At the heart of the exchange is a question that has long troubled sleep researchers: how do you study mortality in a disease as rare as narcolepsy without waiting decades for a prospective cohort to mature? Narcolepsy affects roughly one in two thousand people, and its two main forms, narcolepsy type 1 with cataplexy and narcolepsy type 2 without cataplexy, are distinguished by features that require specialized testing to confirm. The authors argue that for rare conditions, electronic health record based cohort studies remain indispensable precisely because the prospective designs their critics favor are seldom feasible at the scale or duration needed to detect mortality signals. Their study leveraged the Veterans Health Administration’s longitudinal records, which follow patients across decades of care, offering a follow-up window of 25 years that no single-center prospective study could realistically match.

One of the sharpest criticisms concerned bias. Mehri and Finsterer invoked limitations commonly associated with retrospective designs, including recall bias, selection bias, and confounding. The authors’ reply draws a technical distinction here that is worth unpacking. Recall bias, the distortion that arises when patients or investigators remember exposures imperfectly, is a meaningful threat in studies that rely on interviews or chart reviews conducted after the fact. It is not, however, a meaningful threat in administrative electronic health record cohorts, where exposures and outcomes are recorded prospectively at the point of care, independent of anyone’s memory. A diagnosis code entered during a clinic visit and a death recorded in a vital status file are not subject to the fallibilities of recollection. Selection bias and confounding, by contrast, the authors concede, remain legitimate concerns in any observational dataset, and they describe the specific countermeasures they deployed.

Those countermeasures centered on propensity score matching, a statistical technique widely used in non-randomized research to approximate the balance achieved in randomized trials. In the narcolepsy study, patients with narcolepsy were matched to comparator patients from general sleep clinics on age, sex, race and ethnicity, and index diagnosis year, with additional adjustment for body mass index and the Charlson Comorbidity Index, a validated composite measure of baseline disease burden. The authors also align themselves with earlier correspondents, Talari and Goyal, who noted that retrospective designs cannot establish causality. Their own manuscript, they emphasize, explicitly frames the findings as associations rather than causal effects. This is a crucial framing for readers: the study does not claim that narcolepsy itself causes death, but that in real-world clinical practice, patients diagnosed and managed for narcolepsy died at higher rates than comparable sleep clinic patients, even when those comparators carried a substantially heavier comorbidity burden at baseline.

The reliability of the mortality outcome itself became another point of contention, and here the authors marshal quantitative evidence. Vital status data in the Veterans Health Administration have been independently validated against the National Death Index, the United States’ authoritative repository of death records. That validation work, published by Sohn and colleagues, found a sensitivity of 98.3 percent and exact-date agreement of 97.6 percent, figures that support the accuracy and completeness of the VHA’s mortality ascertainment. In practical terms, this means the study’s primary endpoint, all-cause death, was measured against a benchmark that few national datasets can match. The authors argue that these psychometric properties of the data source underpin the reliability of their central finding, whatever reservations may attach to the diagnostic classification of the exposure.

That diagnostic classification was the critics’ most technically substantive objection. Definitive diagnosis of narcolepsy type 1 under the International Classification of Sleep Disorders, third edition, requires objective confirmation: polysomnography followed by a multiple sleep latency test demonstrating shortened sleep latency and early REM onset, or cerebrospinal fluid measurement revealing low hypocretin-1, the neuropeptide whose loss underlies the disorder’s pathophysiology. The critics asked how many participants had cataplexy and whether all had undergone these tests. The authors’ answer is a candid acknowledgment of the limits of administrative data: polysomnography and multiple sleep latency test results are not captured in a standardized format across the national VHA dataset, and cerebrospinal fluid hypocretin testing is rarely performed in routine Veterans Affairs practice. These data simply cannot be ascertained reliably from administrative records, a constraint that shaped the study’s entire design philosophy.

