Generative large language models are all-purpose text analytics engines: text-to-text learning is all your need.

OBJECTIVE: To solve major clinical natural language processing (NLP) tasks using a unified text-to-text learning architecture based on a generative large language model (LLM) via prompt tuning. METHODS: We formulated 7 key clinical NLP tasks as text-to-text learning and solved them using one unified generative clinical LLM, GatorTronGPT, developed using GPT-3 architecture and trained with up to 20 billion parameters. We adopted soft prompts (ie, trainable vectors) with frozen LLM, where the LLM parameters were not updated (ie, frozen) and only the vectors of soft prompts were updated, known as prompt tuning. We added additional soft prompts as a prefix to the input layer, which were optimized during the prompt tuning. We evaluated the proposed method using 7 clinical NLP tasks and compared them with previous task-specific solutions based on Transformer models. RESULTS AND CONCLUSION: The proposed approach achieved state-of-the-art performance for 5 out of 7 major clinical NLP tasks using one unified generative LLM. Our approach outperformed previous task-specific transformer models by ∼3% for concept extraction and 7% for relation extraction applied to social determinants of health, 3.4% for clinical concept normalization, 3.4%-10% for clinical abbreviation disambiguation, and 5.5%-9% for natural language inference. Our approach also outperformed a previously developed prompt-based machine reading comprehension (MRC) model, GatorTron-MRC, for clinical concept and relation extraction. The proposed approach can deliver the "one model for all" promise from training to deployment using a unified generative LLM.

Identifying Predictors of Initial Surgical Failure during Minimally Invasive Endoscopic Intracerebral Hemorrhage Evacuation.

Background and Purpose: Intracerebral hemorrhage (ICH) is a common and severe disease with high rates of morbidity and mortality; however, minimally invasive surgical (MIS) hematoma evacuation represents a promising avenue for treatment. In February of 2019, the MISTIE III study found that stereotactic thrombolysis with catheter drainage did not benefit patients with supratentorial spontaneous ICH but that a clinical benefit may be present when no more than 15 mL of hematoma remains at the end of treatment. Intraoperative CT (iCT) imaging has the ability to assess whether or not this surgical goal has been met in real time, allowing for operations to add additional CT-informed 'evacuation periods' (EPs) to achieve the surgical goal. Here, we report on the frequency and predictors of initial surgical failure on at least one iCT requiring additional EPs in a large cohort of patients undergoing endoscopic minimally invasive ICH evacuation with the SCUBA technique. Methods: All patients who underwent minimally invasive endoscopic evacuation of supratentorial spontaneous ICH in a major health system between December 2015 and October 2018 were included in this study. Patient demographics, clinical and radiographic features, procedural details, and outcomes were analyzed retrospectively from a prospectively collected database. Procedures were characterized as initially successful when the first iCT demonstrated that surgical success had been achieved and initially unsuccessful when the surgical goal was not achieved, and additional EPs were performed. The surgical goal was prospectively identified in December of 2015 as leaving no more than 20% of the preoperative hematoma volume at the end of the procedure. Descriptive statistics and regression analyses were performed to identify predictors of initial failure and secondary rescue. Results: Patients (100) underwent minimally invasive endoscopic ICH evacuation in the angiography suite during the study time period. In 14 cases, the surgical goal was not met on the first iCT and multiple Eps were performed; in 10 cases the surgical goal was not met, and no additional EPs were performed. In 14 cases, the surgical goal was never achieved. When additional EPs were performed, a rescue rate of 71.4% (10/14) was seen, bringing the total percentage of cases meeting the surgical goal to 86% across the entire cohort. Cases in which the surgical goal was not achieved were significantly associated with older patients (68 years vs. 60 years; p = 0.0197) and higher rates of intraventricular hemorrhage (34.2% vs. 70.8%; p = 0.0021). Cases in which the surgical goal was rescued from initial failure had similar levels of IVH, suggesting that these additional complexities can be overcome with the use of additional iCT-informed EPs. Conclusions: Initial and ultimate surgical failure occurs in a small percentage of patients undergoing minimally invasive endoscopic ICH evacuation. The use of intraoperative imaging provides an opportunity to evaluate whether or not the surgical goal has been achieved, and to continue the procedure if the surgeon feels that more evacuation is achievable. Now that level-one evidence exists to target a surgical evacuation goal during minimally invasive ICH evacuation, intraoperative imaging, such as iCT, plays an important role in aiding the surgical team to achieve the surgical goal.

