Background:Septic shock is a severe complication of acute pancreatitis (AP), often associated with poor prognosis. This study aims to analyze the clinical characteristics of patients with acute pancreatitis and develop an interpretable early prediction model for septic shock in these patients using machine learning (ML). The model is intended to assist emergency physicians in resource allocation and medical decision making. Methods:Data were collected from the MIMIC-IV 3.0 database. The dataset was divided into a training set and a test set in a 7:3 ratio. Feature selection was performed using LASSO (Least Absolute Shrinkage and Selection Operator) regression. Subsequently, 10 ML models were developed: Random Forest, Logistic Regression, Gradient Boosting Machine, Neural Network, Extreme Gradient Boosting (XGBoost), K-Nearest Neighbor, Adaptive Boosting, Light Gradient Boosting Machine, Category Boosting, and Support Vector Machine. To enhance and optimize model interpretability, Shapley Additive Explanations (SHAP) were employed. Results:A total of 1032 patients with AP were included in this study, from which 31 variables were selected for model development. By comparing the area under the receiver operating characteristic curve and decision curve analysis results between the training and test sets, the XGBoost model demonstrated a significant advantage over other models. SHAP analysis revealed that white blood cell count, total bilirubin (bilirubin total), and bicarbonate (HCO3 -) levels were the three most critical risk factors for the development of septic shock in patients with AP. Conclusion:ML approaches exhibited promising performance in predicting septic shock in patients with AP. These models may aid in guiding treatment decisions for patients with AP in the emergency department.

Selecting an appropriate population pharmacokinetic (PK) model for individual patients in model-informed precision dosing (MIPD) can be challenging, particularly in the absence of therapeutic drug monitoring (TDM) samples. We developed a machine learning (ML) model to guide individualized PK model selection for a priori MIPD of vancomycin based on routinely recorded patient characteristics. This retrospective analysis included 343,636 vancomycin TDM records, each from a distinct adult patient across 156 healthcare centers, along with a priori predictions from six PK models. A multi-label classification approach was applied, labeling PK model predictions based on whether they fell within 80%-125% of observed TDM values. Various modeling strategies were evaluated using XGBoost as the base algorithm, with binary relevance selected for the final model. At the prediction stage, PK models were ranked and averaged for each patient based on ML-predicted probabilities that predictions would fall within 80%-125% of the observed concentration. Selecting the highest ranked PK model for each patient and ML-based model averaging outperformed all single PK models, body mass index-based selection, and naive averaging. On a population level, these ML approaches resulted in more accurate predictions, a higher proportion of predictions within 80%-125% of observed vancomycin concentrations, and no systematic bias. Predictive performance declined with lower ML-assigned rankings, and selecting the lowest-ranked PK model for each patient resulted in worse performance than the worst-performing single PK model. By guiding the selection of appropriate models and avoiding less suitable ones, ML approaches for a priori MIPD may improve early dosing decisions.

Background and aims Currently, there are still no definitive consensus in the treatment of intrahepatic cholangiocarcinoma (iCCA). This study aimed to build a clinical decision support tool based on machine learning using the Surveillance, Epidemiology, and End Results (SEER) database and the data from the Fifth Medical Center of the PLA General Hospital in China. Methods 4,398 eligible patients from the SEER database and 504 eligible patients from the hospital data, who presented with histologically proven iCCA, were enrolled for modeling by cross-validation based on machine learning. All the models were trained using the open-source Python library scikit-survival version 0.16.0. Shapley additive explanations method was used to help clinicians better understand the obtained results. Permutation importance was calculated using library ELI5. Results All involved treatment modalities could contribute to a better prognosis. Three models were derived and tested using different data sources, with concordance indices of 0.67, 0.69, and 0.73, respectively. The prediction results were consistent with those under actual situations involving randomly selected patients. Model 2, trained using the hospital data, was selected to develop an online tool, due to its advantage in predicting short-term prognosis. Conclusion The prediction model and tool established in this study can be applied to predict the prognosis of iCCA after treatment by inputting the patient's clinical parameters or TNM stages and treatment options, thus contributing to optimal clinical decisions. KEY MESSAGES A prognostic model related to disease staging and treatment mode was conducted using the method of machine learning, based on the big data of multi centers. The online calculator can predict the short-term survival prognosis of intrahepatic cholangiocarcinoma, thus, help to make the best clinical decision. The online calculator built to calculate the mortality risk and overall survival can be easily obtained and applied.

