AI Reads Your Fingertip Pulse to Measure How Sleepy You Really Are

Excessive daytime sleepiness is one of the most common and most dangerous symptoms in medicine, affecting up to 18 percent of the general population and lurking behind countless motor vehicle accidents, workplace injuries, and diminished lives. Yet the gold-standard tests used to measure it objectively—the Multiple Sleep Latency Test (MSLT) and the Maintenance of Wakefulness Test (MWT)—remain locked inside highly specialized sleep laboratories, demanding teams of technicians, elaborate electrode setups, and hours of expert visual scoring. Now a team of researchers from France and Finland reports that a deep learning algorithm can estimate a patient’s objective sleepiness from nothing more than the pulse wave recorded at a fingertip, potentially opening the door to a radically simpler and cheaper way of assessing one of sleep medicine’s central clinical problems.

The study, published in the Annals of Biomedical Engineering, is the first to demonstrate that daytime sleepiness can be objectively assessed using only photoplethysmography, or PPG—the same optical technique found in ordinary pulse oximeters and consumer smartwatches. A PPG sensor shines light into the skin and measures how blood volume pulses with each heartbeat, producing a waveform that carries subtle fingerprints of the autonomic nervous system. As a person drifts toward sleep, heart rate slows, parasympathetic tone rises, and the shape and timing of the pulse wave change in characteristic ways. The researchers hypothesized that a neural network trained to recognize these signatures could detect sleep onset during daytime tests just as reliably as the electroencephalogram-based methods used today.

To build the model, the team turned to the Multimorbidity Apnea Respiratory Failure Sleep (MARS) database maintained at the University Hospital Grenoble Alpes in France. They harvested 2,423 diagnostic overnight polysomnography recordings collected between 2013 and 2023, each of which included a medical-grade fingertip pulse oximeter channel alongside the full complement of brain, eye, and muscle sensors. Sleep experts had manually scored every recording according to American Academy of Sleep Medicine criteria, providing the ground truth labels. After harmonizing the data and filtering the PPG signals—low-pass filtering at 32 hertz, downsampling from 128 to 64 hertz, and applying z-score normalization—2,297 valid overnight recordings were split into training, validation, and test sets for a U-time-based deep learning architecture, a convolutional network well suited to segmenting long biomedical time series.

The crucial test came when this overnight-trained model was applied, without any retraining or adaptation, to daytime recordings: 143 patients who had undergone the MSLT and 127 who had taken the MWT at the Grenoble sleep center. In the MSLT, patients are given four 20-minute opportunities to fall asleep at 9 am, 11 am, 1 pm, and 3 pm; in the MWT, they are instead asked to resist sleep across four 40-minute sessions. The model processed the fingertip pulse wave second by second, producing a continuous probability of sleep, from which an automatic mean sleep latency could be derived and compared against the latencies scored by experienced human technicians using conventional EEG-based rules.

The results were strikingly encouraging for the MSLT. Second-by-second classification of sleep versus wakefulness from the pulse wave alone achieved 81 percent accuracy with a Cohen’s kappa of 0.60, indicating moderate-to-substantial agreement with manual scoring, and the detection of sleep was well balanced between precision (0.71) and recall (0.80). When subjects were classified as objectively sleepy or non-sleepy using the standard clinical threshold of a mean sleep latency below 8 minutes, the fingertip-based method reached 80 percent accuracy, with 65 percent sensitivity and 84 percent specificity. The automatically derived mean sleep latency correlated significantly with the manually scored value (r = 0.61, p < 0.001), and the typical difference between the two methods fell within a few minutes.

Perhaps most compelling were the group-level sleep probability curves. When the researchers averaged the model’s second-by-second sleep probability across all tests for each subject, the sleepy and non-sleepy groups produced visibly distinct trajectories. In the MSLT, the sleepy group—31 subjects across 123 tests—showed a consistently higher and more steeply rising sleep probability than the 112 non-sleepy subjects across 441 tests, with the separation evident from the very beginning of the recording and persisting across most of the 20-minute session. A global permutation test confirmed the difference in profile shape was statistically significant (p = 0.0002). In the MWT, the sleepy group likewise displayed elevated sleep probability, particularly during the middle portion of the 40-minute test, while the non-sleepy group remained low and stable (p = 0.0012).

