COVID ARDS Diagnostic Tool - Age is not best indicator. N=53

[.Scott]

TL;DR

"Computers, Materials & Continua" article: Which COVID patients will suffer deadly ARDS

An article published Monday in "Computers, Materials, & Continua" identifies leading indicators of which patients will go on to develop ARDS (acute respiratory distress syndrome).

From the abstract:

there is an urgent need to augment clinical skills in order to identify from among the many mild cases the few that will progress to critical illness. We present a first step towards building an artificial intelligence (AI) framework, with predictive analytics (PA) capabilities applied to real patient data, to provide rapid clinical decision-making support.

Our results, based on data from two hospitals in Wenzhou, Zhejiang, China, identified features on initial presentation with COVID-19 that were most predictive of later development of ARDS. A mildly elevated alanine aminotransferase (ALT) (a liver enzyme), the presence of myalgias (body aches), and an elevated hemoglobin (red blood cells), in this order, are the clinical features, on presentation, that are the most predictive. The predictive models that learned from historical data of patients from these two hospitals achieved 70% to 80% accuracy in predicting severe cases.

Among the features identified as possible indicators:

Table 4: Feature ranking Predictive features of ARDS in this order using feature selection algorithms described in Section 4.
1. ALT
2. Myalgias
3. Hemoglobin
4. Gender
5. Temperature
6. Na+
7. K+
8. Lymphocyte Count
9. Creatinine
10. Age
11. White Blood Count

A pdf of the full article can be freely downloaded (no paywall).

Additional background information is available in this Yahoo article:
AI tool predicts ...

 

[jedishrfu]

Great find! I know the news likes to focus so much on older people that healthy ones will worry themselves to death.

 

[Vanadium 50]

This paper is almost unimaginably poor. I have seen much better science done at middle school science fairs.

They take 53 patients - 53! - and try and fit the relative strengths of 11 factors from the 5 - five! - who developed ARDS. They try to do this with ill-described machine learning methods, which may be fine, but you are not going to constrain 66 numbers (the relative strengths and the correlation coefficients) with 48 "no" and 5 "yes" responses. They are chasing noise.

Oh, they can't possibly determine how many will develop "deadly ARDS" because all of their patients survived. (To be fair to the authors, they don't claim that. That's on PF)

 

[.Scott]

They take 53 patients - 53! - and try and fit the relative strengths of 11 factors from the 5 - five! - who developed ARDS. They try to do this with ill-described machine learning methods, which may be fine, but you are not going to constrain 66 numbers (the relative strengths and the correlation coefficients) with 48 "no" and 5 "yes" responses. They are chasing noise.

That's why I included "N=53" in the title.
This is exactly the kind and scope of study that you use to justify a larger study - which is what they are using it for.

I would fault them on a couple of points:
1) They are using the term "AI" when, in fact, it is only common statistical analysis.
2) The abstract describes the purpose of their research, but does not make it explicitly clear that this one study does not meet that final objective.

I do not fault them on this point:
1) It is OK that none of the people in the study died, since predicting only whether someone will need a ventilator is of very practical use.

As far as chasing noise is concerned. I agree. In fact, I would make the successful completion of a strong statistics course a prerequisite to any technical course in AI or machine learning. AI can lie. And if you have no background in statistics, you'll be a complete sucker to AI. And this is a very good example of that.

 

[jedishrfu]

The AI part is probably clickbait on their part (for reviewers to approve the grant) because they likely intend to use more sophisticated models once they get larger datasets.

Machine Learning like the @Borg has now assimilated many statistical techniques formerly known as statistical learning into the larger AI / Machine Learning / Deep Learning fields.