Clinical Trial Details
— Status: Not yet recruiting
Administrative data
| NCT number |
NCT05564975 |
| Other study ID # |
2022KYLL053 |
| Secondary ID |
|
| Status |
Not yet recruiting |
| Phase |
|
| First received |
|
| Last updated |
|
| Start date |
October 1, 2022 |
| Est. completion date |
December 1, 2022 |
Study information
| Verified date |
October 2022 |
| Source |
The First People's Hospital of Huzhou |
| Contact |
Pan Huibin licensed doctor, BM |
| Phone |
18767273838 |
| Email |
18767223838[@]126.com |
| Is FDA regulated |
No |
| Health authority |
|
| Study type |
Observational [Patient Registry]
|
Clinical Trial Summary
In the previous investigation, investigators found that when the risk factors of stress
injury in critical patients changed, clinical nurses lacked the awareness of evaluating the
risk of stress injury, and lacked the risk assessment of this link. The stress risk
prediction model is based on etiology. By analyzing the risk factors, the machine learning
algorithm is used to evaluate the risk of pressure damage, and the prediction model of
pressure damage can dynamically and comprehensively evaluate its risk. It is also a risk
assessment tool. At present, there is no research on applying the stress injury risk
prediction model of critical patients to the intensive care information software in China. In
this study, the artificial intelligence algorithm library will be used to construct and apply
the stress injury risk prediction model for critical patients.
Description:
1.1 Epidemiological characteristics of hospital-acquired pressure injuries in critically ill
patients As one of the global health problems, stress injury is a common problem faced by
medical and health institutions, and it is considered as the most serious hospital-acquired
adverse event. Hospital Acquired Pressure Injury (HAPI) refers to the skin and (or) deep
tissue injuries of patients 24 hours after admission. PI can cause pain and worsen the
prognosis of the disease, and in severe cases, it can cause secondary infection and death. At
the same time, it will prolong the hospitalization time, occupy a lot of medical resources
and increase the family financial burden. Critically ill patients are always at high risk of
PI in medical institutions, and its incidence rate is 3.8 times that of non-critically ill
patients. Many risk factors increase the susceptibility of critically ill patients to PI,
such as insufficient blood perfusion, low blood oxygen saturation, use of vasoactive drugs,
mechanical ventilation or renal replacement therapy. The prevalence of PI in critical
patients abroad is 16.9%~23.8%; The prevalence of PI in critically ill patients in China is
5.58% ~ 11.09%; In 2021, Sonia et al. conducted an international multi-center large sample
survey (from 1117 intensive care units and 90 countries/regions), and the prevalence of PI in
critically ill patients was 16.2%. As we all know, PI is a common chronic refractory wound in
clinic, with low cure rate and high treatment cost. Sun Xiaofang followed up 59 PI patients
and found that the wound healing time of stage 3 and above PI was about 37 months, and only
5.1% of the patients' wounds were completely healed, with an average cost of about 6,600
yuan. In 2019, the US spent as much as $2.08 billion on PI for critically ill patients. PI is
also an important reason for the increase of mortality in critically ill patients. As some
critically ill patients are complicated with multiple organ failure, once PI occurs, it is
easier to delay infection and cause sepsis, leading to the death of patients. Therefore, how
to avoid PI in critically ill patients is the primary problem faced by medical workers.
1.2 PI Risk Assessment Tool for Critically Ill Patients Over the years, it has become a
global consensus that prevention of PI is more important than treatment. Risk assessment is
the first step to prevent PI, and the accuracy of its results will directly affect the
selection of preventive measures and preventive effects. At present, there are Braden scale,
Norton scale and Waterlow scale, which are widely used in clinical practice. Norton scale and
Waterlow scale are proved to have the best prediction effect in evaluating elderly patients
and surgical patients. However, when Braden scale is applied to PI risk assessment of
critically ill patients, its predictive ability is moderate (good sensitivity and low
specificity), which leads to the implementation of many non-targeted preventive measures in
clinic and wastes medical resources. As early as 1990s, foreign scholars compiled PI risk
assessment forms for critical patients, including Cubbin &Jackson, SunderLand, SS(Suriadi and
Sanada Scale) and EVARUCI. Ye Tong et al. analyzed the reasons why these scales can't be
widely used. They are compiled according to the characteristics of foreign critical patients,
and some items are not applicable to Chinese population. Secondly, some items need to be
calculated before filling in (such as the range of blood pressure drop), and the steps are
complicated, which increases the workload of medical staff. Some items are subjective
indicators, and the scores of the same patient are quite different, so the ability
requirements of evaluators are strict. At present, there is no special PI evaluation scale
for critically ill patients based on the characteristics of Chinese population. Generally
speaking, the traditional evaluation tools, the evaluation process consumes a lot of
manpower, and the items of the scale also greatly affect the accuracy of the evaluation.