A study of generative large language model for medical research and healthcare.

There are enormous enthusiasm and concerns in applying large language models (LLMs) to healthcare. Yet current assumptions are based on general-purpose LLMs such as ChatGPT, which are not developed for medical use. This study develops a generative clinical LLM, GatorTronGPT, using 277 billion words of text including (1) 82 billion words of clinical text from 126 clinical departments and approximately 2 million patients at the University of Florida Health and (2) 195 billion words of diverse general English text. We train GatorTronGPT using a GPT-3 architecture with up to 20 billion parameters and evaluate its utility for biomedical natural language processing (NLP) and healthcare text generation. GatorTronGPT improves biomedical natural language processing. We apply GatorTronGPT to generate 20 billion words of synthetic text. Synthetic NLP models trained using synthetic text generated by GatorTronGPT outperform models trained using real-world clinical text. Physicians' Turing test using 1 (worst) to 9 (best) scale shows that there are no significant differences in linguistic readability (p = 0.22; 6.57 of GatorTronGPT compared with 6.93 of human) and clinical relevance (p = 0.91; 7.0 of GatorTronGPT compared with 6.97 of human) and that physicians cannot differentiate them (p < 0.001). This study provides insights into the opportunities and challenges of LLMs for medical research and healthcare.

Health system-scale language models are all-purpose prediction engines.

Physicians make critical time-constrained decisions every day. Clinical predictive models can help physicians and administrators make decisions by forecasting clinical and operational events. Existing structured data-based clinical predictive models have limited use in everyday practice owing to complexity in data processing, as well as model development and deployment1-3. Here we show that unstructured clinical notes from the electronic health record can enable the training of clinical language models, which can be used as all-purpose clinical predictive engines with low-resistance development and deployment. Our approach leverages recent advances in natural language processing4,5 to train a large language model for medical language (NYUTron) and subsequently fine-tune it across a wide range of clinical and operational predictive tasks. We evaluated our approach within our health system for five such tasks: 30-day all-cause readmission prediction, in-hospital mortality prediction, comorbidity index prediction, length of stay prediction, and insurance denial prediction. We show that NYUTron has an area under the curve (AUC) of 78.7-94.9%, with an improvement of 5.36-14.7% in the AUC compared with traditional models. We additionally demonstrate the benefits of pretraining with clinical text, the potential for increasing generalizability to different sites through fine-tuning and the full deployment of our system in a prospective, single-arm trial. These results show the potential for using clinical language models in medicine to read alongside physicians and provide guidance at the point of care.

A large language model for electronic health records.

There is an increasing interest in developing artificial intelligence (AI) systems to process and interpret electronic health records (EHRs). Natural language processing (NLP) powered by pretrained language models is the key technology for medical AI systems utilizing clinical narratives. However, there are few clinical language models, the largest of which trained in the clinical domain is comparatively small at 110 million parameters (compared with billions of parameters in the general domain). It is not clear how large clinical language models with billions of parameters can help medical AI systems utilize unstructured EHRs. In this study, we develop from scratch a large clinical language model-GatorTron-using >90 billion words of text (including >82 billion words of de-identified clinical text) and systematically evaluate it on five clinical NLP tasks including clinical concept extraction, medical relation extraction, semantic textual similarity, natural language inference (NLI), and medical question answering (MQA). We examine how (1) scaling up the number of parameters and (2) scaling up the size of the training data could benefit these NLP tasks. GatorTron models scale up the clinical language model from 110 million to 8.9 billion parameters and improve five clinical NLP tasks (e.g., 9.6% and 9.5% improvement in accuracy for NLI and MQA), which can be applied to medical AI systems to improve healthcare delivery. The GatorTron models are publicly available at: https://catalog.ngc.nvidia.com/orgs/nvidia/teams/clara/models/gatortron_og .

