Can real-time spindle detection be fast and accurate enough to support closed-loop sleep stimulation?

Why catching a spindle live is harder than finding one afterward

Sleep spindles are short bursts of brain activity in the 11–16 Hz range that appear during non-REM sleep. They last only briefly and arrive unpredictably, which is exactly what makes them interesting for memory research and awkward for a computer that has to react in the moment [1][2]. Offline, a detector can look at the whole night's recording and decide afterward where the spindles were. In real time, the system must commit to a decision while the burst is still happening, then hand that decision to a stimulator before the window closes [1].

The practical test is therefore not just 'did it find the spindle?' but 'did it find it early enough to act?' A detector that is accurate but slow is useless for closed-loop work, and a fast one that fires on noise wastes stimulation. The real-time spindle detector reported in 2022 was built as a multi-channel signal-processing algorithm specifically to trigger stimulation during spindles in a phase-specific way, meaning it aims at a chosen point in the spindle's waveform rather than just somewhere inside the burst [1].

How accurate are today's real-time detectors?

When the real-time detector was streamed pre-recorded sleep data through a real-time computer system, it reached roughly 83% sensitivity and 78% precision, with an F1 score near 81%, measured against state-of-the-art offline algorithms. Those offline algorithms scored similarly or lower when compared with each other, which puts the real-time version in the same ballpark as the tools researchers already trust [1]. The detected spindles also matched on frequency, duration, amplitude and symmetry, and showed the expected time-frequency profile and centroparietal topography — the pattern over the central and parietal scalp where spindles are typically strongest [1].

A different line of work took a hardware-and-software approach: the Portiloop, a portable, low-cost, open-science closed-loop system built from commercially available parts, with a lightweight neural network tuned by an automatic hyperparameter search. Validated on real-time sleep spindle detection using the MODA dataset's group consensus as a reference, it performed comparably to offline expert annotation [4]. That matters for the question because it shows the speed-accuracy trade-off is not confined to one lab's custom rig; a compact, affordable system can also do the job.

Does the detection hold up when stimulation is actually delivered?

Detection accuracy on replayed data is one thing; a working closed loop is another. In a study of 10 healthy adults across six nights each, spindles were detected and, on half the nights, targeted with auditory stimulation. Stimulation landed before the spindle's offset in 97.6% of detections and did not disturb sleep [3]. That figure is a direct measure of whether the loop closes in time: the system identified the event and delivered the sound while the spindle was still ongoing.

The same study also shows why timing within the spindle matters. Stimulation produced increased sigma activity (11–16 Hz) about one second afterward, but stimulating at the very beginning of a spindle also caused the spindle to end early. And stimulating an already-evoked spindle did not produce additional sigma activity [3]. So a fast, accurate detector is necessary but not sufficient: where in the spindle you stimulate changes the neural outcome, which is why phase-specific triggering was a design goal from the start [1].

The next hurdle: one detector does not fit every sleeper

Spindles vary between individuals, and a detector tuned to one person's brain may not transfer cleanly to another's. A 2025 continual-learning framework addressed this by starting from a model that already classifies sleep stages and detects spindles online, then refining spindle detection for the individual across one or more nights without manual intervention. Fine-tuning alone gave minimal benefit within a single night, but combining it with weight averaging improved performance across multiple nights while limiting catastrophic forgetting — the tendency of a model to lose earlier learning as it adapts to new data [2].

That result suggests personalization is a multi-night project rather than a one-off calibration, at least in this framework. It also points to where the field is heading: detection that adapts to the sleeper rather than a fixed threshold applied to everyone [2]. For now, the evidence supports a straightforward answer to the original question. Real-time spindle detection is fast and accurate enough to drive closed-loop stimulation, demonstrated both in replay validation and in live sleep with auditory stimulation landing before spindle offset in the large majority of detections [1][3][4]. What remains an active design question is not whether the loop can close, but how precisely to aim it — at which phase, in which person, on which night [1][2][3].

Sources used in this answer

Related research questions