Ƶ

PPG-Based AI to Detect Obstructive Sleep Apnea Shows No Bias Based on Skin Color

— No significant differences between AI analysis and gold-standard polysomnography

Ƶ MedicalToday

HOUSTON -- A photoplethysmography (PPG)-based artificial intelligence (AI) system for detection of obstructive sleep apnea (OSA) showed no evidence of bias in performance based on skin color, according to a prospective trial.

There was an average difference of 0.95 points (95% CI 0.66-1.22) between the PPG-based AI system's apnea-hypopnea index (AHI) and the gold-standard polysomnography (PSG)-AHI measure, with 90.4% sensitivity and 71.9% specificity for an AHI of 5 or more, which indicates OSA, reported Chris Fernandez, MS, of EnsoData Research in Madison, Wisconsin.

In addition, the average difference between the PPG-based AI system's total sleep time (TST) measure and the PSG-TST measure was -2.7 minutes (95% CI -4.04 to -1.4), he noted at the annual SLEEP meeting, hosted jointly by the American Academy of Sleep Medicine and the Sleep Research Society.

There were no statistically significant differences observed in AHI performance (P>0.05 for both Munsell Color System-color value and -chroma ordinary least squares analysis) or TST performance (P>0.05 for both measures) in relation to skin color.

The performance of fingertip pulse oximeters during hypoxemia varies for both adult and pediatric patients with darker skin tones, with many popular devices falling short.

Fernandez explained that skin pigmentation can impact the sensitivity and accuracy of PPG, especially for the darkest skin tones. Other factors that may affect PPG performance include low perfusion, severe anemia, high body mass index, nail polish, finger tattoos, low skin temperature, and other demographic characteristics.

However, gaps remain regarding validation of sleep medicine applications, including PPG.

"In the context of clinical trial design, the way that this is traditionally or has traditionally been done most commonly is with ethno-racial categories, basically a survey of nationality and origin," said Fernandez. "This has many challenges in terms of self-identification, how many categories we need to include to accurately capture diversity, [and] challenges with mixed races. In simple terms, people with the same race come in many different skin colors and because skin pigment is what's most important for PPG sensitivity and accuracy, we need to sort of go beyond these surveys."

Fernandez told Ƶ that one of the biggest takeaways from this study is "comfort and confidence that the tools are going to work well for a broad spectrum of patients, for all the types of patients they see."

"I think about both access and inclusivity in creating diagnostic tools that -- probably they work well for all ages, all genders, all skin colors," he added. "They're well suited for all levels of income or any type of insurance carrier and can plug in broadly with ... the PPG pulse oximeter devices that are already out there today. It's really about being able to bring that confidence in inclusivity of the care that they're providing and to feel good about it."

"Standardized skin pigmentation assessments can be readily incorporated into clinical study designs to provide objective evidence and validation that the technology works well for all skin colors," Fernandez said, noting that it is important to "ensure that diagnostic tools can help to address and not reinforce existing disparities and access to OSA testing and care."

The study was conducted across 10 sites in six states and included a total of 235 patients who were referred for diagnostic PSG. Of these patients, 21.5% were between the ages of 60-69, 19% were 30-39, 18.4% were 40-49, 17.1% were 50-59, 15.2% were 18-29, and 7.6% were 70-79; 55.7% were women. Mean AHI was 15.82, mean TST was 5.11, mean skin color value was 6.62, and mean skin color chroma was 3.0.

About 44% had sleep disorders, 36.7% had psychiatric disorders, 20.3% had pulmonary disorders, 16.5% had metabolic or other disorders, 11.4% had neurologic disorders, 7% had cardiac disorders, and 7% had neurodevelopmental disorders.

The AI was trained using a retrospective database of about 2.5 million diagnostic sleep studies and the well-validated Munsell Color System, which assesses both color value, which evaluates the darkness or lightness of skin tone, and color chroma, which evaluates undertones in the skin.

Fernandez pointed to the low cost of the Munsell Color System, as well as its user-friendly design and the fact that it can be easily incorporated into type IV home sleep apnea testing validation design.

PPG signals were collected through wearable devices and during in-lab PSG study recordings, which were scored by three sleep technologists and reviewed by board-certified sleep doctors to ensure quality. PSG scores were then compared with the PPG-based AI system measures.

  • author['full_name']

    Elizabeth Short is a staff writer for Ƶ. She often covers pulmonology and allergy & immunology.

Primary Source

SLEEP

Fernandez C, et al "Prospective clinical validation of AI for PPG-based OSA detection utilizing standardized skin color assessments" SLEEP 2024; Abstract 1352.