Deep Learning Model Detecting Pressure Injury

Overview

In the health care system, pressure injuries, which are among the quality indicators, are a serious patient safety problem that affects the length of hospital stay and the cost of care. Pressure injuries are generally defined as localized injuries caused by pressure on bony prominences or by shear force combined with pressure. This health problem reduces the quality of life of the patient and their family, causes the individual to be socially isolated , requires more intensive and prolonged nursing care, and can cause mortality , morbidity and nosocomial infections if appropriate treatment and care are not provided . systematic staging of pressure injuries positively directs the treatment process and the patient's prognosis . Correct staging of pressure injuries not only affects patient care outcomes but also increases the quality of nursing care provided by providing a common language among nurses.Today, with the increasing use of technology, it is seen that larger data is needed to solve complex problems. In order to meet this need, Convolutional Neural Networks have emerged, which are used in many areas such as object recognition, speech recognition, and natural language processing, and can automatically learn from the symbols of data belonging to images, videos, audio, and texts, instead of learning with coded rules, unlike traditional machine learning methods, based on Artificial Neural Networks. Convolutional Neural Networks are one of the Deep Learning methods, which is a sub-branch of machine learning methods and has the ability to learn from examples. Convolutional Neural Networks are methods that can also learn from raw image or text data and whose prediction accuracy increases according to the size of the data. It has been proven in the literature that artificial intelligence and deep learning models are effective in the risk analysis of pressure injuries. However , no study has been found on the classification of pressure injuries. In light of this information, the study was conducted to develop a deep learning model in the detection and classification of pressure injuries and to determine the effect of the model on the knowledge and satisfaction levels of nurses.

