woman wearing a surgical mask swabbing the inside of an elderly man's nose
COVID-19 Research

Treatment & diagnostics research

The T-Detect COVID Test is a novel technology that assesses the T cell immune response to COVID-19. Information and scientific data that deepen our understanding of SARS-CoV-2 remain important keys to get ahead of this global pandemic.”

–Jeff Shuren, M.D., J.D., director of FDA’s Center for Devices and Radiological Health (opens in new tab)

In March 2020, Microsoft and Adaptive Biotechnologies expanded their existing collaboration to map and measure the immune response to multiple diseases and started applying their combined capabilities to COVID-19. In May, Adaptive started enrollment for a virtual clinical study, ImmuneRACE (opens in new tab), to measure the presence of T cells that identify the disease early on and proliferate to combat the infection. The results of this study (opens in new tab) were published in July 2020, accompanied by a large release of population-level data analyzed to reveal T cell signatures of disease for COVID-19, known as ImmuneCODE (opens in new tab). This was quickly followed by Adaptive’s submission to the FDA in September. The FDA granted emergency use authorization for T-Detect™ COVID in March 2021 as the first T cell-based diagnostic.

That is one example of the development of diagnostic tools and treatments, along with assessments of how these can be applied in practice, that was a focus for much research around the world. This included the application of machine learning and other novel approaches for analyzing large volumes of data to help with triaging COVID-19 patients, predicting outcomes for COVID-19 cases, and estimating treatment effectiveness. The research highlighted here applies such tools to “long Covid”, predictions of death and renal failure, modeling probability for positive SARS-CoV-2 tests, classifying chest X-rays, and predicting the effectiveness of different medical treatments on patients with varying risk factors.

Flowchart of SARS-CoV-2 oubreak in Vo'

Longitudinal analysis of T cell receptor (TCR) repertoires reveals shared patterns of antigen-specific response to SARS-CoV-2 infection

T cells play a prominent role in the immune response to viral diseases, but their role in subsequent immunity to SARS-CoV-2 infection remains poorly understood. This research studied the assortments (repertoires) of TCRs in an Italian population following a SARS-CoV-2 outbreak. An individual’s TCR repertoire changes as immunity is developed following exposure. Samples taken at 2, 9 and 15 months found elevated levels of TCRs associated with SARS-CoV-2, demonstrating the T cells’ central role in mounting a persistent immune defense against SARS-CoV-2.

Venn diagram of T cell test, nAb, EUROIMMUN, and Abbott ARCHITECT

T-cell receptor (TCR) sequencing identifies prior SARS-CoV-2 infection and correlates with neutralizing antibody titers and disease severity

The blood levels of neutralizing antibody titers (nAb) closely correlate with the protection provided by an effective vaccination. But nAb assays are challenging to perform at a large scale. This research instead applied a TCR sequencing assay on a standard blood sample to assess T-cell response to SARS-CoV-2 infection. It found that the magnitude of the SARS-CoV-2-specific T-cell response strongly correlates with nAb titer, as well as with clinical indicators of disease severity including hospitalization, fever, or difficulty breathing, thus demonstrating the utility of a TCR-based assay.

Chest radiography (CXR)

Supporting clinicians to assess COVID-19 severity using AI and Chest X-rays

COVID-19 X-rays have been a recommended procedure for patient triaging and resource management in intensive care units (ICUs) throughout the COVID-19 pandemic. Microsoft Research team worked closely with our clinicians at University Hospitals Birmingham NHS Foundation Trust to model radiological features with a human-interpretable class hierarchy that aligns with the radiological decision process. The model outperformed the clinicians across all hierarchical and multi-class tasks. To better understand the model’s failure patterns, the team employed an error analysis tool in Azure Machine Learning that is not often found in healthcare-related ML studies and is crucial for providing transparency and actionable insights about a model’s behavior.  The analysis may also be useful after deployment if presented as reliability information alongside the model’s predictions.

four scatter plots of various colors

Removing biases from deep learning-based models of COVID-19 chest X-rays

Recent research has proposed creating deep learning-based models that use chest radiographs (CXRs) in a variety of clinical tasks to help manage the COVID-19 crisis. However, due to the small size of existing COVID-19 CXR datasets, data is often pooled from multiple sources under different scenarios. Models trained on such datasets can “overfit” to erroneous features instead of learning pulmonary characteristics. This research adds feature disentanglement to the training process, resulting in better generalization performance on unseen data and outperforming other proposed methods.

Covid research - treatment - sociodemographic research

Sociodemographic and clinical features predictive of SARS-CoV-2 test positivity across healthcare visit-types 

Despite increased testing efforts and the deployment of vaccines, COVID-19 cases and death toll continued to rise at record rates. Health systems routinely collect clinical and non-clinical information in electronic health records (EHR), yet little is known about how the minimal or intermediate spectra of EHR data can be leveraged to characterize patient SARS-CoV-2 pretest probability in support of interventional strategies. This research used machine learning to model patient pretest probability for SARS-CoV-2 test positivity and determined which features most contributed to predicted pretest probability for patients triaged in inpatient, outpatient, and telehealth/drive-up settings. The research found that geographic and sociodemographic factors were more important predictors of SARS-CoV-2 positivity than individual clinical characteristics.

four abstract charts with a blue line

Automatically predicting the effectiveness of heterogeneous treatments for COVID-19 patients

The effects of medical treatments can vary for different patients based on many different underlying risk factors, making independent testing of the effects of multiple treatments unwieldy. This research uses multitask machine learning to automatically predict the heterogeneous (varying) effectiveness of different medical treatments, and trains additive models to estimate personalized treatment benefits.  When applied to mortality risk models of COVID-19 patients, this method uncovered evidence supporting two pathways of mortality: inflammation and thrombosis, and achieved state-of-the-art predictive power and interpretable identification of heterogeneous treatment benefits.

a grey background with pixels of yellow and dark purple

Long-term effects of SARS-CoV-2 and their associations with social determinants of health

Much remains unknown about the complications that can follow SARS-CoV-2 infection (“long Covid”). This research analyzed a medical claims database of over one million COVID-19 survivors to study long-term symptoms and their associations with various social and medical risk factors. It identified the ICD-10 codes whose proportions were significantly increased and included a logistic regression-based association analysis. It found associations of long-term effects with age and gender, but not with race, income and education levels.

a chart with a blue line on a bumpy path up and to the right

Using machine learning to predict death and organ failure in hospitalized patients with COVID-19

To date, COVID-19 prediction models have largely focused on mortality rather than risks for other outcomes such as shock, renal failure or respiratory failure requiring mechanical ventilation. To address these concerns, this research created separate models, based on demographic and clinical information collected upon hospital admission, to predict risks for in-hospital mortality, ICU transfer, shock, and renal replacement therapy, with high accuracy. These models could help improve triage decisions and resource allocation and support clinical trial enrichment.