A real-world exploration into clinical outcomes of direct oral anticoagulant therapy in people with chronic kidney disease: a large hospital-based study.

BACKGROUND: There is limited evidence to support definite clinical outcomes of direct oral anticoagulant (DOAC) therapy in chronic kidney disease (CKD). By identifying the important variables associated with clinical outcomes following DOAC administration in patients in different stages of CKD, this study aims to assess this evidence gap. METHODS: An anonymised dataset comprising 97,413 patients receiving DOAC therapy in a tertiary health setting was systematically extracted from the multidimensional electronic health records and prepared for analysis. Machine learning classifiers were applied to the prepared dataset to select the important features which informed covariate selection in multivariate logistic regression analysis. RESULTS: For both CKD and non-CKD DOAC users, features such as length of stay, treatment days, and age were ranked highest for relevance to adverse outcomes like death and stroke. Patients with Stage 3a CKD had significantly higher odds of ischaemic stroke (OR 2.45, 95% Cl: 2.10-2.86; p = 0.001) and lower odds of all-cause mortality (OR 0.87, 95% Cl: 0.79-0.95; p = 0.001) on apixaban therapy. In patients with CKD (Stage 5) receiving apixaban, the odds of death were significantly lowered (OR 0.28, 95% Cl: 0.14-0.58; p = 0.001), while the effect on ischaemic stroke was insignificant. CONCLUSIONS: A positive effect of DOAC therapy was observed in advanced CKD. Key factors influencing clinical outcomes following DOAC administration in patients in different stages of CKD were identified. These are crucial for designing more advanced studies to explore safer and more effective DOAC therapy for the population.

Evaluation of High-Frequency Measurement Errors from Turned Surface Topography Data Using Machine Learning Methods.

Manufacturing processes in industry applications are often controlled by the evaluation of surface topography. Topography, in its overall performance, includes form, waviness, and roughness. Methods of measurement of surface roughness can be roughly divided into tactile and contactless techniques. The latter ones are much faster but sensitive to external disturbances from the environment. One type of external source error, while the measurement of surface topography occurs, is a high-frequency noise. This noise originates from the vibration of the measuring system. In this study, the methods for reducing high-frequency errors from the results of contactless roughness measurements of turned surfaces were supported by machine learning methods. This research delves into optimizing filtration methods for surface topography measurements through the application of machine learning models, focusing on enhancing the accuracy of surface roughness assessments. By examining turned surfaces under specific machining conditions and employing a variety of digital filters, the study identifies the Gaussian regression filter and spline filter as the most effective methods at a 22.5 µm cut-off. Utilizing neural networks, support vector machines, and decision trees, the research demonstrates the superior performance of SVMs, achieving remarkable accuracy and sensitivity in predicting optimal filtration methods.

Many Models, Little Adoption-What Accounts for Low Uptake of Machine Learning Models for Atrial Fibrillation Prediction and Detection?

(1) Background: Atrial fibrillation (AF) is a major risk factor for stroke and is often underdiagnosed, despite being present in 13-26% of ischemic stroke patients. Recently, a significant number of machine learning (ML)-based models have been proposed for AF prediction and detection for primary and secondary stroke prevention. However, clinical translation of these technological innovations to close the AF care gap has been scant. Herein, we sought to systematically examine studies, employing ML models to predict incident AF in a population without prior AF or to detect paroxysmal AF in stroke cohorts to identify key reasons for the lack of translation into the clinical workflow. We conclude with a set of recommendations to improve the clinical translatability of ML-based models for AF. (2) Methods: MEDLINE, Embase, Web of Science, Clinicaltrials.gov, and ICTRP databases were searched for relevant articles from the inception of the databases up to September 2022 to identify peer-reviewed articles in English that used ML methods to predict incident AF or detect AF after stroke and reported adequate performance metrics. The search yielded 2815 articles, of which 16 studies using ML models to predict incident AF and three studies focusing on ML models to detect AF post-stroke were included. (3) Conclusions: This study highlights that (1) many models utilized only a limited subset of variables available from patients' health records; (2) only 37% of models were externally validated, and stratified analysis was often lacking; (3) 0% of models and 53% of datasets were explicitly made available, limiting reproducibility and transparency; and (4) data pre-processing did not include bias mitigation and sufficient details, leading to potential selection bias. Low generalizability, high false alarm rate, and lack of interpretability were identified as additional factors to be addressed before ML models can be widely deployed in the clinical care setting. Given these limitations, our recommendations to improve the uptake of ML models for better AF outcomes include improving generalizability, reducing potential systemic biases, and investing in external validation studies whilst developing a transparent modeling pipeline to ensure reproducibility.