Abstract Background and aims To date, there is still a lack of consensus on the treatment of intrahepatic cholangiocarcinoma (iCCA). This study aims to build a clinical decision support tool based on machine learning of the Surveillance, Epidemiology, and End Results (SEER) database and the Fifth Medical Center of PLA General Hospital in China. Methods A total of 4,398 eligible patients with pathology-proven iCCA from the SEER database and 504 from the hospital data were enrolled for modeling by cross-validation based on the method of machine learning. All models were trained by the open-source Python library 4scikit-survival version 0.16.0 and explained by SHapley Additive exPlanations. Permutation importance was calculated using the library ELI5. Results All of the involved treatment modalities can contribute to a better prognosis. Three models were derived and tested among different data sources with the concordance index in the test datasets of 0.67, 0.69, and 0.73 respectively. The prediction was also consistent with the actual situation in randomly selected real patients. Model 2 trained by the hospital data was selected to develop an online tool because of the advantage of predicting the short-term prognosis.Conclusion The prediction model and tool of this study can be applied to predicting patients’ prognosis after treatment by inputting the patient’s clinical parameters or TNM stages and the treatment option and thus contribute to the optimal clinical decision.

Background Modern clinical care in intensive care units is full of rich data, and machine learning has great potential to support clinical decision-making. The development of intelligent machine learning–based clinical decision support systems is facing great opportunities and challenges. Clinical decision support systems may directly help clinicians accurately diagnose, predict outcomes, identify risk events, or decide treatments at the point of care. Objective We aimed to review the research and application of machine learning–enabled clinical decision support studies in intensive care units to help clinicians, researchers, developers, and policy makers better understand the advantages and limitations of machine learning–supported diagnosis, outcome prediction, risk event identification, and intensive care unit point-of-care recommendations. Methods We searched papers published in the PubMed database between January 1980 and October 2020. We defined selection criteria to identify papers that focused on machine learning–enabled clinical decision support studies in intensive care units and reviewed the following aspects: research topics, study cohorts, machine learning models, analysis variables, and evaluation metrics. Results A total of 643 papers were collected, and using our selection criteria, 97 studies were found. Studies were categorized into 4 topics—monitoring, detection, and diagnosis (13/97, 13.4%), early identification of clinical events (32/97, 33.0%), outcome prediction and prognosis assessment (46/97, 47.6%), and treatment decision (6/97, 6.2%). Of the 97 papers, 82 (84.5%) studies used data from adult patients, 9 (9.3%) studies used data from pediatric patients, and 6 (6.2%) studies used data from neonates. We found that 65 (67.0%) studies used data from a single center, and 32 (33.0%) studies used a multicenter data set; 88 (90.7%) studies used supervised learning, 3 (3.1%) studies used unsupervised learning, and 6 (6.2%) studies used reinforcement learning. Clinical variable categories, starting with the most frequently used, were demographic (n=74), laboratory values (n=59), vital signs (n=55), scores (n=48), ventilation parameters (n=43), comorbidities (n=27), medications (n=18), outcome (n=14), fluid balance (n=13), nonmedicine therapy (n=10), symptoms (n=7), and medical history (n=4). The most frequently adopted evaluation metrics for clinical data modeling studies included area under the receiver operating characteristic curve (n=61), sensitivity (n=51), specificity (n=41), accuracy (n=29), and positive predictive value (n=23). Conclusions Early identification of clinical and outcome prediction and prognosis assessment contributed to approximately 80% of studies included in this review. Using new algorithms to solve intensive care unit clinical problems by developing reinforcement learning, active learning, and time-series analysis methods for clinical decision support will be greater development prospects in the future.

