ObjectiveTo systematically review the early clinical prediction value of machine learning (ML) for cardiac arrest (CA).MethodsPubMed, EMbase, WanFang Data and CNKI databases were electronically searched to retrieve all ML studies on predicting CA from January 2015 to February 2021. Two reviewers independently screened literature, extracted data and assessed the risk of bias of included studies. The value of each model was evaluated based on the area under receiver operating characteristic curve (AUC) and accuracy.ResultsA total of 38 studies were included. In terms of data sources, 13 studies were based on public database, and other studies retrospectively collected clinical data, in which 21 directly predicted CA, 3 predicted CA-related arrhythmias, and 9 predicted sudden cardiac death. A total of 51 models had been adopted, among which the most popular ML methods included artificial neural network (n=11), followed by random forest (n=9) and support vector machine (n=5). The most frequently used input feature was electrocardiogram parameters (n=20), followed by age (n=12) and heart rate variability (n=10). Six studies compared the ML models with other traditional statistical models and the results showed that the AUC value of ML was generally higher than that in traditional statistical models.ConclusionsThe available evidence suggests that ML can accurately predict the occurrence of CA, and the performance is significantly superior to traditional statistical model in certain cases.
Machine learning-based diagnostic tests have certain differences of measurement indicators with traditional diagnostic tests. In this paper, we elaborate the definitions, calculation methods and statistical inferences of common measurement indicators of machine learning-based diagnosis models in detail. We hope that this paper will be helpful for clinical researchers to better evaluate machine learning diagnostic models.
ObjectiveTo explore the development and application of a novel ventilator alarm management model in critically ill patients receiving invasive mechanical ventilation (MV) in the intensive care unit (ICU) using machine learning (ML) and Internet of Medical Things (IoMT). The study aims to identify alarms’ intervention requirements. MethodsA retrospective cohort study and ML analysis were conducted, including adult patients receiving invasive MV in the ICU at West China Hospital from February 10, 2024, to July 22, 2024. A total of 76 ventilator alarm-related parameters were collected through the IoMT system. Feature selection was performed using a stratified approach, and six ML algorithms were applied: Gaussian Naive Bayes, K-Nearest Neighbors, Linear Discriminant Analysis, Support Vector Machine, Categorical Boosting (CatBoost), and Logistic Regression. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC-ROC). ResultsA total of 107 patients and their associated ventilator alarm records were included. Thirteen highly relevant features were selected from the 76 parameters for model training through stratified feature selection. The CatBoost model demonstrated the best predictive performance, with an AUC-ROC of 0.984 7 and an accuracy of 0.912 3 in the training set. External validation of the CatBoost model yielded an AUC-ROC of 0.805 4. ConclusionThe CatBoost-based ML model successfully constructed in this study has high accuracy and reliability in predicting the ventilator alarms in ICU patients, providing an effective tool for ventilator alarm management. The CatBoost-based ML method exhibited remarkable efficacy in predicting the necessity of ventilator intervention in critically ill ICU patients. Further large-scale multicenter studies are recommended to validate its clinical application value and promote model optimization and implementation.
Currently, the types of kidney stones before surgery are mainly identified by human beings, which directly leads to the problems of low classification accuracy and inconsistent diagnostic results due to the reliance on human knowledge. To address this issue, this paper proposes a framework for identifying types of kidney stones based on the combination of radiomics and deep learning, aiming to achieve automated preoperative classification of kidney stones with high accuracy. Firstly, radiomics methods are employed to extract radiomics features released from the shallow layers of a three-dimensional (3D) convolutional neural network, which are then fused with the deep features of the convolutional neural network. Subsequently, the fused features are subjected to regularization, least absolute shrinkage and selection operator (LASSO) processing. Finally, a light gradient boosting machine (LightGBM) is utilized for the identification of infectious and non-infectious kidney stones. The experimental results indicate that the proposed framework achieves an accuracy rate of 84.5% for preoperative identification of kidney stone types. This framework can effectively distinguish between infectious and non-infectious kidney stones, providing valuable assistance in the formulation of preoperative treatment plans and the rehabilitation of patients after surgery.