Rather than attempting an unattainable gold-standard confirmation, the researchers employed a computable phenotyping algorithm, requiring at least two narcolepsy diagnosis codes separated by 30 to 390 days. This temporal separation requirement is designed to improve diagnostic specificity by filtering out random or provisional miscoding, since a genuine narcolepsy diagnosis tends to persist across multiple encounters while a coding error is less likely to be repeated weeks or months later. The authors note that this approach aligns with prior large electronic health record based studies of narcolepsy. They also push back on the implicit assumption that the diagnostic gold standard is itself flawless, citing work by Trotti and colleagues showing that approximately half of patients with noncataplectic central hypersomnia received a different diagnosis when the multiple sleep latency test was repeated. Even in tertiary referral settings, the boundary between narcolepsy type 2 and idiopathic hypersomnia is unstable, which complicates any demand for perfect phenotypic precision.

The critics also raised the specter of confounding by conditions and medications that mimic or modify hypersomnolence: endocrine disorders such as hypothyroidism, and drug classes including antiepileptics, analgesics, antihistamines, and immunosuppressants. The authors offer a reasoned rebuttal on both fronts. Hypothyroidism and similar conditions, they argue, would more plausibly be misclassified within the general sleep clinic comparator group than within the narcolepsy cohort, where the two-code requirement reduces the likelihood of diagnostic confusion. As for the medication classes, none has a well-established direct effect on the mortality pathways most likely operative in narcolepsy, namely cardiovascular events and accident-related deaths, and adding them would have introduced collinearity without addressing the principal causal hypotheses. The study instead focused on comorbidity groups with established relevance to mortality risk in narcolepsy: cardiovascular, metabolic, pulmonary, neurologic, and psychiatric conditions. The authors caution that expanding the model to every conceivable confounder would reduce interpretability, destabilize the statistical models, and exceed the constraints of available structured data.

The exchange is a window into a broader tension in modern epidemiology. Big data cohorts built from electronic health records offer scale, duration, and real-world generalizability that carefully phenotyped prospective studies cannot match, but they sacrifice diagnostic granularity in return. The authors of the reply do not dispute that prospectively defined cohorts are essential for answering mechanistic questions and refining diagnostic precision; they simply note that such cohorts are logistically difficult to obtain for rare disorders and cannot provide the scale or quarter-century of follow-up available in the VHA population. Their study, they reiterate, was never designed to establish causality or to revalidate diagnostic criteria, but to assess whether narcolepsy, as diagnosed and managed in ordinary clinical practice, is associated with increased mortality and healthcare utilization. On that question, they argue, the answer stands: the elevated mortality persisted even against a comparator group with a higher baseline comorbidity burden, reinforcing the clinical relevance of the finding and the case that real-world evidence, carefully constructed, deserves a seat at the table in rare-disease research.

Subject of Research: Mortality risk and study methodology in narcolepsy research using large electronic health record cohorts

Article Title: Reply to “The question of whether narcolepsy increases mortality should be answered based on reliable studies”

Article References: Sharafkhaneh, A., BaHammam, A. S., Dauvilliers, Y., Azarian, M., Thorpy, M., Han, F., Senel, G. B., Aksu, M., & Razjouyan, J. (2026). Reply to “The question of whether narcolepsy increases mortality should be answered based on reliable studies”. Journal of Clinical Sleep Medicine, 22(1), Article 110. https://doi.org/10.1007/s44470-026-00121-8

Image Credits: AI Generated

DOI: 10.1007/s44470-026-00121-8

Keywords: narcolepsy, mortality, electronic health records, retrospective cohort study, propensity score matching, Veterans Health Administration, sleep medicine, computable phenotyping, polysomnography, multiple sleep latency test, confounding, hypocretin

The original study, published in the Journal of Clinical Sleep Medicine, followed a 25-year propensity-matched cohort of veterans with narcolepsy and compared their outcomes against patients treated at general sleep clinics, most of whom carried diagnoses of obstructive sleep apnea. The findings raised eyebrows because the excess mortality appeared concentrated in patients classified under an “other narcolepsy” category rather than in those with narcolepsy type 1, the classic form of the disease characterized by cataplexy and low cerebrospinal fluid levels of the wakefulness neuropeptide hypocretin. This inversion of expectations prompted Fei and colleagues to ask whether the “other narcolepsy” group was simply a diagnostic dumping ground, a mixture of narcolepsy type 2, unspecified narcolepsy, and misclassified hypersomnolence conditions whose apparent mortality excess might be an artifact of sloppy phenotyping rather than a genuine biological signal.

