A Time-Phased Machine Learning Model for Real-Time Prediction of Sepsis in Critical Care

Transatlantic transferability and replicability of machine-learning algorithms to predict mental health crises

Transferring and replicating predictive algorithms across healthcare systems constitutes a unique yet crucial challenge that needs to be addressed to enable the widespread adoption of machine learning in healthcare. In this study, we explored the impact of important differences across healthcare systems and the associated Electronic Health Records (EHRs) on machine-learning algorithms to predict mental health crises, up to 28 days in advance. We evaluated both the transferability and replicability of such machine learning models, and for this purpose, we trained six models using features and methods developed on EHR data from the Birmingham and Solihull Mental Health NHS Foundation Trust in the UK. These machine learning models were then used to predict the mental health crises of 2907 patients seen at the Rush University System for Health in the US between 2018 and 2020. The best one was trained on a combination of US-specific structured features and frequency features from anonymized patient notes and achieved an AUROC of 0.837. A model with comparable performance, originally trained using UK structured data, was transferred and then tuned using US data, achieving an AUROC of 0.826. Our findings establish the feasibility of transferring and replicating machine learning models to predict mental health crises across diverse hospital systems.

Artificial intelligence and prescription of antibiotic therapy: present and future

INTRODUCTION

In the past few years, the use of artificial intelligence in healthcare has grown exponentially. Prescription of antibiotics is not exempt from its rapid diffusion, and various machine learning (ML) techniques, from logistic regression to deep neural networks and large language models, have been explored in the literature to support decisions regarding antibiotic prescription.

AREAS COVERED

In this narrative review, we discuss promises and challenges of the application of ML-based clinical decision support systems (ML-CDSSs) for antibiotic prescription. A search was conducted in PubMed up to April 2024.

EXPERT OPINION

Prescribing antibiotics is a complex process involving various dynamic phases. In each of these phases, the support of ML-CDSSs has shown the potential, and also the actual ability in some studies, to favorably impacting relevant clinical outcomes. Nonetheless, before widely exploiting this massive potential, there are still crucial challenges ahead that are being intensively investigated, pertaining to the transparency of training data, the definition of the sufficient degree of prediction explanations when predictions are obtained through black box models, and the legal and ethical framework for decision responsibility whenever an antibiotic prescription is supported by ML-CDSSs.

Early sepsis mortality prediction model based on interpretable machine learning approach: development and validation study

Sepsis triggers a harmful immune response due to infection, causing high mortality. Predicting sepsis outcomes early is vital. Despite machine learning's (ML) use in medical research, local validation within the Medical Information Mart for Intensive Care IV (MIMIC-IV) database is lacking. We aimed to devise a prognostic model, leveraging MIMIC-IV data, to predict sepsis mortality and validate it in a Chinese teaching hospital. MIMIC-IV provided patient data, split into training and internal validation sets. Four ML models logistic regression (LR), support vector machine (SVM), deep neural networks (DNN), and extreme gradient boosting (XGBoost) were employed. Shapley additive interpretation offered early and interpretable mortality predictions. Area under the ROC curve (AUROC) gaged predictive performance. Results were cross verified in a Chinese teaching hospital. The study included 27,134 sepsis patients from MIMIC-IV and 487 from China. After comparing, 52 clinical indicators were selected for ML model development. All models exhibited excellent discriminative ability. XGBoost surpassed others, with AUROC of 0.873 internally and 0.844 externally. XGBoost outperformed other ML models (LR: 0.829; SVM: 0.830; DNN: 0.837) and clinical scores (Simplified Acute Physiology Score II: 0.728; Sequential Organ Failure Assessment: 0.728; Oxford Acute Severity of Illness Score: 0.738; Glasgow Coma Scale: 0.691). XGBoost's hospital mortality prediction achieved AUROC 0.873, sensitivity 0.818, accuracy 0.777, specificity 0.768, and F1 score 0.551. We crafted an interpretable model for sepsis death risk prediction. ML algorithms surpassed traditional scores for sepsis mortality forecast. Validation in a Chinese teaching hospital echoed these findings.

Impact of P2Y12 inhibitors on clinical outcomes in sepsis-3 patients receiving aspirin: a propensity score matched analysis

BACKGROUND

Sepsis is a life-threatening disease accompanied by disorders of the coagulation and immune systems. P2Y12 inhibitors, widely used for arterial thrombosis prevention and treatment, possess recently discovered anti-inflammatory properties, raising potential for improved sepsis prognosis.

METHOD

We conducted a retrospective analysis using the data from Medical Information Mart for Intensive Care-IV database. Patients were divided into an aspirin-alone group versus a combination group based on the use of a P2Y12 inhibitor or not. Differences in 30-day mortality, length of stay (LOS) in intensive care unit (ICU), LOS in hospital, bleeding events and thrombotic events were compared between the two groups.