Background There is currently a lack of model for predicting the occurrence of venous thromboembolism (VTE) in patients with lung cancer. Machine learning (ML) techniques are being increasingly adapted for use in the medical field because of their capabilities of intelligent analysis and scalability. This study aimed to develop and validate ML models to predict the incidence of VTE among lung cancer patients. Methods Data of lung cancer patients from a Grade 3A cancer hospital in China with and without VTE were included. Patient characteristics and clinical predictors related to VTE were collected. The primary endpoint was the diagnosis of VTE during index hospitalization. We calculated and compared the area under the receiver operating characteristic curve (AUROC) using the selected best-performed model (Random Forest model) through multiple model comparison, as well as investigated feature contributions during the training process with both permutation importance scores and the impurity-based feature importance scores in random forest model. Results In total, 3,398 patients were included in our study, 125 of whom experienced VTE during their hospital stay. The ROC curve and precision–recall curve (PRC) for Random Forest Model showed an AUROC of 0.91 (95% CI : 0.893–0.926) and an AUPRC of 0.43 (95% CI : 0.363–0.500). For the simplified model, five most relevant features were selected: Karnofsky Performance Status (KPS), a history of VTE, recombinant human endostatin, EGFR-TKI, and platelet count. We re-trained a random forest classifier with results of the AUROC of 0.87 (95% CI : 0.802–0.917) and AUPRC of 0.30 (95% CI : 0.265–0.358), respectively. Conclusion According to the study results, there was no conspicuous decrease in the model’s performance when use fewer features to predict, we concluded that our simplified model would be more applicable in real-life clinical settings. The developed model using ML algorithms in our study has the potential to improve the early detection and prediction of the incidence of VTE in patients with lung cancer.

AbstractBackgroundMost patients with advanced non‐small cell lung cancer (NSCLC) have a poor prognosis. Predicting overall survival using clinical data would benefit cancer patients by allowing providers to design an optimum treatment plan. We compared the performance of nomograms with machine‐learning models at predicting the overall survival of NSCLC patients. This comparison benefits the development and selection of models during the clinical decision‐making process for NSCLC patients.MethodsMultiple machine‐learning models were used in a retrospective cohort of 6586 patients. First, we modeled and validated a nomogram to predict the overall survival of NSCLC patients. Subsequently, five machine‐learning models (logistic regression, random forest, XGBoost, decision tree, and light gradient boosting machine) were used to predict survival status. Next, we evaluated the performance of the models. Finally, the machine‐learning model with the highest accuracy was chosen for comparison with the nomogram at predicting survival status by observing a novel performance measure: time‐dependent prediction accuracy.ResultsAmong the five machine‐learning models, the accuracy of random forest model outperformed the others. Compared with the nomogram for time‐dependent prediction accuracy with a follow‐up time ranging from 12 to 60 months, the prediction accuracies of both the nomogram and machine‐learning models changed as time varied. The nomogram reached a maximum prediction accuracy of 0.85 in the 60th month, and the random forest algorithm reached a maximum prediction accuracy of 0.74 in the 13th month.ConclusionsOverall, the nomogram provided more reliable prognostic assessments of NSCLC patients than machine‐learning models over our observation period. Although machine‐learning methods have been widely adopted for predicting clinical prognoses in recent studies, the conventional nomogram was competitive. In real clinical applications, a comprehensive model that combines these two methods may demonstrate superior capabilities.

