Physiological signal transformation and biomarker estimation for health assessment
Patent Information
- Application Number
- PCT/US2026/024157
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-09-12
- Filing Date
- 2026-04-17
- Publication Date
- 2026-10-22
Smart Images
Figure US2026024157_22102026_PF_FP_ABST
Abstract
Description
[0001] Attorney Docket No 70009- 127-C- WO
[0002] Client case 26590
[0003] PHYSIOLOGICAL SIGNAL TRANSFORMATION AND BIOMARKER ESTIMATION FOR HEALTH ASSESSMENT
[0004] CROSS-REFERENCES TO RELATED APPLICATIONS
[0005] This application is related to U.S. Provisional Application No. 63 / 790,107 filed April 17, 2025, and U.S. Provisional Application No. 63 / 880,450 filed September 12, 2025, the contents of which are incorporated herein by reference.
[0006] BACKGROUND OF THE INVENTION
[0007] This application relates to generating electroencephalography (EEG) representations and other neural -activity' representations from physiological signals, to determining pharmacological exposure from physiological signals, and to assessing disease conditions through cross-physiological signal transformation, with applications including medical screening, diagnosis, monitoring, and treatment, and in particular in contexts related to sleep and neurological disorders.
[0008] Electroencephalography (EEG) is a non-invasive neurophysiological monitoring technique that records electrical activity generated by cortical neurons via electrodes placed on the scalp. EEG signals have been utilized for decades in clinical and research settings to assess functional brain states, owing to their high temporal resolution and ability to capture neural dynamics associated with cognition, sleep, and disease. EEG signals can be used for detection, screening, and monitoring of a wide range of health conditions, including neurological disorders such as Alzheimer's disease and Parkinson's disease, as well as cardiovascular, pulmonary, metabolic, and psychiatric conditions. In sleep medicine, EEG is a primary modality for determining sleep stages and identifying abnormalities in sleep architecture, such as disruptions in slow-wave sleep or rapid eye movement (REM) patterns. EEG-derived features, including spectral power distributions, sleep spindles, and event-related potentials, are widely used as biomarkers in both diagnostic and therapeutic contexts.
[0009] Notwithstanding these advantages, acquisition of EEG signals presents significant technical challenges. EEG measurement typically requires placement of electrodes on the scalp, which can be uncomfortable, time-consuming, and impractical for continuous or large-scale monitoring. Signal quality is highly sensitive to electrode placement, skin impedance, and subject motion, and is frequently degraded by artifacts arising from environmental interference or physiological noise. As a result, reliable EEG acquisition is often limited to clinical or laboratoryAttorney Docket No 70009- 127-C- WO
[0010] Client case 26590
[0011] environments, constraining the accessibility and scalability of EEG-based diagnostic and monitoring systems.
[0012] In contrast, a variety of other physiological signals can be acquired more easily, continuously, and non-invasively using wearable, embedded, or contactless sensing technologies. Such signals include, by way of example, respiratory signals, electrocardiograms (ECG), photopl ethy smograms (PPG), motion signals, and radiofrequency (RF) sensing signals. These signals can be obtained using low-cost sensors and are well-suited for ambulatory and long-term monitoring. However, these nonEEG physiological signals do not directly provide the neural information captured by EEG, and conventional analysis techniques applied to such signals are typically limited in their abi 1 i ty to infer complex neurological states or clinically relevant biomarkers.
[0013] Moreover, although various physiological systems of the body are coupled through underlying biological processes, existing monitoring approaches generally treat different signal modalities independently or require comprehensive multi-modal acquisition, such as polysomnography (PSG), to capture interactions between systems. Such multi-modal systems are resource-intensive and are not readily deployable outside specialized environments. As a result, there remains a technical limitation in the ability to leverage readily available physiological signals to obtain information equivalent to that derived from EEG or to extract clinically actionable insights associated with neural activity.
[0014] In addition, extracting indicators of disease conditions or pharmacological exposure from physiological signals presents further challenges. Conventional approaches often rely on direct measurement of EEG or on large, task-specific datasets to train models for particular conditions. These approaches may not generalize well across populations, may require extensive labeled data, and may fail to capture latent physiological relationships that underlie observable signals.
[0015] Consequently, there is a need for improved techniques that enable robust inference of disease states, biomarkers, or drug effects from accessible physiological measurements.
[0016] Accordingly, there exists a need for systems and methods that (i) enable transformation of non-EEG physiological signals into representations that capture neural or EEG-equivalent information, (ii) facilitate extraction of clinically relevant biomarkers, including indicators of disease or pharmacological exposure, from such representations, and (iii) improve the robustness, scalability, and practicality of physiological monitoring by reducing reliance on direct EEG acquisition.Attorney Docket No 70009- 127-C- WO
[0017] Client case 26590
[0018] SUMMARY OF THE INVENTION
[0019] Disclosed herein are systems and methods for transforming physiological signals into clinically useful representations and assessments using trained neural networks. Five related but independent inventive concepts are described. A first concept concerns generation of an electroencephalography representation from a nonelectroencephalography physiological signal, implemented using a two-network architecture in which a first trained neural network maps the physiological signal to a sequence of codes and a second trained neural network decodes those codes into an electroencephalography representation. A second concept concerns determination of pharmacological exposure from a physiological signal, using a trained neural network to generate a sequence of codes from which an indication of exposure is derived without necessarily reconstructing an electroencephalography signal. A third concept concerns assessment of a risk or condition related to the health of a subject, implemented by obtaining a signal representing a first physiological subsystem and processing it with a trained neural network configured to generate codes representative of activity of a second, different physiological subsystem. A fourth concept concerns training a neural network system using a staged or joint procedure in which an encoder-decoder pair is trained on electroencephalography signals to establish a representation space, and a mapping model is trained to map non-EEG physiological signals into that space using a loss function that may depend on code similarity, reconstruction fidelity, or downstream task performance. A fifth concept concerns improving a degraded electroencephalography signal by processing it jointly with a non-EEG physiological signal using one or more trained neural networks to generate a corrected representation.
[0020] In many clinical and non-clinical monitoring systems, electroencephalography (EEG) provides signal characteristics that are highly informative of neurological state, sleep architecture, and systemic physiological conditions. However, conventional EEG acquisition relies on electrode-based measurement systems that are difficult to deploy outside controlled environments, are susceptible to motion artifacts, impedance variation, and environmental interference, and often provide degraded signal quality in ambulatory or long-term monitoring scenarios. As a result, existing physiological monitoring systems are limited in their ability to obtain reliable neural signal information in wearable, contactless, or large-scale deployment settings.
[0021] At the same time, a variety of physiological signals can be obtained that reflect or encode physiological activity’ of a subject. Such signals may include direct physiological measurements, such as respiratory signals, cardiac signals, or optical signals, as well as indirect or remotely sensed signals that depend on physiologicalAttorney Docket No 70009- 127-C- WO
[0022] Client case 26590
[0023] activity without necessarily constituting direct physiological measurements, such as motion-derived signals or radio-frequency (RF) reflection signals. Such signals may be obtained by direct contact measurement, by proximate non-contact measurement, or by measurement using radio-frequency signals reflected off the subject. These signals may be obtained continuously and non-invasively using wearable, embedded, or contactless devices, but conventional signal processing techniques applied to such signals are generally unable to produce representations that preserve the spectral-temporal structure of neural electrical activity or support EEG-based diagnostic and monitoring systems.
[0024] Accordingly, there is a technical need for signal processing systems that transform accessible physiological measurements into representations that capture neural-information content and that are compatible with EEG-based analysis systems, while improving robustness to noise and enabling deployment in non-clinical environments. There is a further need to extract clinically relevant information, including indicators of disease and pharmacological exposure, from such signals without requiring direct EEG acquisition.
[0025] In one aspect, a first solution addresses these problems by processing a nonelectroencephalography physiological signal generated using at least one sensor and representing a physical measurement of a subject to generate an electroencephalography representation indicative of neural activity. A first trained neural network processes the physiological signal to generate a sequence of codes representing neural dynamics, and a second trained neural network decodes the sequence of codes to generate the electroencephalography representation. As used herein, a ‘‘sequence of codes'’ refers broadly to any ordered set of values, vectors (e.g., multi-dimensional numerical values), indices, embeddings, tokens, or other representations produced by a trained model from an input signal, including without limitation discrete indices selected from a learned codebook, continuous vector embeddings, quantized latent representations, and task-specific feature vectors; the term is intended to encompass all representations that a practitioner in the field of machine learning would recognize as intermediate learned representations of a signal, regardless of whether they are discrete or continuous, fixed-vocabulary or open-ended. When the codes comprise vectors in a multi-dimensional numerical space, each code in the sequence is a real-valued vector of fixed dimension d, where d may range from a few tens to several hundreds of dimensions depending on the model architecture; such vector-valued codes arise naturally in models that employ continuous latent representations, dense embeddings, or codebook entries of non-unit dimension, and the “multi-dimensional numerical space” in which they reside is theAttorney Docket No 70009- 127-C- WO
[0026] Client case 26590
[0027] d -dimensional Euclidean space or a learned submanifold thereof. The sequence of codes is used to generate an electroencephalography representation, such as a timedomain waveform, a time-frequency representation, or a derived feature representation. In various embodiments, the electroencephalography representation may include sampled waveforms, spectrograms or other time-frequency representations, or quantities derived therefrom, such as power in one or more frequency ranges or event-related features.
[0028] In some embodiments, the second trained neural network comprises a decoder trained jointly with a neural network encoder configured to encode electroencephalography signals to produce sequences of codes, such that the decoder leams to reconstruct electroencephalography representations from codes that lie in an EEG-derived representation space. This transformation enables neural-signal representations compatible with EEG-based processing systems to be obtained without direct electrode-based acquisition, thereby improving deployability7and robustness of physiological monitoring systems. In some embodiments, the first trained neural network is trained using pairs of physiological signals and corresponding electroencephalography signals, with training comprising minimizing a loss function dependent on at least one of similarity between codes, similarity7between reconstructed electroencephalography representations, or performance of a downstream task. The electroencephalography representation may be output for use in detection, screening, monitoring, or assessment, including sleep staging, arousal detection, sleep quality assessment, and sleep disorder screening. The physiological signal may comprise a respiratory signal, electrocardiogram (ECG), photoplethysmogram (PPG), electromyogram (EMG), or motion signal, obtained by direct contact measurement, by proximate non-contact measurement, or by measurement using radio-frequency signals reflected off the subject.
[0029] In another aspect, a second solution addresses determination of pharmacological exposure by obtaining, from one or more physiological signals generated using at least one sensor, a representation that encodes physiological dynamics associated with drug effects. A trained neural netw ork generates a sequence of codes or other intermediate representation from the physiological signal, and this representation is processed to determine an indication of presence, absence, class, magnitude, or timing of exposure to one or more pharmacological agents. The indication may comprise a probability, score, class label, or binary output, and may characterize presence or absence of exposure, a class of pharmacological agents, dose range, recency of exposure, onset of exposure, continued exposure, tapering, or cessation of exposure. In certain embodiments, the representation may be converted to anAttorney Docket No 70009- 127-C- WO
[0030] Client case 26590
[0031] electroencephalography representation, while in other embodiments the representation is used directly without reconstructing a time-domain EEG signal. In some embodiments, determining the indication comprises processing physiological signals obtained over a plurality of time points, enabling longitudinal monitoring of pharmacological status. In some embodiments, the trained neural network is trained using labels derived from medication records, prescription data, clinician annotations, or treatment schedules, and may be trained to determine exposure despite absence of exact dose timing labels in the training data. Pharmacological agents of interest include sedatives, antidepressants, opioids, stimulants, anesthetics, antiepileptics, and cardiovascular drugs, among others, and the exposure may comprise therapeutic, subtherapeutic, supratherapeutic, prescribed, or non-prescribed exposure.
[0032] In a further aspect, a third solution addresses assessment of a risk or condition related to the health of a subject by obtaining a signal generated using at least one sensor and representing activity of a first physiological subsystem of the subject, and processing the signal using a trained neural network that has been trained to generate a sequence of codes representative of activity associated with a second physiological subsystem different from the first physiological subsystem. In some embodiments, the sequence of codes is generated in a representation space learned from reference electroencephalography signals, enabling cross-physiological inference in which information derived from one physiological subsystem is used to assess conditions associated with another. Determining the indication may be based at least in part on at least one representation generated using the trained neural network, including all or a subset of the sequence of codes or an intermediate representation from an internal layer of the trained neural network. In some embodiments, determining the indication comprises comparing the sequence of codes with a reference distribution derived from healthy subjects or subjects having a known condition, and the indication may be based on deviation of the codes from a normative representation space. In some embodiments, determining the indication comprises processing sequences of codes obtained from physiological signals acquired at a plurality of time points, enabling longitudinal assessment of disease progression or treatment response. In some embodiments, determining the indication is based further on contextual factors including age, sex, medication status, sleep state, activity state, or subject history. The indication may comprise a probability, severity score, risk score, class label, or binary output, and may represent disease presence, disease severity, disease progression, treatment response, or a future risk of developing a condition. The risk or condition may comprise a disease risk or condition, including neurological, metabolic, cardiovascular, pulmonary, or immune-related conditions.Attorney Docket No 70009- 127-C- WO
[0033] Client case 26590
[0034] In certain embodiments, the representation used across these aspects comprises a sequence of codes that encode latent features of physiological dynamics. The codes may be discrete, continuous, or vector-valued, and may be derived using learned encoders or other transformation models. In some implementations, the representation is used directly for downstream inference without reconstruction of a full EEG waveform, thereby reducing computational complexity and enabling flexible integration with different analy sis systems. In some embodiments, downstream inference is performed using an intermediate representation obtained from an internal layer of the trained neural netw ork rather than from the final sequence of codes, enabling flexible integration with dow nstream inference architectures. In some embodiments, EEG-derived biomarkers, such as spectral pow er features, spindle-related features, coherence-related features, event-related features, or other derived neural features, are extracted from the electroencephalography representation or from intermediate representations.