RESULT

A total of 1701 pairs of matched patients were obtained by propensity score matching. We found that no statistically significant difference in 30-day mortality in aspirin-alone group and combination group (15.3% vs. 13.7%, log-rank p = 0.154). In addition, patients received P2Y12 inhibitors had a higher incidence of gastrointestinal bleeding (0.5% vs. 1.6%, p = 0.004) and ischemic stroke (1.7% vs. 2.9%, p = 0.023), despite having a shorter LOS in hospital (11.1 vs. 10.3, days, p = 0.043). Cox regression showed that P2Y12 inhibitor was not associated with 30-day mortality (HR = 1.14, 95% CI 0.95-1.36, p = 0.154).

CONCLUSION

P2Y12 inhibitors did not provide a survival benefit for patients with sepsis 3 and even led to additional adverse clinical outcomes.

Explainable machine learning model for predicting furosemide responsiveness in patients with oliguric acute kidney injury

BACKGROUND

Although current guidelines didn't support the routine use of furosemide in oliguric acute kidney injury (AKI) management, some patients may benefit from furosemide administration at an early stage. We aimed to develop an explainable machine learning (ML) model to differentiate between furosemide-responsive (FR) and furosemide-unresponsive (FU) oliguric AKI.

METHODS

From Medical Information Mart for Intensive Care-IV (MIMIC-IV) and eICU Collaborative Research Database (eICU-CRD), oliguric AKI patients with urine output (UO) < 0.5 ml/kg/h for the first 6 h after ICU admission and furosemide infusion ≥ 40 mg in the following 6 h were retrospectively selected. The MIMIC-IV cohort was used in training a XGBoost model to predict UO > 0.65 ml/kg/h during 6-24 h succeeding the initial 6 h for assessing oliguria, and it was validated in the eICU-CRD cohort. We compared the predictive performance of the XGBoost model with the traditional logistic regression and other ML models.

RESULTS

6897 patients were included in the MIMIC-IV training cohort, with 2235 patients in the eICU-CRD validation cohort. The XGBoost model showed an AUC of 0.97 (95% CI: 0.96-0.98) for differentiating FR and FU oliguric AKI. It outperformed the logistic regression and other ML models in correctly predicting furosemide diuretic response, achieved 92.43% sensitivity (95% CI: 90.88-93.73%) and 95.12% specificity (95% CI: 93.51-96.3%).

CONCLUSION

A boosted ensemble algorithm can be used to accurately differentiate between patients who would and would not respond to furosemide in oliguric AKI. By making the model explainable, clinicians would be able to better understand the reasoning behind the prediction outcome and make individualized treatment.

Development and Validation of Deep-Learning-Based Sepsis and Septic Shock Early Prediction System (DeepSEPS) Using Real-World ICU Data

BACKGROUND

Successful sepsis treatment depends on early diagnosis. We aimed to develop and validate a system to predict sepsis and septic shock in real time using deep learning.

METHODS

Clinical data were retrospectively collected from electronic medical records (EMRs). Data from 2010 to 2019 were used as development data, and data from 2020 to 2021 were used as validation data. The collected EMRs consisted of eight vital signs, 13 laboratory data points, and three demographic information items. We validated the deep-learning-based sepsis and septic shock early prediction system (DeepSEPS) using the validation datasets and compared our system with other traditional early warning scoring systems, such as the national early warning score, sequential organ failure assessment (SOFA), and quick sequential organ failure assessment.

RESULTS

DeepSEPS achieved even higher area under receiver operating characteristic curve (AUROC) values (0.7888 and 0.8494 for sepsis and septic shock, respectively) than SOFA. The prediction performance of traditional scoring systems was enhanced because the early prediction time point was close to the onset time of sepsis; however, the DeepSEPS scoring system consistently outperformed all conventional scoring systems at all time points. Furthermore, at the time of onset of sepsis and septic shock, DeepSEPS showed the highest AUROC (0.9346).

CONCLUSIONS

The sepsis and septic shock early warning system developed in this study exhibited a performance that is worth considering when predicting sepsis and septic shock compared to other traditional early warning scoring systems. DeepSEPS showed better performance than existing sepsis prediction programs. This novel real-time system that simultaneously predicts sepsis and septic shock requires further validation.

Interpretable machine-learning model for real-time, clustered risk factor analysis of sepsis and septic death in critical care

BACKGROUND AND OBJECTIVE

Interpretable and real-time prediction of sepsis and risk factor analysis could enable timely treatment by clinicians and improve patient outcomes. To develop an interpretable machine-learning model for the prediction and risk factor analysis of sepsis and septic death.