Objective To systematically review prediction models of small for gestational age (SGA) based on machine learning and provide references for the construction and optimization of such a prediction model. Methods The PubMed, EMbase, Web of Science, CBM, WanFang Data, VIP and CNKI databases were electronically searched to collect studies on SGA prediction models from database inception to August 10, 2022. Two researchers independently screened the literature, extracted data, evaluated the risk of bias of the included studies, and conducted a systematic review. Results A total of 14 studies, comprising 40 prediction models constructed using 19 methods, such as logical regression and random forest, were included. The results of the risk of bias assessment from 13 studies were high; the area under the curve of the prediction models ranged from 0.561 to 0.953. Conclusion The overall risk of bias in the prediction models for SGA was high, and the predictive performance was average. Models built using extreme gradient boosting (XGBoost) demonstrated the best predictive performance across different studies. The stacking method can improve predictive performance by integrating different models. Finally, maternal blood pressure, fetal abdominal circumference, head circumference, and estimated fetal weight were important predictors of SGA.
Brain age prediction, as a significant approach for assessing brain health and early diagnosing neurodegenerative diseases, has garnered widespread attention in recent years. Electroencephalogram (EEG), an non-invasive, convenient, and cost-effective neurophysiological signal, offers unique advantages for brain age prediction due to its high temporal resolution and strong correlation with brain functional states. Despite substantial progress in enhancing prediction accuracy and generalizability, challenges remain in data quality and model interpretability. This review comprehensively examined the advancements in EEG-based brain age prediction, detailing key aspects of data preprocessing, feature extraction, model construction, and result evaluation. It also summarized the current applications of machine learning and deep learning methods in this field, analyzed existing issues, and explored future directions to promote the widespread application of EEG-based brain age prediction in both clinical and research settings.
ObjectiveTo systematically evaluate the clinical value of machine learning (ML) for predicting the neurological outcome of out-of-hospital cardiac arrest (OHCA), and to develop a prediction model. MethodsWe searched the PubMed, Web of Science, EMbase, CNKI, Wanfang database from January 1, 2011 to November 24, 2021. Studies on ML for predicting neurological outcomes in OHCA pateints were collected. Two researchers independently screened the literature, extracted the data and evaluated the bias of the included literature, evaluated the accuracy of different models and compared the area under the receiver operating characteristic curve (AUC). ResultsA total of 20 studies were included. Eleven of the studies were from open source databases and nine were from retrospective studies. Sixteen studies directly predicted OHCA neurological outcomes, and four predicted OHCA neurological outcomes after target temperature management. A total of seven ML algorithms were used, among which neural network was the ML algorithm with the highest frequency (n=5), followed by support vector machine and random forest (n=4). Three papers used multiple algorithms. The most frequently used input characteristic was age (n=19), followed by heart rate (n=17) and gender (n=13). A total of 4 studies compared the predictive value of ML with other classical statistical models, and the AUC value of ML model was higher than that of classical statistical models. ConclusionExisting evidence suggests that ML can more accurately predict OHCA nervous system outcomes, and the predictive performance of ML is superior to traditional statistical models in certain situations.