In 2019, Sheng Han et al. collected the data of 278 ICU patients by retrospective
case-control method, and obtained independent risk factors of stress injury by multivariate
analysis: age, edema, hemiplegia, diabetes, acute physiological and chronic health score II,
incontinence, and established a nomogram model of stress injury. The area under ROC curve of
this model is 0.83, which has good stability and discrimination. However, this model has a
small sample size, lacks external verification, and has a high prediction performance bias.
Moreover, diabetes, edema, hemiplegia, incontinence, etc. are all classified variables, which
can't really quantify the relationship between risk factors and stress injuries.
The pathological mechanism of stress injury is complex and dynamic, including internal and
external factors. Dana thinks that tissue tolerance is an intermediate variable between
internal and external factors, which reflects the tolerance of tissue to pressure and oxygen,
and further puts forward the theoretical framework of stress injury prediction
model-patient's own factors (age, comorbidity, gender, body index, etc.) and mechanical
factors (friction/shear force, mobility, mobility, etc.). In 2019, the Prevention and
Treatment of Stress Injury: A Clinical Practice Guide (hereinafter referred to as the "2019
edition of the Clinical Guide for Stress Injury") published by the European Stress Injury
Advisory Committee, the American Stress Injury Advisory Committee and the Pan-Pacific Stress
Injury Alliance proposed that the risk assessment of patients should be Consider the effects
of basic diseases, such as blood perfusion changes caused by diabetes and neuropathy, which
will affect the sensitivity and tolerance of skin. Therefore, when evaluating the risk of
patients' PI, medical workers should comprehensively consider the patients' basic diseases
and various risk factors, and analyze them individually. By analyzing the adverse events of
hospital stress injury in the early stage, it is found that when the condition of critical
patients changes, medical workers' awareness of PI risk assessment is insufficient, which may
be another important reason why critical patients are more prone to PI. How to
comprehensively consider all kinds of risk factors faced by critically ill patients and
evaluate risks in real time, PI risk prediction model brings new hope to solve a difficult
problem.
1.3 the development status of pi risk prediction model Disease risk prediction model has
always been a hot spot in medicine. PI risk prediction model refers to a mathematical model
with stress injury risk factors as variables and machine learning algorithm to predict the
probability of PI. It can quickly, comprehensively and accurately screen out high-risk
patients with PI, and at the same time provide controllable indicators for medical workers,
further promote the implementation of targeted treatment and nursing measures, reduce the
incidence of PI, and reduce medical expenses.
The clinical data of 486 ICU patients in Deng Xiaohong were collected, and the PI risk
prediction model was constructed by using classification regression tree algorithm. The
decision tree model consisted of 4 layers and 11 nodes, and three high-risk groups were
selected: (1) Age > 81 years old; (2) Age ≤8l, with fecal incontinence; (3) Patients with age
≤81 years old, no fecal incontinence, Braden score ≤13, and diastolic blood pressure < 60
mmHg, the area under the ROC curve of this model is 0.82. There is no internal verification
in this study, and the prediction performance of the model may be too high. The selection of
PI risk factors in this study is based on literature review rather than evidence-based, and
it cannot be determined that risk factors are highly correlated with stress injury.
Yu et al. collected the clinical data of inpatients in 2014-2016, took the results of
multi-factor analysis as model prediction variables, and used decision tree, logistic
regression and random forest machine learning algorithm to build prediction models to
evaluate the risk of PI among inpatients. The results showed that the random forest model had
the best prediction performance, and the area under ROC curve was 0.84.
In both studies, a single machine learning algorithm is used to build a stress injury risk
prediction model. The life cycle of machine learning is a process of self-learning and using
known data sets to build a model to predict the occurrence of unknown data events. First, the
best prediction effect can be achieved by selecting an appropriate machine learning algorithm
according to the research objective, variable classification and data set attributes.
