Learning transferable human physiology from two million hours of sleep with SleepFM-2

Motivation

• SleepFM-2 tests whether representations capture information beyond conventional sleep summaries, subjective reports, common health dimensions, and site-specific structure.These questions arise because prior models primarily predicted disease or recording-derived measures, while subjective experience and cross-health-system generalization remained uncertain.Source: Introduction · p. 2

Clinical sleep assessment

• SleepFM-2 extends PSG assessment beyond staging to cortical arousals, limb movements and respiratory events at finer temporal resolution than SleepFM.These events determine clinically used arousal, periodic limb movement and apnea–hypopnea indices, whereas SleepFM used five-second tokens that were too coarse.Source: Introduction · p. 2

Cross-sensor transfer

• SleepFM-2 evaluates whether comprehensive PSG representations transfer to headband and in-ear EEG, wrist PPG, and wrist accelerometry beyond the sleep laboratory.The accelerometry bridge uses respiratory and cardiac surrogates derived from acceleration plus raw axes; disease prediction in UK Biobank is comparable to direct accelerometry pretraining.Source: Introduction · p. 2

Contribution

• SleepFM-2 learns multimodal sleep physiology transferable across clinical outcomes, tasks and sensing modalities from over two million hours of PSG.The frozen encoder is evaluated for sleep staging, event detection, incident disease prediction, subjective sleep endpoints, and transfer beyond the sleep laboratory.Source: Introduction · pp. 2–3

Evaluation scope

• The evaluation includes incident disease prediction in two health systems, one unseen during pretraining, alongside questionnaire-derived subjective sleep and symptom endpoints.All results are benchmarked against interpretable measurements computed from the same input signals.Source: Introduction · p. 2

Model and pretraining

• SleepFM-2 is pretrained on 235,865 clinical PSG recordings spanning more than two million hours, with four modality streams and one-second channel-specific tokens.A shared transformer jointly attends across visible channel-by-time grids using contrastive and masked-autoencoding objectives, after which the encoder is frozen for downstream task heads.Source: Introduction · p. 2, Figure 1 · p. 3

Overview

• SleepFM-2 was pretrained on 235,865 recordings from 24 of 26 PSG cohorts and evaluated on held-out cohorts and sensor families unseen during pretraining.The Harvard HSP, SHHS, UK Biobank accelerometry substudy and wearable datasets support evaluation across an unseen health system, population and sensor family.Source: SleepFM-2 Overview · pp. 3–4
• A single frozen encoder supports one-second sleep staging, cortical-arousal, limb-movement and respiratory-event scoring, phenome-wide disease prediction, subjective-sleep characterization and transfer across clinical and wearable sensors.Datasets are matched to tasks according to available disease histories, expert annotations, questionnaires and non-pretraining sensor recordings.Source: SleepFM-2 Overview · p. 4

Architecture and training

• One-second tokens capture shorter events, retained per-(channel, patch) tokens provide dense and subject-level embeddings, and masked autoencoding complements leave-one-out contrastive learning.The joint objective favors local tasks such as staging and global tasks such as disease prediction, resolving their local–global trade-off.Source: SleepFM-2 Overview · p. 4

Ablation

• In a fixed-corpus component chain, mean disease Harrell’s C-index rose from 0.752 to 0.799, sleep-staging F1macro from 0.748 to 0.771, and age-estimation R2 from 0.586 to 0.796, while parameters fell from 4.83M to 2.57M.These endpoints compare reconstructed SleepFM-1 and SleepFM-2 variants rather than the released model, which uses the full corpus and one-second tokens.Source: SleepFM-2 Overview · p. 4

SleepFM-2 predicts incident disease across the phenome

• SleepFM-2 identified 215 phenotypes meeting the prespecified C-index and significance criterion in both cohorts, with 155 retaining significant PSG-specific information beyond demographics.The highlighted diseases spanned cardiovascular, respiratory, renal, metabolic and neurological conditions across two independent health systems.Source: SleepFM-2 predicts incident disease… · p. 5

SleepFM-2 predictions exceed interpretable baselines

• 0.723/0.711 C-index, versus 0.715/0.689 for 480 interpretable features plus demographics, with SleepFM-2 ahead for 67%/86% of conditions across SSC/HSP.The representation remained stronger at the unseen health system despite the feature bank’s larger cross-cohort decline.Source: SleepFM-2 predictions exceed… · p. 6, Figure 2 · p. 7

