On June 9, 2020, the number of cases with coronavirus disease in France hospitalized in intensive care dropped under the 1,000 threshold for the first time since the start of the outbreak. While on July 10, 2020, this same metric entered the red zone in Florida. As many as 48 hospitals in Florida have reached their full capability in their intensive care units and present zero ICU beds, reported CNN. Another 52 Florida hospitals have less than 10 percent ICU beds available, according to data released by the Agency for Health Care Administration (AHCA). During this period, the sunshine state announced almost 9,000 new coronavirus disease cases. And at least 120 coronavirus disease-related deaths, making the cumulative cases more than 232,000 with over 4,000 coronavirus-related deaths in the state, said the report. Why do situations in France and Florida seem almost opposite, respectively, a slow recovery versus a surge in cases?
In Italy, the nation outside China with the most significant number of patients with COVID-19 until March 29, 2020, up to 12% of all positive cases demanded ICU admission. Some patients are brought to intensive care units (ICU) to be supplied with extra help to fight the coronavirus disease. Boris Johnson happened to be the most high-profile personality to do that. He was initially admitted to the hospital and said that he was “in good spirits.” By the evening, his condition had worsened, and he was transferred to an ICU. In a nutshell, anybody with coronavirus disease can switch from a ‘stable’ state to the need to get into ICU very rapidly. And while the coronavirus disease travels the world, it becomes essential to find new means to anticipate which patients are likely to enter ICU.
Coronavirus disease: around the world in eighty days?
The World Health Organization (WHO) issued its first alert on the novel coronavirus disease SARS-CoV-2 in early January. It announced that the COVID-19 situation was a pandemic on March 11, 2020. In the meantime, travel limitations and containment policies had failed to stop coronavirus disease from becoming a global peril. Most countries reported COVID-19 cases and deaths. To resist the spread of the coronavirus disease and the burden it imposed on healthcare systems, countries throughout the world restricted citizens’ movement in a matter of days. Over a billion souls were trapped at home, wondering when they would resume a normal life. On March 16, 2020, French officials decreed a large scale lockdown to counter the COVID-19 epidemic tide rising in the country, suspending non-essential business, educational, and recreation ventures, maintaining food retailers and healthcare institutions mostly. One month later, the number of new hospitalizations and ICU admissions due to coronavirus disease reached a plateau and began a slow descent in the country. If no control measures had been set up, between March 19 and April 19, 2020, research shows that almost 23% of the French population would have been affected by COVID-19 (14.8 million individuals). Hence, the French lockdown prevented 587,730 hospitalizations and 140,320 ICU admissions at the state level. The total estimate of ICU beds needed to treat patients in critical conditions would have been 104,550, far more than the maximum French ICU capacity.
It seems that the coronavirus disease travels around the world. As a modern time Phileas Fogg. It acts as a wave, creating hotspots from Asia to Europe, Europe to the East coast of the U.S., and then to the West Coast, and Australia. In these hotspots, such as today’s Florida, ICU beds keep filling up.
From coronavirus disease to ICU
Not all critical cases of coronavirus disease are admitted to the ICU. Patients admitted to the hospital start with more simple treatments, such as being given oxygen. ICU admissions depend on the severity of the coronavirus disease and the healthcare system’s ICU capacity. In Italy, the nation outside China with the most significant number of patients with COVID-19 until March 29, 2020, up to 12% of all positive cases demanded ICU admission. Some patients are brought to intensive care units (ICU) to be supplied with extra help to fight the coronavirus disease. Boris Johnson happened to be the most high-profile personality to do that. He was initially admitted to the hospital and said that he was “in good spirits.” By the evening, his condition had worsened, and he was transferred to an ICU.
In a nutshell, anybody with coronavirus disease can switch from a ‘stable’ state to the need to get into ICU very rapidly. And while the coronavirus disease travels the world, it becomes essential to find new means to anticipate which patients are likely to enter ICU.
Can Machine Learning be used to predict ICU entries?
In our first article about coronavirus predictions and Covid-19 infections, we used both statistics and Machine Learning to identify which criteria make people more likely to get infected with the coronavirus. We exploited a public database representing all patients checking into hospitals with or without coronavirus infection. A team of researchers based in Mexico shared this database, among which the medical researcher Omar Yaxmehen Bello-Chavolla is our point of contact. We expect this database to suffer from a sampling bias because it contains only people admitted into hospitals, therefore people showing strong symptoms (as Mexican authorities only test for COVID-19 for highly symptomatic people). Mexico is hit very hard by the coronavirus disease even though it peaked there a while after Europe. Several factors place the population of this country at higher risks, among which a significant prevalence of type 2 diabetes among younger people (under 45) which is unique in the world. Therefore, it becomes all the more important to find means to understand the critical factors causing the infection or the entry into ICU.
We offer to examine the factors affecting ICU admissions in this Mexican public database both statistically and with our Machine Learning tool. Looking up the statistics about overall ICU admissions, we observed that, on average, 2.23% of hospitalized people entered ICU, whether infected with the coronavirus disease or not. Most criteria and co-morbidities logged into the database increased the chances of ICU admission: age, gender, being contaminated with the COVID-19, immunosuppression, pneumonia, diabetes, hypertension, cardiovascular disease, obesity, chronic kidney disease. The astonishments are that asthma has no statistical impact and that pregnant women and people who smoke suffered fewer ICU admissions than the average.
