Researchers input MSLT and preceding PSG data, then report on the results.
In recent years, researchers have sought to improve how we understand the data that providers gather from diagnostic tools like the multiple sleep latency test (MSLT). While the MSLT is the gold standard for diagnosing hypersomnias, including narcolepsy type 1, it doesn’t always detect conditions with arbitrary thresholds that well, such as narcolepsy type 2. This can be a problem, particularly since years of untreated narcolepsy can significantly compromise patient quality of life.
We spoke with Logan Schneider, MD, one of the authors of the study, “Improved Primary CNS Hypersomnia Diagnosis With Statistical Machine Learning” on how machine learning can help clinicians and researchers do a better job of differentiating narcolepsy type 1, narcolepsy type 2, and idiopathic hypersomnia. (This Q&A has been lightly edited for clarity.)
Sleep Review (SR): Why did you find it essential to look into whether we could improve the diagnostic accuracy of hypersomnias like narcolepsy type 2 using machine learning?
Schneider: The MSLT is great at detecting type 1 narcolepsy, and it was primarily devised with type 1 narcolepsy patients (with that distinctive phenotype) in mind.
We pulled the data we have of people who test positive in the MSLT, including those people with untreated sleep apnea, shift work disorders, or even sleep deprivation and put that information into a computer algorithm that not only [looked at] MSLT test data, but also the data on what these patients’ sleep looked like the night before. We thought that might give us information about what is defined as abnormal.
What we did here was an exploratory analysis to provide us with information to understand the disorders and differentiate among the primary hypersomnias we examined.
SR: What made you and your colleagues use statistical machine learning as a tool to differentiate these different conditions better?
Schneider: We used multiple machine learning models, using a slightly less biased approach to put in the data we had, to see what new things these models could teach us about these disorders. [Less biased meaning he did not start with a specific hypothesis in mind.] In this study, we used machine learning with a spirit of curiosity. Some of these conditions are hard to diagnose because they’re arbitrarily defined by the criteria we’ve imposed.
From this exercise, we learned that the model isn’t entirely accurate. The sleep that patients get the night before impacts our interpretation of MSLT results in the next day. So we learned that the thresholds for the MSLT are potentially a little bit more variable. For instance, if the patient had a good night’s sleep the night before and you still look like a narcoleptic, then you might be a type 2 narcoleptic.
SR: How do you see statistical machine learning becoming a useful tool to bridge the gaps we see in diagnostic tools today?
Schneider: Most people will be uncomfortable with the black box that a lot of AI [artificial intelligence] and machine learning lends today. In other words, although machine learning can do something well, we may not understand what exactly it is that they’re doing to some degree, which is an issue.
In this study, we took a semi-supervised approach, but in 20 to 30 years down the road after repeated validations, we hope that we can trust our algorithms enough to automate the data extraction process. But most importantly, what we tell people as researchers and providers needs to be interpretable. In practice, we hope that physicians can use the algorithms that we’ve identified to diagnose patients better and be able to tell the story of what the algorithm conveys to patients.
Ultimately, yes, I think machine learning will change the way we practice medicine and we will be able to use more advanced statistics in an informed way so that we can understand it, interpret it, and tell our patients or our colleagues we did what we did.
SR: Do you see statistical machine learning becoming more integrated into the sleep medicine provider’s toolbox?
Schneider: I think this is something we will see more in the future, and all physicians work on algorithms all the time already. There are lots of tools used to make treatment decisions, particularly in emergency medicine environments. The goal in sleep medicine is to develop a tool that can easily fit into a clinician’s practice to help them better make treatment decisions.
SR: Why do the findings of this article matter to narcolepsy patients: What is the biggest takeaway for them?
Schneider: By using machine learning models that do a great job of the categorization of disorders using more clinical information at different thresholds than the current standard, our hope is that we will be able to shorten the diagnostic process by avoiding opportunities for data misinterpretation that result from false positives. Plus, we want to get better diagnostic certainty at different thresholds, and maybe that involves getting more information to diagnose people better. In the case of our study, we found that our model needs a better variable, more information to assess conditions like idiopathic hypersomnias.
Yoona Ha is a freelance writer and healthcare public relations professional.