SleepFM-2 improves upon SleepFM-1 for disease prediction and sleep scoring

• Mean C-index increased from 0.709 to 0.718 in SSC and from 0.696 to 0.729 in HSP versus SleepFM-1, with the larger gain at the unseen site.SleepFM-2 also improved sleep-staging macro F1 across SSC, MESA, MrOS and SHHS, with additional strong performance in CFS, WSC and HSP.Source: SleepFM-2 improves upon SleepFM-1 for… · pp. 6, 8, Figure 2 · p. 7

SleepFM-2 performs within the range of expert scorers on sleep event annotation

• SleepFM-2 performed within the range of expert scorers, with mean F1 of 0.601 versus 0.546 for arousals, 0.604 versus 0.530 for limb movements and 0.462 versus 0.404 for respiratory events.One-second tokenization enabled temporal localization of events with varying durations.Source: Figure 2 · p. 7, SleepFM-2 performs within the range of… · p. 8

SleepFM-2 predictions organize around a shared axis of physiological risk

• The general-risk axis explained 57.3%/58.2% of risk-score variance in SSC/HSP, associated with reduced sigma-band spatial coupling and increased hypnodensity entropy.It added +0.026/+0.025 C-index to demographics, while condition-specific scores improved 71%/75% of conditions beyond the shared axis.Source: SleepFM-2 predictions organize around… · pp. 8–9

SleepFM-2 predicts disease best from multimodal input

• The multimodal pass improved C-index over demographics by +0.0240 in SSC and +0.0423 in HSP, while no single modality surpassed it.Early fusion outperformed late fusion, whose averaging recovered only 63%/38% of the multimodal gain in SSC/HSP.Source: SleepFM-2 predicts disease best from… · pp. 9–10

Transfer to wrist accelerometry

• Mean C-index was 0.687 for SleepFM-2 versus 0.688 for accelerometry foundation models in UK Biobank, while combining both reached 0.6913 versus 0.6882 for the accelerometry models.The matched test cohort comprised 5,192 participants; the paired SleepFM-2 difference was −0.0015 (95% CI −0.0067 to 0.0033), whereas pooling improved the paired difference by +0.0028 (95% CI +0.0003 to +0.0053).Source: Figure 4 · p. 11, Transfer to wrist accelerometry · pp. 12–13
• Across six accelerometry cohorts, five-class AUROC reached 0.708–0.852 versus 0.590–0.690 for supervised-from-scratch training, improving by 0.106–0.221 on every cohort.The frozen encoder used an LSTM head and surrogate physiological signals derived from raw acceleration, despite accelerometry sharing no PSG channels.Source: Figure 4 · p. 11, Transfer to wrist accelerometry · p. 12

Transfer to wakeful EEG and brain–computer interfaces

• SleepFM-2 exceeded the supervised baseline on all 23 wakeful-EEG datasets, with a mean AUROC gain of 0.163, including a 0.205 gain across clinical and pathological detection tasks.Gains averaged 0.148 for cognitive, affective and phenotyping tasks and 0.062 for motor-imagery BCI; AUROCs reached 0.978 on CHB-MIT, 0.945 on BrainCapture and 0.914 on TUH-Events.Source: Figure 4 · p. 11, Transfer to wakeful EEG and… · p. 12

Transfer to wearable headband and in-ear EEG and PPG

• Macro F1 improved in all four wearable settings: 0.514 to 0.721 on BOAS headband EEG, 0.463 to 0.645 on Wearanize+ headband EEG, 0.239 to 0.746 on EESM19 ear EEG, and 0.175 to 0.531 on Wearanize+ wrist PPG.Transfer therefore extended across reduced-montage headband EEG, anatomically shifted ear EEG and wrist PPG, although PPG had lower absolute staging performance than wearable EEG.Source: Figure 4 · p. 11, Transfer to wearable headband and… · p. 12

Evaluation design

• The study evaluated 25 prespecified AUROC endpoints using frozen SleepFM-2 embeddings against demographics, macro-sleep reports, YASA microstructure features, and their Combined reference.Endpoints covered post-sleep and habitual questionnaires, comorbidities, symptoms, subjective–objective discrepancies, and composite poor-sleep measures.Source: SleepFM-2 predicts poor subjective… · p. 13

Symptoms and comorbidities

• SleepFM-2 was best for all four comorbidities and symptom endpoints, with mean AUROCs of 0.759 versus 0.750 and 0.678 versus 0.657, respectively.Depression improved significantly, as did insomnia, parasomnia, and somatic symptoms; performance was comparable to Combined for fatigue.Source: SleepFM-2 predicts poor subjective… · p. 14

