TTHealthWatch is a weekly podcast from Texas Tech. In it, Elizabeth Tracey, director of electronic media for Johns Hopkins Medicine, and Rick Lange, MD, president of the Texas Tech University Health Sciences Center in El Paso, look at the top medical stories of the week.
This week’s topics include artificial intelligence (AI) and bias; dietary calcium, protein, and falls in the elderly; prescription drugs for mental health during the pandemic; and a selective serotonin reuptake inhibitor (SSRI) for COVID-19.
0:42 Prescription drugs for mental health during the pandemic
1:44 Women had a higher rate of prescriptions
2:44 Increased risk of ED visits
3:44 Habit after pandemic?
4:01 Bias in AI
5:03 Different diagnostic and prognostic models
6:03 Used to create definition
8:10 741 patients treated and about the same control
9:10 Only second medication for those considered at high risk
10:08 Retrospective analysis of those already on SSRIs
10:25 Dietary calcium and protein in older adults
11:25 Additional milk, yogurt, and cheese
12:25 Obligatory calcium and protein requirements
Elizabeth Tracey: What’s happened to the prescription of medicines for mental health disorders during the COVID-19 pandemic?
Rick Lange: Can you change the diet in older adults in residential care to decrease the risk of hip fractures?
Elizabeth: Can a drug called fluvoxamine help to ameliorate the symptoms of COVID-19?
Rick: And learning about machine learning.
Elizabeth: That’s what we’re talking about this week on TTHealthWatch, your weekly look at the medical headlines from Texas Tech University Health Sciences Center in El Paso. I’m Elizabeth Tracey, a Baltimore-based medical journalist.
Rick: I’m Rick Lange, president of Texas Tech University Health Sciences Center in El Paso where I’m also dean of the Paul L. Foster School of Medicine.
Elizabeth: Rick, we know already and we’ve talked a number of times about the rate of mental health issues during the pandemic. Unsurprisingly, of course, they have increased tremendously.
This week we’re talking about a study that’s in JAMA Network Open that’s looking at three different classes of medicines, and what’s happened in the prescription of those during the pandemic in both U.S. men and women.
This study uses data from the Clinformatics Data Mart, which I was unfamiliar with previous to reading the study, which is one of the largest commercial health insurance databases in the U.S. They literally looked at millions of records that were deidentified to assess this issue of how many and who was prescribed benzodiazepines, Z-hypnotics, and serotonergic drugs during this pandemic.
They started with data from January 1, 2018 through March 31, 2021. The upshot of the whole thing is that women had a higher rate of prescriptions for all three drug classes and had larger changes in these prescription rates over time.
If we break them out, we can see that there was an increase in the Z-hypnotics, as well as the SSRIs and the SNRIs [serotonin and norepinephrine reuptake inhibitors] in both men and women, and an increase in benzodiazepine prescriptions in women. The authors interpret this as evidence that there is a substantial mental health impact of COVID-19, particularly for women. They cite other corroboration to that, that women were out of the workplace more often, taking care of kids a lot more often, and had a lot more stress. Of course, we know that women had that before the pandemic got here and also had a higher incidence of mental health disorders.
Rick: Elizabeth, the medications you mentioned are those that are commonly prescribed for people that have insomnia or anxiety or depressive disorders. Previously on our podcast, we have informed our listeners that there has been about a 33% increase in the prevalence of anxiety or depressive symptoms during COVID. Then, obviously, insomnia as a result of not only the mitigation factors, but insomnia is one of the side effects of COVID infection as well.
This also goes along with the fact that we have seen an increased risk of ED visits for these psychiatric symptoms. I’m not terribly surprised at these numbers. One of the great virtues of this study is they studied between 15 and 17 million people between 2018 and 2020.
Elizabeth: I guess one of the things that I’m interested in is what’s going to happen to this prescription drug use as the pandemic hopefully continues its decline. Will people give these things up? Because the other thing that’s concerning, at least to me, is the fact that once people get started on these things it can be really challenging to stop.
Rick: Essentially, because if you look at the waves, there was an increase in benzodiazepine use for women in the first wave and then an increase in the other medicines in both men and women in the second wave. We haven’t seen that in the third wave.
Elizabeth: You think this is a straw man: Don’t worry about it, people are not going to become habituated to using these medicines after the pandemic.
Rick: I think that will be the case. I mean, time will tell.
Elizabeth: Since we are talking about really ginormous datasets, why don’t we turn to The BMJ? This is a study that’s taking a look at artificial intelligence and biases that are inherent when these strategies are used to take a look at really big data sets and outcomes, and then use them as predictive models.
Rick: Elizabeth, our listeners should know that at the beginning of the week, you and I kind of throw around different studies about what we’re going to talk about. By the way, when we talk about them, it’s the first time usually we’ve discussed it. But when you picked this study, I kind of ho-hummed, like why would our listeners want to hear about machine-learning techniques? But then as I read the study, I became more convinced that it was in fact important information.
What happens is that we use prediction models in medicine. We use them to either talk about the probability someone will have a diagnosis or the risk that they will actually develop a particular outcome.
Oftentimes, it’s not just a single variable that predicts that. Well, now we can throw these into computers and use machine learning to do supervised predictions. I think we’re lulled into the fact and concept that, gosh, if we think we throw this information into a computer, we come out with an answer that’s much better than just usual statistics. We tend to believe it.
What these authors did was said, “OK, we have these different diagnostic and prognostic prediction models developed via machine learning. How reliable are they? Or how biased could they be?”
They looked at over 152 different studies. About 40% included a diagnostic prediction model and about 60% of them dealt with a prognostic prediction model.
What they found out was that almost 90% of them had a huge amount of bias. Most of it was based upon the fact that there was just a poor methodological quality that contributed to this high risk of bias. That could be due to the fact there was a small study size, poor handling of data, and a failure to deal with overfitting. That’s taking the analysis a little bit further than it needs to be.
But I was surprised at the fact that nearly 90% of these machine-learning techniques had a high risk of bias. What about you?
Elizabeth: I am so concerned about this, Rick, because everyone is using these things. Everyone is attempting to take these, as I said, ginormous data sets, analyze them and come up with something that’s going to be predictive or definitive. The fact that there is just so much bias that’s inherent in that I find it really concerning.
Rick: Well, that’s why I’m glad you brought this up. Because I think we think again, if it’s machine learning or an artificial intelligence computer, that it must be right. In fact,…