Objectives: The aim of this study was to classify polysomnography (PSG)-based sleep disorders using actigraphy data using a convolutional neural network (CNN). Methods: Actigraphy data, PSG data, and diagnoses were obtained from 214 patients from a single-center sleep clinic. Patients diagnosed with circadian sleep disorders, narcolepsy, or periodic limb movement disorders were excluded. From the actigraphy data, three types of data were selected from the first 5 days, namely, sleep-wake status, activity count, and light exposure per epoch. The data were processed into a two-dimensional array with four instances, namely, 24-hour full-day data and data for 6, 8, and 10 hours timepoints after sleep onset, and then analyzed. Using a CNN, we attempted to classify the processed data into PSG-based diagnoses. Results: Overfitting of the training data was observed. The CNN showed near-perfect accuracy on the test data, but failed to classify the validation data (area under the curve: 24-hour full-day data: 0.6031, 6 hours after sleep onset: 0.5148, 8 hours: 0.6122, and 10 hours: 0.5769). Conclusions The lack and inaccuracy of data were responsible for the results. A higher sampling rate and additional ancillary data, such as PSG or heart rate variability data, are necessary for accurate classification. Additionally, alternative approaches to machine learning, such as transformers, should be considered in future studies.

Objectives: The aim of this study was to classify polysomnography (PSG)-based sleep disorders using actigraphy data using a convolutional neural network (CNN). Methods: Actigraphy data, PSG data, and diagnoses were obtained from 214 patients from a single-center sleep clinic. Patients diagnosed with circadian sleep disorders, narcolepsy, or periodic limb movement disorders were excluded. From the actigraphy data, three types of data were selected from the first 5 days, namely, sleep-wake status, activity count, and light exposure per epoch. The data were processed into a two-dimensional array with four instances, namely, 24-hour full-day data and data for 6, 8, and 10 hours timepoints after sleep onset, and then analyzed. Using a CNN, we attempted to classify the processed data into PSG-based diagnoses. Results: Overfitting of the training data was observed. The CNN showed near-perfect accuracy on the test data, but failed to classify the validation data (area under the curve: 24-hour full-day data: 0.6031, 6 hours after sleep onset: 0.5148, 8 hours: 0.6122, and 10 hours: 0.5769). Conclusions The lack and inaccuracy of data were responsible for the results. A higher sampling rate and additional ancillary data, such as PSG or heart rate variability data, are necessary for accurate classification. Additionally, alternative approaches to machine learning, such as transformers, should be considered in future studies.

Objectives: The aim of this study was to classify polysomnography (PSG)-based sleep disorders using actigraphy data using a convolutional neural network (CNN). Methods: Actigraphy data, PSG data, and diagnoses were obtained from 214 patients from a single-center sleep clinic. Patients diagnosed with circadian sleep disorders, narcolepsy, or periodic limb movement disorders were excluded. From the actigraphy data, three types of data were selected from the first 5 days, namely, sleep-wake status, activity count, and light exposure per epoch. The data were processed into a two-dimensional array with four instances, namely, 24-hour full-day data and data for 6, 8, and 10 hours timepoints after sleep onset, and then analyzed. Using a CNN, we attempted to classify the processed data into PSG-based diagnoses. Results: Overfitting of the training data was observed. The CNN showed near-perfect accuracy on the test data, but failed to classify the validation data (area under the curve: 24-hour full-day data: 0.6031, 6 hours after sleep onset: 0.5148, 8 hours: 0.6122, and 10 hours: 0.5769). Conclusions The lack and inaccuracy of data were responsible for the results. A higher sampling rate and additional ancillary data, such as PSG or heart rate variability data, are necessary for accurate classification. Additionally, alternative approaches to machine learning, such as transformers, should be considered in future studies.