Deploying deep learning models on unseen medical imaging using adversarial domain adaptation.

The fundamental challenge in machine learning is ensuring that trained models generalize well to unseen data. We developed a general technique for ameliorating the effect of dataset shift using generative adversarial networks (GANs) on a dataset of 149,298 handwritten digits and dataset of 868,549 chest radiographs obtained from four academic medical centers. Efficacy was assessed by comparing area under the curve (AUC) pre- and post-adaptation. On the digit recognition task, the baseline CNN achieved an average internal test AUC of 99.87% (95% CI, 99.87-99.87%), which decreased to an average external test AUC of 91.85% (95% CI, 91.82-91.88%), with an average salvage of 35% from baseline upon adaptation. On the lung pathology classification task, the baseline CNN achieved an average internal test AUC of 78.07% (95% CI, 77.97-78.17%) and an average external test AUC of 71.43% (95% CI, 71.32-71.60%), with a salvage of 25% from baseline upon adaptation. Adversarial domain adaptation leads to improved model performance on radiographic data derived from multiple out-of-sample healthcare populations. This work can be applied to other medical imaging domains to help shape the deployment toolkit of machine learning in medicine.

Population scale latent space cohort matching for the improved use and exploration of observational trial data.

A significant amount of clinical research is observational by nature and derived from medical records, clinical trials, and large-scale registries. While there is no substitute for randomized, controlled experimentation, such experiments or trials are often costly, time consuming, and even ethically or practically impossible to execute. Combining classical regression and structural equation modeling with matching techniques can leverage the value of observational data. Nevertheless, identifying variables of greatest interest in high-dimensional data is frequently challenging, even with application of classical dimensionality reduction and/or propensity scoring techniques. Here, we demonstrate that projecting high-dimensional medical data onto a lower-dimensional manifold using deep autoencoders and post-hoc generation of treatment/control cohorts based on proximity in the lower-dimensional space results in better matching of confounding variables compared to classical propensity score matching (PSM) in the original high-dimensional space (P<0.0001) and performs similarly to PSM models constructed by experts with prior knowledge of the underlying pathology when evaluated on predicting risk ratios from real-world clinical data. Thus, in cases when the underlying problem is poorly understood and the data is high-dimensional in nature, matching in the autoencoder latent space might be of particular benefit.

Development and Clinical Validation of a Finite Element Method Model Mapping Focal Intracranial Cooling.

Therapeutic hypothermia (TH) is a common and effective technique to reduce inflammation and induce neuroprotection across a variety of diseases. Focal TH of the brain can avoid the side effects of systemic cooling. The degree and extent of focal TH are a function of cooling probe design and local brain thermoregulation processes. To refine focal TH probe design, with application-specific optimization, we develop precise computational models of brain thermodynamics under intense local cooling. Here, we present a novel multiphysics in silico model that can accurately predict brain response to focal cooling. The model was parameterized from previously described values of metabolic activity, thermal conductivity, and temperature-dependent cerebral perfusion. The model was validated experimentally using data from clinical cases where local cooling was induced intracranially and brain temperatures monitored in real-time with MR thermometry. The validated model was then used to identify optimal design probe parameters to maximize volumetric TH, including considering three stratifications of cooling (mild, moderate, and profound) to produce Volume of Tissue Cooled (VOTC) maps. We report cooling radius increases in a nearly linear fashion with probe length and decreasing probe surface temperature.

Therapeutic hypothermia for intracerebral hemorrhage: Systematic review and meta-analysis of the experimental and clinical literature.