Abstract Background Kidney cancer originates from the urinary tubule epithelial system of the renal parenchyma, accounting for 20% of all urinary system tumors. Approximately 70% of cases are localized at diagnosis, and 30% are metastatic. Most localized kidney cancers can be cured by surgery, but most metastatic patients relapse after surgery and eventually die of kidney cancer. Therefore, accurately predicting patient survival and identifying high‐risk metastatic patients will effectively guide interventions and improve prognosis. Methods This study used the data of 12,394 kidney cancer patients from the surveillance, epidemiology, and end results database to construct a research cohort related to kidney cancer survival and metastasis. Eight machine learning models (including support vector machines, logistic regression, decision tree, random forest, XGBoost, AdaBoost, K‐nearest neighbors, and multilayer perceptron) were developed to predict the survival and metastasis of kidney cancer and six evaluation indicators (accuracy, precision, sensitivity, specificity, F1 score, and area under the receiver operating characteristic [AUROC]) were used to verify, evaluate, and optimize the models. Results Among the eight machine learning models, Logistic Regression has the highest AUROC in both prediction scenarios. For 3‐year survival prediction, the Logistic Regression model had an accuracy of 0.684, a sensitivity of 0.702, a specificity of 0.670, a precision of 0.686, an F1 score of 0.683, and an AUROC of 0.741. For tumor metastasis prediction, the Logistic Regression model had an accuracy of 0.800, a sensitivity of 0.540, a specificity of 0.830, a precision of 0.769, an F1 score of 0.772, and an AUROC of 0.804. Conclusion In this study, we selected appropriate variables from both statistical and clinical significance and developed and compared eight machine learning models for predicting 3‐year survival and metastasis of kidney cancer. The prediction results and evaluation results demonstrated that our model could provide decision support for early intervention for kidney cancer patients.

Background Regional citrate anticoagulation (RCA) is an important local anticoagulation method during bedside continuous renal replacement therapy. To improve patient safety and achieve computer assisted dose monitoring and control, we took intensive care units patients into cohort and aiming at developing a data-driven machine learning model to give early warning of citric acid overdose and provide adjustment suggestions on citrate pumping rate and 10% calcium gluconate input rate for RCA treatment. Methods Patient age, gender, pumped citric acid dose value, 5% NaHCO 3 solvent, replacement fluid solvent, body temperature value, and replacement fluid PH value as clinical features, models attempted to classify patients who received regional citrate anticoagulation into correct outcome category. Four models, Adaboost, XGBoost, support vector machine (SVM) and shallow neural network, were compared on the performance of predicting outcomes. Prediction results were evaluated using accuracy, precision, recall and F1-score. Results For classifying patients at the early stages of citric acid treatment, the accuracy of neutral networks model is higher than Adaboost, XGBoost and SVM, the F1-score of shallow neutral networks (90.77%) is overall outperformed than other models (88.40%, 82.17% and 88.96% for Adaboost, XGBoost and SVM). Extended experiment and validation were further conducted using the MIMIC-III database, the F1-scores for shallow neutral networks, Adaboost, XGBoost and SVM are 80.00%, 80.46%, 80.37% and 78.90%, the AUCs are 0.8638, 0.8086, 0.8466 and 0.7919 respectively. Conclusion The results of this study demonstrated the feasibility and performance of machine learning methods for monitoring and adjusting local regional citrate anticoagulation, and further provide decision-making recommendations to clinicians point-of-care.

BackgroundPatients with non-small cell lung cancer (NSCLC) often have a poor prognosis. Overall survival (OS) prediction through the early diagnosis of cancer has many benefits, such as allowing providers to design the best treatment plan for patients. In this study, we aimed to evaluate the prognostic factors in NSCLC patients, construct a nomogram, and develop machine learning models to predict the OS. We also conducted feature importance analysis to understand how relevant factors of NSCLC patients impact their OS.ResultsMultiple machine learning models were adopted in a retrospective cohort of patients from 2010 to 2015 in the Surveillance, Epidemiology, and End Results (SEER) database. Independent prognostic factors for NSCLC were determined using Cox proportional hazards regression analysis. We modeled OS and vital status as the outcomes and constructed and validated a nomogram to predict the OS of NSCLC. Furthermore, we applied logistic regression, random forest, XGBoost, decision tree, multilayer perceptron, and LightGBM to predict the patients’ vital status. We tested the prediction ability of the models and evaluated their performances using accuracy, sensitivity, specificity, precision, and the area under the receiver operating characteristic curve. A total of 34,567 patients selected from the SEER database that met our criteria were included in this study. The nomogram visualized the OS prediction results of the Cox regression model. Among the classifiers, XGBoost had the best prediction performance, with an area under the curve of 0.733.ConclusionsThe results demonstrated that machine learning-based classifier models are capable of predicting the outcomes of patients with NSCLC. And Cox regression model-based nomogram interpreted the results well and supports potential medical applications.