m1A-Ensem: accurate identification of 1-methyladenosine sites through ensemble models.

BACKGROUND: 1-methyladenosine (m1A) is a variant of methyladenosine that holds a methyl substituent in the 1st position having a prominent role in RNA stability and human metabolites. OBJECTIVE: Traditional approaches, such as mass spectrometry and site-directed mutagenesis, proved to be time-consuming and complicated. METHODOLOGY: The present research focused on the identification of m1A sites within RNA sequences using novel feature development mechanisms. The obtained features were used to train the ensemble models, including blending, boosting, and bagging. Independent testing and k-fold cross validation were then performed on the trained ensemble models. RESULTS: The proposed model outperformed the preexisting predictors and revealed optimized scores based on major accuracy metrics. CONCLUSION: For research purpose, a user-friendly webserver of the proposed model can be accessed through https://taseersuleman-m1a-ensem1.streamlit.app/ .

STW-MD: a novel spatio-temporal weighting and multi-step decision tree method for considering spatial heterogeneity in brain gene expression data.

Gene expression during brain development or abnormal development is a biological process that is highly dynamic in spatio and temporal. Previous studies have mainly focused on individual brain regions or a certain developmental stage. Our motivation is to address this gap by incorporating spatio-temporal information to gain a more complete understanding of brain development or abnormal brain development, such as Alzheimer's disease (AD), and to identify potential determinants of response. In this study, we propose a novel two-step framework based on spatial-temporal information weighting and multi-step decision trees. This framework can effectively exploit the spatial similarity and temporal dependence between different stages and different brain regions, and facilitate differential gene analysis in brain regions with high heterogeneity. We focus on two datasets: the AD dataset, which includes gene expression data from early, middle and late stages, and the brain development dataset, spanning fetal development to adulthood. Our findings highlight the advantages of the proposed framework in discovering gene classes and elucidating their impact on brain development and AD progression across diverse brain regions and stages. These findings align with existing studies and provide insights into the processes of normal and abnormal brain development.

Prediction of Diabetes Using Data Mining and Machine Learning Algorithms: A Cross-Sectional Study.

OBJECTIVES: This study aimed to develop a model to predict fasting blood glucose status using machine learning and data mining, since the early diagnosis and treatment of diabetes can improve outcomes and quality of life. METHODS: This crosssectional study analyzed data from 3376 adults over 30 years old at 16 comprehensive health service centers in Tehran, Iran who participated in a diabetes screening program. The dataset was balanced using random sampling and the synthetic minority over-sampling technique (SMOTE). The dataset was split into training set (80%) and test set (20%). Shapley values were calculated to select the most important features. Noise analysis was performed by adding Gaussian noise to the numerical features to evaluate the robustness of feature importance. Five different machine learning algorithms, including CatBoost, random forest, XGBoost, logistic regression, and an artificial neural network, were used to model the dataset. Accuracy, sensitivity, specificity, accuracy, the F1-score, and the area under the curve were used to evaluate the model. RESULTS: Age, waist-to-hip ratio, body mass index, and systolic blood pressure were the most important factors for predicting fasting blood glucose status. Though the models achieved similar predictive ability, the CatBoost model performed slightly better overall with 0.737 area under the curve (AUC). CONCLUSIONS: A gradient boosted decision tree model accurately identified the most important risk factors related to diabetes. Age, waist-to-hip ratio, body mass index, and systolic blood pressure were the most important risk factors for diabetes, respectively. This model can support planning for diabetes management and prevention.