Subjective–objective discrepancy

• SleepFM-2 predicted subjective–objective discrepancy with AUROCs of 0.646 for |∆TST| and 0.748 for |∆SOL|, consistently exceeding Combined, macro-report, and microstructure baselines.Improvements over Combined were +0.024 for |∆TST| and +0.020 for |∆SOL|; the raw signal therefore contained discrepancy information absent from tabular summaries.Source: SleepFM-2 predicts poor subjective… · p. 14

Composite poor-sleep endpoints

• For composite endpoints, SleepFM-2 achieved AUROC 0.932 versus 0.912 for the non-leaking Objective baseline and 0.774 versus 0.745 for Last-night.The paired improvements were +0.020 for Objective and +0.028 for Last-night, whereas Habitual performance was essentially identical to Combined: 0.680 versus 0.677.Source: SleepFM-2 predicts poor subjective… · p. 14

Interface workflow

• SleepFM-Interface maps uploaded PSG channels to the encoder vocabulary, generates frozen-encoder embeddings, and returns task-specific outputs without programming.Users can inspect and override each processing step, request explanations, or draft a report.Source: An interactive web interface for… · p. 14
• Four tabs—summary, details, recording, and report—support review, with underlying measurements available for verification and drafted reports available for inspection.The summary and recording views are presented as the tabs physicians typically open first.Source: An interactive web interface for… · p. 14

Limitations

• The interface demonstrates a research workflow on a public de-identified CAP Sleep Database sample and is not validated for clinical use; prospective evaluation remains future work.The sample was not part of the SleepFM-2 pretraining corpus, and the screenshots are demonstrations rather than clinical readouts.Source: An interactive web interface for… · pp. 14, 16

Overall contribution

• SleepFM-2 transfers information from multimodal PSG to clinical tasks and wearable sensors, including wrist accelerometry with no shared input channel, while performing within expert-scoring ranges.The findings support a common physiological representation underlying capabilities across clinical and wearable settings.Source: Discussion · pp. 16, 18

Disease prediction

• Disease predictions exceed 480 interpretable same-recording features, with the advantage clearest in the unseen health system, but retrospective associations do not establish causality or clinical utility.Joint multimodal processing suggests relevant information is distributed across physiological systems rather than captured by a single biomarker.Source: Discussion · pp. 16–17

General physiological risk

• A single component explains approximately 57 to 58% of variance in disease-specific predictions and associates with sigma-band spatial coupling and hypnodensity entropy, suggesting a possible general physiological-risk axis.Its meaning should not yet be equated with a specific construct; future work must validate incremental prognostic value, stability across nights, and clinical actionability.Source: Discussion · pp. 16, 18

Subjective sleep

• SleepFM-2 best captures the recorded night, outperforming the clinical and engineered-feature baseline on all six post-sleep questionnaire items, whereas habitual-sleep performance is less consistent.Single-night PSG should not replace longitudinal recording or clinical assessment, particularly for insomnia and other chronic constructs.Source: Discussion · pp. 16–17

Automated PSG interpretation

• SleepFM-2 performs within expert-scoring ranges for cortical arousals, limb movements, and respiratory events, supporting human-in-the-loop automated PSG interpretation rather than autonomous clinical replacement.Uncertainty-guided review and expert override can combine automated annotations with direct signal inspection.Source: Discussion · p. 17

Wearable transfer

• A frozen PSG-trained encoder stages headband, in-ear EEG, and wrist PPG, and predicts disease from wrist accelerometry with performance comparable to directly accelerometry-pretrained models on UK Biobank.Transfer likely reflects channel-wise tokenization and the encoder’s use of physiological information embedded in peripheral movement, but it is not necessarily automatic across new modalities.Source: Discussion · p. 17

Limitations and validation

• Disease analyses are retrospective, primary cohorts are PSG-referred, subgroup performance is incompletely assessed, and wearable transfer uses one accelerometry cohort plus research-grade rather than consumer devices.Prospective studies should evaluate calibration, absolute risk, decision utility, diverse populations, and real-world sensors before establishing clinical benefit.Source: Discussion · p. 17

Future directions

• Paired PSG and wearable recordings could enable one model to learn rich clinical physiology alongside free-living sensing conditions and test calibration beyond populations that typically undergo PSG.This direction would assess whether clinic-learned representations support physiological-risk prediction where sleep actually occurs.Source: Discussion · p. 18

