Wireless wearable sensors can facilitate rapid detection of sleep apnea in hospitalized stroke patients

Abstract

Study objectives: To evaluate wearable devices and machine learning for detecting sleep apnea in patients with stroke at an acute inpatient rehabilitation facility (IRF).

Methods: A total of 76 individuals with stroke wore a standard home sleep apnea test (ApneaLink Air), a multimodal, wireless wearable sensor system (ANNE), and a research-grade actigraphy device (ActiWatch) for at least 1 night during their first week after IRF admission as part of a larger clinical trial. Logistic regression algorithms were trained to detect sleep apnea using biometric features obtained from the ANNE sensors and ground truth apnea rating from the ApneaLink Air. Multiple algorithms were evaluated using different sensor combinations and different apnea detection criteria based on the apnea-hypopnea index (AHI ≥ 5, AHI ≥ 15).

Results: Seventy-one (96%) participants wore the ANNE sensors for multiple nights. In contrast, only 48 participants (63%) could be successfully assessed for obstructive sleep apnea by ApneaLink; 28 (37%) refused testing. The best-performing model utilized photoplethysmography (PPG) and finger-temperature features to detect moderate-severe sleep apnea (AHI ≥ 15), with 88% sensitivity and a positive likelihood ratio (LR+) of 44.00. This model was tested on additional nights of ANNE data achieving 71% sensitivity (10.14 LR+) when considering each night independently and 86% accuracy when averaging multi-night predictions.

Conclusions: This research demonstrates the feasibility of accurately detecting moderate-severe sleep apnea early in the stroke recovery process using wearable sensors and machine learning techniques. These findings can inform future efforts to improve early detection for post-stroke sleep disorders, thereby enhancing patient recovery and long-term outcomes.

Clinical trial: SIESTA (Sleep of Inpatients: Empower Staff to Act) for Acute Stroke Rehabilitation, https://clinicaltrials.gov/study/NCT04254484?term=SIESTA&checkSpell=false&rank=1, NCT04254484.

Keywords: machine learning; obstructive sleep apnea; sleep apnea screening; stroke.

Graphical Abstract

Figure 1.

(i) ApneaLink Air HSAT and (ii) ANNE (Chest and Finger) sensors and ActiWatch, a: ANNE Chest, b: ActiWatch, c: ANNE Finger. (iii) Sensor wear timeline for ActiWatch, ANNE Chest, ANNE finger, and ApneaLink (black with orange stripe). Timeline contains generalized dates, s0-date of admission, s-start of study date, d-date of discharge.

Figure 2.

ApneaLink and ANNE adherence flowchart. * “Successful ApneaLink Screening” based only on ApneaLink producing an AHI score.

Figure 3.

ROC charts depicting the final mean performance across all model evaluations. Left are the models that use all features, middle are finger-based features, right are chest-based features. (i) Considering AHI (< 5) and AHI (≥ 5). (ii) Considering AHI (< 5) and AHI (≥ 15). (iii) Considering AHI (< 15) and AHI (≥ 15).

Figure 4.

(a) Heart rate (PPG) (b) SpO2 median frequencies per AHI value. The bar graphs (left) share the same axis as the scatter plots and depict the two groups distinguished by the model, split into normal and moderate/severe. A significant difference between groups was found for both heart rate (PPG) and SpO2 with the same statistical measure denoted by an asterisk (Mann–Whitney U: U = 30, p = .002).

Figure 5.

Represents the average and standard deviation of the final model predictions on all nights of ANNE per participant. Models were trained excluding the participant whose data is being predicted, meaning a final individual model was trained per participant. Compares the probability of apnea to scaled AHI, where probability values over 0.5 indicate apnea and under 0.5 indicate no apnea. Scaled AHI scales the AHI values such that, normal (0 to 5) are now 0 to 1, mild (5 to 15) are now 2 to 3, moderate (15 to 30) are now 2 to 3, and severe (30+, here capped at 50) are now 3 to 4. (i) Considering AHI (< 5) and AHI (≥ 5), using features from Finger-PPG. (ii) Considering AHI (< 5) and AHI (≥ 15), using features from Finger. (iii) Considering AHI (< 15) and AHI (≥ 15), using features from Finger-PPG.