Sleep apnea causes cardiac arrest, sleep rhythm disorders, nocturnal hypoxia and abnormal blood pressure fluctuations in patients, which eventually lead to nocturnal target organ damage in hypertensive patients. The incidence of obstructive sleep apnea hypopnea syndrome (OSAHS) is extremely high, which seriously affects the physical and mental health of patients. This study attempts to extract features associated with OSAHS from 24-hour ambulatory blood pressure data and identify OSAHS by machine learning models for the differential diagnosis of this disease. The study data were obtained from ambulatory blood pressure examination data of 339 patients collected in outpatient clinics of the Chinese PLA General Hospital from December 2018 to December 2019, including 115 patients with OSAHS diagnosed by polysomnography (PSG) and 224 patients with non-OSAHS. Based on the characteristics of clinical changes of blood pressure in OSAHS patients, feature extraction rules were defined and algorithms were developed to extract features, while logistic regression and lightGBM models were then used to classify and predict the disease. The results showed that the identification accuracy of the lightGBM model trained in this study was 80.0%, precision was 82.9%, recall was 72.5%, and the area under the working characteristic curve (AUC) of the subjects was 0.906. The defined ambulatory blood pressure features could be effectively used for identifying OSAHS. This study provides a new idea and method for OSAHS screening.
ObjectiveTo establish a predictive model of surgical site infection (SSI) following colorectal surgery using machine learning.MethodsMachine learning algorithm was used to analyze and model with the colorectal data set from Duke Infection Control Outreach Network Surveillance Network. The whole data set was divided into two parts, with 80% as the training data set and 20% as the testing data set. In order to improve the training effect, the whole data set was divided into two parts again, with 90% as the training data set and 10% as the testing data set. The predictive result of the model was compared with the actual infected cases, and the sensitivity, specificity, positive predictive value, and negative predictive value of the model were calculated, the area under receiver operating characteristic (ROC) curve was used to evaluate the predictive capacity of the model, odds ratio (OR) was calculated to tested the validity of evaluation with a significance level of 0.05.ResultsThere were 7 285 patients in the whole data set registered from January 15th, 2015 to June 16th, 2016, among whom 234 were SSI cases, with an incidence of SSI of 3.21%. The predictive model was established by random forest algorithm, which was trained by 90% of the whole data set and tested by 10% of that. The sensitivity, specificity, positive predictive value, and negative predictive value of the model were 76.9%, 59.2%, 3.3%, and 99.3%, respectively, and the area under ROC curve was 0.767 [OR=4.84, 95% confidence interval (1.32, 17.74), P=0.02].ConclusionThe predictive model of SSI following colorectal surgery established by random forest algorithm has the potential to realize semi-automatic monitoring of SSIs, but more data training should be needed to improve the predictive capacity of the model before clinical application.
ObjectiveTo develop a machine learning model to identify preoperative, intraoperative, and postoperative high-risk factors of laparoscopic inguinal hernia repair (LHR) and to predict recurrent hernia. Methods The patients after LHR from 2010 to 2018 were included. Twenty-nine characteristic variables were collected, including patient demographic characteristics, chronic medical history, laboratory test characteristics, surgical information, and postoperative status of the patients. Four machine learning algorithms, including extreme gradient boosting (XGBoost), random forest (RF), support vector machine (SVM), and k-nearest neighbor algorithm (KNN), were used to construct the model. We also applied Shapley additive explanation (SHAP) for visual interpretation of the model and evaluated the model using the k-fold cross-validation method, receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA). ResultsA total of 1 178 patients with inguinal hernias were included in the study, including 114 patients with recurrent hernias. The XGBoost algorithm showed the best performance among the four prediction models. The ROC curve results showed that the area under the curve (AUC) value of XGBoost was 0.985 in the training set and 0.917 in the validation set, which showed high prediction accuracy. The K-fold cross-validation method, calibration curve, and DCA curve showed that the XGBoost model was stable and clinically useful. The AUC value in the independent validation set was 0.86, indicating that the XGBoost prediction model has good extrapolation. The results of SHAP analysis showed that mesh size, mesh fixtion, diabetes, hypoproteinemia, obesity, smoking history, low intraoperative percutaneous arterial oxygen saturation (SpO2), and low intraoperative body temperature were strongly associated with recurrent hernia. ConclusionThe predictive model of recurrent hernia after LHR in patients derived from the XGBoost machine learning algorithm in this study can assist clinicians in clinical decision making.