Video-based Automatic Hand Hygiene Detection for Operating Rooms using 3D Convolutional Neural Networks

doi:10.21203/rs.3.rs-2946647/v1

Download PDF

Article

Video-based Automatic Hand Hygiene Detection for Operating Rooms using 3D Convolutional Neural Networks

https://doi.org/10.21203/rs.3.rs-2946647/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 18 Jun, 2024

Read the published version in Journal of Clinical Monitoring and Computing →

Version 1

posted

You are reading this latest preprint version

Hand hygiene among anesthesia personnel is important to prevent hospital-acquired infections in operating rooms; however, an efficient monitoring system remains elusive. In this study, we leverage a deep learning approach based on operating room videos to detect alcohol-based hand hygiene actions of anesthesia providers. The proposed algorithm utilized a two-dimensional (2D) and three-dimensional (3D) convolutional neural networks (CNNs), sequentially. First, multi-person of the anesthesia personnel appearing in the target OR video were detected per image frame using the pre-trained 2D CNNs. Following this, each image frame detection of multi-person was linked and transmitted to a 3D CNNs to classify hand hygiene action. Optical flow was calculated and utilized as an additional input modality. Evaluations of the binary classification of hand-hygiene actions revealed an accuracy of 0.88, a sensitivity of 0.78, a specificity of 0.93, and an area under the operating curve (AUC) of 0.91. The deep learning approach has the potential to be applied in practical clinical scenarios providing continuous surveillance in a cost-effective way.

Health sciences/Health care

Physical sciences/Engineering/Biomedical engineering

Hand hygiene is considered to be the most effective practice to prevent infections, such as hospital-acquired infections (HAIs).[1–3] Several studies support that appropriate hand hygiene significantly reduces the incidence and severity of outbreaks of infection[4 5] and enhances patient safety in hospitals.[6] Despite this notion, compliance rates of hand hygiene remain poor in most countries.[7–11] Anesthesia providers in the operating rooms (OR) who remain un-scrubbed have fewer expectations for hand hygiene compliance and consequently have lower compliance rates.[9 12] In ORs, anesthesia providers continuously provide patient care involving in numerous activities that may induce HAIs, such as intubating patients, accessing intravenous (IV) catheters, injecting anesthetic drugs, and airway management.[13–15] Contaminated hands of anesthesia providers in routine anesthesia activities can be major sources of HAIs in ORs yet the potential risk of HAIs caused by the lack of disinfection of anesthesia providers' hands has been traditionally underestimated.

In an effort to promote safe hand hygiene, healthcare organizations across the world have promoted ‘SAVE LIVES: Clean Your Hands’ campaigns and publicly disseminated guidelines. [16 17] However, those guidelines are suggested in terms of a hospital unit or ward, thus do not fit the scenario in operating rooms (OR). Hand hygiene protocol for anesthesia personnel in ORs mainly relates to the use of alcohol-based hand rubs (ABHR) on gloves since it may be difficult to either change gloves or wash hands during other tasks.[18] Frequent application of ABHR would be required since anesthesia providers touch OR equipment (e.g., the anesthesia machine, monitors, vascular catheters, IV tubing) and patients up to 150 times during routine anesthesia activities.[19]

A study demonstrated that repeated monitoring[10 20] of hand hygiene and positive feedback[21 22] can successfully increase compliance. However, an effective monitoring system for hand hygiene compliance remain elusive. The standard method is direct observation by trained personnel during pre-determined monitoring periods. This method is highly labor-intensive and compliance rates tend to increase only during the monitoring period, as people often modify their behaviors based on the perception of being observed.[23]

Some existing studies have attempted to develop a simple method for measuring hand hygiene compliance in ORs. A mobile application was implemented that enabled instant recording of hand-hygiene compliance via direct observation. [24] This tool helped the collection of quantitative data but could not overcome the sampling bias inherent in direct observation-based methods. A method that estimates the number of ABHR based on hand hygiene container weight was developed. The estimated number exhibited a correlation to the video observation yet only can measure the cumulative frequency, not the moment of hand hygiene use.

