Identification and Prediction of Clinical Phenotypes in Hospitalized Patients With COVID-19: Machine Learning From Medical Records.

BACKGROUND: There is significant heterogeneity in disease progression among hospitalized patients with COVID-19. The pathogenesis of SARS-CoV-2 infection is attributed to a complex interplay between virus and host immune response that in some patients unpredictably and rapidly leads to "hyperinflammation" associated with increased risk of mortality. The early identification of patients at risk of progression to hyperinflammation may help inform timely therapeutic decisions and lead to improved outcomes. OBJECTIVE: The primary objective of this study was to use machine learning to reproducibly identify specific risk-stratifying clinical phenotypes across hospitalized patients with COVID-19 and compare treatment response characteristics and outcomes. A secondary objective was to derive a predictive phenotype classification model using routinely available early encounter data that may be useful in informing optimal COVID-19 bedside clinical management. METHODS: This was a retrospective analysis of electronic health record data of adult patients (N=4379) who were admitted to a Johns Hopkins Health System hospital for COVID-19 treatment from 2020 to 2021. Phenotypes were identified by clustering 38 routine clinical observations recorded during inpatient care. To examine the reproducibility and validity of the derived phenotypes, patient data were randomly divided into 2 cohorts, and clustering analysis was performed independently for each cohort. A predictive phenotype classifier using the gradient-boosting machine method was derived using routine clinical observations recorded during the first 6 hours following admission. RESULTS: A total of 2 phenotypes (designated as phenotype 1 and phenotype 2) were identified in patients admitted for COVID-19 in both the training and validation cohorts with similar distributions of features, correlations with biomarkers, treatments, comorbidities, and outcomes. In both the training and validation cohorts, phenotype-2 patients were older; had elevated markers of inflammation; and were at an increased risk of requiring intensive care unit-level care, developing sepsis, and mortality compared with phenotype-1 patients. The gradient-boosting machine phenotype prediction model yielded an area under the curve of 0.89 and a positive predictive value of 0.83. CONCLUSIONS: Using machine learning clustering, we identified and internally validated 2 clinical COVID-19 phenotypes with distinct treatment or response characteristics consistent with similar 2-phenotype models derived from other hospitalized populations with COVID-19, supporting the reliability and generalizability of these findings. COVID-19 phenotypes can be accurately identified using machine learning models based on readily available early encounter clinical data. A phenotype prediction model based on early encounter data may be clinically useful for timely bedside risk stratification and treatment personalization.

Pediatric sepsis phenotypes for enhanced therapeutics: An application of clustering to electronic health records.

OBJECTIVE: The heterogeneity of pediatric sepsis patients suggests the potential benefits of clustering analytics to derive phenotypes with distinct host response patterns that may help guide personalized therapeutics. We evaluate the relative performance of latent class analysis (LCA) and K-means, 2 commonly used clustering methods toward the derivation of clinically useful pediatric sepsis phenotypes. METHODS: Data were extracted from anonymized medical records of 6446 pediatric patients that presented to 1 of 6 emergency departments (EDs) between 2013 and 2018 and were thereafter admitted. Using International Classification of Diseases (ICD)-9 and ICD-10 discharge codes, 151 patients were identified with a sepsis continuum diagnosis that included septicemia, sepsis, severe sepsis, and septic shock. Using feature sets used in related clustering studies, LCA and K-means algorithms were used to derive 4 distinct phenotypic pediatric sepsis segmentations. Each segmentation was evaluated for phenotypic homogeneity, separation, and clinical use. RESULTS: Using the 2 feature sets, LCA clustering resulted in 2 similar segmentations of 4 clinically distinct phenotypes, while K-means clustering resulted in segmentations of 3 and 4 phenotypes. All 4 segmentations identified at least 1 high severity phenotype, but LCA-identified phenotypes reflected superior stratification, high entropy approaching 1 (eg, 0.994) indicating excellent separation between estimated phenotypes, and differential treatment/treatment response, and outcomes that were non-randomly distributed across phenotypes (P < 0.001). CONCLUSION: Compared to K-means, which is commonly used in clustering studies, LCA appears to be a more robust, clinically useful statistical tool in analyzing a heterogeneous pediatric sepsis cohort toward informing targeted therapies. Additional prospective studies are needed to validate clinical utility of predictive models that target derived pediatric sepsis phenotypes in emergency department settings.

Contact Lens Wear, Corneal Complications, and U.S. Service Member Readiness.