[0035] In some embodiments, the physiological signal is obtained while the subject is in a sleep state, partially or fully. Sleep acquisition is applicable across all three primary' use cases: EEG generation, pharmacological exposure determination, and disease risk or condition assessment. Sleep provides a physiologically stable and reproducible measurement context in which disease-related signatures, pharmacological effects on sleep architecture, and cross-physiological subsystem relationships are particularly clearly expressed. The generated electroencephalography representation, indication of pharmacological exposure, or indication of disease risk or condition obtained from sleep-state signals may be used in downstream applications including sleep staging, arousal detection, sleep quality assessment, sleep disorder screening, neurodegenerative disease screening, metabolic condition monitoring, and pharmacological monitoring.
[0036] In further embodiments, the systems improve operation of signal acquisition and processing systems by reducing artifacts or reconstructing missing signal components. For example, an electroencephalography signal that includes noise, motion artifacts, impedance-related distortions, or missing intervals may be processed together with a non-EEG physiological signal to generate a corrected electroencephalography representation. The corrected representation may be obtained by generating a sequence of codes conditioned on both signals and reconstructing the electroencephalography representation therefrom, thereby restoring physiologically consistent spectral -temporal structure and improving signal quality. The corrected representation may be generated in real time for continuous monitoring applications.Attorney Docket No 70009- 127-C- WO
[0037] Client case 26590
[0038] In some embodiments, multiple physiological signals are processed jointly to improve accuracy and robustness of the generated representations. In certain implementations, the disclosed methods operate in real time or near real time, enabling continuous monitoring and timely feedback.
[0039] In further embodiments, the trained neural networks are configured using training processes that include learning a representation space from electroencephalography signals by mapping EEG signals to sequences of codes and reconstructing EEG representations from those codes, and learning a mapping from non-EEG physiological signals to the representation space using paired datasets. Training may be performed using loss functions based on similarity between codes, similarity between reconstructed representations and reference signals, performance of downstream tasks, or combinations thereof, including time-domain losses, spectral-domain losses, perceptual losses, and adversarial losses. The encoder model and decoder model may be trained before training the mapping model in a staged procedure, or training of the mapping model and decoder may be performed jointly. Training of the mapping model may comprise a masked prediction task, contrastive learning between physiological signals and corresponding electroencephalography signals, or both. The mapping model may be fine-tuned for a target sensor type or target subject population, and training may comprise domain adaptation across different sensor hardware or acquisition environments. In some embodiments, the mapping model is calibrated using subject-specific physiological data. In some embodiments, training comprises compensating for delay between the physiological signals and the electroencephalography signals. In some embodiments, additional physiological signals, demographic information, or medical-history information are incorporated to improve accuracy or generalization.
[0040] The electroencephalography representation generated from the physiological signal may be supplied to a downstream system configured to process naturally acquired EEG signals. Such a downstream system may comprise a machine-learning, algorithmic, rule-based, or heuristic system developed or validated using naturally acquired EEG signals. Because the electroencephalography generation stage may be used with multiple different downstream systems, the same transformed representation can provide immediate utility across multiple detection, screening, prognosticating, or monitoring tasks, including, by way of example, sleep staging, arousal detection, disease assessment, or pharmacological-effect assessment.
[0041] These approaches provide technical improvements to physiological monitoring systems by enabling transformation of accessible physical measurements into neural-activity representations compatible with EEG-based analysis, improving robustness toAttorney Docket No 70009- 127-C- WO
[0042] Client case 26590
[0043] noise and artifacts, enabling reuse of systems trained on EEG data without modification, reducing dependence on direct electrode-based measurements, and facilitating scalable deployment of diagnostic and monitoring technologies in wearable, contactless, and non-clinical environments.
[0044] Each of the described inventive concepts is implemented using specific trained neural network architectures configured with particular learned parameters, specific training procedures that shape the structure of a learned representation space in a defined way, and specific input-output relationships that transform physical sensor measurements into clinically actionable outputs. The first and second trained neural netw orks of the EEG generation concept constitute a particular machine in which the decoder’s weights are specifically constrained by joint training with an EEG encoder, such that the representation space occupied by the codes is not arbitrary but is shaped by and grounded in the structure of naturally acquired EEG signals. The mapping network’s weights are further specifically configured by a training procedure that minimizes loss over code similarity, reconstruction fidelity, or downstream task performance, or combinations thereof, on paired physiological and EEG recordings. These architectural and training constraints are not incidental; they are the technical means by which the transformation of an accessible physiological signal into a clinically useful neural-activity’ representation is achieved, and they distinguish the disclosed systems from prior approaches that treat different physiological signal modalities independently or that require direct EEG acquisition. The resulting trained neural networks are specialized computational structures whose learned parameters encode physiological relationships that took substantial data collection and training to establish, and that produce concrete, physically meaningful outputs, namely electroencephalography representations, pharmacological exposure assessments, and disease risk indicators, suitable for use in clinical monitoring, diagnostic, and therapeutic systems.
[0045] Other features and advantages of the invention are apparent from the following description, and from the claims.
[0046] BRIEF DESCRIPTION OF THE DRAWINGS
[0047] FIG. 1 is signal flow- diagram of a detection, screening, or monitoring system; FIG. 2 is a block diagram of a breathing signal tokenizer;
[0048] FIG. 3 is a block diagram of a EEG generator;
[0049] FIG. 4 is a block diagram illustrating training of an EEG generator;Attorney Docket No 70009- 127-C- WO
[0050] Client case 26590
[0051] FIG. 5 is a block diagram illustrating training of a token transformer;
[0052] FIG. 6 is a table of dataset characteristics;
[0053] FIG. 7 A illustrates EEG data measurement and FIG. 7B is a corresponding EEG spectrogram
[0054] FIGs. 8a-d are representative examples of generated EEG spectrograms;
[0055] FIG. 9A and 9B are tables of Mean Absolute Error (MAE) and SNR respectively; FIGs. lOa-d are graphs of subgroup analysis of EEG generation performance; FIG. 11 includes graphs of comparison of clinical predictive performance between generated and ground truth EEG;
[0056] FIG. 12 includes graphs of comparison of performance between a model’s breathing representation and raw breathing signal across various disease screening and drug effect assessment tasks;
[0057] FIGs. 13a-c show model performance.
[0058] DETAILED DESCRIPTION
[0059] 1 SYSTEM OVERVIEW
[0060] A number of embodiments described below consider physiological time series, like language, as described by a vocabulary' of recurring patterns governed by biological "grammar." and that different signals (e.g., EEG and breathing) may encode overlapping health information using distinct vocabularies. Without being bound to this premise, waveform-conditional and other generative models can translate non-EEG physiological signals into EEG representations, into intermediate representations that encode EEG-relevant information, or into outputs used to determine pharmacological exposure or disease conditions.
[0061] The detailed description below is organized to support several related inventive concepts. A first concept concerns generation of an electroencephalography representation from a non-electroencephalography physiological signal. A second concept concerns determining pharmacological exposure from one or more physiological signals, optionally using an electroencephalography representation or an intermediate representation such as a sequence of codes. A third concept concerns assessing a disease condition from a physiological signal associated with a first physiological modality and a first physiological subsystem by generating aAttorney Docket No 70009- 127-C- WO
[0062] Client case 26590
[0063] representation indicative of activity associated with a second physiological modality and a second physiological subsystem. These concepts may share model components, training procedures, and sensing modalities, but need not all be practiced together.
[0064] Referring to FIG. 1, in one embodiment, a system 100 is shown for estimating an electroencephalography (EEG) signal 165 from a non-EEG physiological signal, such as a respiratory (breathing) signal 135 shown in FIG. 1, and processing the estimated EEG signal using a downstream clinical assessment system for detection, screening, prognostication or monitoring. A subject 120 is monitored by one or more sensors that measure physiological activity using contact and / or non-contact techniques, such as acquisition of a breathing signal 135 representing respiratory7activity. The output of a contact or non-contact sensor is captured by a breathing signal acquisition module 130. In some embodiments, the breathing signal is acquired using a respiratory belt, a nasal airflow sensor, accelerometer data, or a contactless wireless sensor, such as using radio frequency (RF) reflections capturing breathing-related movement of the subject 120. In other embodiments, the breathing signal is derived from other physiological measurements, such as electrocardiogram (ECG)-derived respiration (EDR) or photoplethysmography (PPG)-derived respiration. The acquired breathing signal 135 may be a raw or preprocessed time-series signal representing respiratory motion, airflow, and / or respiratory rate variability.
[0065] In some embodiments, the physiological signal is preprocessed prior to input to the trained neural network. Preprocessing operations may include temporal alignment of the physiological signal with reference electroencephalography data, resampling to a target sampling rate, filtering to suppress out-of-band noise or baseline drift, normalization of signal amplitude or statistics, artifact suppression to remove motion or electrode-related transients, or baseline correction. In some embodiments, the physiological signal is transformed into a time-frequency representation, such as a short-time Fourier transform spectrogram or a wavelet-based representation, prior to generation of the sequence of codes, enabling the neural network to operate directly on spectral features rather than time-domain samples. These preprocessing operations may be applied individually or in combination, and no particular combination is required. In some embodiments, the preprocessing operations are implemented as fixed operations applied prior to the neural network input; in other embodiments, one or more preprocessing operations are implemented as learned differentiable layers within the neural network and are optimized jointly with the network parameters.
[0066] In some embodiments, the acquired non-electroencephalography physiological signal represents a physical measurement of the subject. In some embodiments, the physiological signal is associated with a first physiological modality associated with aAttorney Docket No 70009- 127-C- WO
[0067] Client case 26590
[0068] first physiological subsystem of the subject, such as a respiratory modality associated with a respiratory subsystem, a cardiac modality associated with a cardiovascular subsystem, or a motion-related or RF-derived modality associated with bodily movement or another sensed physiological phenomenon.
[0069] The breathing signal is processed by a breathing signal tokenizer 140 that outputs a sequence of EEG codes 155 corresponding to the input breathing signal 135. In the embodiment shown in FIG. 1 , this mapping is performed in two steps with an intermediate representation of the breathing signal as a sequence of breathing tokens 145, but it should be recognized that a learned mapping from a breathing signal (e.g., time samples of amplitude variation) to a sequence of EEG tokens does not require an a explicit or implicit transformation into breathing tokens. In the illustrated example, the breathing signal 135 is provided to a breathing signal tokenizer 140, which converts the time domain (e.g., time sampled) respiratory signal into a sequence of discrete breathing tokens 145. In one implementation, the tokenizer is a neural network encoder trained to project segments of the breathing signal into a latent token space. The tokenizer may be implemented using convolutional layers followed by vector quantization modules, such as a VQ-VAE or VQ-GAN architecture, and configured using tokenizer parameters 142. In alternative implementations, the tokenizer may, for example, employ hand-crafted feature extraction followed by vector clustering (e.g., k-means) to assign tokens.
[0070] In the illustrated example of FIG. 1, the resulting breathing tokens 145 are input to a token transformer 150. which maps the breathing token sequence to a corresponding sequence of EEG codes 155. The token transformer 150 may be implemented as a transformer-based generative model trained using masked autoencoding techniques, such as a BERT-style or GPT-style architecture. The transformer captures the temporal structure and dependencies between token sequences and is configured using transformer parameters 142. As described more fully below, in this example, the transformer model is trained on synchronized breathing and EEG recordings, learning to predict EEG token sequences conditioned on prior breathing tokens. In alternative implementations, the token transformer may, for example, be a sequence-to-sequence recurrent neural network (RNN), or a convolutional sequence model with dilated causal convolutions. In embodiments without explicit breathing tokenization, the breathing signal (e.g., time samples, or subsequences (“frames"’) of such samples) may be provided to a transformer model directly in the place of the token transformer 150.
[0071] The EEG codes 155 are decoded by an EEG generator 160 to reconstruct a timedomain EEG signal, a spectrogram representation of the EEG, or other EEG features.Attorney Docket No 70009- 127-C- WO
[0072] Client case 26590
[0073] The generator 160 may include a decoder component of a VQ-GAN architecture, which reconstructs EEG spectrograms from latent code sequences. The generator is configured using generator parameters 162 and may be trained to minimize reconstruction error against ground-truth EEG spectrograms. Alternatively, the EEG generator may produce time-domain EEG waveforms directly using autoregressive waveform models or neural vocoders such as WaveNet, GAN-based, or diffusion model based audio decoders.
[0074] The electroencephalography representation generated by the decoder may take a variety7of forms. In some embodiments, the representation comprises a time-domain waveform corresponding to one or more EEG channels. In some embodiments, the representation comprises a spectrogram or other time-frequency representation encoding power as a function of time and frequency. In some embodiments, the representation comprises a plurality of channels corresponding to scalp electrode locations, enabling the generated representation to be supplied directly to downstream systems configured to accept multi-channel EEG input. In some embodiments, the representation corresponds to a reduced-channel montage, such as a single-channel or two-channel derivation, enabling low-complexity downstream processing. The specific form of the electroencephalography representation may be selected based on the requirements of the downstream system or clinical application, and no particular form is required.
[0075] In some embodiments, the output of the disclosed system is an electroencephalography representation. In other embodiments, the output is a sequence of codes, an embedding, a latent representation, or another intermediate representation that encodes neural or physiological dynamics and that may be processed directly, including without reconstructing an electroencephalography representation.
[0076] As used herein, the sequence of codes may comprise discrete codes selected from a finite vocabulary, numerical values, or vectors in a multi-dimensional numerical space. When the codes comprise vectors in a multi-dimensional numerical space, each code is a real-valued vector of fixed dimension d, w here d may range from a few- tens to several hundreds of dimensions depending on the model architecture. Such vectorvalued codes arise naturally in models that employ continuous latent representations, dense codebook embeddings, or learned feature vectors of non-unit dimension. The multi-dimensional numerical space in which such codes reside is the d -dimensional Euclidean space !Rd. and distances or inner products in that space may reflect semantic similarity between the physiological states encoded by the respective codes.Attorney Docket No 70009- 127-C- WO
[0077] Client case 26590
[0078] In various embodiments, the disclosed system performs a structured transformation of a first physical measurement modality into a second physiologically meaningful signal representation through a sequence of signal-processing and generative operations. The non-EEG physiological input, such as respiratory motion, cardiac electrical activity, photoplethysmographic variation, accelerometer-derived movement, or radio-frequency phase reflections, constitutes a time-varying physical measurement obtained from a subject. This physical measurement is converted into one or more intermediate representations, including continuous embeddings, discrete token sequences, or latent feature maps, which encode temporal and spectral structure inherent in the sensed signal.