METHODS

This is a retrospective observational cohort study based on the Medical Information Mart for Intensive Care (MIMIC-IV) dataset; 69,619 patients from the database were screened. The two outcomes include patients diagnosed with sepsis and the death of septic patients. Clinical variables from ICU admission to outcomes were analyzed: demographic data, vital signs, Glasgow Coma Scale scores, laboratory test results, and results for arterial blood gasses (ABGs). Model performance was compared using the area under the receiver operating characteristic curve (AUROC). Model interpretations were based on the Shapley additive explanations (SHAP), and the clustered analysis was based on the combination of K-means and dimensionality reduction algorithms of t-SNE and PCA.

RESULTS

For the analysis of sepsis and septic death, 47,185 and 2480 patients were enrolled, respectively. The XGBoost model achieved a predictive value of area under the curve (AUC): 0.745 [0.731-0.759] for sepsis prediction and 0.8 [0.77, 0.828] for septic death prediction. The real-time prediction model was trained to predict by day and visualize the individual or combined risk factor effects on the outcomes based on SHAP values. Clustered analysis separated the two phenotypes with distinct risk factors among patients with septic death.

CONCLUSION

The proposed real-time, clustered prediction model for sepsis and septic death exhibited superior performance in predicting the outcomes and visualizing the risk factors in a real-time and interpretable manner to distinguish and mitigate patient risks, thus promising immense potential in effective clinical decision making and comprehensive understanding of complex diseases such as sepsis.

The application of artificial intelligence in the management of sepsis

Sepsis is a complex and heterogeneous syndrome that remains a serious challenge to healthcare worldwide. Patients afflicted by severe sepsis or septic shock are customarily placed under intensive care unit (ICU) supervision, where a multitude of apparatus is poised to produce high-granularity data. This reservoir of high-quality data forms the cornerstone for the integration of AI into clinical practice. However, existing reviews currently lack the inclusion of the latest advancements. This review examines the evolving integration of artificial intelligence (AI) in sepsis management. Applications of artificial intelligence include early detection, subtyping analysis, precise treatment and prognosis assessment. AI-driven early warning systems provide enhanced recognition and intervention capabilities, while profiling analyzes elucidate distinct sepsis manifestations for targeted therapy. Precision medicine harnesses the potential of artificial intelligence for pathogen identification, antibiotic selection, and fluid optimization. In conclusion, the seamless amalgamation of artificial intelligence into the domain of sepsis management heralds a transformative shift, ushering in novel prospects to elevate diagnostic precision, therapeutic efficacy, and prognostic acumen. As AI technologies develop, their impact on shaping the future of sepsis care warrants ongoing research and thoughtful implementation.

Early prediction of MODS interventions in the intensive care unit using machine learning

Background

Multiple organ dysfunction syndrome (MODS) is one of the leading causes of death in critically ill patients. MODS is the result of a dysregulated inflammatory response that can be triggered by various causes. Owing to the lack of an effective treatment for patients with MODS, early identification and intervention are the most effective strategies. Therefore, we have developed a variety of early warning models whose prediction results can be interpreted by Kernel SHapley Additive exPlanations (Kernel-SHAP) and reversed by diverse counterfactual explanations (DiCE). So we can predict the probability of MODS 12 h in advance, quantify the risk factors, and automatically recommend relevant interventions.

Methods

We used various machine learning algorithms to complete the early risk assessment of MODS, and used a stacked ensemble to improve the prediction performance. The kernel-SHAP algorithm was used to quantify the positive and minus factors corresponding to the individual prediction results, and finally, the DiCE method was used to automatically recommend interventions. We completed the model training and testing based on the MIMIC-III and MIMIC-IV databases, in which the sample features in the model training included the patients' vital signs, laboratory test results, test reports, and data related to the use of ventilators.

Results

The customizable model called SuperLearner, which integrated multiple machine learning algorithms, had the highest authenticity of screening, and its Yordon index (YI), sensitivity, accuracy, and utility_score on the MIMIC-IV test set were 0.813, 0.884, 0.893, and 0.763, respectively, which were all maximum values of eleven models. The area under the curve of the deep-wide neural network (DWNN) model on the MIMIC-IV test set was 0.960, and the specificity was 0.935, which were both the maximum values of all these models. The Kernel-SHAP algorithm combined with SuperLearner was used to determine the minimum value of glasgow coma scale (GCS) in the current hour (OR = 0.609, 95% CI   0.606-0.612), maximum value of MODS score corresponding to GCS in the past 24 h (OR = 2.632, 95% CI 2.588-2.676), and maximum score of MODS corresponding to creatinine in the past 24 h (OR = 3.281, 95% CI   3.267-3.295) were generally the most influential factors.

Conclusion

The MODS early warning model based on machine learning algorithms has considerable application value, and the prediction efficiency of SuperLearner is superior to those of SubSuperLearner, DWNN, and other eight common machine learning models. Considering that the attribution analysis of Kernel-SHAP is a static analysis of the prediction results, we introduce the DiCE algorithm to automatically recommend counterfactuals to reverse the prediction results, which will be an important step towards the practical application of automatic MODS early intervention.

