KindSleep: Knowledge-Informed Diagnosis of Obstructive Sleep Apnea from Oximetry
Authors: Micky C Nnamdi, Wenqi Shi, Cheng Wan, J. Ben Tamo, Benjamin M Smith, Chad A Purnell, May D Wang
Organizations: Georgia Institute of Technology Atlanta, Georgia, USA · UT Southwestern Medical Center Dallas, Texas, USA · Shriners Children’s Chicago, Illinois, USA
Obstructive sleep apnea (OSA) is a sleep disorder that affects nearly one billion people globally and significantly elevates cardiovascular risk. Traditional diagnosis through polysomnography is resource-intensive and limits widespread access, creating a critical need for accurate and efficient alternatives. In this paper, we introduce KindSleep, a deep learning framework that integrates clinical knowledge with single-channel patient-specific oximetry signals and clinical data for precise OSA diagnosis. KindSleep first learns to identify clinically interpretable concepts, such as desaturation indices and respiratory disturbance events, directly from raw oximetry signals. It then fuses these AI-derived concepts with multimodal clinical data to estimate the Apnea-Hypopnea Index (AHI). We evaluate KindSleep on three large, independent datasets from the National Sleep Research Resource (SHHS, CFS, MrOS; total n = 9,815). KindSleep demonstrates excellent performance in estimating AHI scores (R2 = 0.917, ICC = 0.957) and consistently outperforms existing approaches in classifying OSA severity, achieving weighted F1-scores from 0.827 to 0.941 across diverse populations. By grounding its predictions in a layer of clinically meaningful concepts, KindSleep provides a more transparent and trustworthy diagnostic tool for sleep medicine practices.
Figures & tables
Figure 1. Overview of KindSleep . KindSleep involved two main components: the sleep annotation model, which extracts clinically relevant metrics from raw oximetry signals, and the regression model, which integrates these metrics with processed clinical data to estimate the AHI. (Right) Example of oximetry signals from a mild OSA patient (top; reference AHI = 5.65) and a healthy control (bottom; reference AHI = 0.175), annotated with hypopnea events (green) and desaturations (red), alongside corresponding attention maps from the sleep annotation model that highlight the regions the model concentrates on, and the resulting AHI estimations from the regression model.
Label
Description
ahi_a0h4
(Apneas with no oxygen desaturation threshold used and with or without arousal + hypopneas with > 30% flow reduction and ≥ 4% oxygen desaturation and with or without arousal) / hour of sleep
ahi_a0h4a
Similar to ahi_a0h4, but hypopneas are tallied if they meet either of two conditions: ≥ 4% oxygen desaturation or evidence of arousal.
ahi_c0h3
(Central apneas with no oxygen desaturation threshold used and with or without arousal + hypopneas with > 30% flow reduction and ≥ 3% oxygen desaturation and with or without arousal) / hour of sleep
ahi_c0h4
Central apnea index under the same conditions as above, except that hypopneas are credited if they meet a ≥ 4% desaturation threshold or are associated with arousal.
avgsat
Mean oxygen saturation value calculated over the full sleep period
minsat
Minimum oxygen saturation value recorded during the sleep period
Table 1. Descriptions of Extracted Signal Annotations 2 2 2 https://sleepdata.org/datasets/shhs/variables .
Figure 2. (a) Parity plots, (b) Bland–Altman plots, and (c) confusion matrix results for SHHS1, SHHS2, CFS and MrOS.
Models
SHHS1
SHHS2
CFS
MrOS
MAE
RMSE
MAE
RMSE
MAE
RMSE
MAE
RMSE
Linear
21.526
29.341
24.140
36.512
33.319
45.180
20.320
51.314
Ridge
15.741
21.198
15.204
20.767
21.014
28.612
5.116
11.509
Decision Tree
9.740
14.177
10.049
14.540
11.236
17.108
3.737
8.729
K-Neighbors
9.238
13.805
8.793
13.238
10.363
15.179
8.161
11.095
SVR
8.425
12.710
8.095
12.165
9.744
14.946
8.401
12.078
Table 2. Performance Metrics of SLAM Across Different Datasets. The table shows the MAE and RMSE for SHHS1, SHHS2, CFS, and MrOS, indicating the robustness and generalization capability of SLAM , when compared with other models.
