Table 1.
MESA and MrOS participant demographics in this study.
Fig 1.
SLAMSS network architecture showing sample activity, heart rate mean (HRM) and heart rate standard deviation (HRSD) time series that are the model inputs and sleep stage labels Wake (W), REM (R), and NREM (N) that are the model outputs. Our SLAMSS network implementation operates on 12-epoch input sequences that are passed into a set of CNNs (epoch duration = 30 s). The CNN features go through an attention-guided encoder-decoder system that generates the output labels.
Fig 2.
MESA three-class sleep staging.
Confusion matrices for four classifiers: SLAMSS with activity, HRM, and HRSD inputs (SLAMSS-Act-HR), LSTM with activity, HRM, and HRSD inputs (LSTM-Act-HR), SLAMSS with HRM and HRSD inputs (SLAMSS-HR), and SLAMSS with an activity input (SLAMSS-Act). Each box shows % epochs at the top and the number of epochs below. Columns sum to 100%. Expert manual sleep staging by PSG is used as the ground truth. It should be noted that, for three-class staging, category assignment by random chance would lead to a value of 33.33% for the diagonal elements of these matrices.
Table 2.
Performance comparison of SLAMSS with other publications performing three or four-class sleep staging on activity (Act) and heart rate or heart rate variability (HRV) data.
Table 3.
MESA three-class sleep staging.
Comparison of classifier performance metrics for four classifiers: SLAMSS with activity, HRM, and HRSD inputs (SLAMSS-Act-HR), LSTM with activity, HRM, and HRSD inputs (LSTM-Act-HR), SLAMSS with HRM and HRSD inputs (SLAMSS-HR), and SLAMSS with an activity input (SLAMSS-Act). PSG is used as the ground truth for the computation of all metrics. Subject-wise values are reported as mean(s.d.).
Table 4.
MESA three-class sleep staging.
Comparison of MAE for clinical sleep metrics for four classifiers against PSG: SLAMSS with activity, HRM, and HRSD inputs (SLAMSS-Act-HR), LSTM with activity, HRM, and HRSD inputs (LSTM-Act-HR), SLAMSS with HRM and HRSD inputs (SLAMSS-HR), and SLAMSS with an activity input (SLAMSS-Act). MAE values are provided in the format: mean (s.d.).
Fig 3.
MESA three-class sleep staging.
Comparison of clinical sleep metrics for four classifiers: SLAMSS with activity, HRM, and HRSD inputs (SLAMSS-Act-HR), LSTM with activity, HRM, and HRSD inputs (LSTM-Act-HR), SLAMSS with HRM and HRSD inputs (SLAMSS-HR), and SLAMSS with an activity input (SLAMSS-Act). The orange dotted line corresponds to the PSG (assumed ground truth) value of each metric.
Fig 4.
MrOS three-class sleep staging.
Confusion matrices for four classifiers: SLAMSS with activity, HRM, and HRSD inputs (SLAMSS-Act-HR), LSTM with activity, HRM, and HRSD inputs (LSTM-Act-HR), SLAMSS with HRM and HRSD inputs (SLAMSS-HR), and SLAMSS with an activity input (SLAMSS-Act). Each box shows % epochs at the top and the number of epochs below. Columns sum to 100%. Expert manual sleep staging by PSG is used as the ground truth. It should be noted that, for three-class staging, category assignment by random chance would lead to a value of 33.33% for the diagonal elements of these matrices.
Table 5.
MrOS three-class sleep staging.
Comparison of classifier performance metrics for four classifiers: SLAMSS with activity, HRM, and HRSD inputs (SLAMSS-Act-HR), LSTM with activity, HRM, and HRSD inputs (LSTM-Act-HR), SLAMSS with HRM and HRSD inputs (SLAMSS-HR), and SLAMSS with an activity input (SLAMSS-Act). PSG is used as the ground truth for the computation of all metrics. Subject-wise values are reported as mean(s.d.).
Fig 5.
MrOS three-class sleep staging.