Factors related to suicidal ideation of schizophrenia patients in China: a study based on decision tree and logistic regression model.

This study aimed to investigate the factors associated with suicidal ideation in schizophrenia patients in China using decision tree and logistic regression models. From October 2020 to March 2022, patients with schizophrenia were chosen from Chifeng Anding Hospital and Daqing Third Hospital in Heilongjiang Province. A total of 300 patients with schizophrenia who met the inclusion criteria were investigated by questionnaire. The questionnaire covered general data, suicidal ideation, childhood trauma, social support, depressive symptoms and psychological resilience. Logistic regression analysis revealed that childhood trauma and depressive symptoms were risk factors for suicidal ideation in schizophrenia (OR = 2.330, 95%CI: 1.177 ~ 4.614; OR = 10.619, 95%CI: 5.199 ~ 21.688), while psychological resilience was a protective factor for suicidal ideation in schizophrenia (OR = 0.173, 95%CI: 0.073 ~ 0.409). The results of the decision tree model analysis demonstrated that depressive symptoms, psychological resilience and childhood trauma were influential factors for suicidal ideation in patients with schizophrenia (p < 0.05). The area under the ROC for the logistic regression model and the decision tree model were 0.868 (95% CI: 0.821 ~ 0.916) and 0.863 (95% CI: 0.814 ~ 0.912) respectively, indicating excellent accuracy of the models. Meanwhile, the logistic regression model had a sensitivity of 0.834 and a specificity of 0.743 when the Youden index was at its maximum. The decision tree model had a sensitivity of 0.768 and a specificity of 0.8. Decision trees in combination with logistic regression models are of high value in the study of factors influencing suicidal ideation in schizophrenia patients.

CARDPSoML: Comparative approach to analyze and predict cardiovascular disease based on medical report data and feature fusion approach.

Background and Aims: Diabetes patients are at high risk for cardiovascular disease (CVD), which makes early identification and prompt management essential. To diagnose CVD in diabetic patients, this work attempts to provide a feature-fusion strategy employing supervised learning classifiers. Methods: Preprocessing patient data is part of the method, and it includes important characteristics connected to diabetes including insulin resistance and blood glucose levels. Principal component analysis and wavelet transformations are two examples of feature extraction techniques that are used to extract pertinent characteristics. The supervised learning classifiers, such as neural networks, decision trees, and support vector machines, are then trained and assessed using these characteristics. Results: Based on the area under the receiver operating characteristic curve, sensitivity, specificity, and accuracy, these classifiers' performance is closely evaluated. The assessment findings show that the classifiers have a good accuracy and area under the receiver operating characteristic curve value, suggesting that the suggested strategy may be useful in diagnosing CVD in patients with diabetes. Conclusion: The recommended method shows potential as a useful tool for developing clinical decision support systems and for the early detection of CVD in diabetes patients. To further improve diagnostic skills, future research projects may examine the use of bigger and more varied datasets as well as different machine learning approaches. Using an organized strategy is a crucial first step in tackling the serious problem of CVD in people with diabetes.

Repeated Decision Stumping Distils Simple Rules from Single-Cell Data.

Single-cell data afford unprecedented insights into molecular processes. But the complexity and size of these data sets have proved challenging and given rise to a large armory of statistical and machine learning approaches. The majority of approaches focuses on either describing features of these data, or making predictions and classifying unlabeled samples. In this study, we introduce repeated decision stumping (ReDX) as a method to distill simple models from single-cell data. We develop decision trees of depth one-hence "stumps"-to identify in an inductive manner, gene products involved in driving cell fate transitions, and in applications to published data we are able to discover the key players involved in these processes in an unbiased manner without prior knowledge. Our algorithm is deliberately targeting the simplest possible candidate hypotheses that can be extracted from complex high-dimensional data. There are three reasons for this: (1) the predictions become straightforwardly testable hypotheses; (2) the identified candidates form the basis for further mechanistic model development, for example, for engineering and synthetic biology interventions; and (3) this approach complements existing descriptive modeling approaches and frameworks. The approach is computationally efficient, has remarkable predictive power, including in simulation studies where the ground truth is known, and yields robust and statistically stable predictors; the same set of candidates is generated by applying the algorithm to different subsamples of experimental data.