[0079] The system then applies one or more trained neural network models to transform these intermediate representations into an estimated EEG signal. In embodiments employing tokenization, the non-EEG signal is mapped to a sequence of EEG codes representing discrete spectral-temporal components of neural activity, which are subsequently decoded into a time-domain or time-frequency EEG representation. In other embodiments, the transformation is implemented through direct waveform-to-waveform generation, conditional generative modeling, diffusion-based denoising processes, or other learned mappings that reconstruct an EEG waveform from the input physiological dynamics.
[0080] This sequence of operations constitutes a technological signal transformation rather than a mere abstract data analysis. The system converts a first sensed physical phenomenon (e.g., thoracic displacement, cardiac rhythm, or RF phase modulation caused by physiological motion) into a second signal representation that corresponds to neural electrical activity as conventionally measured by scalp electrodes. The resulting estimated EEG signal is structured to be compatible with downstream EEGbased processing pipelines, including spectral analysis, sleep staging algorithms, disease classifiers, and closed-loop control systems.
[0081] In some embodiments, the transformation further improves signal integrity by suppressing or replacing artifacts in acquired EEG signals using correlated non-EEG physiological inputs. For example, where electrode impedance fluctuations or motion artifacts distort measured EEG power spectra, the system may reconstruct physiologically consistent spectral components conditioned on respiratory or cardiac dynamics. This improves the technical performance of EEG-based monitoring systems by restoring usable neural signal structure under conditions in which conventional acquisition would yield degraded data.Attorney Docket No 70009- 127-C- WO
[0082] Client case 26590
[0083] Accordingly, the disclosed methods implement a concrete transformation of physical measurements within a physiological monitoring system, enabling reliable neural signal reconstruction in environments where direct electrode-based measurement is limited, unreliable, or impractical.
[0084] The disclosed two-stage architecture also constitutes a technical improvement to machine learning systems for cross-modal physiological signal generation. A direct end-to-end mapping from a non-EEG physiological waveform to a reconstructed EEG waveform presents a technically challenging regression problem: the target EEG signal occupies a high-dimensional, highly variable signal space, and the relationship between input and output modalities is not a deterministic function but rather a conditional distribution over plausible neural states consistent with the observed physiological input. The disclosed approach addresses this technical problem by first learning a compact discrete representation space from EEG signals using an encoderdecoder model, thereby constraining the generative target to a finite or structured latent vocabulary that captures spectral -temporal structure of neural activity'. The cross-modal mapping model is then trained to predict within this learned representation space rather than in the raw EEG signal space, converting a highdimensional regression into a more tractable structured prediction task. This architectural choice may improve the convergence, generalization, and computational efficiency of the trained models relative to direct waveform-to-waveform approaches, and represents a specific technical solution to the problem of cross-modal physiological signal generation.
[0085] 2 EEG GENERATION
[0086] In one set of embodiments, the system generates an electroencephalography representation of neural activity of a subject from a non-electroencephalography physiological signal. In some embodiments, the system acquires, using at least one sensor, a non-electroencephalography physiological signal representing a physical measurement of the subject, processes the physiological signal using a first trained neural network to generate a sequence of codes representing neural activity of the subject, decodes the sequence of codes using a second trained neural network to generate an electroencephalography representation, and outputs the electroencephalography representation. The electroencephalography representation may comprise a time-domain waveform, a time-frequency representation, a spectrogram, a derived feature representation, or another representation indicative of neural activity'.Attorney Docket No 70009- 127-C- WO
[0087] Client case 26590
[0088] The first and second trained neural networks are distinct functional components in the described embodiments, and the boundary between them may be implemented in various ways, for example by defining the first model as comprising all processing stages up to and including generation of the sequence of codes and the second model as comprising all decoding stages from the sequence of codes to the electroencephalography representation. Additional intermediate processing stages, such as post-processing of codes or conditioning on auxiliary inputs, may be interposed between the two models without departing from this characterization. The first and second models may share parameters in part, for example through jointly trained embedding layers, while remaining distinct in their primary' functional roles.
[0089] 2.1 Breathing Signal Tokenizer
[0090] Referring to FIG. 2, with reference to the example illustrated in FIG. 1, which makes use of an intermediate tokemzation of the breathing signal, the breathing signal tokenizer 140 receives a time-domain (e.g., time sampled) breathing signal 135 and transforms it into a sequence of discrete breathing tokens 245 that encode relevant temporal features of respiratory activity. In one implementation, the tokenizer is realized using a convolutional neural network-based encoder 240 trained according to its weights that incorporated in the tokenizer parameters 142 to map short, overlapping (or alternatively non-overlapping) segments of the breathing signal into a token 245 in a discrete latent space. The encoder may comprise one or more convolutional layers configured to extract temporal patterns from the input signal, followed by a vector quantization module that assigns each segment to the nearest entry in a learned codebook of representative patterns. The codebook to the quantizer is also part of the tokenizer parameters 142. This architecture may be implemented using a vector-quantized variational autoencoder (VQ-VAE) or a vector-quantized generative adversarial network (VQ-GAN), where the encoder outputs are quantized to the nearest token in a fixed dictionary of latent vectors.
[0091] The breathing signal tokenizer may be trained using a reconstruction-based objective. In a representative training process, a large dataset of breathing signals is divided into training and validation subsets. Each breathing signal is segmented into fixed-length windows (e.g., corresponding to 1-5 seconds of data), and each window is processed by the encoder to produce a latent representation. This representation is then quantized by selecting the nearest codebook vector, resulting in a discrete token index. A breathing decoder network attempts to reconstruct the original breathing window from the tokenized representation. The overall training loss includes a reconstruction loss, ty pically mean squared error between the original andAttorney Docket No 70009- 127-C- WO
[0092] Client case 26590
[0093] reconstructed breathing signal, and a vector quantization loss that encourages efficient and diverse usage of the codebook entries. In the case of VQ-GAN, an adversarial loss may also be employed to promote realistic signal reconstruction.
[0094] In alternative implementations, the tokenizer may utilize unsupervised clustering techniques, such as k-means clustering or Gaussian mixture modeling (GMM), applied to manually or automatically extracted features of the breathing signal. In such cases, the breathing tokens correspond to the cluster assignments of fixed-length signal windows or feature vectors derived therefrom.
[0095] Once trained, the tokenizer may be applied in an inference-only mode shown in FIG. 2, in which the encoder and codebook are fixed. The tokenizer receives a streaming or batch breathing signal and outputs a sequence of token indices that compactly represent the physiological information of the input. These tokens are then used as the input to the downstream generative model, such as the token transformer 150. This tokenized representation allows for standardized input regardless of variations in signal duration, amplitude, or acquisition modality, thereby supporting generalization across datasets and devices.
[0096] 2.2 EEG Generator
[0097] Turning now to the EEG generator 160 later in the data flow show n in FIG. 1, and referring to FIG. 3, the generator is configured to receive as input a sequence of discrete EEG codes 355 and reconstructs an estimated electroencephalography (EEG) signal 165. In some embodiments, the EEG generator is configured to produce a timefrequency representation of the EEG signal, such as a spectrogram, in which each token 265 (i.e., one square in the figure) corresponds to a localized region of time and frequency. In other embodiments (not illustrated in FIG. 3), the generator produces a time-domain waveform representing the EEG signal. The generator may be implemented using a decoder component of a vector-quantized generative adversarial network (VQ-GAN), which maps discrete tokens to spectrogram segments via learned embeddings and a multi-stage neural decoder. The generator operates under control of generator parameters 162, which are trained to minimize reconstruction loss between the generated and ground-truth EEG signals.
[0098] Referring to FIG. 4, during training if the EEG encoder and decoder, the EEG decoder 360 is jointly trained with an associated EEG encoder 440. In the example illustrated in FIG. 4, an encoder 420 converts an EEG spectrogram a latent feature map 425, which includes on feature value or vector of values for each time versus frequency bin of the EEG spectrogram to be estimated. This feature map 425 is thenAttorney Docket No 70009- 127-C- WO
[0099] Client case 26590
[0100] quantized by a quantizer 430 which performs a dictionary lookup using a codebook of discrete embeddings. These dictionary lookups form the EEG codes 445. In alternative examples, the EEG codes are not necessarily from a fixed or finite dictionary, and may comprise latent vector representations, and the encoder 440 is trained on the over ability to estimate the EEG signal from such latent vectors.
[0101] The EEG decoder 360 (using the configured values of the generator parameters 152 not shown in FIG. 4), introduced above with reference to FIG. 3, then reconstructs the EEG signal spectrogram of the estimated EEG signal 465 from these quantized tokens. The training objective includes a reconstruction loss between the input EEG signal 415 and the estimated EEG signal 465. This reconstruction loss is ty pically mean squared error or perceptual loss between the original and reconstructed spectrogram, as well as a vector quantization loss that encourages efficient usage of the codebook of the quantizer 430. In the case of adversarial training, a discriminator may also be trained to distinguish real from reconstructed spectrograms, providing additional gradient signals to improve fidelity.
[0102] In some implementations, the EEG spectrogram for use in the training loss computation is computed from raw EEG recordings using standard preprocessing methods such as filtering, down-sampling, and application of a short-time Fourier transform spectral analysis. The resulting spectrogram is divided into non-overlapping or partially overlapping patches (e.g., 2D segments across time and frequency), each of which is mapped to a token index using the learned encoder. These token sequences form the basis for training the EEG generator decoder, which leams to map back from token sequences to high-fidelity EEG spectrograms. The generator decoder may consist of convolutional layers, transposed convolutions, or transformer-based image decoders, optionally conditioned on positional encodings that preserve temporal structure.
[0103] Once trained, the EEG decoder 360 can be used in inference mode to reconstruct an EEG signal from any valid sequence of EEG codes. In this role, it functions as a fixed component downstream of a EEG code generation system, such as the token transformer 150. The decoder thus provides an important mapping between the EEF coder representations and physiologically meaningful EEG outputs.
[0104] In alternative embodiments, the EEG decoder (i.e., the processor of EEG codes) may produce a time-domain EEG waveform directly’ from code sequences, for example using neural vocoders adapted from audio synthesis domains such as WaveNet, HiFi-GAN, or diffusion models. These architectures predict a raw EEG time signal (e.g., time samples) conditioned on token inputs, enabling continuousAttorney Docket No 70009- 127-C- WO
[0105] Client case 26590
[0106] reconstruction of brain activity signals without the need for intermediate spectrogram representations. Alternatively, where downstream systems operate on frequencydomain features or biomarkers, the decoder may output only partial reconstructions (e.g., specific frequency bands or sleep-stage-relevant features) sufficient for diagnostic or screening purposes.
[0107] The ability of the EEG decoder to accurately reconstruct EEG signals from token sequences enables its use both as a terminal component in the signal synthesis pipeline and as a target model during training of upstream components such as the token transformer. Specifically, during transformer training, EEG code sequences generated from ground-truth EEG data are passed through the decoder to evaluate the fidelity of transformer outputs with respect to physiological EEG signals. This modular arrangement supports flexible integration of multiple upstream code generation models and facilitates use of pretrained EEG generators in transfer learning scenarios.
[0108] 2.3 Breathing Mapping
[0109] Referring back to the example of FIG. 1, breathing signal tokenizer 140 includes the token transformer 150, which receives as input a sequence of discrete breathing tokens 145, generated by the breathing signal tokenizer 140, and produces as output a sequence of discrete EEG codes 155 suitable for input to the EEG generator 160 described above. These EEG codes correspond to token indices within the discrete codebook used by the EEG generator 160 to reconstruct an EEG spectrogram or timedomain signal, as previously described. The token transformer thereby performs a translation from one physiological modality, namely respiration, to another, namely electroencephalographic activity, by mapping token sequences that represent breathing dynamics into token sequences that represent cortical activity.
[0110] In one embodiment, the token transformer is implemented as a transformer-based sequence-to-sequence model that includes multiple layers of multi-head self-attention, position-wise feedforward networks, layer normalization, and positional encoding. The model accepts an input sequence of breathing tokens and produces an output sequence of EEG codes of equal or corresponding length. The transformer may include an embedding layer that maps each discrete breathing token into a continuous vector representation, which is then processed through the transformer architecture to capture temporal dependencies and correlations. Each output vector is then projected onto the EEG token vocabulary' using a final classification layer, yielding a predicted token index at each position in the time-frequency domain of the spectrogram.Attorney Docket No 70009- 127-C- WO
[0111] Client case 26590
[0112] Referring to FIG. 5, the breathing mapper 540 is trained using recordings from a training subject 520 of breathing signals 535 and EEG signals 515. In this example that makes use of intermediate breathing tokens, the breathing signals 535 are preprocessed and tokenized using the trained breathing signal tokenizer 140, while the EEG signals are converted into spectrogram representations and tokenized using the trained encoder 420 and quantizer 430. These paired code sequences 565. 545 serve as input-output pairs for supervised training. The models are trained to minimize a loss function, such as categorical cross-entropy, between the predicted EEG code sequence and the ground-truth EEG code sequence.
[0113] In some embodiments, the token transformer 150 is trained using a masked autoencoding objective, in which a portion of the EEG codes is masked at random during training. The model receives as input the full breathing token sequence along with the partially masked EEG token sequence and is trained to predict the masked EEG codes. This objective encourages the model to leam the conditional distribution of EEG codes given the breathing context and any known EEG context, improving generalization and robustness. The masking approach may be inspired by strategies used in BERT-style or masked image modeling frameworks. In alternative embodiments, a causal autoregressive objective is used instead, wherein the model is trained to predict the next EEG code conditioned on past breathing tokens and previously generated EEG codes.