Supplementary Information

The online version contains supplementary material available at 10.1186/s40537-023-00719-2.

Early prediction of sepsis using double fusion of deep features and handcrafted features

Sepsis is a life-threatening medical condition that is characterized by the dysregulated immune system response to infections, having both high morbidity and mortality rates. Early prediction of sepsis is critical to the decrease of mortality. This paper presents a novel early warning model called Double Fusion Sepsis Predictor (DFSP) for sepsis onset. DFSP is a double fusion framework that combines the benefits of early and late fusion strategies. First, a hybrid deep learning model that combines both the convolutional and recurrent neural networks to extract deep features is proposed. Second, deep features and handcrafted features, such as clinical scores, are concatenated to build the joint feature representation (early fusion). Third, several tree-based models based on joint feature representation are developed to generate the risk scores of sepsis onset that are combined with an End-to-End neural network for final sepsis detection (late fusion). To evaluate DFSP, a retrospective study was conducted, which included patients admitted to the ICUs of a hospital in Shanghai China. The results demonstrate that the DFSP outperforms state-of-the-art approaches in early sepsis prediction.

Dynamic prediction of life-threatening events for patients in intensive care unit

Background

Early prediction of patients’ deterioration is helpful in early intervention for patients at greater risk of deterioration in Intensive Care Unit (ICU). This study aims to apply machine learning approaches to heterogeneous clinical data for predicting life-threatening events of patients in ICU.

Methods

We collected clinical data from a total of 3151 patients admitted to the Medical Intensive Care Unit of Peking Union Medical College Hospital in China from January 1st, 2014, to October 1st, 2019. After excluding the patients who were under 18 years old or stayed less than 24 h at the ICU, a total of 2170 patients were enrolled in this study. Multiple machine learning approaches were utilized to predict life-threatening events (i.e., death) in seven 24-h windows (day 1 to day 7) and their performance was compared.

Results

Light Gradient Boosting Machine showed the best performance. We found that life-threatening events during the short-term windows can be better predicted than those in the medium-term windows. For example, death in 24 h can be predicted with an Area Under Curve of 0.905. Features like infusion pump related fluid input were highly related to life-threatening events. Furthermore, the prediction power of static features such as age and cardio-pulmonary function increased with the extended prediction window.

Conclusion

This study demonstrates that the integration of machine learning approaches and large-scale high-quality clinical data in ICU could accurately predict life-threatening events for ICU patients for early intervention.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12911-022-02026-x.

Machine learning advances the integration of covariates in population pharmacokinetic models: Valproic acid as an example

Background and Aim: Many studies associated with the combination of machine learning (ML) and pharmacometrics have appeared in recent years. ML can be used as an initial step for fast screening of covariates in population pharmacokinetic (popPK) models. The present study aimed to integrate covariates derived from different popPK models using ML.

Methods: Two published popPK models of valproic acid (VPA) in Chinese epileptic patients were used, where the population parameters were influenced by some covariates. Based on the covariates and a one-compartment model that describes the pharmacokinetics of VPA, a dataset was constructed using Monte Carlo simulation, to develop an XGBoost model to estimate the steady-state concentrations (Css) of VPA. We utilized SHapley Additive exPlanation (SHAP) values to interpret the prediction model, and calculated estimates of VPA exposure in four assumed scenarios involving different combinations of CYP2C19 genotypes and co-administered antiepileptic drugs. To develop an easy-to-use model in the clinic, we built a simplified model by using CYP2C19 genotypes and some noninvasive clinical parameters, and omitting several features that were infrequently measured or whose clinically available values were inaccurate, and verified it on our independent external dataset.

Results: After data preprocessing, the finally generated combined dataset was divided into a derivation cohort and a validation cohort (8:2). The XGBoost model was developed in the derivation cohort and yielded excellent performance in the validation cohort with a mean absolute error of 2.4 mg/L, root-mean-squared error of 3.3 mg/L, mean relative error of 0%, and percentages within ±20% of actual values of 98.85%. The SHAP analysis revealed that daily dose, time, CYP2C19*2 and/or *3 variants, albumin, body weight, single dose, and CYP2C19*1*1 genotype were the top seven confounding factors influencing the Css of VPA. Under the simulated dosage regimen of 500 mg/bid, the VPA exposure in patients who had CYP2C19*2 and/or *3 variants and no carbamazepine, phenytoin, or phenobarbital treatment, was approximately 1.74-fold compared to those with CYP2C19*1/*1 genotype and co-administered carbamazepine + phenytoin + phenobarbital. The feasibility of the simplified model was fully illustrated by its performance in our external dataset.