In their reply, Sharafkhaneh and his coauthors, an international roster that includes Yves Dauvilliers of Montpellier, Ahmed BaHammam of King Saud University, Michael Thorpy of Albert Einstein College of Medicine, Fang Han of Peking University, and Turkish sleep specialists Gulcin Benbir Senel and Murat Aksu, do not dispute that heterogeneity exists. Instead, they argue that it is an unavoidable and explicitly acknowledged feature of real-world electronic health record phenotyping. The researchers intentionally defined the other narcolepsy group as a heterogeneous category, and they point out that their design required at least two diagnostic codes recorded 30 to 390 days apart before a patient entered the cohort. That temporal spacing requirement, they explain, was engineered to improve diagnostic specificity by filtering out transient miscoding and diagnostic uncertainty, the kind of single-visit noise that plagues administrative claims data.

The technical crux of the debate revolves around the multiple sleep latency test, or MSLT, the gold-standard physiological instrument for distinguishing narcolepsy type 1 from narcolepsy type 2 and idiopathic hypersomnia. The test measures how quickly patients fall asleep during scheduled daytime naps and whether they enter rapid eye movement sleep abnormally early, a phenomenon known as sleep-onset REM periods. Fei and colleagues highlighted the well-documented poor test-retest reliability of the MSLT in noncataplectic hypersomnia, citing influential work by Lynn Trotti and colleagues showing that patients can flip between diagnostic categories when the test is repeated. The reply authors acknowledge this limitation candidly but note a practical reality: MSLT data were simply not available in a standardized manner across the national Veterans Health Administration dataset, making physiological confirmation impossible for a cohort of this scale.

Rather than abandon physiological rigor altogether, the investigators constructed their cohort using a validated computable phenotyping algorithm analogous to the approach applied by Mayo Clinic researchers in prior narcolepsy cohort studies. This algorithmic strategy, the authors argue, does not eliminate misclassification, but it meaningfully minimizes the single-code phenotyping errors that arise when researchers rely on any one diagnostic entry in isolation. The distinction matters because the entire interpretive weight of the mortality finding depends on what the other narcolepsy group actually contains. If the group is a genuine mixture of undertreated or atypically presenting hypersomnolence patients, the authors contend, one would expect mortality rates equal to or higher than those seen in narcolepsy type 1, precisely the pattern the data revealed. Crucially, they emphasize, their findings highlight this heterogeneity but do not imply that narcolepsy type 2 itself carries an inherently elevated mortality risk.

Fei and colleagues also raised the possibility that positive airway pressure therapy, widely prescribed for obstructive sleep apnea, might have artificially lowered mortality in the general sleep clinic comparator group, citing a recent meta-analysis by Benjafield and colleagues suggesting that PAP use is associated with reduced all-cause and cardiovascular mortality. The reply authors treat this as a plausible but ultimately insufficient explanation. Their data showed obstructive sleep apnea prevalence of 65.4 percent in the general sleep clinic cohort compared with only 33 to 37 percent in the two narcolepsy groups, yet differential PAP exposure cannot explain why mortality differed between the other narcolepsy group and narcolepsy type 1, since neither group received PAP at rates approaching the comparator. The confounding argument, in other words, addresses the wrong contrast.

Perhaps the most striking observation in the exchange concerns comorbidities. The general sleep clinic cohort carried a substantially heavier baseline burden of cardiovascular, metabolic, pulmonary, and psychiatric disease than either narcolepsy group, yet the elevated mortality in narcolepsy persisted despite this seemingly protective asymmetry. If comorbidity profiles alone drove the mortality differences, the general sleep clinic patients should have fared worse. The authors are careful to stress that their study cannot identify causal mechanisms; what it delivers is a significant association that they argue warrants deeper investigation in cohorts enriched with physiologic data and treatment-adherence information, ideally including device-level PAP downloads and pharmacy fill trajectories that would allow researchers to model therapeutic exposure directly.