Machine learning is widely used for applications in hospital environments for video analysis, including automatic hospital entrances, surveillance, and patient monitoring systems. Convolutional neural networks (CNN) has been greatly adapted in the field of image and video-based human action detection. A two-dimensional (2D) CNN-based hand hygiene detection system with depth cameras have been implemented.[25 26] Both studies relied on the location of the hand sanitizer dispenser, which is a strong indicator in the identification of dispenser access, thus lack temporal information of hand hygiene actions. Our study utilizes three-dimensional (3D) CNN-based action detection algorithm which intend to capture the semantic motions of human behaviors so that it can provide a better understanding of the OR scenes.

Such machine learning-based action detection algorithms can be leveraged to assist in the monitoring of hand-hygiene compliance in OR videos, reducing time cost in the direct and video observation. Most existing methods, however, involve the detection of one action performed by a single person or in a distinctive background scene.[27–29] Thus, it is difficult to apply them on datasets comprising complex scenes, such as those captured in ORs. The OR scenes comprise multiple people. Multiple anesthesia personnel who wear scrubs of identical color are involved in each operation depicted in OR videos. In multiple-person scenarios, individual medical personnel is often misaligned or missed between frames, which makes the detection of hand-hygiene actions more challenging. Also, actions of interest in OR videos exhibit high spatial variance of motions (e.g., rubbing hands while walking), unlike actions with low diversity (e.g., standing and talking). The detection of such high diversity actions requires extensive datasets and rich representation of motion.[30]

In this study, a video-based hand hygiene detection algorithm is proposed based on 3D CNN. Our study aimed to develop an algorithm that can automatedly detect hand hygiene actions in OR videos, therefore, can be utilized to enhance the monitoring process for hand hygiene compliance.

Detection of the Anesthesia Personnel

The measured values of AP and AR of anesthesia personnel detection are depicted in Table 1. AP.₅ represents the PASCAL VOC evaluation metric, which uses an IoU threshold of 0.5. AP and AR represent the COCO evaluation metrics, which correspond to the average over multiple outputs with IoU thresholds between 0.5 and 0.95 with a step size of 0.05.

Table 1

Performance of anesthesia personnel detection using Faster R-CNN in terms of frame AP and AR
	Backbone	f-AP.₅	f-AP._5:.95	f-AR._5:.95
Faster R-CNN	ResNet-50-FPN	0.924	0.737	0.798

Table 2

Comparative study of I3D based on different modalities.
Model	RGB	Flow	v-mAP
I3D	✓		0.800
I3D		✓	0.623
I3D	✓	✓	0.803
Model	RGB	Flow	v-mAP.₂
I3D*	✓		0.700
I3D*		✓	0.543
I3D*	✓	✓	0.706

Table 3

Performance in hand hygiene action detection
Model	RGB	Flow	Accuracy	Sensitivity	Specificity	AUC
I3D	✓		0.86	0.81	0.89	0.91
I3D	✓	✓	0.88	0.78	0.93	0.91
I3D*	✓		0.80	0.69	0.87	0.87
I3D*	✓	✓	0.82	0.65	0.91	0.88

Table 4

Performance of anesthesia personnel detection using faster R-CNN in terms of v-mAP
Model	Class	RGB	Flow	C	F	Pre	Rec	F₁score
I3D	Hand Hygiene	✓				0.60	0.90	0.72
I3D	Touching Equipment	✓				0.63	0.82	0.72
I3D	Other action	✓				0.97	0.27	0.42
I3D	Hand Hygiene	✓		✓		0.69	0.89	0.78
I3D	Touching Equipment	✓		✓		0.78	0.73	0.76
I3D	Other action	✓		✓		0.77	0.61	0.68
I3D	Hand Hygiene	✓			✓	0.69	0.86	0.77
I3D	Touching Equipment	✓			✓	0.64	0.88	0.74
I3D	Other action	✓			✓	0.89	0.40	0.55
I3D	Hand Hygiene	✓		✓	✓	0.78	0.82	0.80
I3D	Touching Equipment	✓		✓	✓	0.77	0.83	0.80
I3D	Other action	✓		✓	✓	0.75	0.67	0.71
I3D	Hand Hygiene	✓	✓			0.80	0.82	0.81
I3D	Touching Equipment	✓	✓			0.66	0.95	0.78
I3D	Other action	✓	✓			0.89	0.52	0.66
I3D	Hand Hygiene	✓	✓	✓		0.80	0.80	0.80
I3D	Touching Equipment	✓	✓	✓		0.75	0.63	0.68
I3D	Other action	✓	✓	✓		0.63	0.71	0.67
I3D	Hand Hygiene	✓	✓		✓	0.83	0.76	0.80
I3D	Touching Equipment	✓	✓		✓	0.63	0.98	0.77
I3D	Other action	✓	✓		✓	0.89	0.52	0.66
I3D	Hand Hygiene	✓	✓	✓	✓	0.83	0.80	0.81
I3D	Touching Equipment	✓	✓	✓	✓	0.75	0.76	0.75
I3D	Other action	✓	✓	✓	✓	0.72	0.73	0.73