Conclusion: This study highlighted the bridging role of ML in big data and pharmacometrics, by integrating covariates derived from different popPK models.

Sequences of Events from the Electronic Medical Record and the Onset of Infection

We present a novel model of time-series analysis to learn from electronic health record (EHR) data when infection occurred in the intensive care unit (ICU) by translating methods from proteomics and Bayesian statistics. Using 48,536 patients hospitalized in an ICU, we describe each hospital course as an 'alphabet' of 23 physician actions ('events') in temporal order. We analyze these as k-mers of length 3-12 events and apply a Bayesian model of (cumulative) relative risk (RR). The log2-transformed RR (median=0.248, mean=0.226) supported the conclusion that the events selected were individually associated with increased risk of infection. Selecting from all possible cutoffs of maximum gain (MG), MG>0.0244 predicts administration of antibiotics with PPV 82.0 %, NPV 44.4 %, and AUC 0.706. Our approach holds value for retrospective analysis of other clinical syndromes for which time-of-onset is critical to analysis but poorly marked in EHRs, including delirium and decompensation.

Development and Validation of Machine Learning Models for Real-Time Mortality Prediction in Critically Ill Patients With Sepsis-Associated Acute Kidney Injury

Background

Sepsis-associated acute kidney injury (SA-AKI) is common in critically ill patients, which is associated with significantly increased mortality. Existing mortality prediction tools showed insufficient predictive power or failed to reflect patients' dynamic clinical evolution. Therefore, the study aimed to develop and validate machine learning-based models for real-time mortality prediction in critically ill patients with SA-AKI.

Methods

The multi-center retrospective study included patients from two distinct databases. A total of 12,132 SA-AKI patients from the Medical Information Mart for Intensive Care IV (MIMIC-IV) were randomly allocated to the training, validation, and internal test sets. An additional 3,741 patients from the eICU Collaborative Research Database (eICU-CRD) served as an external test set. For every 12 h during the ICU stays, the state-of-the-art eXtreme Gradient Boosting (XGBoost) algorithm was used to predict the risk of in-hospital death in the following 48, 72, and 120 h and in the first 28 days after ICU admission. Area under the receiver operating characteristic curves (AUCs) were calculated to evaluate the models' performance.

Results

The XGBoost models, based on routine clinical variables updated every 12 h, showed better performance in mortality prediction than the SOFA score and SAPS-II. The AUCs of the XGBoost models for mortality over different time periods ranged from 0.848 to 0.804 in the internal test set and from 0.818 to 0.748 in the external test set. The shapley additive explanation method provided interpretability for the XGBoost models, which improved the understanding of the association between the predictor variables and future mortality.

Conclusions

The interpretable machine learning XGBoost models showed promising performance in real-time mortality prediction in critically ill patients with SA-AKI, which are useful tools for early identification of high-risk patients and timely clinical interventions.

Machine learning model to predict mental health crises from electronic health records

The timely identification of patients who are at risk of a mental health crisis can lead to improved outcomes and to the mitigation of burdens and costs. However, the high prevalence of mental health problems means that the manual review of complex patient records to make proactive care decisions is not feasible in practice. Therefore, we developed a machine learning model that uses electronic health records to continuously monitor patients for risk of a mental health crisis over a period of 28 days. The model achieves an area under the receiver operating characteristic curve of 0.797 and an area under the precision-recall curve of 0.159, predicting crises with a sensitivity of 58% at a specificity of 85%. A follow-up 6-month prospective study evaluated our algorithm’s use in clinical practice and observed predictions to be clinically valuable in terms of either managing caseloads or mitigating the risk of crisis in 64% of cases. To our knowledge, this study is the first to continuously predict the risk of a wide range of mental health crises and to explore the added value of such predictions in clinical practice.

Machine learning applied on electronic health records can predict mental health crises 28 days in advance and become a clinically valuable tool for managing caseloads and mitigating the risk of crisis.

Early prediction of moderate-to-severe condition of inhalation-induced acute respiratory distress syndrome via interpretable machine learning

Background

Several studies have investigated the correlation between physiological parameters and the risk of acute respiratory distress syndrome (ARDS), in addition, etiology-associated heterogeneity in ARDS has become an emerging topic quite recently; however, the intersection between the two, which is early prediction of target conditions in etiology-specific ARDS, has not been well-studied. We aimed to develop and validate a machine-learning model for the early prediction of moderate-to-severe condition of inhalation-induced ARDS.

Methods

Clinical expertise was applied with data-driven analysis. Using data from electronic intensive care units (retrospective derivation cohort) and the three most accessible vital signs (i.e. heart rate, temperature, and respiratory rate) together with feature engineering, we applied a random forest approach during the time window of 90 h that ended 6 h prior to the onset of moderate-to-severe respiratory failure (the ratio of partial pressure of arterial oxygen to fraction of inspired oxygen ≤ 200 mmHg).