SleepFM-2 architecture

• SleepFM-2 jointly encodes brain/eye, cardiac, muscle and respiratory PSG streams, using leave-one-out cross-modal contrastive learning and masked autoencoding on individual channel-patch tokens.The total loss combines contrastive and reconstruction terms with equal weights, λCL = λMAE = 1.Source: SleepFM-2 architecture · pp. 26–27
• Each modality is represented as a channel-by-time token grid, with fixed channel vocabulary alignment, PAD tracking, shared CNN tokenization, and additive 28-way channel-region embeddings.The channel-region lookup supplies anatomical identity while allowing transfer to unseen channel sets without channel-specific weights.Source: SleepFM-2 architecture · p. 26
• Pretraining masks a fraction r of non-PAD tokens using temporal-block, channel-block and random strategies, while excluding PAD positions from masking and loss computation.The strategies encourage recovery from spatial context, neighboring channels and short temporal context while matching the target mask ratio per sample.Source: SleepFM-2 architecture · p. 26
• Stage 1 jointly attends across channels and temporal patches, preserving per-channel, per-patch representations required for masked reconstruction.It uses pre-norm transformer blocks with RMSNorm, SwiGLU, multi-head attention and RoPE over temporal patch indices.Source: SleepFM-2 architecture · p. 26

Inference

• At inference, masking and the MAE branch are removed, while the encoder exposes per-token, per-patch and per-modality embeddings for tasks at corresponding temporal and spatial granularities.Sleep staging uses patch sequences, summary prediction uses modality summaries, and fine-grained event detection can use per-token embeddings.Source: SleepFM-2 architecture · pp. 26–28

SleepFM-2 architecture

• The contrastive branch pools channel-aware Stage 1 outputs spatially and then temporally to produce one L2-normalized summary embedding per modality for cross-modal InfoNCE.Its leave-one-out targets average the other present modalities, and missing modalities are excluded from the relevant positive targets.Source: SleepFM-2 architecture · p. 27
• The parallel MAE decoder reconstructs masked raw-signal patches, re-adds channel identity, suppresses PAD positions, and is discarded after pretraining.Reconstruction uses an L1 loss only at masked, non-PAD positions, with modality terms averaged equally regardless of channel count.Source: SleepFM-2 architecture · p. 27

Optimization

• Optimization uses AdamW with a 1.2×10−3 peak learning rate, cosine decay to 10−6, 5% warmup, weight decay 0.05, gradient clipping 1.0, batch size 64 and mixed precision.A sample contributes to contrastive learning only when at least two modalities are present; pretraining runs on one NVIDIA H100 GPU.Source: SleepFM-2 architecture · p. 27

Ablation protocol

• All ablation-chain models used the same 48,093 recordings and five-second tokenization, with downstream heads and learning rates fixed by task.The chain comprised four cumulative architectural steps followed by fixed-architecture objective comparisons; encoder parameter counts excluded the MAE decoder.Source: Architecture ablations · p. 28

Optimizer

• AdamW improved age R2 by +0.06 and mean disease C-index by +0.02, while mean staging F1macro remained roughly unchanged.This optimizer swap kept the original architecture unchanged and served as baseline modernization.Source: Architecture ablations · p. 28

Channel-region position embedding

• Adding a 28-way channel-region position embedding left disease essentially unchanged, reduced age R2 by 0.05, and left staging essentially unchanged.Because channel pooling precedes the encoder in the earlier architecture, individual channel identity is unavailable for the embedding to use.Source: Architecture ablations · p. 28

Modern transformer backbone

• Despite reducing encoder parameters from 4.83M to 2.56M, the modern backbone improved mean disease C-index by +0.01, age R2 by +0.08, and staging F1macro by +0.01.The six-layer LLaMA-style design uses two-stage cross-channel attention, RMSNorm, SwiGLU, RoPE, and bias-free projections.Source: Architecture ablations · p. 28

Pretraining objectives

• 0.771 staging matched reconstruction-only training, while joint objectives added 0.014 disease C-index and 0.121 age R2 over contrastive-only training.Reconstruction alone achieved disease C-index 0.743 and age R2 0.539; contrastive-only training achieved 0.785 and 0.675, but staging 0.764.Source: Architecture ablations · p. 29

Token size

• One-second tokens left disease prediction unchanged at 0.798, marginally improved staging to 0.773, and reduced age R2 by 0.038 versus five-second tokens.The final model adopted one-second tokens because shorter scoring events cannot be represented at five seconds.Source: Architecture ablations · p. 29

• The frozen encoder produces reusable embeddings so each downstream cohort and task trains only a task head, decoupling encoder cost and keeping comparisons consistent.All embedding-based heads consume the same fixed embeddings, with no encoder optimizer state during downstream training.Source: Downstream adaptation framework · p. 29
• Each recording is divided into non-overlapping 5-minute chunks, yielding an (M, S, dproj) tensor with an (M, S) mask for missing modalities and short trailing chunks.Embeddings use per-modality summary outputs from the contrastive branch; SleepFM-1 v1 embeddings are aggregated to the same resolution.Source: Downstream adaptation framework · p. 29
• LinearProbe mean-pools modalities and chunks before one linear layer, whereas LSTMHead uses learned modality attention, a transformer, a bidirectional LSTM, and an output MLP.For sequence labeling, the LSTM operates per timestep rather than aggregating across time.Source: Downstream adaptation framework · p. 29
• Age, sex and BMI can be concatenated with pooled PSG embeddings, while capacity-matched demographics-only baselines isolate the embeddings’ added value without split leakage.Demographics are training-set z-scored and median-imputed before comparison.Source: Downstream adaptation framework · p. 29