[0114] The transformer may be configured as a single unified model or may adopt an encoder-decoder architecture, in which the encoder processes the breathing token sequence and the decoder predicts the EEG token sequence. In such implementations, the decoder incorporates cross-attention layers that attend to the encoder’s outputs, thereby enabling alignment between respiratory and EEG-related features. The model parameters, denoted as transformer parameters 152, are learned using stochastic gradient descent or other optimization techniques over a large dataset of tokenized physiological recordings.
[0115] In alternative examples, the token transformer may be replaced with or supplemented by other temporal modeling architectures, such as long short-term memory (LSTM) networks, temporal convolutional networks (TCNs), or hybrid attention-based recurrent models. These variants may offer advantages in scenarios with limited training data or where real-time inference is prioritized.
[0116] In example in which no intermediate breathing representation is used, a single transformer is used to map directly from the breathing signal to the EEG codes.Attorney Docket No 70009- 127-C- WO
[0117] Client case 26590
[0118] The trained breathing mapper 540 may be used in inference mode to convert a previously unseen sequence of breathing tokens 145 into a predicted sequence of EEG codes 155, which are then provided to the EEG generator 160 to reconstruct the corresponding EEG signal 165. The modularity of this architecture allows for substitution or retraining of components such as the tokenizer or generator independently, facilitating adaptation to alternative signal modalities or downstream applications.
[0119] 3 REPRESENTATION VARIANTS AND EMBEDDING-FIRST INFERENCE
[0120] In certain embodiments, the representation used in the disclosed methods comprises a sequence of codes that encode latent features of physiological dynamics. The codes may be discrete, continuous, numerical, vector-valued, quantized, or otherwise represented, and may correspond to latent features associated with neural activity or physiological dynamics. The sequence of codes may represent neural activity of the subject, physiological dynamics of the subject, or activity associated with another physiological modality or subsystem.
[0121] In some examples, a minimal projection layer is used to preserve detailed physiological information in the raw input waveform (e.g., the breathing signal). In one embodiment, the raw breathing signal is divided into a sequence of nonoverlapping 4-min segments and each segment is linearly mapped into a continuous embedding space. This results in a sequence of continuous breathing tokens that allows the subsequent Transformer to attend to fine-grained respiratory dynamics. This linear layer is trained together with the subsequent transformer.
[0122] The nature of this raw representation can be significant. Experiments have revealed that utilizing heavy breathing encoders or time-frequency conversions significantly degraded performance, likely by filtering out subtle morphological cues or over-compressing the temporal context required to infer brain states.
[0123] The nature of the tokenization of the EEG spectrogram into discrete tokens can be significant. While the input is preferably preserved in raw form, the target EEG signal is noisy and its structure is distributed across frequencies, making direct waveform generation challenging. To make the search space tractable and capture the underlying EEG structure, a two-step process may be used:
[0124] 1. Spectral Transformation: the time-series single-channel EEG signal is first into a time-frequency spectrogram using the multitaper method. This approach provides Power Spectral Density (PSD) estimates with high spectral resolution andAttorney Docket No 70009- 127-C- WO
[0125] Client case 26590
[0126] minimal leakage, capturing essential sleep EEG features such as posterior dominant rhythms (8-12 Hz), sleep spindles (12-16 Hz), and slow-wave activity- (0.5-1.5 Hz) (Prerau et al.. 2017), as shown in Fig. 3. This data is processed using a window size of 30 seconds: this duration is chosen to align with the standard epoch length defined by the AASM Manual for the Scoring of Sleep and Associated Events.
[0127] 2. Vector Quantization (VQ): The spectrogram is compressed into a sequence of discrete EEG tokens using a VQGAN. This process effectively shrinks the target search space, allowing the translation task to be a tractable classification problem over a finite vocabulary of physiological patterns instead of a complex regression.
[0128] The VQGAN is composed of a encoder E, a decoder D and a learnable codebook. The encoder takes a spectrogram and produces a latent feature grid and then each feature vector is replaced by its nearest neighbor from the codebook to obtain the quantized token sequence. The decoder converts these tokens back to a spectrogram to approximately reconstruct the spectrogram.
[0129] The encoder’s downsampling factor is selected such that each token corresponds to a patch of approximately 4 Hz in frequency and 4 minutes in time. This resolution is strategically chosen to align the latent representation with the physiological semantics of sleep. The 4 Hz frequency resolution naturally segments the spectral energy into canonical EEG bands, isolating Delta (0-4 Hz), Theta (4-8 Hz). Alpha (8-12 Hz), and Sigma (12-16 Hz) into distinct latent rows. The 4-minute temporal resolution aggregates sufficient context to capture stable meso-scale sleep states.
[0130] The VQGAN is pretrained separately. In addition to the standard VQGAN objective, which jointly optimizes reconstruction fidelity-, codebook commitment, and adversarial losses, a correlation regularization term is incorporated that explicitly maximizes the correlation between reconstructed spectrograms and the ground-truth targets. This encourages the tokenizer to preserve intrinsic physiological variability and spectral morphology7, rather than converging to overly smooth or “average” patterns.
[0131] With the asymmetric representations established, long-context raw embeddings for breathing and discrete semantic tokens for EEG, the synthesis is formulated as a sequence-to-sequence translation task. A Transformer-based architecture leveragse a masked generative modeling approach.
[0132] During training, the transformer processes a concatenated sequence of breathing tokens and partially masked EEG tokens. Learnable positional embeddings are added to preserve structure: ID temporal embeddings for the breathing sequence and 2DAttorney Docket No 70009- 127-C- WO
[0133] Client case 26590
[0134] embeddings (encoding time and frequency) for the EEG tokens. The model operates over this joint sequence to reconstruct the discrete tokens at masked positions, optimized via a cross-entropy loss. The masking ratio is sampled from a truncated Gaussian distribution. This variable masking schedule exposes the model to diverse difficulty levels, fostering robust contextual learning between the pulmonary and neurological modalities.
[0135] During inference, the EEG is synthesized from a new breathing input by conditioning the model on breathing tokens con-catenated with a fully masked EEG sequence.The predicted tokens are subsequently decoded by the frozen VQGAN to yield the final EEG spectrogram. Experiments demonstrate that this masked generation strategy7achieves superior reconstruction fidelity7compared to sequential autoregressive generation.
[0136] In some embodiments, the mapping from one or more non-EEG physiological inputs to clinically-useful EEG information is implemented without ever reconstructing a full time-domain EEG waveform or a full time-frequency spectrogram. Instead, the trained model produces one or more continuous or discrete intermediate embeddings, latent representations, or task-specific feature vectors that embody EEG-relevant information. These embeddings may be produced by any module of the model stack (for example a projection layer, the output of a transformer encoder or decoder layer, or the bottleneck representation of an autoencoder) and may be directly consumed by downstream classifiers, regressors, anomaly detectors, or decision logic to perform detection, screening, monitoring, or other assessments. Such embedding-first embodiments may be preferable where downstream tasks have been trained or validated on EEG-derived features or where bandwidth, latency, or energy¬ constraints make full reconstruction undesirable. In these embodiments, the embedding outputs may be converted to diagnostic outputs by simple lightweight classifiers running on the same device or on remote servers, and the embedding dimensionality, statistical distribution, and quantization strategy7may be tuned to optimize performance, storage, and privacy objectives.
[0137] In other embodiments, the system implements direct time-domain synthesis of EEG without passing through a discrete token dictionary. For example, the first trained model may output continuous latent vectors or autoregressive waveform samples that are decoded by neural vocoders, temporal convolutional decoders, diffusion decoders, or other generative waveform models to yield a reconstructed EEG time series. Diffusion-based generative decoders, score-matching decoders, and likelihood-based flows are examples of decoder technologies that can be used in place of or in addition to codebook-based decoders such as VQGANs. These directAttorney Docket No 70009- 127-C- WO
[0138] Client case 26590
[0139] waveform embodiments may be trained end-to-end or in stages, and may employ perceptual, spectral, adversarial, or sample-level reconstruction losses in addition to or instead of token-level cross-entropy losses. Such approaches are particularly useful where precise timing and phase information is important for downstream uses, or where vocoder technologies adapted from audio synthesis yield high-fidelity reconstructions of neural signals.
[0140] In further embodiments, the estimation of the EEG signal may be performed using a conditional diffusion model that directly maps anon-EEG physiological waveform to an EEG waveform without requiring an intermediate tokenization or discrete code representation. In such embodiments, a diffusion-based generative model is trained to iteratively denoise a latent noise process into a time-domain EEG signal while conditioning on one or more input physiological waveforms (for example, breathing, ECG, PPG, motion, or RF-derived signals). During training, paired non-EEG physiological signals and corresponding EEG recordings are used to train the diffusion model under a denoising score-matching or related objective, such that the model leams the conditional distribution of EEG waveforms given the physiological input. At inference time, the model receives the physiological waveform as conditioning information and progressively generates the EEG waveform from an initial noise sample through a sequence of learned denoising steps. In these embodiments, no discrete EEG codebook, spectrogram transformation, or vector quantization stage is required, although such stages may optionally be combined with the diffusion process. This wav eform-to- wav eform conditional diffusion approach may provide improved sample-level fidelity, enhanced modeling of fine temporal structure, and greater flexibility in representing complex or nonstationary neural dynamics, while remaining within the scope of estimating an EEG signal from one or more non-EEG physiological inputs.
[0141] In some embodiments, determining an indication of pharmacological exposure or determining an indication of a disease condition is performed without reconstructing an electroencephalography representation. In such embodiments, a sequence of codes, embedding, latent representation, or task-specific feature vector is processed directly by a classifier, regressor, anomaly detector, or decision logic.
[0142] Finally, a wide variety of implementation choices are compatible with the disclosed methods. Tokenization may employ learned VQ-based codebooks, continuous linear projections, clustering of handcrafted features, or hybrid schemes; decoders may be implemented using convolutional decoders, autoregressive waveform generators, GAN-slyle decoders, diffusion decoders, or combinations thereof; and transformers may be configured as encoder-decoder, decoder-only, orAttorney Docket No 70009- 127-C- WO
[0143] Client case 26590
[0144] encoder-only stacks with varied masking schedules, positional encoding schemes, and attention head allocations. The specification supports these and further variants as embodiments of the inventive concepts described herein, and these embodiments may be combined, interchanged, or refined by those skilled in the art to meet deployment, performance, and regulatory requirements without departing from the scope of the invention.
[0145] 4 PHARMACOLOGICAL EXPOSURE DETERMINATION
[0146] In another set of embodiments, a physiological signal representing a physical measurement of the subject is processed using a trained machine learning model to generate a sequence of codes representing physiological dynamics of the subject, and the sequence of codes is processed to determine an indication of pharmacological exposure. The indication of pharmacological exposure may indicate presence, absence, class, magnitude, timing, degree, or likelihood of exposure to one or more pharmacological agents. The indication may be output for use in detection, screening, monitoring, or assessment of pharmacological exposure.
[0147] The indication of pharmacological exposure may take a variety of forms depending on the clinical or monitoring application. In some embodiments, the indication comprises a probability, score, class label, or binary' output. In some embodiments, the indication identifies the presence or absence of exposure to a pharmacological agent. In some embodiments, the indication identifies a class of pharmacological agents, such as sedatives, antidepressants, opioids, stimulants, anesthetics, antiepileptics, or cardiovascular drugs. In some embodiments, the indication identifies a class of pharmacological agents, such as sedatives, antidepressants, opioids, stimulants, anesthetics, antiepileptics, or cardiovascular drugs. The pharmacologic classes contemplated by the disclosed methods include, without limitation, neuropsychiatric agents, autonomic nervous system agents, bronchodilators, cardiovascular agents, anticoagulants, diuretics, antidiabetic agents, analgesics and opioids, anesthetics, sedatives and hypnotics, anticonvulsants, antipsychotics, beta-blockers, antiarrhythmics, antihypertensives, antihistamines, respiratory tract agents, corticosteroids, and sex hormones, each of which produces characteristic physiological, neurophysiological, or autonomic effects that are reflected in physiological signals or in representations derived therefrom, as described herein. In some embodiments, the indication characterizes one or more of dose range, recency of exposure, onset of exposure, continued exposure, tapering, or cessation of exposure. The pharmacological exposure may comprise therapeutic, subtherapeutic,Attorney Docket No 70009- 127-C- WO
[0148] Client case 26590
[0149] supratherapeutic, prescribed, or non-prescribed exposure, and the trained neural network may be configured to distinguish among these categories.
[0150] In some embodiments, determining the indication of pharmacological exposure comprises processing physiological signals obtained over a plurality of time points, enabling longitudinal assessment of pharmacological status. In such embodiments, the trained neural network or a downstream inference model receives sequences of codes generated from physiological signals acquired across multiple sessions or days, and the resulting indication reflects temporal trends in exposure rather than a single pointin-time estimate. In some embodiments, the method further comprises generating a trend of pharmacological exposure over time, enabling monitoring of changes in drug exposure state across a treatment course. In some embodiments, the output of the method comprises detection of a change in exposure state between successive days or sessions, such as transition from exposure to non-exposure or from therapeutic to supratherapeutic levels.
[0151] In some embodiments, the trained neural network is trained using supervision derived from medication records, prescription data, clinician annotations, or treatment schedules associated with the physiological signals in the training dataset. Such labels enable the model to learn associations between physiological dynamics and pharmacological status without requiring direct pharmacokinetic measurements. In some embodiments, the trained neural network is trained to determine pharmacological exposure despite the absence of exact dose timing labels in the training data, for example by using exposure windows, binary presence labels, or ordinal exposure categories as supervision signals rather than precise timestamps or dose quantities. In embodiments in which the physiological signal is obtained while the subject is in a sleep state, the trained neural network may exploit characteristic pharmacological effects on sleep architecture, such as suppression of slow-wave sleep by benzodiazepines or alteration of REM density by antidepressants, as discriminative features.
