PROJECT SUMMARY
Obstructive Sleep Apnea (OSA) is a common chronic condition that affects more than 1 billion people worldwide.
Currently, presence of OSA and its severity is assessed using the apnea-hypopnea index (AHI) that measures
the frequency/rate of respiratory events. AHI has demonstrable limitations including lack predictive accuracy and
poor relationship to outcomes. Task forces across the American Academy of Sleep Medicine, Sleep Research
Society and others internationally have made an urgent plea to move the diagnosis and management of sleep
apnea beyond the AHI. In response to this, our group has recently developed a physiology guided AI/ML
approach that utilizes the known OSA pathophysiology across ventilatory, hypoxic, and arousal burdens and
combines them into probability scores that provide risk of short- and long-term adverse outcomes of OSA. We
use fully automate, reliable, and interpretable measures for ventilatory, hypoxic, and arousal burden towards our
pursuit of explainable AI. Our preliminary data from 10,952 subjects across three epidemiological cohorts
suggests that the machine-learned sleep apnea probability of sleepiness (SAPs), which combines
ventilatory/hypoxic/arousal burdens, classifies sleepiness with accuracy of 87%. Further, using the same
approach but with different set of weights that were trained to predict cardiovascular (CVD) and all-cause
mortality, the sleep apnea probability of CVD (SAPcv) predicted CVD and all-cause mortality across 4,784
subjects with an accuracy of 81% and 88% respectively. To be translatable into clinical practice, validation of
AI/ML approaches in a sleep clinic population is needed and is the core objective of our proposal. In Aim 1 we
will validate SAPs and SAPcv as probability scores for risk of short- and long-term outcomes across a sleep
clinical population comprising data from a diverse set of subjects (N=6,393) seen at the Mount Sinai Health
System from the Greater New York Area. In Aim 2, we will assess real-world performance of SAPs across
subjects (N=700) see at Mount Sinai, as a tool in the decision-making pipeline to predict improvements in
sleepiness with 3 months of treatment using any modality (e.g., CPAP, oral, positional etc.). Finally, in Aim 3, we
will use causal random forests to assess conditional average treatment effects, thereby identifying subgroup of
subjects who are likely to benefit from treatment with respect to CVD outcomes. Our proposal will offer crucial
evidence needed to translate metrics that go beyond AHI in assessing severity of OSA into clinical practice and
are thus poised to shift the paradigm in clinical management of OSA.