Held-out evaluation and significance testing

• SleepFM-2 disease heads are trained separately within each cohort on 939 phenotypes jointly, with demographics-augmented scores used for C-index and PSG-only scores used for significance testing.Evaluation uses disjoint train, validation and test subjects; reported concordance and significance models are computed afterward with complete held-out risk sets.Source: Disease prediction · p. 30

Interpretable feature bank

• SleepFM-2 is compared with a nonlinear demographics-only baseline and a matched 480-feature interpretable bank spanning spectral EEG, microstructure, coupling, cardiac, respiratory and hypnodensity domains.The feature bank contains 480 validated PSG-derived features, while all interpretable arms use the same nonlinear head and identical subject splits.Source: Disease prediction · pp. 30–31

Held-out evaluation and significance testing

• The primary discrimination criterion requires C ≥0.75 and Bonferroni-corrected P < 0.01 in both cohorts, while a nested likelihood-ratio test measures PSG information beyond age, sex and BMI.The significance analyses use fixed held-out risk scores and do not refit SleepFM-2 on test subjects.Source: Disease prediction · p. 30

The shared axis

• A shared disease-risk axis is defined as the leading principal component of held-out phenotype risk scores after residualizing age, sex and BMI, with variance explained distinguished from component-score correlation.The axis is estimated on the held-out score matrix, and cross-cohort loading correlation confidence intervals are bootstrapped over phenotypes.Source: Disease prediction · p. 31

Representation interpretation

• Frozen-embedding probes recover interpretable PSG features with cross-validated ridge regression, while nested cross-fitted Cox models quantify condition-specific information beyond the shared axis.Feature probes use complete multimodal subjects and out-of-fold R2; condition-specific concordance differences are computed from pooled out-of-fold predictors.Source: Disease prediction · p. 31

Modality ablation

• Modality ablations use one channel per modality in separate frozen-encoder passes, preserving canonical partitions and reporting gains over demographics alongside post-hoc score fusion.The multimodal analysis includes twelve modalities in SSC and eleven in HSP, but its restricted complete-case cohorts make absolute C-indices incomparable with the preceding disease analysis.Source: Disease prediction · pp. 31–32

Sleep scoring

• SleepFM-2 predicts five-class staging, binary cortical arousal, binary limb movement, and four-class respiratory events at every valid one-second timestep.LSTM heads operate on frozen per-token embeddings, while ambiguous or unannotated timesteps are excluded with a valid-label mask.Source: Sleep scoring · p. 32
• Respiratory-event scoring uses cross-attention to pretrained arousal and staging outputs, while all heads consume per-timestep embeddings rather than five-minute summaries.The auxiliary predictions provide additional input without gradient flow through the pretrained models.Source: Sleep scoring · p. 32
• Training uses masked, class-weighted cross-entropy with event-centered sampling for rare events, whereas staging uses representative non-overlapping full-night windows.Evaluation reports 30-second epoch-mode staging, per-second apnea metrics, and per-event arousal and limb-movement metrics.Source: Sleep scoring · p. 32
• Model predictions are compared with expert scorers using identical leave-one-out pseudo-consensus ground truth across WSC arousal and limb-movement recordings and DREEM respiratory recordings.WSC uses event-level F1 after temporal matching, while DREEM reports macro-averaged per-second F1 across obstructive apnea, central apnea, and hypopnea; paired one-sided Wilcoxon tests determine significant differences.Source: Sleep scoring · p. 32

Analysis scope

• These analyses read out the frozen pretrained encoder on SSC recordings seen during pretraining, characterizing encoded content rather than measuring transfer.No encoder weights were updated, and the analyses used seven encoder read-out depths.Source: Representation analysis · p. 33

Feature taxonomy and concept entanglement

• Sparse-autoencoder analyses classified representation features by coverage, alignment, silence, separability, and domain-level entanglement across expansion ratios and encoder depths.The domains included sleep stage, microstructure, demographics, pathology, and disease; concept-direction analyses quantified pairwise entanglement by cosine similarity and sign.Source: Representation analysis · p. 33