[0152] In some embodiments, the determination of pharmacological exposure is performed entirely without generating or reconstructing an electroencephalography representation. In such embodiments, a physiological signal, such as a respiratory signal, ECG, PPG, EMG, or motion signal, is acquired from the subject and processed by a trained neural network to generate a sequence of codes or other intermediate representation encoding physiological dynamics of the subject. A classifier, regressor, or other inference model is applied directly to the sequence of codes to determine an indication of pharmacological exposure, without any intermediate step of decoding the codes into an EEG waveform or spectrogram. For example, a model may beAttorney Docket No 70009- 127-C- WO
[0153] Client case 26590
[0154] trained end-to-end on paired datasets of breathing signals and pharmacological labels, learning to extract features of respiratory dynamics that are predictive of the presence, class, magnitude, or timing of drug exposure, without reference to EEG signals during inference. In other embodiments, the model is first trained using EEG signals to leam a representation space that encodes physiologically relevant dynamics, and the inference model is then applied to codes generated within that representation space, but the EEG representation is used only during training and is not reconstructed at inference time. These embodiments are independent of the EEG generation embodiments described elsewhere herein, and the determination of pharmacological exposure does not require that an EEG signal be generated, estimated, or available. The specific drugs and drug classes described in the embodiments below may be detected independently of one another, and no particular combination of drug targets, physiological input modalities, or model architectures is required.
[0155] In some embodiments, the indication of pharmacological exposure is determined from a sequence of codes or another intermediate representation without reconstructing an electroencephalography representation. In other embodiments, the sequence of codes is used to generate an electroencephalography representation, and the electroencephalography representation or biomarkers derived therefrom are processed to determine the indication of pharmacological exposure.
[0156] Another disclosed EEG processor is configured to detect the presence or influence of pharmaceutical compounds, such as anticholinergics, benzodiazepines, anticoagulants, or insulin. In this embodiment, the processor analyzes the EEG for pharmacodynamic effects characteristic of specific drug classes. For example, benzodiazepines are known to induce increased beta activity and suppress sleep spindle density7, while anticholinergics may reduce theta power and disrupt slow-wave activity. The processor may extract EEG-derived features and compare them to pharmacologically annotated EEG profiles using either a rule-based classifier or a supervised neural network trained on labeled EEG-drug pairs. In some implementations, these classifiers operate on model-derived embeddings obtained from the intermediate layers of the EEG generator or token transformer, rather than on the reconstructed EEG waveform.
[0157] Finally, EEG processors may be designed for pharmacovigilance and therapeutic monitoring applications. These processors may evaluate whether an estimated EEG signal reflects the effects of administered pharmacologic agents, including but not limited to benzodiazepines, anticholinergics, anticoagulants, diuretics, insulin, oral hypoglycemics (e.g., thiazolidinediones), antidepressants, and respiratory stimulantsAttorney Docket No 70009- 127-C- WO
[0158] Client case 26590
[0159] such as xanthines. The processor may generate indicators of drug effect, estimate timing of last exposure, or flag potential adverse response patterns.
[0160] In further embodiments, one or more EEG processors are configured to detect the presence or influence of pharmacological compounds based on characteristic patterns in the estimated EEG signal. For instance, the processor may detect use of central nervous system-active agents such as benzodiazepines, anticholinergics, antipsychotics, antidepressants, or stimulants, by identifying spectral and temporal alterations in the EEG known to correspond to these agents. The processor may also be configured to detect metabolic or cardiovascular medications, including insulin, oral hypoglycemic agents such as thiazolidinediones, anticoagulants, and diuretics. Each pharmacological class may induce distinct neurophysiological signatures, such as increased beta activity (benzodiazepines), suppression of REM-related theta (SSRIs), or enhanced delta power (insulin). These features may be extracted using fixed filters or learned representations, and input to a classifier trained on pharmacologically labeled EEG data. The output may include binary or probabilistic drug detection results, effect size estimates, or temporal confidence intervals indicating the time since exposure.
[0161] 5 DISEASE CONDITION ASSESSMENT
[0162] As used herein, a “physiological subsystem” refers to a functionally distinct component of the body’s overall physiology7that generates or is associated with measurable physical signals, including without limitation a respiratory subsystem associated with breathing and lung activity, a cardiovascular subsystem associated with cardiac electrical activity, blood pressure, and peripheral perfusion, a neural subsystem associated with cortical and subcortical electrical activity, a neuromuscular subsystem associated with the functional interaction between motor neurons and skeletal muscle as reflected in electromyographic activity and motor unit recruitment patterns, a musculoskeletal subsystem associated with mechanical movement, posture, and the structural components of muscle and connective tissue, and an autonomic subsystem associated with regulation of involuntary physiological functions. The first and second physiological subsystems referred to in the methods described herein may be any two distinct subsystems, and the specific examples of a respiratory7subsystem as the first physiological subsystem and a neural subsystem as the second physiological subsystem are non-limiting. Similarly, the specific examples of a breathing signal as the first physiological signal and an EEG signal as the second physiological signal are non-limiting illustrations of the broader concept of cross-physiological-subsystem signal transformation.Attorney Docket No 70009- 127-C- WO
[0163] Client case 26590
[0164] In a further set of embodiments, a physiological signal representing a physical measurement associated with a first physiological modality associated with a first physiological subsystem of the subject is processed using a trained neural network to generate a sequence of codes representing activity associated with a second physiological modality associated with a second physiological subsystem different from the first physiological subsystem. In some such embodiments, the sequence of codes is generated in a representation space learned from reference electroencephalography signals. The sequence of codes, or an electroencephalography representation generated therefrom, is processed to determine an indication of the disease condition, and the indication may be output for use in detection, screening, monitoring, or assessment of the disease condition.
[0165] In some embodiments, the representation space is obtained by encoding electroencephalography signals, and the encoded electroencephalography signals define or constrain the representation space in which the sequence of codes is generated.
[0166] The representation space in which the sequence of codes is generated may be learned from reference electroencephalography signals using any of a variety of learning procedures, and the specific VQ-GAN encoder-decoder architecture described herein is a non-limiting example. In some embodiments, the representation space is defined by training an encoder-decoder model on EEG signals such that the encoder maps EEG signals to sequences of discrete or continuous codes and the decoder reconstructs EEG signals from those codes, as described above. In other embodiments, the representation space is learned using contrastive learning objectives, in which EEG signals from the same subject or physiological state are mapped to nearby codes and EEG signals from different states are mapped to distant codes. In further embodiments, the representation space is learned using supervised objectives, in which codes derived from EEG signals are trained to predict clinical labels such as disease state or pharmacological exposure, thereby organizing the representation space according to clinically relevant dimensions of variation. In still further embodiments, the representation space may be defined by a pre-trained foundation model trained on large corpora of EEG signals, and the cross-modal mapping model is trained to map non-EEG physiological signals into the representation space of that foundation model. In each case, what is common to these embodiments is that the representation space is shaped or constrained by EEG signals, such that codes generated from non-EEG physiological inputs occupy a space that encodes EEG-relevant physiological information. The features, classifierAttorney Docket No 70009- 127-C- WO
[0167] Client case 26590
[0168] architectures, and disease targets described in the embodiments below may be used independently or in any combination, and no particular combination is required.
[0169] In some embodiments, determining the indication of a risk or condition related to the health of the subject comprises comparing the sequence of codes, or another representation generated using the trained neural network, with a reference distribution derived from a population of healthy subjects or subjects having a known disease condition. The trained neural network generates codes that occupy a learned representation space, and deviation of the codes from the normative region of that space as measured by, for example, Mahalanobis distance, likelihood under a learned density model, or a trained anomaly detection model, may be used as an indicator of disease risk or condition severity. In some embodiments, the indication is based on deviation of the sequence of codes from a normative representation space established from healthy reference subjects, enabling detection of pathological states as departures from normal physiological dynamics without requiring explicit disease labels for all training examples. The normative representation space may be constructed by fitting a multivariate Gaussian, a normalizing flow, or another density model to codes generated from a population of healthy subjects, and the deviation of a new subject’s codes from this distribution may be used to compute an anomaly score or risk indicator.
[0170] In some embodiments, determining the indication comprises processing sequences of codes obtained from physiological signals acquired at a plurality of time points, enabling longitudinal assessment of disease progression or treatment response. In such embodiments, the indication may represent disease presence, disease severity, disease progression over time, or response to treatment. In some embodiments, the method further comprises generating an indication of change in disease condition over time, for example by comparing codes or derived features across sessions to identify trends or inflection points. In some embodiments, the indication comprises a future risk of developing a disease condition, for example a risk score representing a likelihood that the subject will develop the condition within a specified future time interval, derived from current physiological dynamics and historical trends.
[0171] In some embodiments, determining the indication is based further on one or more contextual factors associated with the subject, including age, sex, medication status, sleep state, activity state, or subject history. These factors may be incorporated as auxiliary inputs to the trained neural network or to a downstream inference model, enabling the indication to reflect subject-specific physiological baselines and risk profiles. In some embodiments, the trained neural network is adapted using subject-Attorney Docket No 70009- 127-C- WO
[0172] Client case 26590
[0173] specific physiological data, enabling personalized calibration of the model to account for inter-individual variability in physiological dynamics.
[0174] The indication of the risk or condition may take a variety of forms. In some embodiments, the indication comprises a probability, severity score, risk score, class label, or binary output. The indication may represent disease presence, disease severity, disease progression, or treatment response. In some embodiments, the indication comprises a future risk of developing the condition, expressed as a probability or likelihood score over a specified time horizon.
[0175] In some embodiments, the physiological signal used in disease condition assessment is obtained while the subject is in a sleep state, partially or fully. Sleep provides a physiologically rich and reproducible measurement context in which interindividual variability in behavioral state is reduced relative to waking conditions, and in which many disease-related signatures, including disruptions to sleep architecture, slow-wave activity, spindle density, and autonomic regulation, are most clearly expressed. Embodiments directed to sleep-state acquisition are not limited to any particular sleep stage or depth of sleep, and the physiological signal may be acquired across the full night or during a selected interval.
[0176] In some embodiments, the physiological signal used for disease condition assessment comprises a plurality of signals representing the first physiological subsystem acquired simultaneously, enabling the trained neural network to exploit complementary information from multiple sensors or measurement perspectives.
[0177] In some embodiments, the sequence of codes is decoded into a full spectral electroencephalography representation, such as a time-frequency spectrogram representing power across a continuous range of frequencies, rather than a representation limited to energy summaries in discrete frequency bands such as delta, theta, alpha, or beta band power values. The distinction between a full spectral representation and a band-energy summary is technically significant: a full spectral representation preserves the morphological structure of the EEG signal across all frequencies and time points, including fine-grained features such as sleep spindle shape, slow-wave slope, and inter-band relationships, whereas a band-energy summary discards this structure by collapsing power within predefined frequency ranges into scalar values. In some embodiments, the full spectral representation is generated by decoding a sequence of codes using a decoder model trained to reconstruct EEG spectrograms, as described herein. This embodiment is independent of the specific architecture used to generate the sequence of codes, and the full spectral representation may be generated from any sequence of codes that encodesAttorney Docket No 70009- 127-C- WO
[0178] Client case 26590
[0179] sufficient information about neural activity, regardless of whether the codes were produced by a VQ-GAN -based encoder, a contrastive model, a supervised model, or another learned encoder. In embodiments directed to disease condition assessment, the use of a full spectral electroencephalography representation as an intermediate or output representation, rather than band-energy features, constitutes a specific technical feature that distinguishes the described methods from prior approaches that use frequency-band energy summaries derived from EEG as auxiliary training signals.
[0180] The first physiological subsystem may comprise, for example, a respiratory' subsystem, a cardiovascular subsystem, or another subsystem from which a physical measurement can be obtained, and the second physiological subsystem may comprise a neural subsystem or another subsystem whose activity' is reflected in the learned representation. The disease condition may comprise a neurological, cardiovascular, pulmonary, metabolic, immune-related, psychiatric, or sleep-related condition.
[0181] In some embodiments, determining the indication of the disease condition is performed without reconstructing an electroencephalography representation. In other embodiments, the sequence of codes is decoded to generate an electroencephalography representation, and the electroencephalography representation or features derived therefrom are processed to determine the indication of the disease condition.
[0182] In one embodiment, an EEG processor is configured to perform screening for neurodegenerative disease, such as Alzheimer’s disease or Parkinson's disease. In this case, the processor receives an EEG spectrogram or time-domain EEG signal and computes one or more diagnostic features, such as band-specific spectral power (e.g., delta, theta, alpha, beta, gamma bands), peak frequency, spectral entropy, or connectivity7-based features such as coherence between different EEG channels. These features are then input to a machine-learning classifier, such as a convolutional neural network (CNN), random forest, or support vector machine, which outputs a classification score or probability indicative of the presence of a disease condition. For example, reduced alpha power and elevated theta / delta activity may be used as biomarkers of cognitive decline in Alzheimer’s disease, while altered alpha-theta ratios and impaired posterior dominant rhythm (PDR) are frequently associated with Parkinsonian syndromes.
[0183] In alternative embodiments. EEG processors may be configured to analyze EEG signals or EEG-derived features for a variety’ of sleep-related applications beyond conventional staging or event detection. For example, one EEG processor may be designed to classify sleep stages, such as wake, Nl, N2, N3, and REM, on an epoch-Attorney Docket No 70009- 127-C- WO
[0184] Client case 26590
[0185] by-epoch basis using spectro-temporal features extracted from the estimated EEG spectrogram. These features may include relative band power, sleep spindle density, slow-wave activity, and temporal continuity of stage transitions. The classification may be implemented using a neural network trained on labeled sleep EEG data and evaluated against manually scored polysomnography. The output of this processor may include sleep stage labels, transition events, and computed sleep architecture metrics, such as total sleep time, sleep efficiency, or REM latency.