Results

The trained random forest classifier was validated using two independent validation cohorts, with an area under the curve of 0.9127 (95% confidence interval 0.8713–0.9542) and 0.9026 (95% confidence interval 0.8075–1), respectively. A Stable and Interpretable RUle Set (SIRUS) was used to extract rules from the RF to provide guidelines for clinicians. We identified several predictive factors, including resp_96h_6h_min < 9, resp_96h_6h_mean ≥ 16.1, HR_96h_6h_mean ≥ 102, and temp_96h_6h_max > 100, that could be used for predicting inhalation-induced ARDS (moderate-to-severe condition) 6 h prior to onset in critical care units. (‘xxx_96h_6h_min/mean/max’: the minimum/mean/maximum values of the xxx vital sign collected during a 90 h time window beginning 96 h prior to the onset of ARDS and ending 6 h prior to the onset from every recorded blood gas test).

Conclusions

This newly established random forest‑based interpretable model shows good predictive ability for moderate-to-severe inhalation-induced ARDS and may assist clinicians in decision-making, as well as facilitate the enrolment of patients in prevention programmes to improve their outcomes.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12890-022-01963-7.

Artificial Intelligence in Critical Care Medicine

This article is one of ten reviews selected from the Annual Update in Intensive Care and Emergency Medicine 2022. Other selected articles can be found online at https://www.biomedcentral.com/collections/annualupdate2022 . Further information about the Annual Update in Intensive Care and Emergency Medicine is available from https://link.springer.com/bookseries/8901 .

A Standardized Step-by-Step Approach for the Diagnosis and Treatment of Sepsis

Sepsis is the major culprit of death among critically ill patients who are hospitalized in intensive care units (ICUs). Although sepsis-related mortality is steadily declining year-by-year due to the continuous understanding of the pathophysiological mechanism on sepsis and improvement of the bundle treatment, sepsis-associated hospitalization is rising worldwide. Surviving Sepsis Campaign (SSC) guidelines are continuously updating, while their content is extremely complex and comprehensive for a precisely implementation in clinical practice. As a consequence, a standardized step-by-step approach for the diagnosis and treatment of sepsis is particularly important. In the present study, we proposed a standardized step-by-step approach for the diagnosis and treatment of sepsis using our daily clinical experience and the latest researches, which is close to clinical practice and is easy to implement. The proposed approach may assist clinicians to more effectively diagnose and treat septic patients and avoid the emergence of adverse clinical outcomes.

Efficacy prediction of noninvasive ventilation failure based on the stacking ensemble algorithm and autoencoder

Background

Early prediction of noninvasive ventilation failure is of great significance for critically ill ICU patients to escalate or change treatment. Because clinically collected data are highly time-series correlated and have imbalanced classes, it is difficult to accurately predict the efficacy of noninvasive ventilation for severe patients. This paper aims to precisely predict the failure probability of noninvasive ventilation before or in the early stage (1–2 h) of using it on patients and to explain the correlation of the predicted results.

Methods

In this paper, we proposed a SMSN model (stacking and modified SMOTE algorithm of prediction of noninvasive ventilation failure). In the feature generation stage, we used an autoencoder algorithm based on long short-term memory (LSTM) to automatically extract time series features. In the modelling stage, we adopted a modified SMOTE algorithm to address imbalanced classes, and three classifiers (logistic regression, random forests, and Catboost) were combined with the stacking ensemble algorithm to achieve high prediction accuracy.

Results

Data from 2495 patients were used to train the SMSN model. Among them, 80% of 2495 patients (1996 patients) were randomly selected as the training set, and 20% of these patients (499 patients) were chosen as the testing set. The F1 of the proposed SMSN model was 79.4%, and the accuracy was 88.2%. Compared with the traditional logistic regression algorithm, the F1 and accuracy were improved by 4.7% and 1.3%, respectively.

Conclusions

Through SHAP analysis, oxygenation index, pH and H1FIO₂ collected after 1 h of noninvasive ventilation were the most relevant features affecting the prediction.

Early Prediction of Sepsis Based on Machine Learning Algorithm

Sepsis is an organ failure disease caused by an infection resulting in extremely high mortality. Machine learning algorithms XGBoost and LightGBM are applied to construct two processing methods: mean processing method and feature generation method, aiming to predict early sepsis 6 hours in advance. The feature generation methods are constructed by combining different features, including statistical strength features, window features, and medical features. Miceforest multiple interpolation method is applied to tackle large missing data problems. Results show that the feature generation method outperforms the mean processing method. XGBoost and LightGBM algorithms are both excellent in prediction performance (AUC: 0.910∼0.979), among which LightGBM boasts a faster running speed and is stronger in generalization ability especially on multidimensional data, with AUC reaching 0.979 in the feature generation method. PTT, WBC, and platelets are the key risk factors to predict early sepsis.