BACKGROUND: Intracerebral hemorrhage remains the deadliest form of stroke worldwide, inducing neuronal death through a wide variety of pathways. Therapeutic hypothermia is a robust and well-studied neuroprotectant widely used across a variety of specialties. AIMS: This review summarizes results from preclinical and clinical studies to highlight the overall effectiveness of therapeutic hypothermia to improve long-term intracerebral hemorrhage outcomes while also elucidating optimal protocol regimens to maximize therapeutic effect. SUMMARY OF REVIEW: A systematic review was conducted across three databases to identify trials investigating the use of therapeutic hypothermia to treat intracerebral hemorrhage. A random-effects meta-analysis was conducted on preclinical studies, looking at neurobehavioral outcomes, blood brain barrier breakdown, cerebral edema, hematoma volume, and tissue loss. Several mixed-methods meta-regression models were also performed to adjust for variance and variations in hypothermia induction procedures. Twwenty-one preclinical studies and five human studies were identified. The meta-analysis of preclinical studies demonstrated a significant benefit in behavioral scores (ES = -0.43, p = 0.02), cerebral edema (ES = 1.32, p = 0.0001), and blood brain barrier (ES = 2.73, p ≤ 0.00001). Therapeutic hypothermia was not found to significantly affect hematoma expansion (ES = -0.24, p = 0.12) or tissue loss (ES = 0.06, p = 0.68). Clinical study outcome reporting was heterogeneous; however, there was recurring evidence of therapeutic hypothermia-induced edema reduction. CONCLUSIONS: The combined preclinical evidence demonstrates that therapeutic hypothermia reduced multiple cell death mechanisms initiated by intracerebral hemorrhage; yet, there is no definitive evidence in clinical studies. The cooling strategies employed in both preclinical and clinical studies were highly diverse, and focused refinement of cooling protocols should be developed in future preclinical studies. The current data for therapeutic hypothermia in intracerebral hemorrhage remains questionable despite the highly promising indications in preclinical studies. Definitive randomized controlled studies are still required to answer this therapeutic question.

Stimulating the Facial Nerve to Treat Ischemic Stroke: A Systematic Review.

Acute ischemic stroke (AIS) is a common devastating disease that has increased yearly in absolute number of cases since 1990. While mechanical thrombectomy and tissue plasminogen activator (tPA) have proven to be effective treatments, their window-of-efficacy time is very short, leaving many patients with no viable treatment option. Over recent years there has been a growing interest in stimulating the facial nerves or ganglions to treat AIS. Pre-clinical studies have consistently demonstrated an increase in collateral blood flow (CBF) following ganglion stimulation, with positive indications in infarct size and neurological scores. Extensive human trials have focused on trans-oral electrical stimulation of the sphenopalatine ganglion, but have suffered from operational limitations and non-significant clinical findings. Regardless, the potential of ganglion stimulation to treat AIS or elongate the window-of-efficacy for current stroke treatments remains extremely promising. This review aims to summarize results from recent trial publications, highlight current innovations, and discuss future directions for the field. Importantly, this review comes after the release of four important clinical trials that were published in mid 2019.

Scalable, effective, and rapid decontamination of SARS-CoV-2 contaminated N95 respirators using germicidal ultraviolet C (UVC) irradiation device.