[0186] Another EEG processor may be configured to detect signs of sleep fragmentation or obstructive sleep apnea by analyzing irregularities in the EEG signal associated with cortical arousals. In this embodiment, the processor identifies transient bursts of high-frequency activity, abrupt shifts in EEG spectral composition, or reductions in slow-wave amplitude, which are characteristic of micro-arousals or apneic events. Such features may be quantified across the night and used to derive indices of arousal burden or apnea-hypopnea index (AHI) estimates, particularly when combined with breathing signal features from upstream components.
[0187] In additional embodiments, EEG processors may be used to extract biomarkers relevant to neurodegenerative disease, such as decreased delta power, diminished spindle activity, or disrupted REM patterns, which may be indicative of early cognitive decline or parkinsonian syndromes. These features may be computed from the estimated EEG spectrogram using signal processing techniques such as spectral decomposition, wavelet analysis, or complexity measures, and provided to a classifier trained to detect disease signatures.
[0188] Yet another EEG processor may operate in conjunction with event-related potential (ERP) paradigms to assess cortical responsiveness during sleep or transitions between consciousness states. In such cases, the processor identifies time-locked EEG components (e.g., P300, MMN) in response to auditory or tactile stimuli and compares these to normative templates, enabling assessment of residual cognitive function or sleep depth.
[0189] Additional EEG processors may integrate posture or movement information — derived from auxiliary sensors or from other physiological signal channels — with EEG-derived sleep staging to enable posture-aware sleep diagnostics. For instance, combining supine-position detection with REM stage identification may help quantify posture-related risk in sleep apnea or nocturnal seizure susceptibility. In further embodiments, EEG processors may support closed-loop stimulation systems in which real-time EEG signals are monitored for specific sleep stages or transitions, and upon detection, auditory, tactile, or vestibular stimuli are triggered to modulate sleep depthAttorney Docket No 70009- 127-C- WO
[0190] Client case 26590
[0191] or improve sleep continuity. These systems rely on fast and accurate EEG processing to deliver stimulation during precisely defined neurophysiological states, such as enhancing slow-wave activity during N3 sleep or suppressing undesired awakenings.
[0192] Further extending these embodiments, the EEG processors may be configured to detect non-neurological conditions that are indirectly reflected in EEG characteristics. For example, changes in EEG dynamics may be indicative of congestive heart failure, chronic obstructive pulmonary disease, diabetes, or immune-mediated diseases such as rheumatoid arthritis, lupus, and inflammatory bowel disease. Such conditions can manifest through sleep disturbance, cortical slowing, or autonomic imbalance that subtly modulates EEG rhythms. In these cases, the EEG processor may derive features reflecting these broader physiological disruptions and apply a diseasespecific classifier to detect or screen for such conditions.
[0193] In some embodiments, the EEG processors are configured to detect physiological or pathological states associated with conditions beyond those affecting the central nervous system. For example, the processors may be trained to detect or screen for systemic diseases including congestive heart failure (CHF), chronic obstructive pulmonary disease (COPD), diabetes mellitus, or immune-mediated inflammatory diseases such as rheumatoid arthritis (RA), lupus erythematosus, and inflammatory bowel disease (IBD). These conditions may manifest indirectly in the EEG through autonomic dysregulation, sleep architecture disruption, or cortical slowing. The processor may compute time-frequency features, connectivity indices, or nonlinear signal metrics from the estimated EEG signal, and apply a trained classifier to generate a diagnostic score or categorical prediction indicating the likelihood of the condition.
[0194] 6 DOWNSTREAM PROCESSING AND EEG PROCESSORS
[0195] The EEG processors 170A and 170B are configured to receive an estimated EEG signal 165 and produce task-specific outputs, referred to as results 175 A and 175B. Each EEG processor may be implemented as a signal-processing pipeline, a machinelearning classifier, a heuristic rule-based system, or a combination thereof, depending on the clinical objective and the target physiological or pathological state. These processors are configured using respective processor parameters 172A and 172B. which may include model weights, thresholds, tuning coefficients, or conditionspecific feature specifications.
[0196] The generated EEG signal 165 is then processed by one or more EEG processors 170A and 170B. Each processor receives the estimated EEG signal and appliesAttorney Docket No 70009- 127-C- WO
[0197] Client case 26590
[0198] analysis procedures tailored to a specific diagnostic, prognostication, or monitoring task. In some embodiments, the EEG processor implements a machine-learning classifier trained on EEG spectrograms to detect the presence of neurological, cardiovascular, respiratory, or metabolic diseases, such as Alzheimer's disease.
[0199] Parkinson's disease, congestive heart failure (CHF), or chronic obstructive pulmonary disease (COPD). The processor may also detect patterns associated with pharmaceutical intake, including the presence of drugs such as antidepressants, anticholinergics, anticoagulants, benzodiazepines, or insulin. The EEG processors operate using respective processor parameters 172A and 172B, which may include model weights, thresholds, or rule sets.
[0200] In alternative embodiments, the EEG processors may make use of specific EEG-derived biomarkers that are either explicit in the EEG signal or derived from the EEG signal, such as spectral power in defined frequency bands, sleep spindle characteristics, coherence measures, or event-related potentials (ERPs), and use these as inputs to the classification task. The processors may also operate on intermediate token representations or model embeddings, bypassing reconstruction of the full time domain EEG signal. In some cases, the processors support real-time or near real-time feedback to a clinician or to the subject via external monitoring systems.
[0201] The outputs of the EEG processors are results 175A and 175B, which may be diagnostic decisions, alerts, scores, or inferred physiological states. These results may¬ be used for real-time decision-making, longitudinal health monitoring, or clinical documentation. In embodiments that implement multiple downstream tasks, the same estimated EEG signal 165 may be routed to multiple EEG processors operating in parallel or in sequence.
[0202] Collectively, these EEG processor embodiments extend the utility7of the estimated EEG signal across a wide spectrum of diagnostic, monitoring, and therapeutic applications. Each processor may be trained independently and operate in parallel with others, thereby enabling multimodal physiological inference from a single, non-invasive respiratory input.
[0203] The electroencephalography representation generated from the physiological signal may be supplied to a downstream system configured to process naturally acquired EEG signals. Such a downstream system may comprise a machine-learning, algorithmic, rule-based, or heuristic system developed or validated using naturally acquired EEG signals. Because the electroencephalography generation stage may be used with multiple different downstream systems, the same transformedAttorney Docket No 70009- 127-C- WO
[0204] Client case 26590
[0205] representation can provide utility across multiple detection, screening, prognosticating, or monitoring tasks.
[0206] 7 RADIO-FREQUENCY A D I DIRECT PHYSIOLOGICAL SENSING
[0207] In the discussion above, use of radio-frequency sensing is introduced. As one example, radio reflections from a subject’s body may represent non-EEG physiological processes. One such manifestation of such processes is through movement of the subject, such as chest movement associated with breathing and / or heart beating. While the description above may focus on determining a '‘breathing” or a '‘heat beat” signal, the non-EEG physiological signal may depend on a combination of physiological signals, such as breathing and heart beating, and it is not necessary' to attempt to extract those signals explicitly. That is, the reflection RF signal, or a processing of the signal, for example, to extract relative phase variation with respect to a transmitted RF signal, may be used as input to a mapper that produces EEG codes in a manner analogous to the use to the breathing signal tokenizer 140 of FIG. 1.
[0208] A further set of embodiments explicitly contemplates using raw sensor modalities that are not conventional physiological waveforms as the direct input to the estimator. For example, radio-frequency (RF) reflections, radar phase traces, WiFi channel state information, ultra-wideband impulse responses, Doppler spectrograms, or other contactless sensing outputs may be provided directly to the mapping model either in raw form or after lightweight preprocessing. In these embodiments the system need not explicitly compute a breathing time series; instead the RF or other contactless data stream is processed, optionally tokenized or embedded, and conditioned on by the transformer or other sequence model to yield EEG tokens, embeddings, or reconstructions. RF-input embodiments may leverage sensor-specific preprocessing such as clutter suppression, beamforming, phase unwrapping, range gating, or coherent demodulation; alternatively, these preprocessing steps may be subsumed into learned front-end encoders trained jointly or sequentially with the downstream mapping modules. This class of embodiments covers both explicitly decoded breathing signals derived from RF (and then used as the input modality) and embodiments where the RF data itself is handled as the input modality without an intervening explicit breathing estimate.
[0209] In these and related embodiments, the physiological signal may reflect or encode physiological activity without necessarily constituting a direct physiological measurement. The input may therefore comprise a signal that depends on breathing, heart beating, motion, or another physiological activity, and the disclosed mapping orAttorney Docket No 70009- 127-C- WO
[0210] Client case 26590
[0211] inference operations may be performed directly on that signal or on a processed version thereof.
[0212] As used herein, "‘proximate non-contact measurement” refers to measurement of a physiological signal using a sensor positioned near the subject's body, for example, within a few centimeters to a few tens of centimeters, without requiring physical contact between the sensor and the subject’s skin. Examples include capacitive or inductive sensors embedded in a mattress or chair that measure respiratory motion or cardiac activity from proximity, optical sensors positioned near the body that measure photoplethysmographic signals without skin contact, millimeter-wave radar sensors worn at close range, and near-field electromagnetic sensors embedded in wearable housings that detect physiological signals through a thin air gap. Proximate noncontact measurement is distinguished from RF reflection-based remote sensing, in which the sensor may be positioned at a distance of one meter or more from the subject, and from contact measurement, in which electrodes or transducers are placed directly on the subject’s skin or body surface. The terms “in contact with or proximate to the subject,” as used in the claims, are intended to encompass both contact and proximate non-contact measurement modalities, as distinguished from remote sensing modalities such as RF reflection at distance.
[0213] 8 MULTIPLE INPUT SIGNALS
[0214] In some embodiments, the system is configured to receive multiple non-EEG physiological signals as inputs and to generate an estimated EEG signal based on a combined representation of these signals. Such an approach may be referred to as a multi -modality input. Referring again to FIG. 1, the signal acquisition module may acquire a plurality of non-EEG biosignals, including but not limited to a breathing signal 112, a cardiac signal 113 (such as an electrocardiogram (ECG) or photoplethysmogram (PPG)), an electromyogram (EMG) signal, and a motion or posture signal 116 (e.g., derived from an accelerometer, gyroscope, or radiofrequency sensor). Each of these signals may be individually preprocessed, tokenized, or otherwise encoded into discrete representations using respective signal tokenizers. and subsequently passed to a shared token transformer 150 that integrates information across modalities. The transformer outputs a sequence of EEG codes 155, which are decoded by the EEG generator 160 into an estimated EEG signal 165.
[0215] In this configuration, the system leverages complementary physiological information encoded in each input signal. For example, the respiratory signal may contain information about sleep stage-dependent breathing patterns, while the cardiacAttorney Docket No 70009- 127-C- WO
[0216] Client case 26590
[0217] signal reflects autonomic tone, heart rate variability, and other neurocardiac correlates of cortical arousal or sleep depth. Posture and movement signals may be indicative of sleep-wake transitions or state fragmentation, and may also disambiguate context in which certain EEG features are expected or suppressed. By integrating these signals within a unified transformer architecture, the model can resolve ambiguities or missing information that may arise when only a single input modality is available.
[0218] In one such embodiment, the breathing signal is tokenized by a breathing signal tokenizer 137 as previously described, the cardiac signal is tokenized by a cardiac signal tokenizer 141 (which may, for instance, quantize ECG waveforms into morphology- or rhythm-based tokens), and the motion signal is tokenized by a motion signal tokenizer 140. The resulting token sequences are concatenated or otherwise combined and passed through the token transformer 150. The output sequence of EEG codes 155 is decoded to produce a spectrogram or waveform representation of the estimated EEG signal 165. In some implementations, the transformer includes modality-specific attention heads or cross-attention blocks that allow it to attend to each input stream selectively, depending on the task or context.
[0219] This multi-signal configuration enables more robust and accurate EEG synthesis, particularly in cases where any individual input signal may be degraded or insufficient to recover all relevant EEG patterns. For example, during obstructive apneic events where respiratory signals may be compromised, cardiac and motion data may provide indirect indicators of cortical arousal. Likewise, during motionless periods, cardiac variability' may reveal sleep-stage transitions that influence EEG morphology. The resulting estimated EEG signal may thus reflect a more complete and artifact-resistant reconstruction of brain activity, supporting improved performance in downstream applications such as sleep staging, cognitive monitoring, or disease detection.
[0220] To enhance the model’s ability to generalize across varied sensor availability', a modality dropout strategy may be applied during training. In this approach, one or more non-EEG input signals, such as breathing, posture, or cardiac signals, are randomly omitted from the model’s input with a predetermined probability during each training step. This trains the model to be robust to missing or degraded modalities and enables the system to flexibly operate under different sensor configurations at inference time. For example, by training the model to occasionally predict EEG using only breathing input, the system can maintain performance in wearable or home-use environments where fewer sensors are available or where certain signals may be unreliable.Attorney Docket No 70009- 127-C- WO
[0221] Client case 26590
[0222] In some implementations, the system is configured to generate an estimated EEG signal using a variable set of input modalities at inference time. For example, the system may operate using breathing signal input alone in one setting, and using a combination of breathing and cardiac signals in another. The architecture accommodates such variability by training with multi-modal inputs and by learning shared latent representations that support generalization across modality combinations. During inference, the system dynamically selects or attends to available input channels and generates an EEG output based on the observed signals. This flexibility enables deployment of the same model across diverse devices and use cases, including mobile, ambulatory7, or in-clinic settings with different sensor capabilities.
[0223] In an embodiments, the model may receive posture information as an auxiliary input for estimating EEG or for refining EEG-based predictions. Posture may be derived from accelerometer data, bed-based pressure sensors, or RF -based spatial localization systems, and may be discretized into classes such as supine, left-lateral, right-lateral, or prone. The model incorporates posture as a contextual feature to modulate EEG generation or interpretation, recognizing that certain EEG features, such as REM atonia or arousal profiles, may vary systematically with sleep posture. In some cases, posture input alone may suffice to constrain the model to predict classtypical EEG patterns, particularly in well-characterized behavioral regimes such as supine REM sleep.
