MambaSleepCVD

A diagnostic sleep study records eight continuous hours of brain, eye and cardiac activity, and then almost all of it is discarded: cardiovascular risk is read off a handful of summary indices such as total sleep time and minutes of slow-wave sleep. Models that keep the raw waveforms hit a different wall: a full night is far too long for quadratic self-attention, so existing systems fuse modalities inside short windows and stitch the night together afterwards with a recurrent head, leaving interactions that unfold over hours structurally invisible.

MambaSleepCVD, developed at the University of Brescia with collaborators in Lugano and Zurich and posted as a preprint in September 2026, keeps the whole recording in play. It is a modular framework in three stages: frozen modality-specific encoders embed each raw stream, a Coupled Mamba block fuses those sequences continuously across the whole night at linear cost, and a multi-label survival head emits a risk score for each of six cardiovascular endpoints.

The encoders are pretrained U-Sleep instances — one per modality, trained to score sleep stages from that single signal, with every Sleep Heart Health Study record excluded from their pretraining cohort — and they are held fixed throughout. That design makes the central claim testable: supervised sleep-staging representations, integrated by a sufficiently expressive fusion engine, match or beat the task-agnostic multimodal self-supervised pretraining used by SleepFM on this prognostic task.

#Key Features

• Frozen staging encoders as physiological experts: Embeddings come from the dense segmentation layer of a sleep-staging U-Sleep instance, before its five-class projection, so representations are task-aligned to sleep architecture rather than to waveform morphology.
• Continuous cross-modal state coupling: Coupled Mamba gives each modality its own selective state-space chain whose update depends on the preceding states of all chains, so cardiac and cortical routes inform one another throughout the night rather than only after per-window summaries are formed.
• Multi-label survival objective: One head predicts cumulative hazard for angina, CVD death, congestive heart failure, coronary heart disease death, myocardial infarction and stroke under a summed Cox proportional-hazards loss, weighted inversely to event frequency.
• Swappable modality branches: Because the encoders are fixed feature extractors, the ECG branch can be replaced — the authors substitute the self-supervised Self-DANA cardiac encoder — with no monolithic backbone to retrain.

#Technical Details

Development used SHHS1 (5,774 subjects after exclusions), retaining the SleepFM test partition of 1,991 recordings as the internal test set. SHHS2, a follow-up visit of the same parent cohort with 2,618 recordings, served as a within-cohort temporal validation set. The montage is two bipolar EEG channels, two EOG channels and one bipolar ECG lead; U-Sleep emits a six-dimensional embedding per second, concatenated across five seconds into one token. In the no-demographics setting on the SHHS1 test set, U-Sleep plus Coupled Mamba reached a macro-average Harrell C-index of 0.73 against 0.72 for a reimplemented SleepFM-plus-LSTM baseline; a variant trained end-to-end from raw signals underperformed both, evidence that pretrained representations, not the fusion engine alone, carry the result. Under temporal shift on SHHS2, adding EEG to ECG raised the C-index by 0.027 over U-Sleep ECG alone. In multivariable Cox models adjusted for age, sex, BMI, blood pressure, medication and comorbidities, risk scores remained associated with CVD death, CHD death and heart failure at hazard ratios of roughly 1.25–1.6 per standard deviation.

#Applications

The framework targets opportunistic risk stratification rather than standalone screening: where a PSG has already been acquired for a sleep complaint, the same recording yields a score that flags patients for closer cardiovascular follow-up. Its modularity suits groups wanting to reuse an existing single-modality expert inside a prognostic pipeline without assembling a new multimodal pretraining corpus. Calibration at a five-year horizon is imperfect, so outputs are meant for ranking and stratification, not individualized event probabilities.

#Impact

A more consequential contribution is negative. Chronological age acts as a dominant statistical shortcut in SHHS: it is inversely correlated with time-to-event, and both the Cox loss and the C-index reward ranking older subjects ahead of younger ones, so a model handed age near the decision layer can score well without extracting any physiological signal — an age-and-gender multilayer perceptron reproduces much of a full pipeline's reported performance. Age also remains partly recoverable from no-demographics embeddings (R² of 0.52 for U-Sleep and 0.53 for SleepFM features), so excluding it reduces explicit confounding without establishing age-independent value. Validation is single-cohort — SHHS2 is a later visit by the same participants, not an independent population — and ECG alone accounts for most of the discriminative signal on the internal test set. The authors conclude that reliable CVD screening will require independent external cohorts and explicit confounder control. The preprint has not been peer reviewed, the code carries a non-commercial licence, and no trained weights have been published.