.aucc.2018.06.008. [DOI] [PubMed] [Google Scholar]
  • 30.Montrucchio G., Sales G., Catozzi G., Bosso S., Scanu M., Vignola T.V., et al. Effectiveness of an active and continuous surveillance program for intensive care units infections based on the EPIC III (extended prevalence of infection in intensive care) approach. J Clin Med. 2022;11(9):2482. doi: 10.3390/jcm11092482. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31.Mosch L., Sümer M., Flint A.R., Feufel M., Balzer F., Mörike F., et al. Alarm management in intensive care: qualitative triangulation study. JMIR Hum Factors. 2024;11 doi: 10.2196/55571. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32.Nyarko B.A., Yin Z.Z., Chai X.Y., Yue L.Q. Nurses' alarm fatigue, influencing factors, and its relationship with burnout in the critical care units: a cross-sectional study. Aust Crit Care. 2024;37(2):273–280. doi: 10.1016/j.aucc.2023.06.010. [DOI] [PubMed] [Google Scholar]
  • 33.Yang S.M., Galvagno S., Badjatia N., Stein D., Teeter W., Scalea T., et al. A novel continuous real-time vital signs viewer for intensive care units: design and evaluation study. JMIR Hum Factors. 2024;11 doi: 10.2196/46030. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 34.Peršolja M. Effects of nurses' schedule characteristics on fatigue: an integrative review. Nurs Manage. 2023;54(4):28–37. doi: 10.1097/01.NUMA.0000921904.11222.11. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35.Zhang H.Y., Liu Z.X., Liu J.P., Feng Y.J., Zou D.D., Zhao J., et al. Factors influencing nurse fatigue during COVID-19: regression vs. fuzzy-set qualitative comparative analysis. Front Public Health. 2023;11 doi: 10.3389/fpubh.2023.1184702. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 36.Festor P., Nagendran M., Gordon A.C., Faisal A.A., Komorowski M. Safety of human-AI cooperative decision-making within intensive care: a physical simulation study. PLoS Digit Health. 2025;4(2) doi: 10.1371/journal.pdig.0000726. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37.Almukadi W.S., Alrowais F., Saeed M.K., Yahya A.E., Mahmud A., Marzouk R. Deep feature fusion with computer vision driven fall detection approach for enhanced assisted living safety. Sci Rep. 2024;14(1) doi: 10.1038/s41598-024-71545-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 38.Chen Y.H., Xu J.L. Applying artificial intelligence to predict falls for inpatient. Front Med (Lausanne) 2023;10 doi: 10.3389/fmed.2023.1285192. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 39.Davoudi A., Shickel B., Tighe P.J., Bihorac A., Rashidi P. Potentials and challenges of pervasive sensing in the intensive care unit. Front Digit Health. 2022;4 doi: 10.3389/fdgth.2022.773387. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 40.Pyun K.R., Kwon K., Yoo M.J., Kim K.K., Gong D., Yeo W.H., et al. Machine-learned wearable sensors for real-time hand-motion recognition: toward practical applications. Natl Sci Rev. 2023;11(2):nwad298. doi: 10.1093/nsr/nwad298. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 41.Angelucci A., Greco M., Cecconi M., Aliverti A. Wearable devices for patient monitoring in the intensive care unit. Intensive Care Med Exp. 2025;13(1):26. doi: 10.1186/s40635-025-00738-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 42.Park Y., Chang S.J., Kim E. Artificial intelligence in critical care nursing: a scoping review. Aust Crit Care. 2025;38(4) doi: 10.1016/j.aucc.2025.101225. [DOI] [PubMed] [Google Scholar]
  • 43.Al Aseri Z., Côté M., Baharoon S., Alabdulaali M.K., Altamimi S., Arntfield R. AI in critical care: a narrative review of prospective applications and future potential in KSA’s health transformation 2030. J Taibah Univ Med Sci. 2025;20(3):359–364. doi: 10.1016/j.jtumed.2025.05.007. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 44.Mao Z., Liu C., Li Q.L., Cui Y.T., Zhou F.H. Intelligent intensive care unit: current and future trends. Intensive Care Res. 2023:1–7. doi: 10.1007/s44231-023-00036-5. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Multimedia component 1
mmc1.docx (12.4KB, docx)

Data Availability Statement

The datasets generated during the current study are included in the article.

Articles from International Journal of Nursing Sciences are provided here courtesy of Chinese Nursing Association

RESOURCES