[0224] Multi-modal embodiments are supported in which any arbitrary7subset of available physiological channels may be tokenized or embedded and provided to a shared integrative model. The integrative model may include modality -specific encoders and cross-attention or fusion modules that selectively weight and combine information across breathing, cardiac (ECG or PPG), electromyogram (EMG), movement / posture (accelerometer / gyroscope or bed sensors), oximetry, temperature, and RF channels. During training, modality dropout strategies may be applied so that the model leams to operate robustly under variable sensor availability7; at inference time the model adapts to whichever subset of inputs is present. Modality-specific tokenizers and embedding projections may be learned jointly with the integrative model or pre-trained and frozen, and cross-modal attention heads, gating mechanisms, or late fusion classifiers may be used to improve interpretability and performance in heterogeneous deployment contexts.Attorney Docket No 70009- 127-C- WO
[0225] Client case 26590
[0226] 9 END-TO-END TRAINING AND REPRESENTATION-SPACE LEARNING
[0227] In some embodiments, a method for training a neural network system to generate a neural-activity' representation from physiological signals comprises receiving electroencephalography signals, training an encoder model to map the electroencephalography signals to sequences of codes, training a decoder model to reconstruct electroencephalography representations from the sequences of codes, receiving physiological signals and corresponding electroencephalography signals, training a mapping model to generate sequences of codes from the physiological signals based on a loss function dependent on at least one of similarity between codes, similarity between reconstructed electroencephalography representations, or performance of a downstream task, and configuring the mapping model to generate neural-activity representations from physiological signals.
[0228] In some embodiments, the electroencephalography signals and the physiological signals are received as paired or corresponding training examples, and the trained mapping model, encoder model, and decoder model are configured such that the mapping model generates neural-activity representations from physiological signals during runtime use.
[0229] In some embodiments, the system illustrated in FIG. 1 is trained, or ‘Tine-tuned’' with incremental training for already trained separate components, in an end-to-end manner, such that the breathing signal tokenizer 137, the token transformer 150, and the EEG generator 160 are optimized jointly. This integrated training strategy enables the entire model stack to learn a unified representation that translates breathing signals into estimated EEG signals 165 with high fidelity, while also preserving features relevant for clinical or physiological interpretation. Training is conducted using synchronized breathing and EEG recordings acquired from a diverse set of human subjects, and may optionally include additional labels for downstream tasks such as sleep staging, disease classification, or pharmacological status.
[0230] The training data may consist of a large-scale, multi-cohort collection of corecorded breathing and EEG signals gathered under standardized polysomnographic (PSG) protocols. In one embodiment, the training dataset includes tens of thousands of overnight recordings from subjects aged across the lifespan (e.g., 3 to 100+ years), with variation in sex, race, and health status. These datasets may be aggregated from institutional sleep clinics, multi-site clinical trials, or open-access research repositories, and may span multiple sensor modalities. The breathing signal may be obtained directly using a respiratory belt, nasal cannula, accelerometer, or contactless RF sensor, or indirectly derived from electrocardiography (ECG-derived respiration,Attorney Docket No 70009- 127-C- WO
[0231] Client case 26590
[0232] EDR) or photoplethysmography (PPG-derived respiration). The EEG signal is typically obtained using standard scalp electrode placements, such as C4-A1 or 01-A2 montages, consistent with American Academy of Sleep Medicine (AASM) guidelines.
[0233] To enhance model generalization and robustness, a variety of data augmentation techniques may be applied during training. These may include amplitude scaling, temporal jittering, time- warping, random masking of signal intervals, synthetic noise injection, and spectral filtering. Augmentation strategies may be applied to the breathing signals, the EEG signals, or both, and may be targeted to simulate real-world recording variability' such as movement artifacts, sensor dropout, or electrode impedance changes. Such augmentations enable the model to leam invariant features and prevent overfitting to specific devices, sites, or populations.
[0234] Training is typically performed using stochastic optimization methods, such as Adam or AdamW, applied over mini-batches of synchronized breathing-EEG pairs. Each training sample is tokenized by the breathing signal tokenizer 137, producing a sequence of discrete breathing tokens 145. These tokens are processed by the token transformer 150, which outputs a predicted sequence of EEG codes 155. The EEG tokens are decoded by the EEG generator 160 into an EEG spectrogram or waveform 165. A primary training loss, such as mean squared error (MSE) or perceptual loss, is computed between the generated EEG and the corresponding ground-truth EEG signal. Additional losses may include vector quantization commitment loss (in VQ-style architectures), cross-entropy over token indices (for the transformer), and adversarial losses (if a discriminator is used in the EEG generator).
[0235] In embodiments where downstream EEG processors 170A and 170B are included in the training loop, the training loss further incorporates one or more task-specific objectives, such as classification accuracy for disease detection or agreement with expert sleep stage labels. This allows the model to optimize not only for EEG fidelity , but also for clinical utility. For example, the system may be trained to maximize the area under the receiver operating characteristic curve (AUC) for detecting Parkinson’s disease or to minimize classification error for sleep stage prediction.
[0236] Validation and evaluation protocols are employed to ensure generalizability and guard against overfitting. In one embodiment, the dataset is partitioned into training, validation, and test sets using stratified sampling at the subject level, such that all recordings from a given individual reside in a single split. Internal validation is conducted using k-fold cross-validation, in which training is repeated across k partitions and performance metrics are averaged. External validation is performedAttorney Docket No 70009- 127-C- WO
[0237] Client case 26590
[0238] using held-out datasets acquired from independent sites, sensor platforms, or demographic cohorts, thereby confirming that the model generalizes beyond the training distribution. The model evaluation may be conducted using stratified analysis across distinct demographic and clinical subpopulations, including age. sex, disease class, and comorbidity groupings. This stratified evaluation enables the detection of performance disparities and supports equitable generalization. In addition, calibration techniques may be applied to align the statistical properties of the model-generated EEG with those of true EEG recordings. These techniques may include histogram matching, distributional regularization, or adaptive scaling, and may be used to improve fi del i ty in power spectral density, amplitude distribution, or sleep stagespecific metrics. Such calibration enhances interpretability and facilitates integration with downstream clinical decision systems.
[0239] Evaluation metrics include both signal-level measures, such as mean absolute error (MAE), signal-to-noise ratio (SNR), and spectral fidelity, as well as task-specific metrics such as classification accuracy, AUC, precision-recall curves, or confusion matrices. These metrics are calculated across a range of subpopulations (e.g., age, sex, disease class) to assess performance equity and identify potential model biases. In some embodiments, calibration techniques may be applied to align the statistical properties of the model-generated EEG with those of the ground truth, thereby improving reliability in downstream applications.
[0240] This end-to-end training approach enables the system to leam cohesive and biologically plausible mappings from respiratory' dynamics to neural activity, while ensuring that the generated EEG signals are usable across a wide range of clinical, research, or consumer health monitoring applications.
[0241] In some embodiments, the mapping model is trained using only a downstream task loss, without any reconstruction loss or code-similarity' loss. In such embodiments, the mapping model receives physiological signals as input and is trained to optimize performance on one or more downstream tasks, such as disease classification, pharmacological exposure detection, or sleep staging, using labeled training data. The mapping model leams to generate sequences of codes or intermediate representations that are maximally informative for the downstream task, without being constrained to reproduce a particular EEG waveform or spectrogram. This approach may be preferred where labeled downstream task data is available in sufficient quantity, where reconstruction fidelity is less important than task performance, or where end-to-end optimization of the full pipeline yields better taskspecific results than staged training. The loss functions recited in the training method described herein, in particular similarity between codes, similarity betweenAttorney Docket No 70009- 127-C- WO
[0242] Client case 26590
[0243] reconstructed electroencephalography representations, and performance of a downstream task, may each be used alone or in any combination, and no particular combination is required.
[0244] In some embodiments, the encoder model and decoder model are trained in a first stage before training of the mapping model begins, such that the representation space is established prior to learning the cross-modal mapping. This staged training procedure ensures that the mapping model is trained to generate codes that conform to an already-stable representation space, potentially improving convergence and reducing the risk that the representation space collapses or degenerates during joint optimization. In other embodiments, training of the mapping model and training of the decoder model are performed jointly, enabling the decoder to adapt to the distribution of codes produced by the mapping model and potentially improving end-to-end reconstruction fidelity.
[0245] In some embodiments, training of the mapping model comprises a masked prediction task, in which a subset of the code positions in the target EEG code sequence is masked, and the mapping model is trained to predict the masked codes from the physiological signal input and the unmasked context. This objective encourages the mapping model to leam globally coherent representations of physiological dynamics rather than relying on local or positional cues.
[0246] In some embodiments, training comprises contrastive learning between physiological signals and corresponding electroencephalography signals, in which pairs of physiological and EEG signals recorded from the same subject during the same interval are treated as positive pairs, and signals from different subjects or intervals are treated as negative pairs. Contrastive objectives encourage the mapping model to generate codes for physiological signals that are similar in the representation space to codes generated by the EEG encoder for the corresponding EEG signals, without requiring exact reconstruction.
[0247] In some embodiments, the physiological signals and corresponding electroencephalography signals used fortraining are temporally synchronized, with alignment performed to account for any recording delay or latency between the two signal streams. In some embodiments, training the mapping model comprises explicitly compensating for a delay between the physiological signals and the electroencephalography signals, for example by applying a learned or fixed temporal offset during alignment of training pairs. Failure to account for such delays can degrade training signal quality and reduce the fidelity of the learned mapping.Attorney Docket No 70009- 127-C- WO
[0248] Client case 26590
[0249] In some embodiments, the electroencephalography representations are segmented into windows of fixed or variable duration during training, and the mapping model is trained to generate codes corresponding to each window independently or in sequence. Window segmentation enables the model to process arbitrarily long recordings during inference by applying the trained mapping model in a sliding or non-overlapping window fashion.
[0250] In some embodiments, the reconstruction loss used during training comprises a time-domain loss, a spectral-domain loss, a perceptual loss computed in a learned feature space, an adversarial loss, or any combination thereof. The specific form of the reconstruction loss may be selected based on the fidelity requirements of the dow nstream application.
[0251] In some embodiments, the mapping model is fine-tuned for a target sensor ty pe or a target subject population after initial training on a broader dataset. Fine-tuning enables adaptation of the model to the specific signal characteristics, noise properties, and physiological distributions of a particular deployment context without requiring retraining from scratch. In some embodiments, training comprises domain adaptation across different sensor hardw are or acquisition environments, enabling the mapping model to generalize across devices, electrode configurations, or clinical sites. In some embodiments, configuring the mapping model comprises calibrating the model using subject-specific physiological data, enabling personalized adjustment of the model’s mapping to account for inter-individual differences in physiological signal characteristics.
[0252] Architectural variants and training regimes are described herein without limiting the claims to any particular variant. Masked autoencoding objectives, autoregressive likelihood objectives, contrastive objectives, triplet or metric learning losses, adversarial losses, perceptual losses computed in feature spaces, and downstream task losses may be combined in multi-objective training regimes. In some embodiments the mapping model is trained as a foundation or pretraining stage using large paired corpora of non-EEG and EEG data and then fine-tuned with smaller, task-specific labeled datasets; in other embodiments the mapping and downstream classification objectives are optimized jointly in a multi-task learning setup so as to maximize endtask utility. Training data augmentation methods — such as amplitude scaling, timewarping, masking, additive noise, synthetic artifact injection, and sensor-specific distortions — may be applied to increase robustness and to reduce overfitting to particular acquisition devices or sites.Attorney Docket No 70009- 127-C- WO
[0253] Client case 26590
[0254] 10 DENOISING
[0255] In some embodiments, a method for improving an electroencephalography signal of a subject comprises acquiring, using at least one sensor, an electroencephalography signal including noise, artifacts, or missing intervals, acquiring, using at least one sensor, a non-electroencephalography physiological signal representing a physical measurement of the subject, processing the electroencephalography signal and the physiological signal using one or more trained machine learning models to generate a sequence of codes, generating, from the sequence of codes, a corrected electroencephalography representation, and outputting the corrected electroencephalography representation having reduced artifacts or reconstructed signal components. In some embodiments, generating the sequence of codes compnses conditioning on the physiological signal.
[0256] In some embodiments, the corrected electroencephalography representation has improved signal-to-noise ratio, improved spectral fidelity, or both, relative to the acquired electroencephalography signal including noise, artifacts, or missing intervals.
[0257] The noise, artifacts, or missing intervals present in the electroencephalography signal may comprise motion artifacts, muscle artifacts, electrode pop artifacts, line noise, baseline drift, or missing channels. In some embodiments, the missing intervals comprise temporally contiguous missing segments of the electroencephalography signal, for example resulting from electrode disconnection, signal saturation, or data loss during transmission. In some embodiments, the corrected electroencephalography representation is generated in real time, enabling continuous monitoring applications in which artifact-affected EEG signals are corrected as they are acquired. In some embodiments, the corrected electroencephalography representation is generated offline from stored recordings, enabling retrospective improvement of historical data.
[0258] In certain embodiments, a denoising operation is performed to enhance the quality of EEG signals by leveraging concurrently acquired breathing signals.
[0259] Referring again to FIG. 1, this denoising process may be implemented within or downstream of the EEG generator 110, whereby the generator receives as inputs both a noisy EEG signal and a corresponding breathing signal, and outputs a denoised EEG signal. The goal of this configuration is to reduce or eliminate noise artifacts in the EEG recording, such as those arising from poor electrode contact, movement artifacts, or ambient electrical interference, by utilizing the information-rich, less noise-prone respiratory signal to constrain and regularize the reconstruction of the brain signal. Such embodiments therefore can suppress electrode impedance artifacts, motionAttorney Docket No 70009- 127-C- WO
[0260] Client case 26590
[0261] artifacts, or environmental electrical interference present in an acquired EEG signal by leveraging temporal correlations between the non-EEG physiological signal and the EEG signal.
[0262] In one implementation, the EEG generator 110 comprises a neural network model trained to generate a clean EEG signal conditioned on both the noisy EEG input and a tokenized representation of the breathing signal. The model may include an encoder that maps the noisy EEG and breathing signal into a joint latent representation, and a decoder that reconstructs a cleaner EEG from this representation. The joint latent representation facilitates the isolation of physiological patterns that are consistent between the EEG and respiratory' domains, thereby suppressing components of the EEG that are not explained or corroborated by the respiratory signal.
