Abstract:Background and Objectives: Sepsis is one of the common causes of death in intensive care units. A reliable prognostic model based on patients' data acquired at the intensive care unit (ICU) would enable clinicians to make treatment decisions to improve clinical outcomes for septic patients. This study aims to develop a machine-learning framework for building such prognostic tools by exploring the class-imbalanced longitudinal data of a group of septic patients.Methods: A feature-represented input dataset is devised in the form of concatenated triples to increase the data size relative to the dimension of the feature space. Each concatenated triplet consists of a patient's static data, the k-day consecutively collected longitudinal data, and the clinical outcome (k=2,3,4,5). The structured input data are then used to train classifiers in combination with appropriate feature engineering techniques. The trained classifiers are tested on a new set of septic patients to ensure their clinical efficacy. We implement the modeling approach using five classifiers: K nearest neighbors, Logistic Regression, Support Vector Machine, Random Forest (RF), and Extreme Gradient Boosting (XGBoost) coupled with a set of feature engineering techniques. AUROC and a new metric, $\gamma$, made up of the F1 score on the external validation set, are used to assess the efficacy of the models.Results: Five prognostic models are built on the feature-represented input dataset accounting for 10 selected dynamic features from the patient medical data. Our research shows that the XGBoost (AUROC=0.777, F1 score=0.694) and RF (AUROC=0.769, F1 score=0.647) model combined with the ensemble under-sampling strategy outperform all other models in the external validation or testing. For example, the improvement in AUROC and overfitting are (6.66\%, 54.96\%) and (0.52\%, 77.72\%) for the RF and XGBoost model with the sampling strategy compared to the same models without using the sampling strategy, respectively. This indicates that the machine-learning framework can greatly improve the accuracy and generalizability of standard classifiers. Conclusion:A new modeling framework is devised to develop prognostic tools for treatment outcomes of septic patients using small, class-imbalanced, and high-dimensional datasets. It enables standard classifiers to use small datasets to achieve relatively high predictability by engineering new structured datasets encoded with temporal features, sampling strategies, and dimension reduction techniques, providing clinically useful prognostic models and setting an example for applying machine learning methods to small data problems in medicine.

Development of Machine Learning Classifiers for Blood-based Diagnosis and Prognosis of Suspected Acute Infections and Sepsis

Using machine learning and centrifugal microfluidics at the point-of-need to predict clinical deterioration of patients with suspected sepsis within the first 24 h.

Aiding clinical assessment of neonatal sepsis using hematological analyzer data with machine learning techniques

Development and validation of a machine learning model integrated with the clinical workflow for early detection of sepsis

The application of metagenomics, radiomics and machine learning for diagnosis of sepsis

Machine-Learning-Enabled Prognostic Models for Sepsis

Development and utility assessment of a machine learning bloodstream infection classifier in pediatric patients receiving cancer treatments

Predicting bloodstream infection outcome using machine learning

Machine learning model for the prediction of gram-positive and gram-negative bacterial bloodstream infection based on routine laboratory parameters

Using machine learning to predict bacteremia in urgent care patients on the basis of triage data and laboratory results

Machine Learning Interpretability Methods to Characterize the Importance of Hematologic Biomarkers in Prognosticating Patients with Suspected Infection

Machine learning for early detection of sepsis: an internal and temporal validation study

Machine Learning with Alpha Toxin Phenotype to Predict Clinical Outcome in Patients with Staphylococcus aureus Bloodstream Infection

Using interpretable machine learning to predict bloodstream infection and antimicrobial resistance in patients admitted to ICU: Early alert predictors based on EHR data to guide antimicrobial stewardship

Machine learning algorithms for early sepsis detection in the emergency department: A retrospective study

Detection of Bacteremia in Surgical In-Patients Using Recurrent Neural Network Based on Time Series Records: Development and Validation Study (Preprint)

Early diagnosis of bloodstream infections in the intensive care unit using machine-learning algorithms

Development and Validation of a Machine Learning Model for Early Detection of Untreated Infection

Detection of Bacteremia in Surgical In-Patients Using Recurrent Neural Network Based on Time Series Records: Development and Validation Study

Early Prediction of Sepsis in EMR Records Using Traditional ML Techniques and Deep Learning LSTM Networks

Machine-Learning for Phenotyping and Prognostication of Myocardial Infarction and Injury in Suspected Acute Coronary Syndrome