The MWT told a more cautionary tale. Because participants in the MWT are actively trying to stay awake, a staggering 98 percent of the second-by-second data consists of wakefulness, creating an extreme class imbalance that renders raw accuracy figures misleading. Although overall accuracy reached 88 percent, precision for detecting sleep collapsed to just 0.10, meaning that when the model labeled a moment as sleep, it was usually wrong. Cohen’s kappa fell to a weak 0.15, and the misclassification of objectively sleepy subjects rose to 38 percent. The authors attribute this to a fundamental physiological ambiguity: during quiet wakefulness in the MWT, transient parasympathetic relaxation or reduced movement can mimic the cardiovascular signature of sleep onset in the fingertip pulse wave, fooling a model trained exclusively on overnight physiology.

This limitation points to the study’s most important methodological caveat: the model was trained on nocturnal polysomnography and deployed zero-shot onto daytime recordings, without any domain adaptation. Autonomic physiology during nighttime sleep differs meaningfully from that during daytime naps, and even more so from the tense, motionless wakefulness of a patient instructed to resist sleep. The researchers acknowledge that transfer learning or fine-tuning on daytime-specific data will likely be essential before the approach can be trusted in wake-maintenance testing, and they note that some large outliers—differences of 20 to 40 minutes in estimated sleep latency—appeared in the MWT dataset where manual scoring suggested very long latencies that the pulse-wave model dramatically underestimated.

Even so, the broader implications are considerable. The MSLT is embedded in the diagnostic criteria for narcolepsy types 1 and 2 and idiopathic hypersomnia, and both the MSLT and MWT are used to evaluate residual sleepiness in CPAP-treated obstructive sleep apnea patients, to assess fitness for safety-critical professions such as commercial driving, and to measure the effects of wake-promoting medications. Yet excessive sleepiness persists in 9 to 22 percent of sleep apnea patients even when CPAP therapy is well tolerated, and objective testing is rarely performed outside specialized centers because of its cost and labor intensity. Notably, the fingertip-based sleep latency correlated with subjective sleepiness on the Epworth Sleepiness Scale to a degree statistically indistinguishable from the manually scored latency, suggesting the simple signal captures clinically meaningful information about the patient’s experienced state.

The authors are careful to frame this as a first step rather than a finished clinical tool. The analysis was retrospective, the model has not been shown to track sleep onset in real time, and a pulse oximeter-only test conducted without concurrent brain monitoring may not be feasible under current MWT protocols, in which patients must be prevented from sleeping between sessions. Future work, they write, should optimize the MSLT and MWT protocols for pulse-wave analysis, develop daytime-adapted training strategies, and validate the approach prospectively across diverse patient groups. Still, the vision is clear: a cheap, passive, wrist-worn or fingertip-based sensor paired with a neural network could one day bring objective sleepiness assessment out of the sleep laboratory and into routine clinical practice—transforming how narcolepsy is diagnosed, how tired drivers are evaluated, and how millions of chronically sleepy patients are monitored over the course of their treatment.

Subject of Research: Deep learning assessment of daytime sleepiness from fingertip photoplethysmography during MSLT and MWT testing

Article Title: Deep Learning-Based Assessment of Sleepiness from Fingertip Pulse Wave Analysis During Objective Daytime Testing

Article References: Rusanen, M., Kainulainen, S., Myllymaa, S., Leppänen, T., Tamisier, R., Baillieul, S., Bailly, S., & Pepin, J.-L. (2026). Deep Learning-Based Assessment of Sleepiness from Fingertip Pulse Wave Analysis During Objective Daytime Testing. Annals of Biomedical Engineering. https://doi.org/10.1007/s10439-026-04393-2

Image Credits: AI Generated

DOI: 10.1007/s10439-026-04393-2

Keywords: deep learning, photoplethysmography, daytime sleepiness, Multiple Sleep Latency Test, Maintenance of Wakefulness Test, sleep medicine, sleep apnea, narcolepsy, pulse wave analysis, sleep staging, biomedical engineering, Epworth Sleepiness Scale