Representation-level characterization

• AUROC reached 0.94 for five-stage sleep classification and 0.92 for cortical-arousal detection, while spindles, slow waves, and eye movements ranged from 0.90 to 0.92.Separability increased with encoder depth, and all probes used frozen per-token embeddings with subject-level train/test separation.Source: Representation analysis · p. 33

Sleep-geometry atlas

• The sleep-geometry atlas arranged stages along a continuous Wake-to-N3 axis with REM as a separate mode, while microstructure events occupied expected clinical regions.Spindles and slow waves mapped over N2/N3, eye movements over REM, and arousals over lighter transitional stages.Source: Representation analysis · p. 33

Wakeful EEG and BCI transfer

• SleepFM-2 transfers across 23 wakeful-EEG datasets after fixed-window preprocessing, signal scaling, anatomical channel mapping, and masking of unmatched channels.Signals are sampled at 128 Hz, clamped to ±5 × 10−5 V, divided by 10−5, and mapped to the pretrained channel-region vocabulary.Source: Wakeful EEG and BCI transfer · p. 34
• Read-out depth is fixed by dataset group: Stage-2 spatially pooled embeddings serve clinical, pathological, cognitive, affective, and phenotyping tasks, while flattened input tokens preserve layout for motor-imagery BCI.This choice is made before comparison and is not adjusted per model.Source: Wakeful EEG and BCI transfer · p. 34
• The evaluation compares a frozen SleepFM-2 encoder with L2-regularized linear probing against a randomly initialized architecture trained end to end, using subject-grouped five-fold AUROC.Performance is reported as mean ± s.d. across test folds; because metrics are computed over windows, they represent window-level rather than subject-level discrimination.Source: Wakeful EEG and BCI transfer · p. 34

Wearable EEG and PPG sleep-staging transfer

• Wearable transfer used harmonized five-class PSG stages from BOAS, Wearanize+ and EESM19, with participant-level five-fold splits and standardized 128-Hz signals.Invalid, movement, unknown and unscored epochs were excluded; no band-pass filtering, re-referencing or PPG feature extraction was applied.Source: Wearable EEG and PPG sleep-staging… · p. 34
• SleepFM-2 mapped headband, ear-EEG and infrared PPG channels anatomically into its pretrained vocabulary, masking unavailable channels without reconstructing bipolar montages.BOAS and Wearanize+ headband channels mapped to frontal EEG, EESM19 ear channels to temporal EEG, and Wearanize+ PPG to the PPG region.Source: Wearable EEG and PPG sleep-staging… · p. 34
• Frozen SleepFM-2 embeddings were mean-pooled into 30-second representations and classified by a two-layer bidirectional LSTM using 100-epoch contexts.The encoder processed 300-second windows as one-second patches, while the comparator used the same architecture with random end-to-end initialization.Source: Wearable EEG and PPG sleep-staging… · p. 34

Wrist accelerometry transfer

• Wrist accelerometry transfer derived respiratory and cardiac surrogate channels from three axes, routed them through existing PSG respiratory and cardiac channel types, and masked unused channels.UK Biobank additionally routed raw axes through the encoder’s position channel type, producing eight channels instead of five.Source: Wrist accelerometry transfer · p. 35
• For wrist sleep staging, 30 one-second patch embeddings were mean-pooled per scoring epoch and passed to SleepFM-2’s learned contrastive-branch readout.The head used per-channel attention pooling, a positional-encoded transformer layer and a two-layer bidirectional LSTM; the random-initialized comparator used the same architecture.Source: Wrist accelerometry transfer · p. 35
• Wrist sleep staging evaluated five-class 30-second predictions across six cohorts—TBI, DREAMT, STAGES, Newcastle, Amazfit and SleepAccel—using subject-level five-fold cross-validation and macro one-versus-rest AUROC.Approximately 453 subjects remained after quality control, and evaluated wrist participants were disjoint from participants contributing pretraining polysomnography.Source: Wrist accelerometry transfer · p. 35
• UK Biobank disease prediction used frozen 300-second accelerometry embeddings concatenated with age, sex and body mass index in a multi-outcome Cox model.The design followed Dige et al., using 390 outcomes, participant-disjoint train-validation-test splits, three-day windows and averaged participant-level log-hazards.Source: Wrist accelerometry transfer · pp. 35–36
• Disease predictions were evaluated on 5,192 overlapping test participants with Dige et al.’s unmodified scoring code, participant-bootstrap confidence intervals and Benjamini–Hochberg correction across 101 outcomes.Comparisons were paired across models, with 1,000 bootstrap replicates.Source: Wrist accelerometry transfer · p. 36