Deep learning based decision tree ensembles for incomplete medical datasets.

BACKGROUND: In practice, the collected datasets for data analysis are usually incomplete as some data contain missing attribute values. Many related works focus on constructing specific models to produce estimations to replace the missing values, to make the original incomplete datasets become complete. Another type of solution is to directly handle the incomplete datasets without missing value imputation, with decision trees being the major technique for this purpose. OBJECTIVE: To introduce a novel approach, namely Deep Learning-based Decision Tree Ensembles (DLDTE), which borrows the bounding box and sliding window strategies used in deep learning techniques to divide an incomplete dataset into a number of subsets and learning from each subset by a decision tree, resulting in decision tree ensembles. METHOD: Two medical domain problem datasets contain several hundred feature dimensions with the missing rates of 10% to 50% are used for performance comparison. RESULTS: The proposed DLDTE provides the highest rate of classification accuracy when compared with the baseline decision tree method, as well as two missing value imputation methods (mean and k-nearest neighbor), and the case deletion method. CONCLUSION: The results demonstrate the effectiveness of DLDTE for handling incomplete medical datasets with different missing rates.

Profile of cardiovascular diseases and physiotherapeutic intervention in a hospital emergency service / Perfil das doenças cardiovasculares e atuação fisioterapeutica em um serviço de emergência hospitalar

Abstract Introduction Cardiovascular disease (CVD) is the lead-ing cause of death globally, with a high proportion of hospitalizations and costs. In view of this, it is essential to understand the main CVDs in patients admitted to hospital emergency services and the role of physiotherapists, in order to plan and direct health services, and to denote participation and encourage specific physiotherapy training in the context of tertiary care. Objective To outline the profile of cardiovascular emergencies and to evaluate physiotherapy in adult patients in the emergency department of a hospital in the interior of the state of São Paulo. Methods This was an observational study which analyzed 1,256 on-call records over a period of eight months. The data collected included age, gender, cardiovascular diagnostic hypothesis and physiotherapy treatment carried out. Results A total of 75 patients with cardiovascular emergencies were included, the most prevalent of which were: heart failure (n = 21), acute coronary syndrome (n = 14), acute myocardial infarction (n = 13), bradyarrhythmia (n = 6) and hypertensive crisis (n = 5). Regarding physiotherapeutic actions and their applications, the most frequent were invasive mechanical ventilation management (n = 34), lung re-expansion maneuvers (n = 17), orotracheal intubation assistance (n = 17), non-invasive mechanical ventilation (n = 14), bronchial hygiene maneuvers (n = 12), kinesiotherapy (n = 10) and sedation (n = 10). Conclusion Heart failure and acute coronary syndrome were the cardiovascular diseases that caused the most admissions to the hospital emergency department and that the procedures with an emphasis on the respiratory system were the most applied.

Resumo Introdução As doenças cardiovasculares (DCV) repre-sentam a principal causa de morte global, destacando-se em internações e gastos. Diante disso, é essencial compreender as principais DCV em pacientes admitidos em serviços de emergência hospitalar e a atuação do fisioterapeuta para planejamento e direcionamento dos serviços de saúde e para denotar a participação e incentivar formações fisioterapêuticas específicas no contexto da atenção terciária. Objetivo Traçar o perfil de emergências cardiovasculares e avaliar a atuação fisioterapêutica em pacientes adultos de serviço de emergência de um hospital no interior do estado de São Paulo. Métodos Trata-se de um estudo observacional, em que foram analisadas 1.256 fichas de passagem de plantão, no período de oito meses. Os dados coletados foram idade, sexo, hipótese diagnóstica cardiovascular e tratamento fisioterapêutico realizado. Resultados Foram incluídos 75 pacientes que apresentavam o perfil de emergências cardiovasculares, sendo as mais prevalentes: insuficiência cardíaca (n = 21), síndrome corona-riana aguda (n = 14), infarto agudo do miocárdio (n = 13), bradarritmia (n = 6) e crise hipertensiva (n = 5). Em relação à atuação fisioterapêutica e suas aplicações, as mais frequentes foram manejo da ventilação mecânica invasiva (n = 34), manobras de reexpansão pulmonar (n = 17), auxílio a intubação orotraqueal (n = 17), ventila-ção mecânica não invasiva (n = 14), manobras de higiene brônquica (n = 12), cinesioterapia (n = 10) e sedestação (n = 10). Conclusão A insuficiência cardíaca e a síndrome coronária aguda foram as doenças cardiovasculares que mais ocasionaram internação no serviço de emergência hospitalar e as condutas com ênfase no aparelho respiratório foram as mais aplicadas.