Comparison of clinical sleep metrics for four classifiers: SLAMSS with activity, HRM, and HRSD inputs (SLAMSS-Act-HR), LSTM with activity, HRM, and HRSD inputs (LSTM-Act-HR), SLAMSS with HRM and HRSD inputs (SLAMSS-HR), and SLAMSS with an activity input (SLAMSS-Act). The orange dotted line corresponds to the PSG (assumed ground truth) value of each metric.
Table 6.
MrOS three-class sleep staging.
Comparison of MAE for clinical sleep metrics for four classifiers against PSG: SLAMSS with activity, HRM, and HRSD inputs (SLAMSS-Act-HR), LSTM with activity, HRM, and HRSD inputs (LSTM-Act-HR), SLAMSS with HRM and HRSD inputs (SLAMSS-HR), and SLAMSS with an activity input (SLAMSS-Act). MAE values are provided in the format: mean (s.d.).
Fig 6.
Confusion matrices for the SLAMSS model without MESA pretraining (i.e., direct training) and with MESA pretraining (i.e., transfer learning) with activity, HRM, and HRSD inputs for the Apple Watch dataset.
The corresponding accuracies reported in [46] are provided for reference.
Fig 7.
MESA four-class sleep staging.
Confusion matrices for four-class sleep staging using SLAMSS with an inverse-frequency-weighted cross-entropy loss function (SLAMSS-IF) and SLAMSS with a real-world-weighted cross-entropy loss function (SLAMSS-RW). It should be noted that, for four-class staging, category assignment by random chance would lead to a value of 25% for the diagonal elements of these matrices.
Table 7.
MESA four-class sleep staging.
Comparison of classifier performance metrics for four-class sleep staging using SLAMSS with an inverse-frequency-weighted cross-entropy loss function (SLAMSS-IF) and SLAMSS with a real-world-weighted cross-entropy loss function (SLAMSS-RW). PSG is used as the ground truth for the computation of all metrics. Subject-wise values are reported as mean(s.d.).
Table 8.
MESA four-class sleep staging.
Comparison of MAE for clinical sleep metrics for four-class sleep staging using SLAMSS with an inverse-frequency-weighted cross-entropy loss function (SLAMSS-IF) and SLAMSS with a real-world-weighted cross-entropy loss function (SLAMSS-RW) against PSG. MAE values are provided in the format: mean (s.d.).
Fig 8.
MESA four-class sleep staging.
Comparison of clinical sleep metrics for four-class sleep staging using SLAMSS with an inverse-frequency-weighted cross-entropy loss function (SLAMSS-IF) and SLAMSS with a real-world-weighted cross-entropy loss function (SLAMSS-RW). The orange dotted line corresponds to the PSG (assumed ground truth) value of each metric.
Fig 9.
MrOS four-class sleep staging.
Confusion matrices for four-class sleep staging using SLAMSS with an inverse-frequency-weighted cross-entropy loss function (SLAMSS-IF) and SLAMSS with a real-world-weighted cross-entropy loss function (SLAMSS-RW). It should be noted that, for four-class staging, category assignment by random chance would lead to a value of 25% for the diagonal elements of these matrices.
Table 9.
MrOS four-class sleep staging.
Comparison of classifier performance metrics for four-class sleep staging using SLAMSS with an inverse-frequency-weighted cross-entropy loss function (SLAMSS-IF) and SLAMSS with a real-world-weighted cross-entropy loss function (SLAMSS-RW). PSG is used as the ground truth for the computation of all metrics. Subject-wise values are reported as mean(s.d.).
Fig 10.
MrOS four-class sleep staging.
Comparison of clinical sleep metrics for four-class sleep staging using SLAMSS with an inverse-frequency-weighted cross-entropy loss function (SLAMSS-IF) and SLAMSS with a real-world-weighted cross-entropy loss function (SLAMSS-RW). The orange dotted line corresponds to the PSG (assumed ground truth) value of each metric.
Table 10.
MrOS four-class sleep staging.
Comparison of MAE for clinical sleep metrics for four-class sleep staging using SLAMSS with an inverse-frequency-weighted cross-entropy loss function (SLAMSS-IF) and SLAMSS with a real-world-weighted cross-entropy loss function (SLAMSS-RW) against PSG. MAE values are provided in the format: mean (s.d.).