Comprehensive Analysis of Machine Learning Models for Five-Class Sleep Stage Classification Using PPG Signals

The Nighttime Detective: Why We Need to Know How We Sleep

Imagine your brain is a busy city that never truly shuts down, even when you close your eyes. Instead of going dark, the city switches between different districts: a bustling downtown where you're wide awake, a quiet library where you're drifting off, a deep underground bunker where you're in a heavy, restorative slumber, and a surreal, dream-filled amusement park. Scientists call these different districts "sleep stages." Knowing exactly which district your brain is in is crucial because if the city gets stuck in the wrong zone—like staying in the busy downtown when it should be in the deep bunker—it can lead to health problems like heart trouble, diabetes, or just feeling groggy all day.

For a long time, the only way to map this nighttime city was to send in a team of experts with a massive, tangled backpack of wires. This "gold standard" method, called Polysomnography (PSG), involves sticking electrodes to your head, eyes, and chin to read your brainwaves and muscle twitches. It's incredibly accurate, but it's also expensive, uncomfortable, and usually requires you to sleep in a lab, which is hardly a natural way to catch some Zs.

Enter the "wearable" revolution. You've probably seen people wearing smartwatches that track their heartbeats. These devices use a sensor called Photoplethysmography (PPG). Instead of reading brainwaves directly, PPG shines a tiny light into your skin to watch your blood pulse with every heartbeat. The idea is simple: your heart beats differently depending on whether you are awake, dreaming, or in deep sleep. So, if we can listen to the rhythm of your heart, can we guess which "district" of the sleep city you are in? This is the big question researchers are trying to answer, hoping to turn a simple wristband into a sleep detective that works right in your own bedroom.

The Paper's Mission: Can a Simple Watch Tell the Whole Story?

In this study, a team of researchers decided to see if they could use these heart-beat sensors to solve the hardest version of the sleep puzzle: distinguishing all five distinct sleep stages (Wake, Light Sleep 1, Light Sleep 2, Deep Sleep, and REM/Dreaming). While many previous studies tried to simplify the problem by grouping stages together (like saying "just tell me if I'm asleep or awake"), the authors knew that merging these stages loses important details about your health. They wanted to see if a standard machine learning model could tell the difference between all five, using only the heart data from a wearable device, without needing the heavy, expensive brain-wave equipment.

To do this, they didn't build a new super-computer; instead, they acted like master chefs using a standard set of ingredients. They took data from 48 people who had already slept in a lab with full monitoring equipment (so the "correct" answer was already known). They grabbed the heart signals (PPG) from these people and chopped them into 3-minute chunks. Why 3 minutes? The authors reasoned that while sleep is often measured in 30-second slices, some heart-rate clues are like slow-cooking stews; they need a longer time to simmer and reveal their true flavor. They then used these chunks to teach several different types of "traditional" computer models—think of them as different styles of detectives, including Random Forest, k-Nearest Neighbors, and XGBoost—to recognize the patterns of each sleep stage.

The researchers tested two main strategies: one where the computer looked only at the shape of the heart signal, and another where it looked at the shape plus the time gaps between heartbeats (called Inter-Beat Intervals or IBIs). It's like trying to identify a song by listening to the melody alone versus listening to the melody and the rhythm together.

The Findings: The Heartbeat Rhythm Wins

The results were quite promising, but with a few caveats. The study found that combining the shape of the signal with the rhythm gaps (IBIs) was the secret sauce. When the models used both types of clues, they got much better at guessing the sleep stages.

The star of the show was the Random Forest model. Imagine a committee of 100 different experts, each asking a slightly different question about the data, and then taking a vote on the answer. This "committee" approach worked best, achieving an accuracy of 72.2% and a score called Cohen's Kappa of 0.65. This means the model was correct about three-quarters of the time, which is a solid performance for such a difficult task. The k-Nearest Neighbors model (which guesses based on what the "closest" similar sleep chunks were) came in a very close second with 71.6% accuracy.

However, the paper is careful to point out where the detectives struggled. The models were great at spotting when you were awake or in deep sleep, but they had a hard time distinguishing the "Light Sleep" stages (specifically Stage 2). It's like trying to tell the difference between two shades of blue that look almost identical; the heart signals for these stages are so similar that even the best models got confused.

The authors also explicitly ruled out the idea that you need a massive, complex "deep learning" super-computer to get these results. They showed that these simpler, traditional models could perform just as well as the fancy neural networks found in other studies, but with the added benefit of being easier to understand and requiring less computing power. This is a big deal because it suggests that your next smartwatch might be able to do this analysis right on the device, without needing to send your data to a giant server farm.

The Bottom Line

So, what does this all mean? The study suggests that we don't necessarily need to strap ourselves to a hospital bed to get a detailed map of our sleep. By using a simple heart sensor and smart, traditional computer models, we can get a pretty good picture of our sleep stages, including the tricky ones. While the system isn't perfect yet—it still gets confused by the lighter sleep stages—it proves that a wearable device could one day give us a detailed, five-star report on our sleep quality right from our wrists. The authors conclude that this approach is a strong, practical step toward making sleep monitoring affordable and comfortable for everyone, even if it's not quite the "magic bullet" that solves every sleep mystery overnight.