Particulate respirators such as N95s are an essential component of personal protective equipment (PPE) for front-line workers. This study describes a rapid and effective UVC irradiation system that would facilitate the safe re-use of N95 respirators and provides supporting information for deploying UVC for decontamination of SARS-CoV-2 during the COVID-19 pandemic. To assess the inactivation potential of the proposed UVC germicidal device as a function of time by using 3 M 8211-N95 particulate respirators inoculated with SARS-CoV-2. A germicidal UVC device to deliver tailored UVC dose was developed and test coupons (2.5 cm2) of the 3 M-N95 respirator were inoculated with 106 plaque-forming units (PFU) of SARS-CoV-2 and were UV irradiated. Different exposure times were tested (0-164 s) by fixing the distance between the lamp and the test coupon to 15.2 cm while providing an exposure of at least 5.43 mWcm-2. Primary measure of outcome was titration of infectious virus recovered from virus-inoculated respirator test coupons after UVC exposure. Other measures included the method validation of the irradiation protocol, using lentiviruses (biosafety level-2 agent) and establishment of the germicidal UVC exposure protocol. An average of 4.38 × 103 PFU ml-1 (SD 772.68) was recovered from untreated test coupons while 4.44 × 102 PFU ml-1 (SD 203.67), 4.00 × 102 PFU ml-1 (SD 115.47), 1.56 × 102 PFU ml-1 (SD 76.98) and 4.44 × 101 PFU ml-1 (SD 76.98) was recovered in exposures 2, 6, 18 and 54 s per side respectively. The germicidal device output and positioning was monitored and a minimum output of 5.43 mW cm-2 was maintained. Infectious SARS-CoV-2 was not detected by plaque assays (minimal level of detection is 67 PFU ml-1) on N95 respirator test coupons when irradiated for 120 s per side or longer suggesting 3.5 log reduction in 240 s of irradiation, 1.3 J cm-2. A scalable germicidal UVC device to deliver tailored UVC dose for rapid decontamination of SARS-CoV-2 was developed. UVC germicidal irradiation of N95 test coupons inoculated with SARS-CoV-2 for 120 s per side resulted in 3.5 log reduction of virus. These data support the reuse of N95 particle-filtrate apparatus upon irradiation with UVC and supports use of UVC-based decontamination of SARS-CoV-2 during the COVID-19 pandemic.

Federated learning for predicting clinical outcomes in patients with COVID-19.

Federated learning (FL) is a method used for training artificial intelligence models with data from multiple sources while maintaining data anonymity, thus removing many barriers to data sharing. Here we used data from 20 institutes across the globe to train a FL model, called EXAM (electronic medical record (EMR) chest X-ray AI model), that predicts the future oxygen requirements of symptomatic patients with COVID-19 using inputs of vital signs, laboratory data and chest X-rays. EXAM achieved an average area under the curve (AUC) >0.92 for predicting outcomes at 24 and 72 h from the time of initial presentation to the emergency room, and it provided 16% improvement in average AUC measured across all participating sites and an average increase in generalizability of 38% when compared with models trained at a single site using that site's data. For prediction of mechanical ventilation treatment or death at 24 h at the largest independent test site, EXAM achieved a sensitivity of 0.950 and specificity of 0.882. In this study, FL facilitated rapid data science collaboration without data exchange and generated a model that generalized across heterogeneous, unharmonized datasets for prediction of clinical outcomes in patients with COVID-19, setting the stage for the broader use of FL in healthcare.

An Hepatozoon americanum-like protozoan in crab-eating (Cerdocyon thous) and grey pampean (Lycalopex gymnocercus) foxes from Uruguay.

In South America, apicomplexan parasites of the genus Hepatozoon have been sporadically detected in mammals. Previous studies in wild canids from Brazil and Argentina demonstrated infections by species genetically related to Hepatozoon americanum. The aim of the present work was to detect the presence of Hepatozoon in road-killed foxes encountered in Uruguayan highways. Blood samples from 45 crab-eating (Cerdocyon thous) and 32 grey pampean (Lycalopex gymnocercus) foxes were analyzed by PCR for Hepatozoon 18S rRNA gene. Eight foxes (10.4%) were found to be infected with an H. americanum-like protozoan, an Hepatozoon closely related to H. americanum. Bayesian and maximum-likelihood phylogenetic analyses revealed that the sequences obtained in this study cluster with H. americanum from the United States, and with an H. americanum-like species from dog and foxes from Brazil and Argentina. In the Unites States, H. americanum causes severe disease in dogs. In addition to this, an increasing habitat overlap between dogs and foxes makes the presence of H. americanum-like protozoan in foxes acquires veterinary relevance. This work represents the first report of L. gymnocercus infected with an H. americanum-like protozoan, and of wild canids infected with Hepatozoon in Uruguay.