Secondly, whether the missing data can be handled correctly will also affect the prediction
performance of the model, because the final effect of the model depends on the amount of data
and the amount of useful information contained in the data. Finally, the verification of the
model is the most critical. Internal verification can prevent the model from over-fitting, so
as to obtain a more reliable and accurate evaluation value. External verification mainly
evaluates the prediction performance of the model through clinical application, and analyzes
and optimizes the model by using newly collected data, so that the model can be continuously
updated dynamically, which provides a basis for the future wide-range use of the model. In
this study, the leading artificial energy intelligent models in the world are used to form an
algorithm library. Based on the standard data established in the early stage of the project,
multiple models in the model library are trained by random and repeated sampling, and the
best integrated prediction model is mined. Artificial intelligence model is an integrated
learning model (it is a big category of machine learning, which is different from single
machine learning model). It completes learning tasks by building multiple learners, mainly
including single model training and multi-model fusion. Ensemble learning can fully make up
for the deficiency of single machine learning algorithm, and its main advantages are as
follows: (1) There is no limit to the types of sub-learners, and the integrated total
learners can be homogeneous (same type of sub-learners) or heterogeneous (different types of
sub-learners), which improves the performance and compatibility of ensemble learning; (2)
Generally, the generalization performance can be significantly superior to that of a single
learner through integration. If the sub-learner is weak, this superiority will be even more
obvious. PI artificial intelligence model mainly transforms clinical data through six steps:
data extraction and cleaning, data set formation, model training, model evaluation and
verification, model interpretation and formation of clinical decision information, which will
make full use of the useful information in the data. This study will summarize the risk
factors of stress injury in critical patients through evidence-based medicine and expert
correspondence, and make use of the advantages of integrated learning model to establish a PI
artificial intelligence prediction model to ensure its high prediction performance.
1.4 PI artificial intelligence model's clinical application prospect In 2020, the Notice of
Further Strengthening Nursing Work in Medical Institutions issued by the National Health and
Health Commission pointed out that it is necessary to further promote the development of
nursing informatization, promote the deep integration of information technology and nursing
work, and establish a nursing management platform with different functions. With the
continuous development of information technology, the application of machine learning
algorithms to collect and utilize big data has become a development trend. Machine learning
algorithm is the core of artificial intelligence. In PI management, by building a PI risk
prediction model, not only It can evaluate the PI risk of critically ill patients, realize
automatic data analysis and promote the development of PI electronic information management.
Park et al. used retrospective research method to collect clinical data of 14,907 inpatients
with electronic medical record database, and established cox risk regression model to predict
patients' PI risk. The area under ROC curve of this model is 0.95, and its prediction effect
is much better than that of Braden scale (area under ROC curve is 0.82). The electronic
medical record system contains a lot of patient information (personal information of
patients, medical records, test results, medication information, etc.). The PI artificial
intelligence prediction model is embedded in the electronic medical record system, which can
transform the patient's records and data into valuable medical information. It is expected to
provide a new direction for early prevention, continuously monitor the risk of patients'
stress injury, and promote medical workers to introduce effective treatment and care. It can
provide appropriate medical advice for patients and their families, facilitate the
implementation of clinical decision-making based on evidence, and ultimately reduce the
occurrence of stress injuries in critically ill patients, improve nursing quality and reduce
medical expenses.It can evaluate the PI risk of critically ill patients, realize automatic
data analysis and promote the development of PI electronic information management. Park et
al. used retrospective research method to collect clinical data of 14,907 inpatients with
electronic medical record database, and established cox risk regression model to predict
patients' PI risk. The area under ROC curve of this model is 0.95, and its prediction effect
is much better than that of Braden scale (area under ROC curve is 0.82). The electronic
medical record system contains a lot of patient information (personal information of
patients, medical records, test results, medication information, etc.). The PI artificial
intelligence prediction model is embedded in the electronic medical record system, which can
transform the patient's records and data into valuable medical information. It is expected to
provide a new direction for early prevention, continuously monitor the risk of patients'
stress injury, and promote medical workers to introduce effective treatment and care. It can
provide appropriate medical advice for patients and their families, facilitate the
implementation of clinical decision-making based on evidence, and ultimately reduce the
occurrence of stress injuries in critically ill patients, improve nursing quality and reduce
medical expenses.