Dominant flood types in mountains catchments: Identification and change analysis for the landscape planning.

The classification of floods may be a supporting tool for decision-makers in regard to water management, including flood protection. The main objective of this work is the classification of flood generation mechanisms in 28 catchments of the upper Vistula basin. A significant innovation in this study lies in the utilization of decision trees for flood classification. The methodology has so far been applied in the Alpine region. The analysis reveals that peak daily precipitation in the catchments mainly occurs in summer, particularly from June to August. Maximal daily snowmelt typically happens at the end of winter (March to April) and occasionally in November. Winter peaks are observed in March to April and, in some areas, in November to December, while summer peaks occur in May and, in specific catchments, in October. Higher peak flows for annual floods are noted in March to April and June to August. Most annual floods in the Upper Vistula basin are classified as Rain-on-Snow Floods (RoSFs) or Lowland River Floods (LRFs). LRFs contribute from 19% to almost 72%, while RoSFs range from 18% to 75%. In Season 1 (summer), most seasonal floods are identified as LRFs (51%-100%), with very few as RoSFs (0%-46.9%). In Season 2 (winter), the opposite pattern is observed, with most RoSFs (48.4%-97.9%) and fewer LRFs (0%-20.6%). While there are changes in flood patterns, they are not statistically significant. Conducted studies and obtained results can be useful for the preparation of flood prevention documentation and for flood management in general.

Corrigendum: Therapy by physician-pharmacist combination and economic returns for cancer pain management in China: a cost-effectiveness analysis.

[This corrects the article DOI: 10.3389/fphar.2023.1073939.].

On the Relationship between Feature Selection Metrics and Accuracy.

Feature selection metrics are commonly used in the machine learning pipeline to rank and select features before creating a predictive model. While many different metrics have been proposed for feature selection, final models are often evaluated by accuracy. In this paper, we consider the relationship between common feature selection metrics and accuracy. In particular, we focus on misorderings: cases where a feature selection metric may rank features differently than accuracy would. We analytically investigate the frequency of misordering for a variety of feature selection metrics as a function of parameters that represent how a feature partitions the data. Our analysis reveals that different metrics have systematic differences in how likely they are to misorder features which can happen over a wide range of partition parameters. We then perform an empirical evaluation with different feature selection metrics on several real-world datasets to measure misordering. Our empirical results generally match our analytical results, illustrating that misordering features happens in practice and can provide some insight into the performance of feature selection metrics.

Association between biochemical and hematologic factors with COVID-19 using data mining methods.

BACKGROUND AND AIM: Coronavirus disease (COVID-19) is an infectious disease that can spread very rapidly with important public health impacts. The prediction of the important factors related to the patient's infectious diseases is helpful to health care workers. The aim of this research was to select the critical feature of the relationship between demographic, biochemical, and hematological characteristics, in patients with and without COVID-19 infection. METHOD: A total of 13,170 participants in the age range of 35-65 years were recruited. Decision Tree (DT), Logistic Regression (LR), and Bootstrap Forest (BF) techniques were fitted into data. Three models were considered in this study, in model I, the biochemical features, in model II, the hematological features, and in model II, both biochemical and homological features were studied. RESULTS: In Model I, the BF, DT, and LR algorithms identified creatine phosphokinase (CPK), blood urea nitrogen (BUN), fasting blood glucose (FBG), total bilirubin, body mass index (BMI), sex, and age, as important predictors for COVID-19. In Model II, our BF, DT, and LR algorithms identified BMI, sex, mean platelet volume (MPV), and age as important predictors. In Model III, our BF, DT, and LR algorithms identified CPK, BMI, MPV, BUN, FBG, sex, creatinine (Cr), age, and total bilirubin as important predictors. CONCLUSION: The proposed BF, DT, and LR models appear to be able to predict and classify infected and non-infected people based on CPK, BUN, BMI, MPV, FBG, Sex, Cr, and Age which had a high association with COVID-19.