Combining Initial Radiographs and Clinical Variables Improves Deep Learning Prognostication in Patients with COVID-19 from the Emergency Department.

PURPOSE: To train a deep learning classification algorithm to predict chest radiograph severity scores and clinical outcomes in patients with coronavirus disease 2019 (COVID-19). MATERIALS AND METHODS: In this retrospective cohort study, patients aged 21-50 years who presented to the emergency department (ED) of a multicenter urban health system from March 10 to 26, 2020, with COVID-19 confirmation at real-time reverse-transcription polymerase chain reaction screening were identified. The initial chest radiographs, clinical variables, and outcomes, including admission, intubation, and survival, were collected within 30 days (n = 338; median age, 39 years; 210 men). Two fellowship-trained cardiothoracic radiologists examined chest radiographs for opacities and assigned a clinically validated severity score. A deep learning algorithm was trained to predict outcomes on a holdout test set composed of patients with confirmed COVID-19 who presented between March 27 and 29, 2020 (n = 161; median age, 60 years; 98 men) for both younger (age range, 21-50 years; n = 51) and older (age >50 years, n = 110) populations. Bootstrapping was used to compute CIs. RESULTS: The model trained on the chest radiograph severity score produced the following areas under the receiver operating characteristic curves (AUCs): 0.80 (95% CI: 0.73, 0.88) for the chest radiograph severity score, 0.76 (95% CI: 0.68, 0.84) for admission, 0.66 (95% CI: 0.56, 0.75) for intubation, and 0.59 (95% CI: 0.49, 0.69) for death. The model trained on clinical variables produced an AUC of 0.64 (95% CI: 0.55, 0.73) for intubation and an AUC of 0.59 (95% CI: 0.50, 0.68) for death. Combining chest radiography and clinical variables increased the AUC of intubation and death to 0.88 (95% CI: 0.79, 0.96) and 0.82 (95% CI: 0.72, 0.91), respectively. CONCLUSION: The combination of imaging and clinical information improves outcome predictions.Supplemental material is available for this article.© RSNA, 2020.

Deep anomaly detection of seizures with paired stereoelectroencephalography and video recordings.

Real-time seizure detection is a resource intensive process as it requires continuous monitoring of patients on stereoelectroencephalography. This study improves real-time seizure detection in drug resistant epilepsy (DRE) patients by developing patient-specific deep learning models that utilize a novel self-supervised dynamic thresholding approach. Deep neural networks were constructed on over 2000 h of high-resolution, multichannel SEEG and video recordings from 14 DRE patients. Consensus labels from a panel of epileptologists were used to evaluate model efficacy. Self-supervised dynamic thresholding exhibited improvements in positive predictive value (PPV; difference: 39.0%; 95% CI 4.5-73.5%; Wilcoxon-Mann-Whitney test; N = 14; p = 0.03) with similar sensitivity (difference: 14.3%; 95% CI - 21.7 to 50.3%; Wilcoxon-Mann-Whitney test; N = 14; p = 0.42) compared to static thresholds. In some models, training on as little as 10 min of SEEG data yielded robust detection. Cross-testing experiments reduced PPV (difference: 56.5%; 95% CI 25.8-87.3%; Wilcoxon-Mann-Whitney test; N = 14; p = 0.002), while multimodal detection significantly improved sensitivity (difference: 25.0%; 95% CI 0.2-49.9%; Wilcoxon-Mann-Whitney test; N = 14; p < 0.05). Self-supervised dynamic thresholding improved the efficacy of real-time seizure predictions. Multimodal models demonstrated potential to improve detection. These findings are promising for future deployment in epilepsy monitoring units to enable real-time seizure detection without annotated data and only minimal training time in individual patients.

Stereoscopic virtual reality does not improve knowledge acquisition of congenital heart disease.