The value of AP.₅ was observed to be 0.924, which demonstrates the viability of leveraging a 2D person detector not trained on the dataset of interest for the task of action detection. The detection results obtained by Faster R-CNN on our dataset are visualized in Fig. 1. Each row indicates the bounding box predictions of random samples over time. The ground truths are labeled in red and the bounding box results are labeled in cyan. As is evident from the figure, most of the failure cases occurred when two members of the anesthesia personnel wearing scrubs of identical colors were overlapping or crossing-over. The remaining failure cases were observed to correspond to cases of occlusion of the anesthesia personnel by OR equipment, e.g., the OR table or the IV stand.

Action Detection on the OR dataset

The action detection performances of the I3D network equipped with different modalities are presented in Table 2. The experiment with the ground truth is adopted as the baseline, and “*” is used to denote experiments using the 2D person detector, Faster R-CNN. In both experiments, the combined model trained on RGB and optical flow modality results exhibited the best performance among the three models, with v-mAPs of 0.803 and 0.706, respectively. By pooling the RGB model with optical flow modality, small increments of 0.003 and 0.006 in v-mAP were observed in the two experiments, implying that the effect of incorporating the optical flow stream is negligible in the case of our dataset. The models trained using solely optical flow inputs were considered to have failed in learning distinct representations of each action, producing v-mAPs of 0.623 and 0.543 in the two experiments, respectively. Figure 2 depicts the confusion matrix of the action detection performance of each model. The confusion matrix of the optical flow-based model exhibited a high confusion, especially between the “touching equipment” and “other actions” classes.

Next, we analyzed the action detection results from the perspective of action tubelet creation. The overall performance was observed to deteriorate on using Faster R-CNN, exhibiting a decrease of 0.097 in v-mAP with an IoU threshold of 0.2. The performance degradation can be attributed to the misalignment of bounding boxes on the clip-level and the mis-assignment of labels in action tubelets. Finally, we analyzed the result in terms of the binary classification of hand-hygiene actions. In this evaluation, the action of touching equipment was considered to belong to the “other actions” class. The accuracy, sensitivity, specificity, and AUC of results obtained using the RGB-based and the combined models using the ground truth and 2D person detector, Faster R-CNN, were evaluated. The results, as presented in Table 3, indicate that the combined model using the ground truth performs the best, with scores of 0.88, 0.78, 0.93 and 0.91, respectively.

Effects of Upper Body Cropping and Temporal Smoothing

Ablation experiments were conducted to validate the contributions of upper body cropping ("C") and spatial extent fixing ("S") to the resolution of the problem of spatial discrepancy. The performances of the RGB-based and combined models using the ground truth were compared. As presented in Table 4, in the RGB model, upper body cropping improved the performance more than spatial extent fixing, with increments of 0.06, 0.04, and 0.26 in the F1-score corresponding to the three classes of actions in the former case compared to 0.05, 0.02, and 0.13 in the latter. Thus, the “other actions” class was most affected by both techniques. On the other hand, neither upper body cropping nor spatial extent fixing induced a significant improvement in the performance of the combined model. In particular, upper body cropping induced a significant reduction of 0.10 in the F₁score corresponding to the “touching equipment” class in the case of the combined model. On the other hand, the combination of spatial extent fixing and upper body cropping induced an increase of 0.07 for the combined model compared to the experiment without either technique.

