Facebook claims that AI can use X-rays to predict COVID-19 results

Researchers at Facebook and New York University (NYU) claim to have developed three machine learning models to help doctors predict how the condition of COVID-19 patients progresses. All open source models require only X-ray sequences and, on the surface, predict patient deterioration up to 4 days in advance and the amount of oxygen supply (if any) that the patient may need.

The new coronavirus pandemic continues to reach amazing new heights in the United States and around the world. For the first time since the beginning of the health crisis last week in the United States, the number of deaths per day exceeded 4,000. With hundreds of thousands of record-breaking infections per day straining the national healthcare system, states like California struggle to maintain excessive intensive care unit space.

Huiying Medical, Alibaba, RadLogics, lunitos, DarwinAI, inference, Qure.ai, Others have developed an AI algorithm for superficially diagnosing COVID-19 from X-rays with high accuracy. However, the difference between the approaches used by Facebook and NYU is that they try to predict the long-term clinical trajectory.Developed by Stanford, Mount Sinai, and electronic health record vendors Epic and Sarner model This shows the risk score of the patient’s chances of dying or needing a ventilator, but few make these predictions from a single scan or electronic medical record.

As part of an ongoing collaboration with NYU Langone Health’s Predictive Analytics Unit and Department of Radiology, Facebook researchers used self-managed learning techniques to AI on two large public chest X-ray datasets, MIMIC-CXR-JPG and CheXpert. Pre-trained the system. It is called Momentum Contrast (MoCo). Self-supervised learning allows MoCo models to be trained from X-ray scans in datasets, even if labels that describe those scans are not available.

The next step was to fine-tune the MoCo model using the extended version of the published NYUCOVID-19 dataset. Researchers created a classifier using 26,838 x-ray images from 4,914 patients to see if the patient’s condition deteriorated within 24, 48, or 72 hours of the problem scan. Annotated to indicate. One classifier predicts patient deterioration based on a single x-ray, and the other classifier uses a series of aggregated x-rays.

Researchers claim that a series of radiographically-dependent classifiers outperform human experts in predicting ICU needs, mortality, and adverse events up to 96 hours in advance. .. The results are not necessarily applicable to other hospitals with unique datasets, but researchers believe that relatively few resources, perhaps a single GPU, can build new classifiers from the MoCo model.

“It is also the first to be able to predict whether patients will need oxygen resources, which may help hospitals decide how to allocate resources in the coming weeks and months. COVID-19 As cases grow again around the world, hospitals need tools to anticipate and prepare for future spikes when planning resource allocations, “the Facebook team wrote in a blog post. I will. “These predictions help doctors avoid getting patients at risk to return home quickly and help hospitals more accurately predict the demand for oxygen supplements and other limited resources.”

Recently the study According to a study by the University of Toronto, Vector Institute, and MIT, chest x-ray datasets (such as MIMIC-CXR and CheXpert) used to train diagnostic models show imbalances and are specific gender, socioeconomic, and racial groups. Is biased against. Female patients suffer from the highest levels of inequality, even though the proportion of women in the dataset is slightly lower than that of men. Caucasian patients (the majority, who make up 67.6% of all x-ray images) are the most preferred subgroup, and Hispanic patients are the least preferred.

Researchers at Facebook and NYU say they addressed this bias by pre-training non-COVID data and carefully selecting each test sample. However, earlier last year, the Centers for Disease Control and Prevention were in the American College of Radiology (ACR), Canada, New Zealand, and Australia. This is because even the best AI systems may not tell the difference between COVID-19 and common lung infections such as bacterial pneumonia and viral pneumonia.

With the release of code, datasets, and technologies ahead, much of the data used today to train AI algorithms for diagnosing illness can perpetuate inequality.A team of British scientists found Almost all eye disease datasets are from patients in North America, Europe, and China, and it is uncertain whether eye disease diagnostic algorithms will work well in racial groups in underrated countries. In another study, researchers at Stanford University claimed that most of the US data on studies, including medical use of AI, came from California, New York, and Massachusetts. A UnitedHealth Group Algorithm Research We have determined that half of the number of black patients in need of more care may be underestimated. And A series of work to grow Since AI models are primarily trained on images of light-skinned patients, it suggests that skin cancer detection algorithms tend to be less accurate when used in black patients.

Determining the reliability of Facebook and NYU algorithms may require thorough testing in multiple diverse healthcare systems around the world — With the consent of the patient. A study published in Nature Machine Intelligence found that the successful deployment of a COVID-19 degradation model in Wuhan, China, resulted in better results than rolling dice when applied to a sample of patients in New York. It was. Careful tweaking may help the Facebook and NYU algorithms avoid the same fate, but it is impossible to predict where the bias will occur. This indicates that you need to audit before deploying at any scale.