Advances in virtual reality have made it possible for clinicians and trainees to interact with 3D renderings of hearts with congenital heart disease in 3D stereoscopic vision. No study to date has assessed whether this technology improved instruction compared to standard 2D interfaces. The purpose of this study was to assess whether stereoscopic virtual reality improves congenital heart disease anatomy education. Subjects in a prospective, blinded, randomized trial completed a pre-test assessing factual and visuospatial knowledge of common atrioventricular canal and were randomized to an intervention or control group based on their score. The intervention group used a 3D virtual reality (VR) headset to visualize a lecture with 3D heart models while the control group used a desktop (DT) computer interface with the same models. Subjects took a post-test and provided subjective feedback. 51 subjects were enrolled, 24 in the VR group & 27 in the DT group. The median score difference for VR subjects was 12 (IQR 9-13.3), compared to 10 (IQR 7.5-12) in the DT group. No difference in score improvement was found (p = 0.11). VR subjects' impression of the ease of use of their interface was higher than DT subjects (median 8 vs 7, respectively, p = 0.01). VR subjects' impression of their understanding of the subject matter was higher than desktop subjects (median 7 vs 5, respectively, p = 0.01). There was no statistically significant difference in the knowledge acquisition observed between the stereoscopic virtual reality group and the monoscopic desktop-based group. Participants in virtual reality reported a better learning experience and self-assessment suggesting virtual reality may increase learner engagement in understanding congenital heart disease.

Predicting adult neuroscience intensive care unit admission from emergency department triage using a retrospective, tabular-free text machine learning approach.

Early admission to the neurosciences intensive care unit (NSICU) is associated with improved patient outcomes. Natural language processing offers new possibilities for mining free text in electronic health record data. We sought to develop a machine learning model using both tabular and free text data to identify patients requiring NSICU admission shortly after arrival to the emergency department (ED). We conducted a single-center, retrospective cohort study of adult patients at the Mount Sinai Hospital, an academic medical center in New York City. All patients presenting to our institutional ED between January 2014 and December 2018 were included. Structured (tabular) demographic, clinical, bed movement record data, and free text data from triage notes were extracted from our institutional data warehouse. A machine learning model was trained to predict likelihood of NSICU admission at 30 min from arrival to the ED. We identified 412,858 patients presenting to the ED over the study period, of whom 1900 (0.5%) were admitted to the NSICU. The daily median number of ED presentations was 231 (IQR 200-256) and the median time from ED presentation to the decision for NSICU admission was 169 min (IQR 80-324). A model trained only with text data had an area under the receiver-operating curve (AUC) of 0.90 (95% confidence interval (CI) 0.87-0.91). A structured data-only model had an AUC of 0.92 (95% CI 0.91-0.94). A combined model trained on structured and text data had an AUC of 0.93 (95% CI 0.92-0.95). At a false positive rate of 1:100 (99% specificity), the combined model was 58% sensitive for identifying NSICU admission. A machine learning model using structured and free text data can predict NSICU admission soon after ED arrival. This may potentially improve ED and NSICU resource allocation. Further studies should validate our findings.

Federated Learning of Electronic Health Records to Improve Mortality Prediction in Hospitalized Patients With COVID-19: Machine Learning Approach.

BACKGROUND: Machine learning models require large datasets that may be siloed across different health care institutions. Machine learning studies that focus on COVID-19 have been limited to single-hospital data, which limits model generalizability. OBJECTIVE: We aimed to use federated learning, a machine learning technique that avoids locally aggregating raw clinical data across multiple institutions, to predict mortality in hospitalized patients with COVID-19 within 7 days. METHODS: Patient data were collected from the electronic health records of 5 hospitals within the Mount Sinai Health System. Logistic regression with L1 regularization/least absolute shrinkage and selection operator (LASSO) and multilayer perceptron (MLP) models were trained by using local data at each site. We developed a pooled model with combined data from all 5 sites, and a federated model that only shared parameters with a central aggregator. RESULTS: The LASSOfederated model outperformed the LASSOlocal model at 3 hospitals, and the MLPfederated model performed better than the MLPlocal model at all 5 hospitals, as determined by the area under the receiver operating characteristic curve. The LASSOpooled model outperformed the LASSOfederated model at all hospitals, and the MLPfederated model outperformed the MLPpooled model at 2 hospitals. CONCLUSIONS: The federated learning of COVID-19 electronic health record data shows promise in developing robust predictive models without compromising patient privacy.