We demonstrated that hand-hygiene actions of the anesthesia personnel in the OR can be automatically detected using 3D CNN. The success of the 2D person detector trained on a large but irrelevant dataset is promising in circumventing cumbersome training using a custom dataset. Most of the failure cases were observed when multiple persons wearing scrubs of identical colors were overlapping each other or were occluded by OR equipment. We demonstrated that failures in 2D person detection resulted in a decrease of 0.097 in v-mAP during the action detection task. We also demonstrated that the contribution of optical flow is negligible in the case of the considered dataset. In particular, the optical flow-based model was incapable of appropriately predicting the actions of the anesthesia personnel, with m-vAPs of 0.623 and 0.543, respectively, in two separate experiments. The pooled model trained using both RGB and optical flow inputs exhibited small increments in v-mAP by 0.003 and 0.006 in each combined model using ground truth and 2D person detector. This contradicts the results from previous studies claiming that the two-stream architecture significantly improved performance on multiple benchmark datasets.[27] This dissonance in conclusions can be attributed to the difference in motion information between distinct actions of interests. Actions involving large movements, such as golf-swing and running/jogging, yield better performances if the optical flow model is incorporated.[31–33] On the other hand, the performance was observed to degrade remarkably with respect to actions involving small movements, such as touching object, reading, and watching TV. In the dataset constructed in this study, actions belonging to all three categories involve subtle motions of hands and thus little information for learning rich representations using optical flow modality. In addition, occasional cases of overlapping individuals are included, which generates additional confusing optical flow features.

In the evaluation of a binary classification of hand-hygiene actions, we observed that both experiments using ground truth and Faster R-CNN exhibit a high specificity of 0.93 and 0.91. The high specificity of hand-hygiene action is worthwhile in practice since the compliance rate is measured by detecting the moment when hand-hygiene should be performed but not.

We observed that upper body cropping and spatial extent fixing can reduce spatial discrepancy in action tubelets. In particular, the performance of the RGB-based model was improved by using upper body cropping. Corresponding to certain action categories, alterations in body regions or bounding box ratios are important during the prediction of actions. For instance, the ratio of bounding boxes corresponding to the action “getting up” gradually increase with respect to the frames, and this can be used as a distinctive characteristic of the action. However, in our dataset, all actions of interest were primarily concentrated in the upper body and changes in the ratio of the body were caused by occlusions induced by other persons or surgical equipment rather than the natural characteristics of the actions. Spatial extent fixing also improved the performance of the RGB-based model by enabling the use of other action tubelets. Finally, the combination of both techniques was observed to exhibit the best performance.

There are several limitations in this study. First, the evaluation was conducted on the videos recorded in the same OR. Second, the number of samples of hand rubbing actions was not sufficient to enable complete training on its own. Additional video recording in a different OR settings is required to examine the generalizability of the proposed model. Future works should address the understanding moments to follow hand hygiene protocol in ORs, such as after touching a patient.

In this study, we proposed a 3D CNN-based hand hygiene detection algorithm in ORs. Using a pretrained 2D person detector, the ROIs of all anesthesia personnel appearing in OR videos were detected at the frame-level. Upper body cropping and spatial extent fixing were conducted after creating action tubelets to reduce spatial discrepancy at the clip-level. The experimental results demonstrated that the I3D network with two streams narrowly outperformed the RGB-based one. Further, the utilization of optical flow was demonstrated to be not always helpful, especially during the detection of actions with sparse motion information.

An overall architecture of our approach to detect hand hygiene action in the OR is shown in Fig. 3. Given a video clip, each member of the anesthesia personnel appearing in the video clip was detected using the per-frame 2D person detector, Faster R-CNN.[34] The regions of interest (ROIs) on the upper body of each person were automatically cropped and spatial extent fixing was conducted on them. Following this, the ROIs were transmitted to a 3D CNN, I3D,[27] and classified into three categories: hand hygiene, touching equipment, and other actions. In the evaluation in terms of a binary classification of hand-hygiene actions, the action of touching equipment was considered to belong to the “other actions” class.

Data Acquisition

The videos were collected over a period of four months (November, 2018 to February, 2019) at a single OR in Asan Medical Center, South Korea. Video recording was conducted throughout multiple consecutive surgeries at the aforementioned OR everyday. The Intel® RealSense™ D415 camera (used RGB images only) was used during the recordings. The videos were recorded at 15 frames per second (fps) at a resolution of 640 × 480. Due to the lack of the collected dataset, additional data was simulated and added to it. These deficiencies can be attributed to missing annotations and the confined field-of-view of the camera. We recorded an hour-long video under a simulated situation that was a close approximation of actual surgical conditions, in which eight non-medical personnel, accompanied by an anesthesiologist, repeatedly rubbed their hands with hand sanitizer from dispensers installed within the OR. The camera was installed on one side of the walls and was verified to be capable of capturing most of the activities performed around the OR table. Figure 4 depicts the location of camera and alcohol-based hand sanitizers in the OR. The Institutional review board for human investigations at Asan Medical Center approved this retrospective study based on waivers of informed consent.