Treatment exposure emerges as a genuinely promising mechanistic thread in the reply. In the real-world VA data, patients with narcolepsy type 1 received sodium oxybate, stimulants, and wake-promoting agents more frequently than the other narcolepsy group, a pattern the authors attribute to greater diagnostic certainty and the established therapeutic indication for cataplexy rather than to disparities in clinical access. Fei and colleagues noted that low-sodium oxybate formulations are associated with a more favorable cardiovascular profile, referencing a 2025 study in Hypertension comparing the blood pressure effects of high- versus low-sodium oxybate in narcolepsy patients. The reply authors agree that differential pharmacologic exposure deserves further research, and they go further, suggesting that unequal treatment exposure may itself represent an important mechanism contributing to poorer outcomes in the heterogeneously composed other narcolepsy group, a hypothesis with direct implications for how clinicians manage patients whose diagnoses remain uncertain.

The broader literature on narcolepsy mortality remains stubbornly inconsistent, and the reply situates the debate within that larger landscape. A recent Taiwanese cohort study by Hsu and colleagues, published in JAMA Network Open, reported no excess all-cause or cause-specific mortality among narcolepsy patients after adjustment for comorbidities, standing in apparent contradiction to the VA findings of elevated mortality relative to a clinically complex sleep-clinic comparator. Reconciling these divergent results, the authors argue, will require harmonized phenotyping standards, cause-specific death ascertainment, and treatment-exposure modeling across diverse health systems, since a diagnosis of narcolepsy may carry different meanings in a national claims database, a single-payer registry, and an integrated veteran health system. The authors express hope that their work, together with the constructive critique from Fei and colleagues, will catalyze such multi-system efforts.

What makes this exchange more than an academic squabble is what it reveals about the epistemology of modern sleep medicine. As ever-larger electronic health record datasets become the primary engine of epidemiological discovery, researchers must navigate a fundamental tension between scale and precision: cohorts large enough to detect mortality signals over decades are rarely cohorts in which every diagnosis has been physiologically confirmed. The VA team’s reply models one way of handling that tension, defending algorithmic phenotyping while transparently cataloging its limits and inviting the very scrutiny it received. For patients with narcolepsy type 2 and related hypersomnolence disorders, the practical message is nuanced but important: the apparent mortality excess in non-type 1 narcolepsy may reflect diagnostic mixture and undertreatment rather than the disease itself, which means that clearer diagnostic pathways and more aggressive therapeutic engagement for diagnostically ambiguous patients could be among the most consequential interventions sleep medicine can offer.

Subject of Research: Mortality risk and diagnostic heterogeneity in narcolepsy type 2 and other hypersomnolence disorders using electronic health record phenotyping

Article Title: Reply to “Unexpectedly higher mortality in non‑type 1 narcolepsy: diagnostic heterogeneity or unmeasured confounders?”

Article References: Sharafkhaneh, A., Dauvilliers, Y., BaHammam, A. S., Azarian, M., Thorpy, M., Han, F., Senel, G. B., Aksu, M., & Razjouyan, J. (2026). Reply to “Unexpectedly higher mortality in non‑type 1 narcolepsy: diagnostic heterogeneity or unmeasured confounders?”. Journal of Clinical Sleep Medicine, 22(1), Article 124. https://doi.org/10.1007/s44470-026-00120-9

Image Credits: AI Generated

DOI: 10.1007/s44470-026-00120-9

Keywords: narcolepsy, narcolepsy type 1, narcolepsy type 2, mortality, electronic health records, multiple sleep latency test, sodium oxybate, obstructive sleep apnea, positive airway pressure, hypersomnolence, sleep medicine, Veterans Affairs cohort

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