Characterization of long COVID temporal sub-phenotypes by distributed representation learning from electronic health record data: a cohort study.

Background: Characterizing Post-Acute Sequelae of COVID (SARS-CoV-2 Infection), or PASC has been challenging due to the multitude of sub-phenotypes, temporal attributes, and definitions. Scalable characterization of PASC sub-phenotypes can enhance screening capacities, disease management, and treatment planning. Methods: We conducted a retrospective multi-centre observational cohort study, leveraging longitudinal electronic health record (EHR) data of 30,422 patients from three healthcare systems in the Consortium for the Clinical Characterization of COVID-19 by EHR (4CE). From the total cohort, we applied a deductive approach on 12,424 individuals with follow-up data and developed a distributed representation learning process for providing augmented definitions for PASC sub-phenotypes. Findings: Our framework characterized seven PASC sub-phenotypes. We estimated that on average 15.7% of the hospitalized COVID-19 patients were likely to suffer from at least one PASC symptom and almost 5.98%, on average, had multiple symptoms. Joint pain and dyspnea had the highest prevalence, with an average prevalence of 5.45% and 4.53%, respectively. Interpretation: We provided a scalable framework to every participating healthcare system for estimating PASC sub-phenotypes prevalence and temporal attributes, thus developing a unified model that characterizes augmented sub-phenotypes across the different systems. Funding: Authors are supported by National Institute of Allergy and Infectious Diseases, National Institute on Aging, National Center for Advancing Translational Sciences, National Medical Research Council, National Institute of Neurological Disorders and Stroke, European Union, National Institutes of Health, National Center for Advancing Translational Sciences.

Evaluation of Machine Learning-based Patient Outcome Prediction Using Patient-specific Difficulty and Discrimination Indices.

Given the extensive use of machine learning in patient outcome prediction, and the understanding that the challenging nature of predictions in this field may considerably modify the performance of predictive models, research in this area requires some forms of context-sensitive performance metrics. The area under the receiver operating characteristic curve (AUC), precision, recall, specificity, and F1 are widely used measures of performance for patient outcome prediction. These metrics have several merits: they are easy to interpret and do not need any subjective input from the user. However, they weight all samples equally and do not adequately reflect the ability of predictive models in classifying difficult samples. In this paper, we propose the Difficulty Weight Adjustment (DWA) algorithm, a simple method that incorporates the difficulty level of samples when evaluating predictive models. Using a large dataset of 139,367 unique ICU admissions within the eICU Collaborative Research Database (eICU-CRD), we show that the classification difficulty and the discrimination ability of samples are critical aspects that need to be considered when comparing machine learning models that predict patient outcomes.

Novel Feature Selection for Artificial Intelligence Using Item Response Theory for Mortality Prediction.

Feature selection is a critical component in supervised machine learning classification analyses. Extraneous features introduce noise and inefficiencies into the system leading to a need for feature reduction techniques. Many feature reduction models use the end-classification results in the feature reduction process, committing a circular error. Item Response Theory (IRT) examines the characteristics of features independent of the end-classification results, and provides high levels of information regarding feature utility. A two-parameter dichotomous IRT model was used to analyze 18 features from an intensive care unit data set with 2520 cases. The classification results showed that the features selected via IRT were comparable to that using more traditional machine learning approaches. Strengths and limitations of the IRT selection protocol are discussed.