A Human-Machine Coupled System for Efficient Sleep Spindle Detection by Iterative Revision

Abstract Sleep spindles are characteristic events in EEG signals during non-REM sleep, and are known to be important biological markers. Manually labeling spindles by visual inspection, however, has proved to be a tedious task. Automatic detection algorithms generalize weakly for versatile spindle forms, and machine-learning methods require large datasets to train, which are unfeasible to acquire particularly for experimental animal groups. Here, a novel, integrated system based on a process of iterative “Selection-Revision” (iSR) is introduced to aid in the efficient detection of spindles. By coupling low-threshold automatic detection of spindle events based on selected parameters with manual “Revision,” the human task is effectively simplified from searching across signal traces to binary verification. Convergence was observed between resulting spindle sets through iSR, largely independent of their initial labeling, demonstrating the robustness of the method. Although possible breakdown of the revised spindle sets could be seen after multiple rounds of Revision, due to overfitting of the revised set to the initial human labeling, this could be compensated for by a Selection scheme tolerant to higher False-Negative rates of the machine labeling relative to the standard set. It was also found that iSR is generalizable to different datasets, and that initial human labeling could be substituted by low-threshold machine detection. Overall, this human-machine coupled approach allows for fast labeling to obtain consistent spindle sets, which can also be used to train machine-learning models in the future. The principle of iSR may also be applied for many different data types to assist with other pattern detection tasks. Significance Statement Electroencephalography (EEG) recordings are widely adopted in brain research. Abnormalities in the occurrence of particular EEG waveforms, such as sleep spindles, can be used to diagnose psychiatric diseases. Traditionally, human experts have labeled EEG traces for sleep spindles, a time consuming process; automated detection algorithms, however, often yield inaccurate results. This study introduces a new method for efficient sleep spindle detection with a human-machine coupled system that can iteratively revise labeled datasets, enabling convergence towards a robust, accurate spindle labeling. This system eases large-scale sleep spindle detection, which can yield datasets for both biological analyses and for training machine-learning models. Furthermore, the underlying method of iterative revision can be used to analyze other types of patterns efficiently.