Bounding Box Annotations on the Anesthesia Personnel

We used a bounding box to represent the locations of the anesthesia personnel as ground truth. We followed a hybrid approach similar to the process of person annotation in the AVA dataset.[30] To simplify the process, a baseline set of bounding boxes was generated using a deep learning-based object segmentation tool, instead of drawing them from scratch. Mask R-CNN[35] was used to generate the initial set of bounding box predictions with a confidence score exceeding 0.9 to reduce the number of false-positives. Subsequently, incorrectly defined bounding boxes were manually corrected by two human annotators using Labelme,[36] which is a graphical image annotation tool.

Action Annotation on Anesthesiologists

The OR scenes before and after surgery are selected in untrimmed videos depicting intense involvement of anesthesiologists. The average duration of these scenes is observed to be 5.7 min. Then, they are divided into video clips of shorter duration—between 1 and 22 s (16 to 323 frames)—depicting actions related to hand hygiene protocols (rubbing hands, wearing/removing gloves, and touching equipment). The actions not included in the aforementioned categories were annotated as other actions. Each video clip consisted of one to six actions with a mean and standard deviation of 2.82 and 1.13, respectively. Detailed statistics of these video clips is stated in Statistics of the Data used.

Detection of the Anesthesia Personnel

For this task, we utilized a 2D person detector trained on a large dataset, Microsoft COCO,[37] instead of training it on the OR dataset. A well-known object detection network, Faster R-CNN,[34] which is known to exhibit high accuracies and low error rates on several benchmark datasets was used in this study. Given a video clip comprising a sequence of L frames, the person detector outputted the bounding boxes of the individuals detected in each frame. The detection accuracy of the 2D person detector was estimated by comparing its outputs with the ground truth bounding boxes. The performance of Faster R-CNN at the frame-level was also evaluated in terms of average precision (AP) and average recall (AR). The ground truth bounding boxes of the anesthesia personnel were annotated by human annotators. AP and AR were measured by following the PASCAL VOC[38] and the COCO evaluation protocols.[37] In PASCAL VOC evaluation, a detection is deemed to be correct if its intersection over union (IoU) with a ground truth is greater than the threshold, 0.5. In COCO evaluation, the average value over 10 instances, corresponding to 10 IoU thresholds between 0.5 and 0.95 at intervals of 0.05, is used. The maximum detection number was set to be 10 during the calculation of AR.

Action Linking

Given the bounding box of L frames, action tubelets were created by linking the detected person instances on the frame-level throughout each video clip. Links were drawn between each detected instance of the same person throughout the L frames corresponding to a single constant action. IoUs of the bounding boxes of all adjacent pairs of frames were evaluated and the detected person instance was assigned to the frame exhibiting the highest IoU score with its subsequent neighbor. False assignments were manually corrected after completion of the linking process. During the experiment using the 2D person detector, false assignments were ignored and missing instances of detection in the L frames were filled via linear interpolation.

Upper Body Cropping and Temporal Smoothing

Two techniques were used to reduce the spatial discrepancy across the L frames in the action tubelets. First, upper body cropping was performed on the bounding boxes. To this end, the bottom portion of each bounding box depicting a full body region with a height-to-width ratio greater than 1.5 was cropped out. In addition, spatial extent fixing was conducted on each action tubelet by selecting the largest 3D bounding box enclosing it completely. This guaranteed that each of the L frames within a video clip exhibited the same spatial extent. This was accomplished by considering the pixels corresponding to the extrema of the range of the bounding box results along the X and Y axes. The effect of each technique was evaluated and compared in terms of video mean AP (v-mAP)—an action detection is deemed to be correct if its label is true and the mean frame AP of the video clip is greater than 0.2.