Diagnosis of kidney insufficiency by using the pressure waveforms of wrist-type sphygmomanometers: toward a convenient point-of-care device.

OBJECTIVES: Digital sphygmomanometers have been used for more than 40 years in Western medicine for accurately measuring systolic and diastolic blood pressures, which are vital signs observed for the diagnosis of different diseases. Similarly, traditional Chinese medicine (TCM) has been using wrist pulse diagnosis for thousands of years. Some studies have combined digital wrist pulse signals and the diagnosis method of TCM to quantify pulse waves and identify diseases. However, the effectiveness of this approach is limited because of scattered methods and complex pathological features. Moreover, the literature on TCM does not provide quantitative data or objective indicators. METHODS: In this prospective study, we developed a diagnostic system that contains a modified sphygmomanometer. In addition, we designed a procedure for analyzing pulse waves with 156 features of harmonic modes and a decision tree method for diagnosing kidney insufficiency. RESULTS: In the decision tree method, at least three features of harmonic modes can achieve an accuracy of 0.86, a specificity of 0.91, and a Cohen's kappa coefficient of 0.72. By comparison, the random forest method can achieve an accuracy of 0.99, a specificity of 0.99, and a Cohen's kappa coefficient of 0.94 within 200 trees. The results of this study indicated that even in patients with kidney insufficiency and complex etiology, common features can be distinguished by identifying changes in pulse waveforms. CONCLUSION: By using the modified sphygmomanometer to measure blood pressure, people can monitor their health status and take care of it in advance by simply measuring their blood pressure.

Predictive model for temporomandibular disorder in adolescents: Decision tree.

BACKGROUND: Temporomandibular disorders (TMD) do not only occur in adults but also in adolescents, with negative impacts on their development. AIM: To propose a predictive model for TMD in adolescents using a decision tree (DT) analysis and to identify groups at high and low risk of developing TMD in the city of Recife, PE, Brazil. DESIGN: This cross-sectional study was conducted in Recife on 1342 schoolchildren of both sexes aged 10-17 years. The analyses were performed using Pearson's chi-squared test and Fisher's exact test, as well as the CHAID algorithm for the construction of the DT. The SPSS statistical program was used. RESULTS: The prevalence of TMD was 33.2%. Statistically significant associations were observed between TMD and sex, depression, self-reported orofacial pain, and orofacial pain on clinical examination. The DT consisted of self-reported orofacial pain, orofacial pain on physical examination, and depression, with an overall predictive power of 73.0%. CONCLUSION: The proposed tree has a good predictive capacity and permits to identify groups at high risk of developing TMD among adolescents, such as those with self-reported orofacial pain or orofacial pain on examination associated with depression.

A novel method for continuous measurements of clinical practice guideline adherence.

Introduction: Clinical practice guidelines (hereafter 'guidelines') are crucial in providing evidence-based recommendations for physicians and multidisciplinary teams to make informed decisions regarding diagnostics and treatment in various diseases, including cancer. While guideline implementation has been shown to reduce (unwanted) variability and improve outcome of care, monitoring of adherence to guidelines remains challenging. Real-world data collected from cancer registries can provide a continuous source for monitoring adherence levels. In this work, we describe a novel structured approach to guideline evaluation using real-world data that enables continuous monitoring. This method was applied to endometrial cancer patients in the Netherlands and implemented through a prototype web-based dashboard that enables interactive usage and supports various analyses. Method: The guideline under study was parsed into clinical decision trees (CDTs) and an information standard was drawn up. A dataset from the Netherlands Cancer Registry (NCR) was used and data items from both instruments were mapped. By comparing guideline recommendations with real-world data an adherence classification was determined. The developed prototype can be used to identify and prioritize potential topics for guideline updates. Results: CDTs revealed 68 data items for recording in an information standard. Thirty-two data items from the NCR were mapped onto information standard data items. Four CDTs could sufficiently be populated with NCR data. Conclusion: The developed methodology can evaluate a guideline to identify potential improvements in recommendations and the success of the implementation strategy. In addition, it is able to identify patient and disease characteristics that influence decision-making in clinical practice. The method supports a cyclical process of developing, implementing and evaluating guidelines and can be scaled to other diseases and settings. It contributes to a learning healthcare cycle that integrates real-world data with external knowledge.