Background Unfractionated heparin is widely used in the intensive care unit as an anticoagulant. However, weight-based heparin dosing has been shown to be suboptimal and may place patients at unnecessary risk during their intensive care unit stay. Objective In this study, we intended to develop and validate a machine learning–based model to predict heparin treatment outcomes and to provide dosage recommendations to clinicians. Methods A shallow neural network model was adopted in a retrospective cohort of patients from the Multiparameter Intelligent Monitoring in Intensive Care III (MIMIC III) database and patients admitted to the Peking Union Medical College Hospital (PUMCH). We modeled the subtherapeutic, normal, and supratherapeutic activated partial thromboplastin time (aPTT) as the outcomes of heparin treatment and used a group of clinical features for modeling. Our model classifies patients into 3 different therapeutic states. We tested the prediction ability of our model and evaluated its performance by using accuracy, the kappa coefficient, precision, recall, and the F1 score. Furthermore, a dosage recommendation module was designed and evaluated for clinical decision support. Results A total of 3607 patients selected from MIMIC III and 1549 patients admitted to the PUMCH who met our criteria were included in this study. The shallow neural network model showed results of F1 scores 0.887 (MIMIC III) and 0.925 (PUMCH). When compared with the actual dosage prescribed, our model recommended increasing the dosage for 72.2% (MIMIC III, 1240/1718) and 64.7% (PUMCH, 281/434) of the subtherapeutic patients and decreasing the dosage for 80.9% (MIMIC III, 504/623) and 76.7% (PUMCH, 277/361) of the supratherapeutic patients, suggesting that the recommendations can contribute to clinical improvements and that they may effectively reduce the time to optimal dosage in the clinical setting. Conclusions The evaluation of our model for predicting heparin treatment outcomes demonstrated that the developed model is potentially applicable for reducing the misdosage of heparin and for providing appropriate decision recommendations to clinicians.

Background Analgesia and sedation therapy are commonly used for critically ill patients, especially mechanically ventilated patients. From the initial nonsedation programs to deep sedation and then to on-demand sedation, the understanding of sedation therapy continues to deepen. However, according to different patient’s condition, understanding the individual patient’s depth of sedation needs remains unclear. Methods The public open source critical illness database Medical Information Mart for Intensive Care III was used in this study. Latent profile analysis was used as a clustering method to classify mechanically ventilated patients based on 36 variables. Principal component analysis dimensionality reduction was used to select the most influential variables. The ROC curve was used to evaluate the classification accuracy of the model. Results Based on 36 characteristic variables, we divided patients undergoing mechanical ventilation and sedation and analgesia into two categories with different mortality rates, then further reduced the dimensionality of the data and obtained the 9 variables that had the greatest impact on classification, most of which were ventilator parameters. According to the Richmond-ASS scores, the two phenotypes of patients had different degrees of sedation and analgesia, and the corresponding ventilator parameters were also significantly different. We divided the validation cohort into three different levels of sedation, revealing that patients with high ventilator conditions needed a deeper level of sedation, while patients with low ventilator conditions required reduction in the depth of sedation as soon as possible to promote recovery and avoid reinjury. Conclusion Through latent profile analysis and dimensionality reduction, we divided patients treated with mechanical ventilation and sedation and analgesia into two categories with different mortalities and obtained 9 variables that had the greatest impact on classification, which revealed that the depth of sedation was limited by the condition of the respiratory system.

