Sleep Detection Task

Overview

The objective of sleep detection is to identify periods of sustained sleep over the course of several days or weeks. The de facto standard for long-term, ambulatory sleep detection is actigraphy, which is a method of monitoring gross motor activity using an accelerometer. However, actigraphy is not a reliable method for sleep detection as it employs very simple heuristics to determine sleep. Often actigraphy can misclassify periods of inactivity or even not-worn as detected sleep.

In this task, we look to leverage a light-weight model that can outperform actigraphy similarly using only data from an IMU. For more advanced sleep analysis, refer to the Sleep Stage Classification. By leveraring Ambiq’s ultra-low-power microcontroller along with an ultra-low-power IMU, an efficient AI enabled actigraphy or fitness band will be able to run for weeks off a single charge.

Wrist-based Sleep Detection

---

Model Zoo

The following table provides the latest performance and accuracy results for pre-trained models. Additional result details can be found in Model Zoo → Detect.

NAME	LOCATION	# CLASSES	MODEL	PARAMS	FLOPS	ACCURACY	AP
SD-2-TCN-SM	Wrist	2	TCN	9K	1.5M/hr	95.9%	94.5%

---

Target Classes

Below outlines the classes available for sleep detect classification. When training a model, the number of classes, mapping, and names must be provided.

CLASS	LABELS
0	AWAKE
1	SLEEP

Class Mapping

Below is an example of a class mapping for a 2-class sleep detect model. The class map keys are the original class labels and the values are the new class labels. Any class not included will be skipped.

JSON example

{
    "num_classes": 2,
    "class_names": ["AWAKE", "SLEEP"],
    "class_map": {
        "0": 0,
        "1": 1
    }
}

---

References

• AI-Driven sleep staging from actigraphy and heart rate