On Complexity of Deterministic and Nondeterministic Decision Trees for Conventional Decision Tables from Closed Classes.

In this paper, we consider classes of conventional decision tables closed relative to the removal of attributes (columns) and changing decisions assigned to rows. For tables from an arbitrary closed class, we study the dependence of the minimum complexity of deterministic and nondeterministic decision trees on the complexity of the set of attributes attached to columns. We also study the dependence of the minimum complexity of deterministic decision trees on the minimum complexity of nondeterministic decision trees. Note that a nondeterministic decision tree can be interpreted as a set of true decision rules that covers all rows of the table.

A Hybrid Decision Tree and Deep Learning Approach Combining Medical Imaging and Electronic Medical Records to Predict Intubation Among Hospitalized Patients With COVID-19: Algorithm Development and Validation.

BACKGROUND: Early prediction of the need for invasive mechanical ventilation (IMV) in patients hospitalized with COVID-19 symptoms can help in the allocation of resources appropriately and improve patient outcomes by appropriately monitoring and treating patients at the greatest risk of respiratory failure. To help with the complexity of deciding whether a patient needs IMV, machine learning algorithms may help bring more prognostic value in a timely and systematic manner. Chest radiographs (CXRs) and electronic medical records (EMRs), typically obtained early in patients admitted with COVID-19, are the keys to deciding whether they need IMV. OBJECTIVE: We aimed to evaluate the use of a machine learning model to predict the need for intubation within 24 hours by using a combination of CXR and EMR data in an end-to-end automated pipeline. We included historical data from 2481 hospitalizations at The Mount Sinai Hospital in New York City. METHODS: CXRs were first resized, rescaled, and normalized. Then lungs were segmented from the CXRs by using a U-Net algorithm. After splitting them into a training and a test set, the training set images were augmented. The augmented images were used to train an image classifier to predict the probability of intubation with a prediction window of 24 hours by retraining a pretrained DenseNet model by using transfer learning, 10-fold cross-validation, and grid search. Then, in the final fusion model, we trained a random forest algorithm via 10-fold cross-validation by combining the probability score from the image classifier with 41 longitudinal variables in the EMR. Variables in the EMR included clinical and laboratory data routinely collected in the inpatient setting. The final fusion model gave a prediction likelihood for the need of intubation within 24 hours as well. RESULTS: At a prediction probability threshold of 0.5, the fusion model provided 78.9% (95% CI 59%-96%) sensitivity, 83% (95% CI 76%-89%) specificity, 0.509 (95% CI 0.34-0.67) F1-score, 0.874 (95% CI 0.80-0.94) area under the receiver operating characteristic curve (AUROC), and 0.497 (95% CI 0.32-0.65) area under the precision recall curve (AUPRC) on the holdout set. Compared to the image classifier alone, which had an AUROC of 0.577 (95% CI 0.44-0.73) and an AUPRC of 0.206 (95% CI 0.08-0.38), the fusion model showed significant improvement (P<.001). The most important predictor variables were respiratory rate, C-reactive protein, oxygen saturation, and lactate dehydrogenase. The imaging probability score ranked 15th in overall feature importance. CONCLUSIONS: We show that, when linked with EMR data, an automated deep learning image classifier improved performance in identifying hospitalized patients with severe COVID-19 at risk for intubation. With additional prospective and external validation, such a model may assist risk assessment and optimize clinical decision-making in choosing the best care plan during the critical stages of COVID-19.