As this article is being written, the current death toll from COVID-19 stands at 1,77,000 worldwide, and going by the trend, a few more would have taken place by the time this article is published. Weeks after it was declared a pandemic, all efforts to find a cure or stop its spread by health and government authorities has gone in vain. Indeed, lack of protective gears, supportive medical machines such as ventilators, and lack of respect for social distancing guidelines have made the operation tough for doctors.
What is worse, some experts claim that many positive COVID-19 cases may be asymptomatic in nature. If some people do not exhibit symptoms, how can one know if they are infected? In a recent paper published by Mayo Clinic and a startup called Nference, researchers claim that AI can zero in on symptoms that can indicate if a person is actually suffering from the disease.
Mayo Clinic & Nference’s AI claim
Mayo Clinic, along with Nference, analysed tests from biomedical reports, and claimed that they have used AI to segregate the phenotypes characteristic of the novel virus. As per the team of researchers, the first few symptoms that arose between 4 to 7 days are a combination of cough, diarrhoea, excessive sweating, and anosmia.
To begin their analysis, the team used a natural language processing system that has been designed to automate the identification of diseases, drugs, phenotypes and other entities. This initial step was taken to measure the robust connection between those entities and tag that connection under categories such as positive, negative and other. It also uses Google’s Transformer architecture containing neurons that are arranged in layers and transmit from signals data to adjust the robustness of each connection. Although all AI models learn to predict this way, the Transformer carries a different approach by connecting every output element to the input element.
The system was ingested with 82,29,092 clinical notes of electronic medical records derived from the Mayo Clinic for PCR tested patients amounting to a total number of 14,967. Symptoms – along with presumed symptoms – were noted down before the PCR test as well as after a few weeks from the test date.
The information extracted by AI revealed that 43 COVID-19 positive patients had diarrhoea in the week before the test, whereas only 822 COVID-19 negative patients had diarrhoea. Other symptoms such as excessive sweating, fatigue, and headache amount to 31, 37 and 35 patients respectively. Moving to fever, the system showed only 24.6% of positive patients showed the symptom before the test, whereas 18.6% were tested negative.
Further analysis of data revealed that 251 conjunctions out of 27 phenotypes for a positive and negative patient, two phenotypes-mainly coughs and diarrhoea, and sweating and diarrhoea were found to be significant. Cough and diarrhoea simultaneously existed in 13.2% of patients, whereas 486 patients did not have COVID-19, which indicated a four-fold amplification. On the other hand, 21 COVID-19 patients were found to be suffering from diaphoresis and diarrhoea, whereas 204 patients facing the same symptoms did not test positive.
As per the team, the latest developments that were found from the EHR analysis of coronavirus progression can assist human pathophysiology sanctioned synopsis of the exploratory therapies that are currently being investigated for COVID-19.
“A caveat of relying solely on electronic medical record inference is that mild phenotypes that may not lead to a presentation for clinical care, such as anosmia, may go unreported in otherwise asymptomatic patients. As at-home serology-based tests for COVID-19 with high sensitivity and specificity are approved, capturing these symptoms will become increasingly important to facilitate the continued development and refinement of disease models. EHR-integrated digital health tools may help address this need,” concluded the co-authors.