What has Machine Learning to say about this? Our tool generates each model by selecting seven criteria to make a prediction, expressed as a simple equation. For each question asked to the model, i.e., who is likely to enter ICU?, we produce one hundred models. Then we count the usage of each criterion. In a first set of simulations where the patients’ COVID status was excluded, we have measured a model accuracy of 83%. It means that 83 times out of 100, the model predicted correctly which patients were going into ICU and which patients were not. It is an excellent performance.
Moreover, we have observed pneumonia is the number one criteria used in 100% of models, it is followed by:
- age in 22% of models,
- gender in 20%,
- smoking in 19% of models
- hypertension in 17% of models,
- chronic kidney disease in 17% of models,
- cardiovascular disease in 16% of models,
- obesity in 15% of models,
- asthma in 14%,
- being immunosuppressed in 13%,
- diabetes in 13%.
The two surprises in these results come from smokers and, asthma which seem to have much influence statistically and yet show up with a high ranking in the models. But since pneumonia is used in 100% of models, we estimated that it might overshadow other criteria and sway the results. Therefore, we generated another series of one hundred models, asking once more: “Who is likely to enter ICU?”. But this time we included the covid-status information and we excluded the pneumonia information. The models’ performance declined slightly to 69% accuracy, which is still a fascinating figure for the medical world. With this approach, a medical doctor can anticipate in over 2 cases out of 3 correctly among the patients checking into the hospital, which ones will move to ICU. It is interesting because it gives a means to focus on these patients and to anticipate the ICU bed requirements a few days ahead. The criteria the models used were:
- age in 51% of models,
- COVID status in 37% of models,
- gender in 26%,
- diabetes in 20%,
- obesity in 20%,
- chronic kidney disease in 19%,
- smoking in 18%,
- hypertension in 15%,
- cardiovascular disease in 14%.
Again smoking shows up as a discriminating factor while it did not appear in the statistics. Smoking and coronavirus disease stays a controversial topic. Although heavily discussed, the current trend is to consider that smokers are less likely to be infected but suffer harder complications when they do.
We analyzed both statistically and using our Machine Learning tool ICU entrances for people without the coronavirus disease. Through statistics, we observe that 1.48% of these patients enter ICU. Age, being immunosuppressed, having pneumonia, diabetes, hypertension, cardiovascular diseases, obesity, and chronic kidney disease, increase the chances of entering ICU. For these non-COVID patients investigated statistically, gender, pregnancy, asthma, and smoking seem to have no impact on ICU entrance.
Machine Learning in Action
Using Machine Learning, and including pneumonia, we obtained excellent accuracy results, i.e., 86%. And again, because pneumonia was used in 100% of models, we feared it overshadowed the other criteria. So we generated another batch of Machine Learning models, excluding pneumonia in the requirements. The accuracy decreased to 75%, which is still a significant figure. The criteria used were: gender in 41% of models, age in 40% of models, diabetes in 31% of models, being immunosuppressed in 22% of models, hypertension in 19% of models, obesity in 19% of models, pregnancy in 16% of models and smoking in 14% of models. What pops up here is the importance of diabetes.
And last, we examined the statistics and the Machine Learning models for patients with the coronavirus disease. 3.67% of those patients entered ICU (twice as much as the non-COVID patients). Among the people contaminated with the coronavirus disease, the statistics tell us that gender, age, pneumonia, diabetes, hypertension, cardiovascular disease, obesity, and chronic kidney disease worsen things. While pregnancy, asthma, and smoking appear to be neutral.
When we look at the same data set with the perspective of Machine Learning, we obtain models with a 75% accuracy. Here again, pneumonia is used in 100% of the models, and we think pneumonia overshadows other criteria. So we generated another set of models, excluding pneumonia. The accuracy dropped to 65%, which is still very correct in this context. The criteria used were: age in 74% of models, gender in 33% of models, diabetes in 33% of models, hypertension in 22% of models, obesity in 21% of models, pregnancy in 18% of models, being immunosuppressed in 16% of models. Once again, diabetes stands out as a significant risk factor behind the well-known gender and age for COVID patients.
Take away and next steps
While the coronavirus disease becomes a part of our everyday life, just like influenza, we need to find means to handle it to minimize casualties. In a previous article, we have shown how Machine Learning can help forecast the patients at risk of infection. In this article, we see that Machine Learning delivers insights that are consistent with the general statistical approach but adds the twist of identifying the significant criteria among a varied group. Since the model comes out as an equation, the most impacting criteria show up in total transparency. It makes the predictions both interpretable, i.e., diabetes has a significant impact on entering ICU both for COVID and non-COVID patients, for instance. It also makes the situation explainable. This transparency is critical in supporting the medical research efforts.
In the next article, we will analyze the criteria impacting the likelihood of intubation. Stay tuned!