Chapter 7 — Read Critically: DeepSleepNet and One More Architecture
Think of it like this: two chefs given the exact same ingredients (raw EEG) and the exact same dish to make (a sleep-stage label every 30 seconds) can still end up with completely different recipes. Neither chef is wrong. They just made different trade-offs. Reading architecture papers critically means asking why a chef chose their specific steps — not just memorizing the recipe.
DeepSleepNet: Two Lenses, One Photo
DeepSleepNet (Supratak, Dong, Wu & Guo, 2017, IEEE TNSRE) is TinySleepNet's bigger, more elaborate relative. Instead of Workshop 1's single stamp, it uses two parallel CNN branches with different filter sizes at once — like photographing the same scene with two different lenses: one lens zoomed in tight, built to catch sharp, brief details (a spindle's crisp edge); one lens zoomed out wide, built to catch slow, rolling shapes (a whole delta wave). The two photos get combined into one feature per epoch, richer than either lens alone.
Those combined features then feed a 2-layer bidirectional LSTM with a residual connection. "Bidirectional" is the key new idea here: instead of Workshop 2's journal reading only forward (past → present), this one reads the whole night twice — once forward, once backward — and combines both readings. Knowing what happens after an epoch genuinely helps classify it (an epoch followed by unmistakable N3 was probably already heading toward N3). The "residual connection" is a shortcut hallway: a way for information to skip around parts of a deep network instead of being forced through every single room, which helps very deep networks train without their earlier insights getting diluted on the way through.
Training happens in two separate steps: first the CNN half alone, then the whole network fine-tuned together with real sequences. More machinery, more accuracy headroom, more to get right.
U-Sleep: No Journal At All
U-Sleep (Perslev, Darkner, Kempfner, Nikolic, Jennum & Igel, 2021, npj Digital Medicine) takes a genuinely different path: it drops the LSTM journal entirely. It's fully convolutional — an encoder that progressively zooms out, capturing bigger and bigger patterns, then a decoder that zooms back in, step by step, assigning a label to every moment along the way. No memory mechanism, no reading in order — the whole recording gets processed more like one giant image than a sequence of 30-second chapters.
What that buys you: it was trained and validated on 15,660 people across 16 different clinical studies — an enormous, genuinely diverse dataset — and it generalizes remarkably well to hospitals, devices, and populations it never saw during training, matching specialist systems built specifically for that data. It can also label sleep stages at a much finer resolution than the usual 30-second window, something a epoch-by-epoch journal design can't naturally do.
Try It: Real-Time Ready?
Both of these are real, published, genuinely strong designs — and neither is naturally built for the thing Dreamento actually needs: an answer while someone is still asleep, not afterward. This game drills exactly that distinction before you write anything.
Write It Down: A Half-Page Comparison
Don't stare at a blank page. Answer these four questions, in your own words, a few sentences each — that's your half page:
- What's the single biggest architectural difference between DeepSleepNet and U-Sleep?
- Which of the three designs in this course (TinySleepNet, DeepSleepNet, U-Sleep) could genuinely run live, one epoch at a time, as someone sleeps? Why specifically?
- Which one would you expect to score highest on an offline accuracy benchmark, and why might that not be the most important factor for Dreamento specifically?
- If you had to pick one of the three to actually wire into Dreamento's real-time scorer today, which would you pick — and what are you deliberately trading away by not picking the others?
DONE WHEN
You can argue, with real reasons (not just a gut feeling), why one design suits a real-time wearable like ZMax/Dreamento better than another — and you can name specifically what gets sacrificed by that choice.
References
Supratak, A., Dong, H., Wu, C., & Guo, Y. (2017). DeepSleepNet: A model for automatic sleep stage scoring based on raw single-channel EEG. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 25(11), 1998–2008.
Perslev, M., Darkner, S., Kempfner, L., Nikolic, M., Jennum, P.J., & Igel, C. (2021). U-Sleep: resilient high-frequency sleep staging. npj Digital Medicine, 4, 72.