Statistics of the Data used

A total of 45 untrimmed videos, with an average duration of eight hours, depicting the target action—the rubbing of hands— were acquired. Each untrimmed video included one to eight occurrences of the target action. In aggregate, the videos comprised 309 clips with 873 action tubelets, of which 252, 109, and 512 action tubelets were labeled to depict hand-hygiene actions, touching equipment, and other actions, respectively. The means (median) of the numbers of image frames in clips depicting actions from the three aforementioned categories were observed to be 30.16 (22.0), 51.15 (34.0), and 36.69 (25.0), respectively as shown in Fig. 5. We divided the video clips in the ratio 6:2:2 to form the training, validation, and testing datasets, respectively. All dataset did not share video clips recorded on the same date. All the action tubelets retained the ratio of each class by temporal augmentation.

Action Detection

We trained 3D CNN on the action tubelets in each video clip for hand hygiene action classification. To this end, I3D,[27] an Inception-v1 architecture[39] that inflates 2D convolutions into 3D ones, was selected. We took advantage of the network pretrained on a large-scale video dataset, Kinetics-400, comprising 400 categories, including 9 hand-related human actions: air drumming, applauding, clapping, nail trimming, doing nails, finger drumming, finger snapping, fist pumping, and hand washing. We adapted a two-stream design as suggested in [27]. The two networks for RGB and optical flow inputs were jointly trained on the training dataset. Then, we conducted an ablation study to explore the effect of the optical flow stream. To this end, we compared the performances of the three models: RGB, optical flow (Flow), and the joint model (RGB + Flow). A training experiment using the RGB model was assumed to be the baseline and the performances of the optical flow-based model and the combined model were compared with it. All methods were evaluated on each video clip in the testing dataset. The optical flow was computed using a TV-L1 algorithm implemented in the OpenCV library. The pretrained I3D was fine-tuned using an additional 1 × 1 × 1 convolution following the final convolutional network and a softmax output was used to account for the multi-classification of the three action categories: hand hygiene, touching equipment, and other actions. All the layers preceding the average pooling layer were frozen during training. The number of input frames was selected to be 16.

Implementation details

Training was performed using standard SGD with momentum set to 0.9, with synchronous parallelization across 2 GPUs of NVIDIA® Tesla® P40. We trained models for 100 epochs, with a learning rate of 1e^− 3 and a weight decay of 1e^− 5. We set a class weight of 0.75 for "other actions" class while 1 for "hand hygiene" and "touching equipment" class.

Data augmentation was performed spatially and temporally. In spatial augmentation, a real-time multi-scale cropping was applied around the bounding box. The random scale was chosen ranging from 0.75 to 1.10 times its original size to include the original bounding boxes. In addition, a random horizontal flipping with a probability of 0.5 and brightness jittering with an extent from − 0.1 to 0.1 was applied. All of the input was scaled to 224 pixel size and center-cropped. In temporal augmentation, we adapted a sliding window fashion with different size of stride per class. A number of 1, 1, and 3 which are inversely proportional to the average number of image frames for hand hygiene, touching equipment, and other actions, respectively.

Author contributions

M.K., J.C. and N.K. wrote the main manuscript text. M.K. and J.C. performed experiments and prepared figures. M.K. and J.J cleansed the dataset. J.J, W.K., and S.S. prepared the dataset and J.J confirmed the hand hygiene activities. All authors reviewed the manuscript.

Corresponding authors

Correspondence to Namkug Kim.

Data availability statement

The datasets generated and/or analyzed during the current study are not publicly available due study participant privacy but are available from the corresponding author on reasonable request.

Additional Information

Funding

This study was supported by a grant of the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health & Welfare, Republic of Korea (HI18C2383).

Ethical approval

Asan Medical Center approved the experiments, and all experiments were performed in accordance with the Declaration of Helsinki.