Cohort and endpoints

• BioSerenity included 74,616 subjects aged 16–100 years, split into 66,704 training, 1,295 validation and 6,617 test subjects.Effective sample size varied by endpoint because analyses included only test subjects carrying that endpoint’s label.Source: BioSerenity endpoints · p. 36

Modeling and comparison

• SleepFM-2 used an LSTMHead on frozen sequence embeddings, while tabular arms used nonlinear DemoOnlyMLP heads on median-imputed, standardized single-subject feature vectors.Both pipelines used 20 epochs, batch size 16, learning rate 10−4, weight decay 10−4 and dropout 0.1; preprocessing used training-split statistics only.Source: BioSerenity endpoints · p. 36
• SleepFM-2 and each baseline were compared with 1,000-iteration paired subject bootstrapping on identical endpoint-specific subject sets.This design evaluated both methods on the same resampled subjects and dropped feature columns lacking training-split observations.Source: BioSerenity endpoints · p. 36

Data availability

• The study used 26 polysomnography cohorts and one wrist-accelerometry cohort for pretraining and evaluation, plus evaluation-only wearable and non-sleep EEG datasets.Recording counts per cohort are provided in Supplementary Table 1, and access differs by cohort.Source: Data availability · p. 36
• Twelve cohorts are available through the National Sleep Research Resource after registration and acceptance of its data use agreement.These include ABC, BestAIR, CCSHS, CHAT, HomePAP, Cleveland Family Study, MESA, MrOS, SOF, Sleep Heart Health Study, STAGES and Wisconsin Sleep Cohort.Source: Data availability · p. 36
• Sleep-EDF Database Expanded and the Haaglanden Medisch Centrum sleep staging database [101] are openly available through PhysioNet.Both are distributed via https://physionet.org.Source: Data availability · p. 36
• The HSP held-out evaluation subsample and Stanford Sleep Cohort are available through Human Sleep Project version 3.0 under credentialing, training, and health-data-use requirements.The constituent sites are de-identified and referred to by identifiers in that repository.Source: Data availability · p. 36
• UK Biobank wrist-accelerometry recordings and linked outcomes were obtained under Application Number 62249 and are available to bona fide researchers through its application process.The study was not permitted to redistribute these data.Source: Data availability · p. 36
• Remaining clinical polysomnography cohorts were contributed by collaborating sleep centers and are available on reasonable request, subject to institutional agreements and local ethics approval.These include AHC, CNC, FHC, IHC, KHC, DCSM, CDH, Jazz and RASP.Source: Data availability · p. 37
• BioSerenity recordings, comprising most pretraining data and all subjective-sleep analyses, are proprietary clinical data that cannot be shared.They were accessed under a research agreement between BioSerenity and Stanford University; evaluation-only datasets were obtained from their original publications under described access terms.Source: Data availability · p. 37

Code availability

• Code for SleepFM-2 is available at https://github.com/zou-group/sleepfm-v2-public.Source: Code availability · p. 37

Cohort composition and splits

• SleepFM-2 was pretrained on 235,865 recordings with participant-level train, validation and test splits, while downstream heads remained frozen-encoder evaluations and held-out wearable, wakeful-EEG and accelerometry cohorts were evaluated separately.The pretraining set includes prediction-head training recordings, whereas validation and test recordings were excluded from pretraining; UK Biobank used wrist actigraphy rather than PSG.Source: Table 1 · p. 38, A Supplementary Material · p. 38

Configuration and ablations

• The ablation chain progressively adds architectural changes and compares contrastive, masked-autoencoding, and joint objectives, with the MAE decoder discarded at downstream inference.The shared ablation protocol evaluates frozen encoders for disease prediction, sleep staging, and age estimation without demographic features; the released model uses finer tokenization and a larger corpus than the ablation-chain configuration.Source: Table 4 · p. 40, Table 5 · p. 40

Disease phenotypes

• The phecode analysis defines 155 conditions that achieved Harrell’s C-index of at least 0.75 in both cohorts, Bonferroni-corrected P < 0.01 in both cohorts, and PSG-only likelihood-ratio-test q < 0.05 in both cohorts.Rows are ordered by mean gain over demographics, with confidence intervals obtained from paired subject bootstrap resampling.Source: Table 6 · p. 41

Supplementary conditions

• The supplement also lists additional qualifying conditions across cardiovascular, neurologic, psychiatric, endocrine, respiratory, gastrointestinal and sensory categories.Examples include cerebrovascular disease, Parkinson’s disease, dementia, anxiety disorders, acute bronchospasm, cholelithiasis, tinnitus and iron deficiency anemia.Source: A Supplementary Material · pp. 43, 44, 46, 47

