Although machine learning has become a powerful tool to augment doctors in clinical analysis, the immense amount of labeled data that is necessary to train supervised learning approaches burdens each development task as time and resource intensive. The vast majority of dense clinical information is stored in written reports, detailing pertinent patient information. The challenge with utilizing natural language data for standard model development is due to the complex and unstructured nature of the modality. In this research, a model pipeline was developed to utilize an unsupervised approach to train an encoder-language model, a bidirectional recurrent neural network, to generate document encodings; which then can be used as features passed into a decoder-classifier model that requires magnitudes less labeled data than previous approaches to differentiate between fine-grained disease classes accurately. The language model was trained on unlabeled radiology reports from the Massachusetts General Hospital Radiology Department (n=218,159) and terminated with a loss of 1.62 and a word prediction accuracy of 62%. The classification models were trained on three labeled datasets of head CT studies of reported patients, presenting large vessel occlusion (n=1403), acute ischemic strokes (n=331), and intracranial hemorrhage (n=4350), to identify a variety of different findings directly from the radiology report data; resulting in AUCs of 0.98, 0.95, and 0.99, respectively, for the large vessel occlusion, acute ischemic stroke, and intracranial hemorrhage datasets. The output encodings are able to be used in conjunction with imaging data, to create models that can process a multitude of different modalities. The ability to automatically extract relevant features from textual data allows for faster model development and integration of Preprint. Under review.

Regular eye examinations to screen for the initial signs of diabetic retinopathy (DR) are crucial for preventing vision loss. Teleretinal imaging (TRI) offered in a primary care setting provides a means to improve adherence to DR screening, particularly for patients who face challenges in visiting eye care providers regularly. The present study evaluates the utilization of TRI to screen for DR in an outpatient, hospital-based primary care clinic. Patients with diabetes mellitus (DM) but without DR were eligible for point-of-care screening facilitated by their primary care provider, utilizing a non-mydriatic, handheld fundus camera. Patient demographics and clinical characteristics were extracted from the electronic medical record. Patients who underwent TRI were more likely to be male, non-White, and have up-to-date monitoring and treatment measures, including hemoglobin A1c (HbA1c), microalbumin, and low-density lipoprotein (LDL) levels, in accordance with Healthcare Effectiveness Data and Information Set (HEDIS) guidelines. Our findings demonstrate that TRI can reduce screening costs compared to a strategy where all patients are referred for in-person eye examinations. A net present value (NPV) analysis indicates that a screening site reaches the break-even point of operation within one year if an average of two patients are screened per workday.

The COVID-19 global pandemic has caused unprecedented worldwide changes in healthcare delivery. While containment and mitigation approaches have been intensified, the progressive increase in the number of cases has overwhelmed health systems globally, highlighting the need for anticipation and prediction to be the basis of an efficient response system. This study demonstrates the role of population health metrics as early warning signs of future health crises. We retrospectively collected data from the emergency department of a large academic hospital in the northeastern United States from 01/01/2019 to 08/07/2021. A total of 377,694 patient records and 303 features were included for analysis. Departing from a multivariate artificial intelligence (AI) model initially developed to predict the risk of high-flow oxygen therapy or mechanical ventilation requirement during the COVID-19 pandemic, a total of 19 original variables and eight engineered features showing to be most predictive of the outcome were selected for further analysis. The temporal trends of the selected variables before and during the pandemic were characterized to determine their potential roles as early warning signs of future health crises. Temporal analysis of the individual variables included in the high-flow oxygen model showed that at a population level, the respiratory rate, temperature, low oxygen saturation, number of diagnoses during the first encounter, heart rate, BMI, age, sex, and neutrophil percentage demonstrated observable and traceable changes eight weeks before the first COVID-19 public health emergency declaration. Additionally, the engineered rule-based features built from the original variables also exhibited a pre-pandemic surge that preceded the first pandemic wave in spring 2020. Our findings suggest that the changes in routine population health metrics may serve as early warnings of future crises. This justifies the development of patient health surveillance systems, that can continuously monitor population health features, and alarm of new approaching public health crises before they become devastating.