The impact of recency and adequacy of historical information on sepsis predictions using machine learning

Sepsis is a major public and global health concern. Every hour of delay in detecting sepsis significantly increases the risk of death, highlighting the importance of accurately predicting sepsis in a timely manner. A growing body of literature has examined developing new or improving the existing machine learning (ML) approaches for timely and accurate predictions of sepsis. This study contributes to this literature by providing clear insights regarding the role of the recency and adequacy of historical information in predicting sepsis using ML. To this end, we implemented a deep learning model using a bidirectional long short-term memory (BiLSTM) algorithm and compared it with six other ML algorithms based on numerous combinations of the prediction horizons (to capture information recency) and observation windows (to capture information adequacy) using different measures of predictive performance. Our results indicated that the BiLSTM algorithm outperforms all other ML algorithms and provides a great separability of the predicted risk of sepsis among septic versus non-septic patients. Moreover, decreasing the prediction horizon (in favor of information recency) always boosts the predictive performance; however, the impact of expanding the observation window (in favor of information adequacy) depends on the prediction horizon and the purpose of prediction. More specifically, when the prediction is responsive to the positive label (i.e., Sepsis), increasing historical data improves the predictive performance when the prediction horizon is short-moderate.

Explainable machine learning model to predict refeeding hypophosphatemia

BACKGROUND & AIMS

Refeeding syndrome (RFS) is a disease that occurs when feeding is restarted and metabolism changes from catabolic to anabolic status. RFS can manifest variously, ranging from asymptomatic to fatal, therefore it may easily be overlooked. RFS prediction using explainable machine learning can improve diagnosis and treatment. Our study aimed to propose a machine learning model for RFS prediction, specifically refeeding hypophosphatemia, to evaluate its performance compared with conventional regression models, and to explain the machine learning classification through Shapley additive explanations (SHAP) values.

METHODS

A retrospective study was conducted including 806 patients, with 2 or more days of nothing-by-mouth prescription, and with phosphate (P) level measurements within 5 days of refeeding were selected. We divided the patients into hypophosphatemia (n = 367) and non-hypophosphatemia groups (n = 439) at a P level of 0.8 mmol/L. Among the features examined within 48 h after admission, we reviewed laboratory test results and electronic medical records. Logistic, Lasso, and ridge regressions were used as conventional models, and performances were compared with our extreme gradient boosting (XGBoost) machine learning model using the area under the receiver operating characteristic curve. Our model was explained using the SHAP value.

RESULTS

The areas under the curve were 0.950 (95% confidence interval: 0.924-0.975) for our XGBoost machine learning model and surpassed the performance of conventional regression models; 0.760 (0.707-0.813) for logistic regression, 0.751 (0.694-0.807) for Lasso regression, and 0.758 (0.701-0.809) for ridge regression. According to the SHAP values in the order of importance, low initial P, recent weight loss, high creatinine, diabetes mellitus with insulin use, low haemoglobin A1c, furosemide use, intensive care unit admission, blood urea nitrogen level of 19-65, parenteral nutrition, magnesium below or above the normal range, low potassium, and older age were features to predict refeeding hypophosphatemia.

CONCLUSIONS

The machine learning model for predicting RFS has a substantially higher effectiveness than conventional regression methods. Creating an accurate risk assessment tool based on machine learning for early identification of patients at risk for RFS can enable careful nutrition management planning and monitoring in the intensive care unit, towards reducing the incidence of RFS-related morbidity and mortality.

A review of machine learning for cardiology

This paper reviews recent cardiology literature and reports how Artificial Intelligence Tools (specifically, Machine Learning techniques) are being used by physicians in the field. Each technique is introduced with enough details to allow the understanding of how it works and its intent, but without delving into details that do not add immediate benefits and require expertise in the field. We specifically focus on the principal Machine Learning based risk scores used in cardiovascular research. After introducing them and summarizing their assumptions and biases, we discuss their merits and shortcomings. We report on how frequently they are adopted in the field and suggest why this is the case based on our expertise in Machine Learning. We complete the analysis by reviewing how corresponding statistical approaches compare with them. Finally, we discuss the main open issues in applying Machine Learning tools to cardiology tasks, also drafting possible future directions. Despite the growing interest in these tools, we argue that there are many still underutilized techniques: while Neural Networks are slowly being incorporated in cardiovascular research, other important techniques such as Semi-Supervised Learning and Federated Learning are still underutilized. The former would allow practitioners to harness the information contained in large datasets that are only partially labeled, while the latter would foster collaboration between institutions allowing building larger and better models.

