Today we’re speaking with Ivan Kairatov, a biopharma expert with a keen eye for how technology is reshaping research and development. He’ll be helping us unpack a groundbreaking new study on SleepFM, an AI model that analyzes sleep data to predict future disease, potentially turning a simple overnight study into a powerful crystal ball for our long-term health. We’ll explore how this model learns the hidden language of our bodies during sleep, its startling accuracy in predicting conditions like Parkinson’s disease, and the significant hurdles that remain before such a tool can be used in your local hospital.
How does a foundation model like SleepFM unlock the hidden potential of overnight sleep studies, and what specific physiological patterns is it learning that traditional analysis might miss? Please elaborate on the process.
It’s a fantastic question because it gets to the heart of what makes this approach so revolutionary. For decades, we’ve used polysomnography, or PSG, as the gold standard, but we’ve been looking at it with a very narrow lens, typically focusing on specific events like sleep apnea. What SleepFM does is completely different. By training on a massive dataset—we’re talking about 585,000 hours of sleep recordings—it uses a self-supervised learning approach. This means it’s not told what to look for; instead, it learns the fundamental grammar and syntax of sleep itself. It discovers the incredibly complex, moment-to-moment interplay between brainwaves, eye movements, muscle tone, and heart rhythms. These are subtle, multimodal patterns that a human expert would never be able to consciously track or codify, representing the deep, underlying physiological state of an individual.
SleepFM predicted Parkinson’s disease with high accuracy years in advance. Can you walk us through how the model translates raw sleep signals into a specific risk score for such a complex neurodegenerative disease, and what makes it so effective?
This is one of the most striking findings. The model achieved an area under the curve, or AUROC, of 0.93 for Parkinson’s over a six-year window, which is truly remarkable. The key is that the model isn’t “diagnosing” Parkinson’s from the sleep study. Instead, it’s detecting the very early, subtle physiological signatures that precede clinical diagnosis. The pre-training process teaches the model to create a rich, internal representation of a person’s sleep physiology. When this representation is then fine-tuned on health records, it learns to associate certain patterns with a future diagnosis. For a neurodegenerative disease like Parkinson’s, these could be minute changes in motor control during REM sleep or shifts in autonomic nervous system regulation that are invisible to the naked eye but are captured in the PSG data. The model essentially translates this incredibly complex, high-dimensional data into a single, potent risk score, acting as an early warning system long before symptoms become obvious.
This model outperformed baselines using demographics or raw PSG data, especially for predicting mortality. What does this reveal about pre-trained sleep patterns, and in what clinical scenarios would this added accuracy be most impactful for patient care?
This highlights the profound power of pre-training. A model that just looks at demographics like age and BMI, or even one trained on raw PSG data from scratch for a specific task, is starting with a blank slate. SleepFM, on the other hand, comes to the table with a deep, foundational understanding of what healthy and unhealthy sleep physiology looks like. When predicting something as complex as all-cause mortality, this pre-trained knowledge is a massive advantage. It’s not just looking at obvious risk factors; it’s identifying a kind of systemic frailty or a breakdown in physiological resilience that is written into the language of sleep. The model achieved an AUROC of 0.85 for mortality, compared to 0.78 for the other models. That 7-point jump isn’t just a number; in a clinical setting, it could mean correctly identifying a high-risk patient who looks perfectly healthy on paper, allowing for early intervention, more aggressive lifestyle counseling, or closer monitoring.
The model generalized well to an external dataset for cardiovascular outcomes. What are the key technical and data hurdles to overcome before deploying such a tool across different hospitals with unique patient populations and equipment? Please provide specific examples.
This is the critical “bench-to-bedside” question. While the success on the external SHHS dataset is very encouraging, it’s just the first step. The biggest hurdle is data heterogeneity. One hospital might use a certain brand of PSG machine with specific sensor settings, while another uses something completely different. Patient populations vary wildly; a model trained primarily at a tertiary care center might not perform as well in a community hospital with a different demographic mix. To deploy this safely, you need to conduct extensive, multi-site validation to ensure the model is robust to these variations and isn’t harboring hidden biases. Furthermore, you have to solve the logistical nightmare of data standardization. Every hospital information system is different. Creating a seamless, secure pipeline to extract PSG data, run it through the model, and deliver an interpretable result back to the clinician’s electronic health record is a massive technical challenge that requires deep collaboration between data scientists, IT departments, and clinicians.
A key challenge noted is the model’s “black box” nature. How does this lack of interpretability affect clinical trust, and what steps are being taken to make the model’s reasoning transparent to doctors and patients? Please share some insights.
The “black box” problem is arguably the single greatest barrier to clinical adoption. A physician simply cannot act on a recommendation—especially a serious one about risk for cancer or dementia—without understanding the rationale behind it. Trust is the cornerstone of medicine, and if a doctor has to say, “The computer says you’re at risk, but I can’t tell you why,” that trust is immediately eroded. The model’s learned representations are incredibly complex mathematical constructs, not easily translated into human-readable physiology. The next wave of research in this field is focused squarely on “explainable AI,” or XAI. These are techniques designed to peer inside the black box and highlight which specific features in the sleep recording—perhaps a particular pattern in heart rate variability during a specific sleep stage—drove the model’s prediction. Until we can provide that level of transparency, models like SleepFM will remain powerful research tools rather than standard clinical instruments.
The model’s age estimation was less accurate for older adults. What might this reveal about the physiological signals of aging in sleep, and how could this specific finding be clinically relevant or further investigated for geriatric care?
I find this to be a fascinating result, not as a failure of the model, but as a potential discovery. The model’s mean absolute error was about 7.33 years, but this error was larger in older adults. This suggests that as we age, our sleep physiology doesn’t degrade in a simple, linear fashion. It likely becomes more heterogeneous—more varied from person to person—due to the accumulation of different health conditions, medications, and life experiences. This variability makes it harder for the model to pin down a single “chronological age.” Clinically, this could be incredibly powerful. We could move from thinking about chronological age to “biological sleep age.” For instance, if the model predicts a 75-year-old’s sleep age is 65, that could be a powerful biomarker of healthy aging. Conversely, if their sleep age is 85, it could be an early, non-invasive indicator of underlying frailty, warranting a closer look from their geriatrician.
What is your forecast for the integration of sleep-based AI diagnostics into routine clinical practice over the next decade?
I believe we’re at the beginning of a paradigm shift. Over the next five years, I foresee models like SleepFM being used primarily as advanced risk-stratification tools within specialized sleep and neurology clinics. They will help clinicians identify which patients require more urgent or in-depth follow-up for conditions ranging from heart failure to cognitive impairment. Looking out over the next decade, as the models become more explainable and are validated across diverse populations, I expect them to move into broader clinical practice. The ultimate goal is to integrate these insights not just from clinical PSGs, but eventually from sophisticated at-home wearables. Imagine a future where your primary care physician reviews your “longitudinal sleep health score” alongside your blood pressure and cholesterol, enabling a truly proactive and personalized approach to preventing disease before it ever takes root.