Today, with the increasing use of technology, it is seen that larger data is needed to solve complex problems. In order to meet this need, Convolutional Neural Networks have emerged, which are used in many areas such as object recognition, speech recognition, and natural language processing, and can automatically learn from the symbols of data belonging to images, videos, audio, and texts, instead of learning with coded rules, unlike traditional machine learning methods, based on Artificial Neural Networks. Convolutional Neural Networks are one of the Deep Learning methods, which is a sub-branch of machine learning methods and has the ability to learn from examples. Convolutional Neural Networks are methods that can also learn from raw image or text data and whose prediction accuracy increases according to the size of the data. It has been proven in the literature that artificial intelligence and deep learning models are effective in the risk analysis of pressure injuries. However , no study has been found on the classification of pressure injuries. Raju , Su, Patrician et al. (2015), provided more accurate and faster prediction of Braden risk scale scores with the deep learning model they developed as a result of a four-year follow-up in a military hospital . Alderden , Pepper , Wilson et al. (2018) developed a deep learning model that reveals the risk analysis of pressure injuries in intensive care patients, and provided more accurate and meaningful pressure injury risk analysis with more sensitive measurements for intensive care patients who are considered high risk according to risk assessment tools. Demircan, Yücedağ , Toz et al. (2016) developed a mathematical model that analyzes the risk factors in the formation process of pressure injuries, and ensured that pressure injuries were detected at an early stage. The use of these innovative applications, which are included and used in the world literature , is limited in our country. In an environment where technology is rapidly developing and consumed, not remaining indifferent to innovative initiatives and integrating technology into nursing practices will increase the visibility of our profession by training innovative nurses. In light of this information, the study was conducted to develop a deep learning model in the detection and classification of pressure injuries and to determine the effect of the model on the knowledge and satisfaction levels of nurses. Aim A randomized controlled experimental study was conducted to develop a deep learning model for the detection and classification of pressure injuries and to determine the effect of the model on the knowledge and satisfaction levels of nurses . Hypotheses of the Research; H1: Deep Learning Model Provides Detection of Pressure Sores. H2: Deep Learning Model Provides Classification of Pressure Sores. H3: Mobile Application Developed with Deep Learning Model Plays an Active Role in Pressure Sore Treatment and Care. Variables of the Study; Independent variables of the study; deep learning model, braden risk assessment Dependent variables of the study; pressure sore knowledge level, satisfaction level with education Method Research is a randomized controlled experimental study. Universe and Sample of the Research consisted of 80 nurses working in the intensive care , internal medicine and surgery clinics of a foundation university hospital between March 2021 and June 2022 . In order to determine the number of nurses constituting the research sample, a similar study was taken as an example in the relevant literature and the power analysis G\* Power 3.1 software was used. In this study, in order to reach a power level of 95% at a 0.5 effect size and 5% error level, the sample size was calculated as 56, with 28 participants in each group. Considering the high power of the test and the losses in the study , a total of 60 people were reached, 30 in each group. The nurses in the sample group were randomly assigned to the control (n=30) and experimental groups (n=30) by the researcher using a simple number table. The randomization table was created by using the website http://stattrek.com/statistics/random-number-generator.aspx for randomization . As a result, the sample of the study consisted of a total of 60 nurses, 30 participants in each nurse group, who were informed about the purpose of the study and who were allowed to participate in the study and who met the sample criteria of the study. The sample criteria were ; the nurse was over 18 years old, worked as an intensive care or clinic nurse, and accepted to participate in the study verbally and in writing. Data Collection Tools Research data, "Structured Nurse Introduction Form", " Modified Pieper Pressure Sore Knowledge Test", " Braden Risk Assessment Scale" and "Nursing Satisfaction Scale" were used. Structured Nurse Introduction Form: literature information on the subject and includes information about the introductory characteristics of nurses. Modified Pieper Pressure Sore Knowledge Test: As a result of the research, the Modified Scale was developed by Pieper and Mott in 1995, modified by Lawrence , and its validity and reliability were determined by Asiye Gül and her colleagues in 2017. Pieper Pressure Wound Knowledge Test was used. This test consists of 49 items. The scale is divided into three sub-dimensions. The general knowledge score can be up to 49 points, the prevention knowledge score can be up to 33 points, the staging knowledge score can be up to 9 points, and the wound identification score can be up to 7 points. Modified Permission was requested from Prof. Dr. Asiye Gül for the Pieper Pressure Sore Knowledge Test. A reliability analysis was performed to determine the reliability level of the scale used in the study and the Chronbach alpha coefficient of the experimental group was obtained as 0.838 and that of the control group as 0.812. Braden Risk Assessment Scale: Braden and Bergstrom was conducted by Oğuz in 1997 in Turkey. A total score ranging from 6 to 23 is obtained from the scale. According to the total score, 12 points and below are considered high risk, 13-14 points are considered risky, and 15-16 points are considered low risk. Nurse Satisfaction Survey: Likert -type questions prepared by researchers in line with the literature to determine the satisfaction levels of nurses with pressure injury training . The lowest score is 0 and the highest is 25. Reliability analysis was performed to determine the reliability level of the questionnaire used in the study and the Chronbach alpha coefficient was obtained as 0.95. The Deep Learning Model We Developed for Detection and Classification of Pressure Injuries The "Deep Learning Model for Detection and Classification of Pressure Injuries (BYT-CNN model)" that we developed was first trained on 175 sample patient images and showed a classification prediction success rate of approximately 97%. Since the BYT-CNN model was thought to perform better with more patient images; it was continued to be developed using 500 patient and 500 non-patient image data to solve the classification problems of both estimating whether the disease is present or not and the stage of the disease (Figure 1). This model will provide risk analysis of pressure injuries and their classification according to the "NPUAP Pressure Injury Classification System" in cases where they develop . A mobile application that includes nursing interventions recommended by NPUAP according to the stage determined by the deep learning model was developed and the treatment and care of the pressure injury was also carried out (Figure 2). In order to detect and classify pressure injuries using the BYT-CNN model; a picture of the pressure injury or the area at risk is taken and connected to the main computer where deep learning is located via Bluetooth or Wi- Fi. Then, whether the patient has a pressure injury and, if so, what stage it is, is reported to the nurse's phone within "3 seconds". The mobile application, prepared in line with NPUAP 2016 pressure injury treatment and care recommendations, informs the nurse about which care to provide for which stage. The Method Followed in Collecting Data Before starting to collect data, the researchers developed a "Deep Learning Model for Detection and Classification of Pressure Injuries (BYT-CNN model). All nurses participating in the study were given a 4-hour theoretical lesson on "Pressure Injuries" in a classroom environment at the same time. One week after all nurses received theoretical training, they were divided into experimental (Deep Learning Model) and control (Traditional Method) groups. Both groups were given a "Structured Nurse Introduction Form" and a "Pre-Test ( Modified "Pieper Pressure Sore Knowledge Test)" was applied. Application in Control Group: After the theoretical lesson, the nurses in the control group determined and classified the pressure injuries in their patients using the " Braden Risk Assessment Scale", which has been accepted as valid and reliable. The nurses in the control group, who determined the pressure injuries using the scale, were given training on the determination and classification of pressure injuries using written material, the content of which was prepared by the researchers. After the training, the "Satisfaction with the Training Method Survey" was applied to the nurses. One week after the training, the nurses were given the "Post-Test ( Modified "Pieper Pressure Sore Knowledge Test)" was applied. After the completion of the application, volunteer nurses from the control group were subjected to pressure injury detection and classification with the deep learning model and trained with the mobile application. Application in the Experimental Group: After the theoretical course, the nurses in the experimental group detected and classified pressure injuries in their patients with the "Deep Learning Model". In the experimental group, a mobile application developed by the researchers was installed on the phones of the nurses who detected pressure injuries using the deep learning model and training was applied. Thus, the nurses were provided with the patient's care and treatment according to the developed mobile application according to the pressure injury stage detected by the deep learning model. After the training, the "Satisfaction Survey with the Training Method" was applied to the nurses. 1 week after the training, the nurses were given the "Post-Test ( Modified "Pieper Pressure Sore Knowledge Test)" was applied Ethical and Legal Aspects of Research In order to conduct the research, the necessary interviews and correspondence were made and institutional permissions and ethics committee approval were obtained. Ethics committee permission was obtained from the Istinye University Human Research Ethics Committee with decision number 2104 dated 27.01.2021 . Institutional permission; In order to conduct the research, an application permit was obtained from the Istanbul Beykent University Hospital Chief Physician's Office on 08.02.2022 with document number 14. Informed consent forms were obtained from the individuals participating in the research. Three of the five researchers who completed the implementation phase of the research work in the Department of Computer Engineering. The other two researchers work in the Department of Nursing (two researchers are faculty members in Computer Engineering, one researcher is a faculty member in Fundamentals of Nursing, one researcher is a software engineer, and one researcher is an intensive care nurse). Analysis of Data The data obtained in the study were obtained using SPSS (Statistical Package for Social Sciences ) for Windows 25.0 program was used for analysis. Descriptive statistical methods were used for data evaluation as number, percentage, mean, standard deviation. Differences between proportions of categorical variables in independent groups were analyzed using Chi-Square and Fisher It was analyzed with exact tests. T-test was used to compare quantitative continuous data between two independent groups. Dependent groups t-test was used to compare within-group measurements.