Federated Learning used for predicting outcomes in SARS-COV-2 patients.

'Federated Learning' (FL) is a method to train Artificial Intelligence (AI) models with data from multiple sources while maintaining anonymity of the data thus removing many barriers to data sharing. During the SARS-COV-2 pandemic, 20 institutes collaborated on a healthcare FL study to predict future oxygen requirements of infected patients using inputs of vital signs, laboratory data, and chest x-rays, constituting the "EXAM" (EMR CXR AI Model) model. EXAM achieved an average Area Under the Curve (AUC) of over 0.92, an average improvement of 16%, and a 38% increase in generalisability over local models. The FL paradigm was successfully applied to facilitate a rapid data science collaboration without data exchange, resulting in a model that generalised across heterogeneous, unharmonized datasets. This provided the broader healthcare community with a validated model to respond to COVID-19 challenges, as well as set the stage for broader use of FL in healthcare.

Scalable, effective, and rapid decontamination of SARS-CoV-2 contaminated N95 respirators using germicidal ultra-violet C (UVC) irradiation device.

IMPORTANCE: Particulate respirators such as N95 masks are an essential component of personal protective equipment (PPE) for front-line workers. This study describes a rapid and effective UVC irradiation system that would facilitate the safe re-use of N95 respirators and provides supporting information for deploying UVC for decontamination of SARS-CoV-2 during the COVID19 pandemic. OBJECTIVE: To assess the inactivation potential of the proposed UVC germicidal device as a function of time by using 3M 8211 - N95 particulate respirators inoculated with SARS-CoV-2. DESIGN: A germicidal UVC device to deliver tailored UVC dose was developed and snippets (2.5cm2) of the 3M-N95 respirator were inoculated with 106 plaque-forming units (PFU) of SARS-CoV-2 and were UV irradiated. Different exposure times were tested (0-164 seconds) by fixing the distance between the lamp (10 cm) and the mask while providing an exposure of at least 5.43 mWcm-2. SETTING: The current work is broadly applicable for healthcare-settings, particularly during a pandemic such as COVID-19. PARTICIPANTS: Not applicable. Main Outcome(s) and Measure(s): Primary measure of outcome was titration of infectious virus recovered from virus-inoculated respirator pieces after UVC exposure. Other measures included the method validation of the irradiation protocol, using lentiviruses (biosafety level-2 agent) and establishment of the germicidal UVC exposure protocol. RESULTS: An average of 4.38x103 PFUml-1(SD 772.68) was recovered from untreated masks while 4.44x102 PFUml-1(SD 203.67), 4.00x102 PFUml-1(SD 115.47), 1.56x102 PFUml-1(SD 76.98) and 4.44x101 PFUml-1(SD 76.98) was recovered in exposures 2s,6s,18s and 54 seconds per side respectively. The germicidal device output and positioning was monitored and a minimum output of 5.43 mWcm-2 was maintained. Infectious SARS-CoV-2 was not detected by plaque assays (minimal level of detection is 67 PFUml-1) on N95 respirator snippets when irradiated for 120s per side or longer suggesting 3.5 log reduction in 240 seconds of irradiation. CONCLUSIONS AND RELEVANCE: A scalable germicidal UVC device to deliver tailored UVC dose for rapid decontamination of SARS-CoV-2 was developed. UVC germicidal irradiation of N95 snippets inoculated with SARS-CoV-2 for 120s per side resulted in 100% (3.5 log in total) reduction of virus. These data support the reuse of N95 particle-filtrate apparatus upon irradiation with UVC and supports use of UVC-based decontamination of SARS-CoV-2 virus during the COVID19 pandemic.