Mathur P. Hand hygiene: Back to the basics of infection control. Indian J Med Res 2011;134:611–20 doi: 10.4103/0971-5916.90985[published Online First: Epub Date]|.
Nystrom B. Impact of Handwashing on Mortality in Intensive Care: Examination of the Evidence. Infection Control & Hospital Epidemiology 1994;15(7):435–36 doi: 10.2307/30148491[published Online First: Epub Date]|.
Simmons B, Bryant J, Neiman K, Spencer L, Arheart K. The Role of Handwashing in Prevention of Endemic Intensive Care Unit Infections. Infection Control & Hospital Epidemiology 1990;11(11):589–94 doi: 10.2307/30148433[published Online First: Epub Date]|.
Daniels IR, Rees BI. Handwashing: simple, but effective. Ann R Coll Surg Engl 1999;81(2):117–18
Kampf G, Kramer A. Epidemiologic Background of Hand Hygiene and Evaluation of the Most Important Agents for Scrubs and Rubs. Clinical Microbiology Reviews 2004;17(4):863 doi: 10.1128/CMR.17.4.863-893.2004[published Online First: Epub Date]|.
Allegranzi B, Pittet D. Role of hand hygiene in healthcare-associated infection prevention. Journal of Hospital Infection 2009;73(4):305–15 doi: 10.1016/j.jhin.2009.04.019[published Online First: Epub Date]|.
Barrera L, Zingg W, Mendez F, Pittet D. Effectiveness of a hand hygiene promotion strategy using alcohol-based handrub in 6 intensive care units in Colombia. American Journal of Infection Control 2011;39(8):633 – 39 doi: https://doi.org/10.1016/j.ajic.2010.11.004[published Online First: Epub Date]|.
Loftus MJ, Guitart C, Tartari E, et al. Hand hygiene in low- and middle-income countries. International Journal of Infectious Diseases 2019;86:25–30 doi: https://doi.org/10.1016/j.ijid.2019.06.002[published Online First: Epub Date]|.
Munoz-Price LS, Birnbach DJ. Hand Hygiene and Anesthesiology. International Anesthesiology Clinics 2013;51(1):79–92 doi: 10.1097/AIA.0b013e31826f2db6[published Online First: Epub Date]|.
Pittet D, Simon A, Hugonnet S, Pessoa-Silva CL, Sauvan V, Perneger TV. Hand Hygiene among Physicians: Performance, Beliefs, and Perceptions. Annals of Internal Medicine 2004;141(1):1–8 doi: 10.7326/0003-4819-141-1-200407060-00008[published Online First: Epub Date]|.
Voss A, Widmer AF. No Time for Handwashing!? Handwashing Versus Alcoholic Rub Can We Afford 100% Compliance? Infection Control & Hospital Epidemiology 1997;18(3):205–08 doi: 10.2307/30141985[published Online First: Epub Date]|.
Dallolio L, Raggi A, Sanna T, et al. Surveillance of Environmental and Procedural Measures of Infection Control in the Operating Theatre Setting. Int J Environ Res Public Health 2017;15(1) doi: 10.3390/ijerph15010046[published Online First: Epub Date]|.
Loftus R, Muffly M, Brown J, et al. Hand Contamination of Anesthesia Providers Is an Important Risk Factor for Intraoperative Bacterial Transmission. Survey of Anesthesiology 2011;55:314 – 15 doi: 10.1097/01.SA.0000407050.32901.1c[published Online First: Epub Date]|.
Smith B. Anesthesia as a Risk for Health Care Acquired Infections. Perioperative Nursing Clinics 2010;5:427–41 doi: 10.1016/j.cpen.2010.07.005[published Online First: Epub Date]|.
Zorrilla-Vaca A, Escandón-Vargas K. The importance of infection control and prevention in anesthesiology. Revista Colombiana de Anestesiología 2017;45:69–77
Boyce JM, Pittet D, Healthcare Infection Control Practices Advisory Committee. Society for Healthcare Epidemiology of America. Association for Professionals in Infection Control. Infectious Diseases Society of America. Hand Hygiene Task F. Guideline for Hand Hygiene in Health-Care Settings: recommendations of the Healthcare Infection Control Practices Advisory Committee and the HICPAC/SHEA/APIC/IDSA Hand Hygiene Task Force. Infect Control Hosp Epidemiol 2002;23(12 Suppl):S3-40 doi: 10.1086/503164[published Online First: Epub Date]|.
Organization WH. WHO Guidelines on Hand Hygiene in Health Care: First Global Patient Safety Challenge Clean Care Is Safer Care. Geneva, 2009.