Regression Model
Dataset
F1 Score
Precision
Sensitivity
Specificity
R2
ICC
MAE
RMSE
MLP-Regressor
SHHS1
0.827±0.030
0.832±0.030
0.826±0.034
0.935±0.013
0.917
0.957
3.101
4.677
SHHS2
0.756±0.016
0.763±0.017
0.754±0.017
0.899±0.008
0.873
0.934
3.731
5.589
CFS
0.786±0.032
0.803±0.031
0.778±0.034
0.921±0.017
0.884
0.938
3.930
6.743
MrOS
0.777±0.026
0.782±0.025
0.775±0.025
0.916±0.010
0.861
0.929
3.486
6.038
Table 3. Evaluation Metrics Across Various Datasets: The weighted F1 Score, precision, sensitivity, specificity, R2 , ICC , MAE and RMSE across various datasets for the regression model considered. All metrics are weighted, meaning they are calculated by averaging values proportionally to class sizes to account for class imbalances in the datasets.
Figure 3. Outcome comparison across varying proportions of knowledge-informed metrics.
Figure 4. Radar charts comparing various performance metrics of our KindSleep model against two baseline multimodal integration methods on the (a) SHHS1, (b) SHHS2, (c) CFS.
Figure 5. Attention mechanism employed by the SLAM model across oximetry signals, with events (e.g., desaturation, apnea, and artifacts) identified from ground truth annotations. The top section displays the global signal over the full duration (0–25,200 seconds), highlighting areas of high activation that correspond to physiologically relevant events, such as desaturation and apnea, while effectively ignoring artifact-prone regions. The middle section provides a focused view of the signal between 5,000 and 9,000 seconds, where the model demonstrates precise attention on desaturation events and hypopneas. The bottom section zooms into a segment from 20,000 to 25,000 seconds, showcasing the model’s ability to neglect low-amplitude, artifact-dominated regions devoid of clinically significant activities.
Figure 6. MLP-Regressor : SHAP results show the ranking of predicted knowledge-informed metrics against the patient clinical data.
Figure 7. Relationship between the F1 scores, MAE, and RMSE as errors are intercepted. We observed that identifying and adjusting these errors before passing them to the regression model during training significantly improves the system architecture’s performance.
Annotations
F1 Score
Precision
Sensitivity
Specificity
R2
ICC
MAE
RMSE
incorrect annotations
0.223±0.036
0.174±0.032
0.313±0.038
0.749±0.020
0.125
0.245
10.831
15.226
incorrect + 5% correct annotations
0.637±0.039
0.647±0.040
0.643±0.040
0.871±0.017
0.832
0.910
4.854
6.679
correct annotations
0.827±0.030
0.832±0.030
0.826±0.034
0.935±0.013
0.917
0.957
3.101
4.677
Table 4. Impact of Incorrect, Partially Correct, and Fully Correct Annotations on Pipeline Performance. The results indicate that KindSleep performance decreases with incorrect annotations.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Description
Train
Validation
Test
Recording (Counts)
3539
1531
4745
Duration (Hours)
24,773
10,717
33,215
Gender(Male/Female)
47.40/52.67
48.07/51.93
57.74/42.26
Age (Years)
63.27±11.23
63.47±10.87
65.57±15.42
Body Mass Index (BMI)
28.22±5.10
27.99±4.97
28.46±5.66
Ethnicity (Non-Hispanic/Hispanic)
95.51/4.49
94.91/5.09
97.01/2.99
Appendix
Table 5. Overview of dataset details, including training, validation, and testing splits, participant counts, demographics (age, BMI, gender, ethnicity, and race), and AHI categories (Healthy, Mild, Moderate, Severe) with corresponding mean and standard deviation.
Hong Kong Xsmart Century Technology Co., Ltd., Hong Kong, China · Urban Mobility Lab, Massachusetts Institute of Technology, Cambridge, MA 02139, USA · Singapore-MIT Alliance for Research and Technology (SMART) Centre, Singapore 138602, Singapore +2
School of Electronic and Electrical Engineering, Shanghai University of Engineering Science, Shanghai 200438, China · Tencent Youtu Lab, Shanghai 200233, China · ENT Institute and Department of Otorhinolaryngology, Eye & ENT Hospital of Fudan University, Shanghai 200031, China +1