KindSleep: K nowledge- In formed D iagnosis of Obstructive Sleep Apnea from Oximetry Conference: ; ; CCS: Computing methodologies Machine learning approaches CCS: Applied computing Health informatics
Micky C. Nnamdi, Wenqi Shi, Cheng Wan, J. Ben Tamo, Benjamin Smith, Chad Purnell, May D. Wang · 2026
Read KindSleep: K nowledge- In formed D iagnosis of Obstructive Sleep Apnea from Oximetry Conference: ; ; CCS: Computing methodologies Machine learning approaches CCS: Applied computing Health informatics by Micky C. Nnamdi, Wenqi Shi, Cheng Wan, J. Ben Tamo, Benjamin Smith, Chad Purnell, May D. Wang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-04755.txt.
Introduction
Nnamdi and colleagues offer KindSleep, a deep learning framework spotting blocked sleep apnea from oxygen. Sleep apnea affects nearly one billion people and raises heart risk, yet sleep labs limit access. KindSleep first learns clear concepts, like oxygen dips and breathing pauses, from raw oxygen signals. It fuses these concepts with medical data from several sources to gauge the Apnea-Hypopnea Index. Tested on three large sets, KindSleep gauges AHI scores with an R-squared of 0.917 and beats existing severity tests. Grounding calls in medical concepts supports sound care.
Catalog record
| CREATOR | Micky C. Nnamdi, Wenqi Shi, Cheng Wan, J. Ben Tamo, Benjamin Smith, Chad Purnell, May D. Wang |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 48840 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.04755
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
KindSleep: K nowledge- In formed D iagnosis of Obstructive Sleep Apnea from Oximetry Conference: ; ; CCS: Computing methodologies Machine learning approaches CCS: Applied computing Health informatics Micky C. Nnamdi, Wenqi Shi, Cheng Wan, J. Ben Tamo, Benjamin Smith, Chad Purnell, May D. Wang Abstract . 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 outperformed 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. 1. Introduction Obstructive sleep apnea (OSA) is a significant global health concern, affecting approximately one billion individuals worldwide between the ages of 30 and 69 ( 26 ) . OSA involves repeated episodes of partial or complete airway obstruction during sleep, increasing the risk of hypertension, cardiovascular disease, stroke, and significantly impairing overall quality of life ( 28 ; 58 ; 39 ) . Therefore, accurate and timely diagnosis is essential for effective management and reduction of associated health complications. 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. Polysomnography (PSG), the gold standard for diagnosing OSA ( 6 ) , requires overnight monitoring of multiple physiological parameters, including electroencephalogram (EEG), electrocardiogram (ECG), electromyogram (EMG), electrooculogram (EOG), oronasal airflow, respiratory effort, and oxygen saturation levels. Each of these signals provides complementary information: EEG identifies sleep stages, airflow and oxygen saturation detect breathing disruptions, and EMG/EOG capture muscle tone and eye movements essential for event classification. Accurately computing the apnea hypopnea index (AHI) requires integrating these heterogeneous signals in a time-synchronized manner, a process that is technically demanding due to noise, inter-signal variability, and the need for expert scoring. Clinicians rely on this integrated signal analysis measurements to compute the AHI, classifying OSA severity into normal (AHI < 5), mild (5 \leq AHI < 15), moderate (15 \leq AHI < 30), and severe (AHI \geq 30) ( 17 ) . However, PSG is complex, resource-intensive, and expensive, severely restricting its accessibility, especially in resource-limited regions. Given these limitations, researchers have increasingly focused on automated diagnostic methods leveraging fewer physiological channels, notably oximetry signals ( 8 ; 16 ; 2 ; 3 ; 23 ; 49 ; 56 ; 46 ) . Recently, multimodal approaches combining oximetry signals with clinical data have emerged, aiming to enhance diagnostic accuracy by incorporating comprehensive patient-specific data ( 22 ; 13 ) . Oximetry provides crucial insights into respiratory disturbances during sleep through measurements of blood oxygen saturation fluctuations ( 29 ) . Concurrently, clinical data offer valuable demographic, clinical, and lifestyle information, including critical risk factors and comorbidities linked to OSA ( 20 ; 43 ) . Body mass index (BMI), in particular, is strongly correlated with OSA prevalence, significantly higher among overweight and obese populations ( 1 ; 45 ; 55 ; 37 ; 27 ; 12 ) . Such findings highlight the importance of integrating clinical and demographic data into OSA diagnostic models. However, existing automated OSA diagnostic methods using oximetry often face challenges, including limited accuracy, generalizability, and transparency in decision-making. Such "black-box" AI approaches hinder clinical trust, preventing clinicians from understanding, validating, and confidently using these tools in clinical practice ( 44 ; 54 ; 15 ; 40 ; 30 ; 42 ; 48 ; 41 ) . To address these issues, we introduce KindSleep , an AI-based framework designed to integrate clinical expertise with oximetry and clinical data for accurate and reliable OSA diagnosis. KindSleep employs the concept bottleneck model (CBM) paradigm ( 21 ; 31 ) , leveraging clinically interpretable metrics derived from oximetry data, such as oxygen desaturation indices, oxygen saturation levels, and hypopnea event frequencies, as intermediate features to guide model predictions ( 14 ) . In this study, we validated KindSleep using extensive data from three independent cohorts (SHHS, CFS, MrOS) encompassing 9,815 sleep recordings from the National Sleep Research Resource (NSRR). KindSleep demonstrated strong predictive performance, achieving a coefficient of determination ( R^{2} ) of 0.917 and an intraclass correlation coefficient (ICC) of 0.957 , significantly outperforming existing methods. Notably, our framework exhibited high generalizability across diverse patient populations, achieving weighted F1-scores ranging from 0.827 to 0.941. 2. KindSleep 2.1. Overview We propose KindSleep , a clinically grounded machine learning framework designed to estimate AHI using a combination of oximetry signals and clinical data. The key innovation lies in introducing an intermediate layer of clinically interpretable features—knowledge-informed metrics—derived from the raw oximetry signal. These metrics simulate the annotations a sleep technologist would typically provide and help bridge the gap between raw physiological input and meaningful clinical prediction. KindSleep comprises two key components: a SLeep Annotation Model ( SLAM ) that derives interpretable concepts and a regressor model (primarily implemented as MLP-Regressor ) that leverages these concepts—together with clinical data—to produce the final AHI prediction. Table 1 . Descriptions of Extracted Signal Annotations 2 2 2 https://sleepdata.org/datasets/shhs/variables . Label Description ahi_a0h4 (Apneas with no oxygen desaturation threshold used and with or without arousal + hypopneas with > 30% flow reduction and \geq 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: \geq 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 \geq 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 \geq 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 rdi0p Apneas with no oxygen desaturation threshold used and with or without arousal + hypopneas with > 30%