Birnbach D, Thiesen T, McKenty N, et al. Targeted Use of Alcohol-Based Hand Rub on Gloves During Task Dense Periods: One Step Closer to Pathogen Containment by Anesthesia Providers in the Operating Room. Anesthesia & Analgesia 2019;129:1 doi: 10.1213/ANE.0000000000004107[published Online First: Epub Date]|.
Munoz-Price LS, Weinstein R. Fecal Patina in the Anesthesia Work Area. Anesthesia and analgesia 2015;120:703–5 doi: 10.1213/ANE.0000000000000542[published Online First: Epub Date]|.
Maury E, Moussa N, Lakermi C, Barbut F, Offenstadt G. Compliance of health care workers to hand hygiene: Awareness of being observed is important [8]. Intensive care medicine 2007;32:2088–9 doi: 10.1007/s00134-006-0398-9[published Online First: Epub Date]|.
Lenglet A, van Deursen B, Viana R, et al. Inclusion of Real-Time Hand Hygiene Observation and Feedback in a Multimodal Hand Hygiene Improvement Strategy in Low-Resource Settings. JAMA Netw Open 2019;2(8):e199118 doi: 10.1001/jamanetworkopen.2019.9118[published Online First: Epub Date]|.
Fuller C, Michie S, Savage J, et al. The Feedback Intervention Trial (FIT)--improving hand-hygiene compliance in UK healthcare workers: a stepped wedge cluster randomised controlled trial. PLoS One 2012;7(10):e41617 doi: 10.1371/journal.pone.0041617[published Online First: Epub Date]|.
23(12):974 doi: 10.1136/bmjqs-2014-003080[published Online First: Epub Date]|.
Bellaard-Smith ER, Gillespie EE. Implementing hand hygiene strategies in the operating suite. Healthcare Infection 2012;17(1):33–37 doi: 10.1071/HI12002[published Online First: Epub Date]|.
Haque A, Guo M, Alahi A, et al. Towards Vision-Based Smart Hospitals: A System for Tracking and Monitoring Hand Hygiene Compliance. 2017
Vision-Based Hand Hygiene Monitoring in Hospitals. AMIA; 2016.
Carreira J, Zisserman A. Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset, 2017.
Hara K, Kataoka H, Satoh Y. Can Spatiotemporal 3D CNNs Retrace the History of 2D CNNs and ImageNet?, 2018.
Ji S, Xu W, Yang M, Yu K. 3D Convolutional Neural Networks for Human Action Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 2013;35(1):221–31 doi: 10.1109/TPAMI.2012.59[published Online First: Epub Date]|.
Gu C, Sun C, Vijayanarasimhan S, et al. AVA: A Video Dataset of Spatio-temporally Localized Atomic Visual Actions. 2017
Soomro K, Zamir A, Shah M. UCF101: A Dataset of 101 Human Actions Classes From Videos in The Wild. CoRR 2012
Kuehne H, Jhuang H, Garrote E, Poggio T, Serre T. HMDB51: A Large Video Database for Human Motion Recognition, 2011.
Kay W, Carreira J, Simonyan K, et al. The Kinetics Human Action Video Dataset. 2017
Ren S, He K, Girshick R, Sun J. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks, 2016.
He K, Gkioxari G, Dollar P, Girshick R. Mask R-CNN. IEEE Transactions on Pattern Analysis and Machine Intelligence 2018;PP:1–1 doi: 10.1109/TPAMI.2018.2844175[published Online First: Epub Date]|.
Torralba A, Russell B, Yuen J. LabelMe: Online Image Annotation and Applications. Proceedings of the IEEE 2010;98:1467-84 doi: 10.1109/JPROC.2010.2050290[published Online First: Epub Date]|.
Lin T-Y, Maire M, Belongie S, et al. Microsoft COCO: Common Objects in Context. 2014
Hoiem D, Divvala S, Hays J. Pascal VOC 2008 Challenge. 2009
Szegedy C, Liu W, Jia Y, et al. Going Deeper with Convolutions. 2014

No competing interests reported.

Download PDF

Journal Publication

published 18 Jun, 2024

Read the published version in Journal of Clinical Monitoring and Computing →

Version 1

posted

You are reading this latest preprint version

Video-based Automatic Hand Hygiene Detection for Operating Rooms using 3D Convolutional Neural Networks

Status:

Journal Publication

Version 1

Abstract

Figures

Introduction

Experiments and RESULTS

DISCUSSION

CONCLUSION

METHODS

Declarations

References

Additional Declarations

Status:

Journal Publication

Version 1