Tensor learning of pointwise mutual information from EHR data for early prediction of sepsis

Early detection of sepsis can facilitate early clinical intervention with effective treatment and may reduce sepsis mortality rates. In view of this, machine learning-based automated diagnosis of sepsis using easily recordable physiological data can be more promising as compared to the gold standard rule-based clinical criteria in current practice. This study aims to develop such a machine learning framework that demonstrates the quantification of heterogeneity within the tabular electronic health records (EHR) data of clinical covariates to capture both linear relationships and nonlinear correlation for the early prediction of sepsis. Here, the statistics of pairwise association for each hour-covariate pair within the EHR data for every 6-hours window-duration with selected 24 covariates is described using pointwise mutual information (PMI) matrix. This matrix gives the heterogeneity of data as a two-dimensional map. Such matrices are fused horizontally along the z-axis as vertical slices in the xy plane to form a 3-way tensor for each record with the corresponding Length of Stay (L). Tensor factorization of such fused tensor for every record is performed using Tucker decomposition, and only the core tensors are retained later, excluding the 3 unitary matrices to provide the latent feature set for the prediction of sepsis onset. A five-fold cross-validation scheme is employed wherein the obtained 120 latent features from the reshaped core tensor, are fed to Light Gradient Boosting Machine Learning models (LightGBM) for binary classification, further alleviating the involved class imbalance. The machine-learning framework is designed via Bayesian optimization, yielding an average normalized utility score of 0.4519 as defined by challenge organizers and area under the receiver operating characteristic curve (AUROC) of 0.8621 on publicly available PhysioNet/Computing in Cardiology Challenge 2019 training data. The proposed tensor decomposition of 3-way fused tensor formulated using PMI matrices leverages higher-order temporal interactions between the pairwise associations among the clinical values for early prediction of sepsis. This is validated with improved risk prediction power for every hour of admission to the ICU in terms of utility score, AUROC, and F1 score. The results obtained show a significant improvement particularly in terms of utility score of ~1.5-2% under a 5-fold cross-validation scheme on entire training data as compared to a top entrant research study that participated in the challenge.

Interpretable deep learning for automatic diagnosis of 12-lead electrocardiogram

Electrocardiogram (ECG) is a widely used reliable, non-invasive approach for cardiovascular disease diagnosis. With the rapid growth of ECG examinations and the insufficiency of cardiologists, accurate and automatic diagnosis of ECG signals has become a hot research topic. In this paper, we developed a deep neural network for automatic classification of cardiac arrhythmias from 12-lead ECG recordings. Experiments on a public 12-lead ECG dataset showed the effectiveness of our method. The proposed model achieved an average F1 score of 0.813. The deep model showed superior performance than 4 machine learning methods learned from extracted expert features. Besides, the deep models trained on single-lead ECGs produce lower performance than using all 12 leads simultaneously. The best-performing leads are lead I, aVR, and V5 among 12 leads. Finally, we employed the SHapley Additive exPlanations method to interpret the model's behavior at both the patient level and population level.

•

We develop a deep learning model for the automatic diagnosis of ECG

•

We present benchmark results of 12-lead ECG classification

•

We find out the top performance single lead in diagnosing ECGs

•

We employ the SHAP method to enhance clinical interpretability

Machine Learning for the Prediction of Red Blood Cell Transfusion in Patients During or After Liver Transplantation Surgery

Aim: This study aimed to use machine learning algorithms to identify critical preoperative variables and predict the red blood cell (RBC) transfusion during or after liver transplantation surgery. Study Design and Methods: A total of 1,193 patients undergoing liver transplantation in three large tertiary hospitals in China were examined. Twenty-four preoperative variables were collected, including essential population characteristics, diagnosis, symptoms, and laboratory parameters. The cohort was randomly split into a train set (70%) and a validation set (30%). The Recursive Feature Elimination and eXtreme Gradient Boosting algorithms (XGBOOST) were used to select variables and build machine learning prediction models, respectively. Besides, seven other machine learning models and logistic regression were developed. The area under the receiver operating characteristic (AUROC) was used to compare the prediction performance of different models. The SHapley Additive exPlanations package was applied to interpret the XGBOOST model. Data from 31 patients at one of the hospitals were prospectively collected for model validation. Results: In this study, 72.1% of patients in the training set and 73.2% in the validation set underwent RBC transfusion during or after the surgery. Nine vital preoperative variables were finally selected, including the presence of portal hypertension, age, hemoglobin, diagnosis, direct bilirubin, activated partial thromboplastin time, globulin, aspartate aminotransferase, and alanine aminotransferase. The XGBOOST model presented significantly better predictive performance (AUROC: 0.813) than other models and also performed well in the prospective dataset (accuracy: 76.9%). Discussion: A model for predicting RBC transfusion during or after liver transplantation was successfully developed using a machine learning algorithm based on nine preoperative variables, which could guide high-risk patients to take appropriate preventive measures.