Supplementary Table 7

• 151 phecodes met the requirements for adequate discrimination, demographic-model improvement, and PSG-specific likelihood-ratio significance in both cohorts.Rows are ordered by mean gain across cohorts; the table defines Cdemo, C, ΔC, follow-up case counts, and bootstrap confidence intervals.Source: Table 7 · p. 45

Supplementary conditions

• 0.080 and 0.119 C-index gains were observed for pain, increasing SleepFM-2 from 0.636 to 0.715 and from 0.591 to 0.710 across the two cohorts.The corresponding condition is listed as pain, with muscle symptoms also shown in the same supplementary entry.Source: A Supplementary Material · p. 46
• 0.041 and 0.147 C-index gains were observed for hypoventilation, increasing SleepFM-2 from 0.644 to 0.686 and from 0.552 to 0.699 across cohorts.The same entry lists complex regional/central pain syndrome alongside the hypoventilation results.Source: A Supplementary Material · p. 46
• 0.060 and 0.118 C-index gains were observed for chronic skin ulcer, while hypocalcemia is listed in the same supplementary entry.SleepFM-2 increased C-index from 0.687 to 0.747 and from 0.713 to 0.830 in the two cohorts.Source: A Supplementary Material · p. 46

Disease prediction

• SleepFM-2’s multimodal disease-prediction advantage varies by disease category, while no single physical channel consistently outperforms the multimodal encoder across categories and cohorts.The modality analysis compares individual channels with multimodal passes and also evaluates joint encoding against post hoc score-combination strategies.Source: Figure 1 · p. 49, Figure 2 · p. 57, Figure 3 · p. 58

Scope

• SleepFM-2’s supplementary evaluations span disease prediction, sleep staging, transfer across EEG and wearable sensors, subjective sleep scores, and latent sleep-biology structure.The material includes held-out cohort staging, outside-sleep EEG, wearable EEG/PPG and wrist accelerometry, UK Biobank outcome comparisons, BioSerenity subjective endpoints, and encoder-depth analyses.Source: Table 8 · p. 49, Table 9 · p. 50, Table 11 · p. 51

Sensor transfer

• SleepFM-2 transfers to sleep staging from wearable EEG, PPG and wrist accelerometry, with supplementary tables comparing its frozen encoder against end-to-end random-initialization models.The staging evaluations report Cohen’s κ, macro F1 or window-level AUROC across held-out cohorts, depending on the sensor modality.Source: Table 8 · p. 49, Table 11 · p. 51, Table 12 · p. 52

Disease prediction

• The UK Biobank comparison reports SleepFM-2 and accelerometry models across 101 outcomes, with no individual difference reaching significance after Benjamini–Hochberg correction.Both models include age, sex and BMI; outcomes are ordered by SleepFM-2 Harrell’s C on a matched test cohort of n = 5,192.Source: Table 13 · p. 53

Subjective sleep

• The supplementary subjective-sleep analyses define balanced Somatic, Last-night, Habitual and Objective scores and report per-endpoint SleepFM-2 AUROC and paired differences versus a Combined baseline.The overall score averages the three sub-scores, while the Objective poor-sleep comparison excludes microstructure from the baseline because it would create a definitional leak.Source: Table 14 · p. 55, Table 15 · p. 56

Representation analysis

• Deeper encoder layers sharpen sleep concepts: five-class staging reaches AUROC 0.94, cortical arousal 0.92, spindles, slow waves and eye movements 0.90–0.92, while respiratory apnea rises from 0.70 to 0.80.The representation organizes stages along a continuous Wake-to-N3 depth axis, with REM as a separate mode and event densities located in their clinically corresponding stages or transitional regions.Source: Figure 7 · p. 61, Figure 8 · p. 62

• The supplementary taxonomy found substantial entanglement in narrow dictionaries that decreased as dictionaries widened, whereas most features aligned with no labeled domain across widths.This indicates that narrow-dictionary entanglement reflects dictionary capacity rather than an intrinsic encoder property.Source: Figure 9 · p. 62
• Same-group concepts were more entangled than different-group concepts (mean cosine 0.20 vs 0.04), while disease labels were co-directed (mean cosine 0.55).Expected pairings included slow wave with N3 (+0.90), N2 anti-aligned with wake (−0.70), atrial fibrillation with heart failure (+0.82), and stroke with hypertension (+0.79).Source: A Supplementary Material · p. 63
• The direction analysis is coarser than the main-text phecode analysis because it uses fewer disease-presence labels and measures shared structure directly in the encoder.The disease labels represent presence rather than incidence.Source: A Supplementary Material · p. 63