BACKGROUND:Blood pressure is a critical therapeutic goal in intensive care unit (ICU). One important factor influencing blood pressure are analgesia and sedation. Analgesic and sedative drugs are commonly used in critically ill patients. These drugs affect blood pressure by reducing the tension of the venous system, the cardiac preload, and cardiac output and inhibiting cardiac functions. Consequently, vasoactive agents are commonly used to increase blood pressure. The indications for the usage of vasoactive agents are unequivocal. However, opinions on when to stop raising blood pressure vary. This study explored the relationship between blood pressure and sedation.METHODS:Patients in the Multiparameter Intelligent Monitoring in Intensive Care-III (MIMIC) database who had received mechanical ventilation, had been administered sedative analgesics during their ICU stay, and met the inclusion criteria were included in this study. The mean arterial pressure (MAP) tendency patterns were identified using spectral clustering and visualized using the t-distributed Stochastic Neighbor Embedding (t-SNE) algorithm. The 28-day mortality rates of patients with different MAP patterns during their first 24 hours in the ICU and their sedation levels were calculated in the crosstab.RESULTS:Fourteen thousand seven hundred and eighty-five patients from the MIMIC-III database were included in this study. Three MAP patterns were identified by spectral clustering. The median MAP of the low, moderate, and high MAP groups was 71.2, 80.4, and 97.6 mmHg, respectively. The 28-day mortality rate of patients in the moderate MAP group (13.0%) was lower than that of patients in the low (16.6%) and high (15.6%) MAP groups. No difference was found in the 28-day mortality rate between the low and high MAP groups. Dynamic changes in blood pressure at different sedation depths were also examined. Notably, compared with light and moderate sedation level, patients in the deep sedation group, especially those in the high MAP group (48.5%), had a higher 28-day mortality rate (36.5%).CONCLUSIONS:Low MAP in the first 24 hours in ICU indicates a high possibility of poor prognosis for critically ill patients on mechanical ventilation. For patients under deep sedation, maintaining a high mean arterial pressure also indicates poor prognosis. A personalized MAP target should be determined according to the severity of illness and level of sedation for each patient.

Background Although protective mechanical ventilation (MV) has been used in a variety of applications, lung injury may occur in both patients with and without acute respiratory distress syndrome (ARDS). The purpose of this study is to use machine learning to identify clinical phenotypes for critically ill patients with MV in intensive care units (ICUs). Methods A retrospective cohort study was conducted with 5013 patients who had undergone MV and treatment in the Department of Critical Care Medicine, Peking Union Medical College Hospital. Statistical and machine learning methods were used. All the data used in this study, including demographics, vital signs, circulation parameters and mechanical ventilator parameters, etc., were automatically extracted from the electronic health record (EHR) system. An external database, Medical Information Mart for Intensive Care III (MIMIC III), was used for validation. Results Phenotypes were derived from a total of 4009 patients who underwent MV using a latent profile analysis of 22 variables. The associations between the phenotypes and disease severity and clinical outcomes were assessed. Another 1004 patients in the database were enrolled for validation. Of the five derived phenotypes, phenotype I was the most common subgroup (n = 2174; 54.2%) and was mostly composed of the postoperative population. Phenotype II (n = 480; 12.0%) led to the most severe conditions. Phenotype III (n = 241; 6.01%) was associated with high positive end-expiratory pressure (PEEP) and low mean airway pressure. Phenotype IV (n = 368; 9.18%) was associated with high driving pressure, and younger patients comprised a large proportion of the phenotype V group (n = 746; 18.6%). In addition, we found that the mortality rate of Phenotype IV was significantly higher than that of the other phenotypes. In this subgroup, the number of patients in the sequential organ failure assessment (SOFA) score segment (9,22] was 198, the number of deaths was 88, and the mortality rate was higher than 44%. However, the cumulative 28-day mortality of Phenotypes IV and II, which were 101 of 368 (27.4%) and 87 of 480 (18.1%) unique patients, respectively, was significantly higher than those of the other phenotypes. There were consistent phenotype distributions and differences in biomarker patterns by phenotype in the validation cohort, and external verification with MIMIC III further generated supportive results. Conclusions Five clinical phenotypes were correlated with different disease severities and clinical outcomes, which suggested that these phenotypes may help in understanding heterogeneity in MV treatment effects.