INTRODUCTION: Ulcerative keratitis (UK), or corneal ulcer, is a sight-threatening and readiness-lowering medical condition that begins with a corneal infiltrative event (CIE). Contact lens (CL) wear poses a particular risk for a CIE and therefore is restricted for most active duty service members (SMs). In this study, we explored a large Department of Defense/Veterans Affairs (DoD/VA) database to estimate the prevalence of UK and CIE and their association with CL wear. MATERIALS AND METHODS: The DoD/VA Defense and Veterans Eye Injury Vision Registry, an initiative of the DoD/VA Vision Center of Excellence, was explored using natural language processing software to search for words and diagnostic codes that might identify cornea injuries and CL wear. The effect of UK and CIE on readiness was explored by evaluating the duration between the first and final visits noted in the database. RESULTS: A total of 213 UK cases were identified among the 27,402 SMs for whom data were recorded in Defense and Veterans Eye Injury Vision Registry. The odds ratios of UK and CIE being associated with CL wear were 13.34 and 2.20, respectively. A less specific code (superficial corneal injury) was found to be the most commonly used diagnosis in the database, and the odds ratio of CL wearers having that diagnosis was 2.25. CL-wearing patients with corneal disease also required more clinic encounters than those who did not wear CLs. CONCLUSIONS: This study supports the current restriction on CL wear among nonpilot active duty SMs and quantifies the significantly enhanced risk of developing corneal ulcers posed by that habit.

The Relationship Between Severe Visual Acuity Loss, Traumatic Brain Injuries, and Ocular Injuries in American Service Members From 2001 to 2015.

INTRODUCTION: Although traumatic brain injury (TBI) is known to cause many visual problems, the correlation between the extent of severe visual acuity loss (SVAL) and severity of TBI has not been widely explored. In this retrospective analysis, combined information from Department of Defense (DoD)/Veterans Affairs ocular injury and TBI repositories were used to evaluate the relationship between chronic SVAL, TBI, ocular injuries, and associated ocular sequelae for U.S. service members serving between 2001 and 2015. MATERIALS AND METHODS: The Defense and Veterans Eye Injury and Vision Registry (DVEIVR) is an initiative led by the DoD and Veterans Affairs that consists of clinical and related data for service members serving in theater since 2001. The Defense and Veterans Brain Injury Center (DVBIC) is the DoD's office for tracking TBI data in the military and maintains data on active-duty service members with a TBI diagnosis since 2000. Longitudinal data from these 2 resources for encounters between February 2001 and October 2015 were analyzed to understand the relation between SVAL, and TBI while adjusting for ocular covariates such as open globe injury (OGI), disorders of the anterior segment and disorders of the posterior segment in a logistic regression model. TBI cases in DVEIVR were identified using DVBIC data and classified according to International Statistical Classification of Diseases criteria established by DVBIC. Head trauma and other open head wounds (OOHW) were also included. SVAL cases in DVEIVR were identified using both International Statistical Classification of Diseases criteria for blindness and low vision as well as visual acuity test data recorded in DVEIVR. RESULTS: Data for a total of 25,193 unique patients with 88,996 encounters were recorded in DVEIVR from February, 2001 to November, 2015. Of these, 7,217 TBI and 1,367 low vision cases were identified, with 638 patients experiencing both. In a full logistic model, neither UTBI nor differentiated TBI (DTBI, ie, mild, moderate, severe, penetrating, or unclassified) were significant risk factors for SVAL although ocular injuries (disorders of the anterior segment, disorders of the posterior segment, and OGI) and OOHW were significant. CONCLUSION: Any direct injury to the eye or head risks SVAL but the location and severity will modify that risk. After adjusting for OGIs, OOHW and their sequelae, TBI was found to not be a significant risk factor for SVAL in patients recorded in DVEIVR. Further research is needed to explore whether TBI is associated with more moderate levels of vision acuity loss.

Towards Reliable ARDS Clinical Decision Support: ARDS Patient Analytics with Free-text and Structured EMR Data.

In this work, we utilize a combination of free-text and structured data to build Acute Respiratory Distress Syndrome(ARDS) prediction models and ARDS phenotype clusters. We derived 'Patient Context Vectors' representing patientspecific contextual ARDS risk factors, utilizing deep-learning techniques on ICD and free-text clinical notes data. The Patient Context Vectors were combined with structured data from the first 24 hours of admission, such as vital signs and lab results, to build an ARDS patient prediction model and an ARDS patient mortality prediction model achieving AUC of 90.16 and 81.01 respectively. The ability of Patient Context Vectors to summarize patients' medical history and current conditions is also demonstrated by the automatic clustering of ARDS patients into clinically meaningful phenotypes based on comorbidities, patient history, and presenting conditions. To our knowledge, this is the first study to successfully combine free-text and structured data, without any manual patient risk factor curation, to build real-time ARDS prediction models.

Open Globe Injury Patient Identification in Warfare Clinical Notes.

The aim of this study is to utilize the Defense and Veterans Eye Injury and Vision Registry clinical data derived from DoD and VA medical systems which include documentation of care while in combat, and develop methods for comprehensive and reliable Open Globe Injury (OGI) patient identification. In particular, we focus on the use of free-form clinical notes, since structured data, such as diagnoses or procedure codes, as found in early post-trauma clinical records, may not be a comprehensive and reliable indicator of OGIs. The challenges of the task include low incidence rate (few positive examples), idiosyncratic military ophthalmology vocabulary, extreme brevity of notes, specialized abbreviations, typos and misspellings. We modeled the problem as a text classification task and utilized a combination of supervised learning (SVMs) and word embeddings learnt in a unsupervised manner, achieving a precision of 92.50% and a recall of89.83%o. The described techniques are applicable to patient cohort identification with limited training data and low incidence rate.