BACKGROUND Heparin is one of the most commonly used medications in intensive care units. In clinical practice, the use of a weight-based heparin dosing nomogram is standard practice for the treatment of thrombosis. Recently, machine learning techniques have dramatically improved the ability of computers to provide clinical decision support and have allowed for the possibility of computer generated, algorithm-based heparin dosing recommendations. OBJECTIVE The objective of this study was to predict the effects of heparin treatment using machine learning methods to optimize heparin dosing in intensive care units based on the predictions. Patient state predictions were based upon activated partial thromboplastin time in 3 different ranges: subtherapeutic, normal therapeutic, and supratherapeutic, respectively. METHODS Retrospective data from 2 intensive care unit research databases (Multiparameter Intelligent Monitoring in Intensive Care III, MIMIC-III; e–Intensive Care Unit Collaborative Research Database, eICU) were used for the analysis. Candidate machine learning models (random forest, support vector machine, adaptive boosting, extreme gradient boosting, and shallow neural network) were compared in 3 patient groups to evaluate the classification performance for predicting the subtherapeutic, normal therapeutic, and supratherapeutic patient states. The model results were evaluated using precision, recall, F1 score, and accuracy. RESULTS Data from the MIMIC-III database (n=2789 patients) and from the eICU database (n=575 patients) were used. In 3-class classification, the shallow neural network algorithm performed the best (F1 scores of 87.26%, 85.98%, and 87.55% for data set 1, 2, and 3, respectively). The shallow neural network algorithm achieved the highest F1 scores within the patient therapeutic state groups: subtherapeutic (data set 1: 79.35%; data set 2: 83.67%; data set 3: 83.33%), normal therapeutic (data set 1: 93.15%; data set 2: 87.76%; data set 3: 84.62%), and supratherapeutic (data set 1: 88.00%; data set 2: 86.54%; data set 3: 95.45%) therapeutic ranges, respectively. CONCLUSIONS The most appropriate model for predicting the effects of heparin treatment was found by comparing multiple machine learning models and can be used to further guide optimal heparin dosing. Using multicenter intensive care unit data, our study demonstrates the feasibility of predicting the outcomes of heparin treatment using data-driven methods, and thus, how machine learning–based models can be used to optimize and personalize heparin dosing to improve patient safety. Manual analysis and validation suggested that the model outperformed standard practice heparin treatment dosing.

ABSTRACTBackgroundWith the outbreak of coronavirus disease 2019 (COVID-19), a sudden case increase in late February 2020 led to deep concern globally. Italy, South Korea, Iran, France, Germany, Spain, the US and Japan are probably the countries with the most severe outbreaks. Collecting epidemiological data and predicting epidemic trends are important for the development and measurement of public intervention strategies. Epidemic prediction results yielded by different mathematical models are inconsistent; therefore, we sought to compare different models and their prediction results to generate objective conclusions.MethodsWe used the number of cases reported from January 23 to March 20, 2020, to estimate the possible spread size and peak time of COVID-19, especially in 8 high-risk countries. The logistic growth model, basic SEIR model and adjusted SEIR model were adopted for prediction. Given that different model inputs may infer different model outputs, we implemented three model predictions with three scenarios of epidemic development.ResultsWhen comparing all 8 countries’ short-term prediction results and peak predictions, the differences among the models were relatively large. The logistic growth model estimated a smaller epidemic size than the basic SERI model did; however, once we added parameters that considered the effects of public health interventions and control measures, the adjusted SERI model results demonstrated a considerably rapid deceleration of epidemic development. Our results demonstrated that contact rate, quarantine scale, and the initial quarantine time and length are important factors in controlling epidemic size and length.ConclusionsWe demonstrated a comparative assessment of the predictions of the COVID-19 outbreak in eight high-risk countries using multiple methods. By forecasting epidemic size and peak time as well as simulating the effects of public health interventions, the intent of this paper is to help clarify the transmission dynamics of COVID-19 and recommend operation suggestions to slow down the epidemic. It is suggested that the quick detection of cases, sufficient implementation of quarantine and public self-protection behaviors are critical to slow down the epidemic.