[0263] Alternatively, the denoising may be formulated as a conditional generation task in which a transformer-based model, such as the token transformer described above, is trained to predict clean EEG tokens from a sequence of noisy EEG tokens and aligned breathing tokens. The transformer leams to mask and replace noisy EEG segments that are inconsistent with respiratory dynamics. During training, pairs of noisy-clean EEG signals and breathing data are used, enabling the model to learn mappings from corrupted to clean EEG conditioned on the corresponding breathing sequence.
[0264] In a further embodiment, the EEG spectrogram derived from the noisy signal may be passed through a tokenization module (such as a VQGAN encoder) to obtain a sequence of noisy EEG tokens. Simultaneously, the breathing signal is tokenized via the breathing signal tokenizer. These two token sequences are then input to a generative transformer that is trained to reconstruct the clean EEG token sequence. The output token sequence may then be decoded by a VQGAN decoder to produce a denoised EEG spectrogram. In some embodiments, the transformer is trained using a masked autoencoding strategy that encourages the model to fill in masked segments of the EEG sequence in a manner consistent with both the observed noisy EEG and the breathing signal.
[0265] This approach takes advantage of the finding that Al-generated EEG signals derived from breathing signals can exhibit improved signal fidelity relative to groundtruth EEGs in the presence of recording artifacts. By incorporating the respiratory' signal during denoising, the system can identify and reconstruct EEG patterns — such as posterior dominant rhythms, sleep spindles, or slow wave activity — that may be missing or corrupted in the raw EEG but are nonetheless inferable from breathing patterns. In particular, this enables restoration of sleep-related EEG features relevantAttorney Docket No 70009- 127-C- WO
[0266] Client case 26590
[0267] to downstream analyses such as sleep staging, disease detection, or pharmacological effect assessment.
[0268] Through these mechanisms, the denoising process improves the quality’ and interpretability of EEG signals and may enhance the performance of downstream EEG processors 170A-170B in clinical tasks such as disease screening, drug detection, or cognitive state estimation.
[0269] In certain embodiments, the denoising approach described above is applied in the context of wearable or consumer electronic devices that acquire EEG signals of limited quality’ due to hardware constraints or deployment conditions. For example, earbud-based EEG systems acquire EEG signals using electrodes positioned in or around the ear canal, which are subject to substantial motion artifacts, poor electrode contact, and mechanical noise arising from jaw movement, ambient vibration, and subject activity^. Similarly, dry -electrode headbands, behind-the-ear sensors, over-ear or on-ear headphones, and other form-factor-constrained EEG acquisition systems produce recordings that are frequently degraded relative to clinical-grade scalp EEG. Such devices are used, for example, in sleep monitoring, biofeedback, and consumer brain-computer interface applications. In such embodiments, one or more non-EEG physiological signals acquired from the same or a different wearable device are used to condition the denoising model. For example, a photoplethysmogram (PPG) signal acquired from an earbud, a smartwatch, or a finger-worn sensor, a respiratory signal derived from a chest-wom accelerometer or smartwatch motion sensor, an electromyogram (EMG) acquired from a wearable sensor, or a cardiac signal derived from a smartwatch ECG electrode, may each serve as the conditioning physiological signal. The conditioning signal need not be acquired from the same device as the EEG signal; signals from separate wearable devices worn simultaneously by the subject are expressly contemplated. The trained denoising model uses temporal correlations between the conditioning physiological signal and the underlying neural activity to reconstruct physiologically consistent EEG spectral structure, replacing or suppressing artifact-corrupted intervals and restoring clinically or analy tically useful EEG features such as sleep spindles, slow-wave activity7, or posterior dominant rhythms. Such embodiments enable high-quality EEG analysis from consumer-grade wearable platforms that would otherwise produce signals of insufficient quality for clinical or research use.
[0270] Denoising and calibration embodiments are also contemplated. In some configurations, an available noisy or artifact-contaminated EEG recording is provided together with one or more non-EEG signals and the model is trained or applied to reconstruct a cleaned EEG. This may be implemented as a conditional generation taskAttorney Docket No 70009- 127-C- WO
[0271] Client case 26590
[0272] in which the noisy EEG and the non-EEG modalities are jointly encoded and a cleaner EEG output is produced, or as a masking / reconstruction task in which corrupted EEG tokens are replaced using contextual information supplied by breathing, cardiac, motion, or RF channels. Calibration embodiments, in which statistical properties (for example average power, spectral baseline, or dynamic range) of the generated EEG are adjusted to match those of a reference population or to align with the expected input distributions of downstream processors, are likewise supported. Calibration techniques may include histogram matching, per-band scaling, adaptive gain control, distributional regularization during training, or learned calibration heads trained to minimize downstream task loss.
[0273] 11 ALTERNATIVES, APPLICATIONS, AND APPLICATIONS
[0274] In some embodiments, downstream inference for disease condition assessment or pharmacological exposure determination is performed using an intermediate representation obtained from an internal layer of the trained neural network, rather than from the sequence of codes at the output of the mapping model or from a fully reconstructed electroencephalography representation. The trained neural netw ork generates a hierarchy of internal representations as it processes the input physiological signal, and any of these representations, including representations at layers below' the final code output, may encode physiologically or clinically relevant information suitable for downstream inference. In such embodiments, the downstream inference model receives as input one or more internal layer activations of the trained neural network, rather than the final sequence of codes, and generates the indication of disease condition or pharmacological exposure therefrom. This approach may reduce computational complexity relative to full EEG reconstruction, may improve taskspecific performance by exploiting features that are intermediate between raw physiological dynamics and EEG-structured codes, and may enable flexible integration with downstream inference architectures that are not constrained to accept a fixed-length code sequence. The specific internal layer or layers used for downstream inference may be selected based on task performance, and no particular layer is required.
[0275] In some embodiments, the physiological signal used as input to any of the methods described herein is obtained while the subject is in a sleep state, partially or fully. Sleep acquisition is applicable across all three primary use cases described herein: generation of an electroencephalography representation, determination of pharmacological exposure, and assessment of a risk or condition related to the health of the subject. Sleep provides a physiologically stable and reproducible measurementAttorney Docket No 70009- 127-C- WO
[0276] Client case 26590
[0277] context with reduced voluntary motion, and many clinically relevant physiological signatures, including disruptions to sleep architecture, altered spectral power distributions, changes in autonomic regulation, and pharmacological effects on sleep stages, are most clearly expressed during sleep. In some embodiments, the acquired physiological signal spans a full overnight recording; in other embodiments, only a selected interval of sleep, such as a period of confirmed non-REM sleep or a period following sleep onset, is used. The electroencephalography representation, indication of pharmacological exposure, or indication of disease risk or condition generated from sleep-state signals may be used in any of the downstream applications described herein, including sleep staging, arousal detection, sleep quality assessment, sleep disorder screening, neurodegenerative disease screening, metabolic condition monitoring, and pharmacological monitoring.
[0278] In some embodiments, a plurality of physiological signals associated with the first physiological subsystem are obtained simultaneously and processed jointly by the trained neural netw ork. Joint processing of multiple simultaneous signals may improve the accuracy and robustness of the generated sequence of codes by providing complementary information about physiological dynamics from different sensors or measurement perspectives. For example, a respiratory signal and a cardiac signal obtained simultaneously from the same subject may jointly encode autonomic interactions that are not fully captured by either signal alone.
[0279] In some embodiments directed to pharmacological exposure determination, the pharmacological agents of interest include sedatives, antidepressants, opioids, stimulants, anesthetics, antiepileptics, and cardiovascular drugs, among others.
[0280] Sedatives, including benzodiazepines and non-benzodiazepine sleep aids, are associated with characteristic effects on sleep architecture, including suppression of slow-wave sleep and alteration of spindle density. Antidepressants, including selective serotonin reuptake inhibitors and tricyclic antidepressants, are associated with suppression of REM sleep and changes in theta band dynamics. Opioids are associated with suppression of slow-w ave sleep and respiratory depression that may be reflected in cross-modal physiological signals. Stimulants are associated with reduced total sleep time and altered spectral composition during sleep. Anesthetics and antiepileptics produce distinct spectral signatures that may be detected in both EEG-derived and physiological representations.
[0281] Cardiovascular agents, including beta-blockers, antiarrhythmics, antihypertensives, and related agents, alter cardiac rate, rhythm, contractility, vascular tone, or conduction in ways that produce characteristic signatures in cardiac-derived and respiratory physiological signals. Anticonvulsants affect neuronal excitability andAttorney Docket No 70009- 127-C- WO
[0282] Client case 26590
[0283] produce characteristic alterations in EEG background activity and spectral composition that may be reflected in EEG representations generated by the disclosed methods. Autonomic nervous system agents, including anticholinergic agents and adrenergic agents, modulate the balance between sympathetic and parasympathetic tone, producing detectable effects on heart rate variability, respiratory rate, and related physiological dynamics. Bronchodilators, including beta-2 adrenergic agonists and xanthine derivatives such as theophylline, affect respiratory mechanics and airway dynamics in ways that are reflected in respiratory waveform characteristics; xanthine-class bronchodilators additionally produce central nervous system stimulation that may be reflected in EEG spectral activity7.
[0284] Antidiabetic agents, including insulins and oral hypoglycemic agents such as thiazolidinediones, affect glucose metabolism and autonomic homeostasis in ways that may be reflected in cardiovascular and respiratory7physiological dynamics and in EEG-derived representations. Antihistamines, particularly first-generation agents with central nervous system penetrance, produce sedation, anticholinergic effects, and characteristic changes in EEG spectral composition including increased slow-wave activity and reduced alpha power detectable in the representations generated by the disclosed methods. Respiratory' tract agents, including agents acting on airway smooth muscle, mucosal secretions, or bronchial tone, alter respiratory waveform morphology and dynamics in ways detectable in respiratory' physiological signals. Corticosteroids affect cortisol-mediated circadian regulation and produce characteristic effects on sleep architecture including suppression of slow-wave sleep and reduction of REM sleep, which are reflected in EEG spectral features and in physiological signals acquired during sleep. Sex hormones, including estrogens, progestins, and androgens, modulate sleep architecture, autonomic tone, and cortical activity7; progesterone in particular has GABAergic properties associated with increased slow -wave activity and altered spindle density7, and estrogenic and androgenic effects on sleep and autonomic regulation produce physiological signatures detectable in cardiac and respiratory signals and in EEG-derived representations.
[0285] The pharmacological exposure may comprise therapeutic, subtherapeutic, supratherapeutic, prescribed, or non-prescribed exposure, and the trained neural network may be configured to distinguish among these categories based on training data derived from subjects with known exposure status. In some embodiments, the trained neural network is trained using supervision derived from medication records, prescription data, clinician annotations, or treatment schedules, enabling the model to leam exposure-associated physiological signatures without requiring direct pharmacokinetic measurements or precise dose timing information.Attorney Docket No 70009- 127-C- WO
[0286] Client case 26590
[0287] In some embodiments directed to disease condition assessment, the normative representation space against which a subject’s codes are compared may be constructed by fitting a density model, such as a multivariate Gaussian, a mixture model, or a normalizing flow, to sequences of codes generated from a reference population of healthy subjects. The deviation of a new subject’s codes from the normative distribution, as measured by a distance metric or likelihood score, may be used as an indicator of disease risk or condition severity. This approach enables screening for pathological states as departures from normative physiological dynamics without requiring explicit disease labels for all training subjects, and supports generalization to disease conditions that may not have been represented in labeled training data.
[0288] Real-time, streaming, and low-latency embodiments are contemplated. Models may be implemented with causal architectures, windowed processing pipelines, incremental token generation, or chunked inference strategies that permit streaming input and near-real-time output suitable for closed-loop control. To support embedded or wearable deployment, model compression, quantization, pruning, knowledge distillation, and optimized inference kernels may be applied to reduce memory footprint and computational load. Where latency and determinism are important, hardware accelerators such as GPUs, NPUs. FPGAs, or ASICs may be used to perform tokenization, attention, and decoding operations with guaranteed timing characteristics.
[0289] Privacy, security, and data management embodiments are contemplated.
[0290] Intermediate embeddings or EEG tokens may be encry pted, obfuscated, or quantized prior to transmission to remote servers. Federated learning, split learning, and on-device fine-tuning embodiments may be used to update model parameters without centrally aggregating raw subject data; secure aggregation and differential privacy mechanisms may be applied to minimize disclosure risk. Furthermore, embodiments that provide only binary or probabilistic health outputs and that suppress or never reconstruct full EEG waveforms may be used to reduce privacy exposure where required by clinical or regulatory policies.
[0291] Additional practical embodiments include closed-loop therapeutic systems in which estimated EEG or derived biomarkers trigger actuators or interventions. Such systems may detect target neurophysiological states and generate stimulation, auditory cues, haptic feedback, or environmental adjustments in response, with timing, intensity, and safety logic governed by task-specific rules or learned controllers. Safety-critical embodiments may incorporate redundancy, watchdog modules, human-in-the-loop confirmations, and conservative thresholds to avoid unintended actuation.Attorney Docket No 70009- 127-C- WO
[0292] Client case 26590
[0293] In certain embodiments, the system is configured to operate in real-time, such that the breathing signal is continuously tokenized and processed by the token transformer to produce EEG token sequences, which are then decoded into estimated EEG signals without perceptible latency. The architecture supports low-latency inference by enabling causal or stream-based processing of token sequences, with minimal computational overhead. This configuration allows the system to be deployed in real-time monitoring applications, including ambulatory' physiological tracking, sleep intervention systems, and brain-computer interface (BCI) scenarios.
[0294] In certain embodiments, the estimated EEG signal is used to drive closed-loop control actions or therapeutic interventions. For example, the system may detect transitions into specific sleep stages, such as slow-wave sleep or REM, and trigger auditory stimulation, tactile feedback, or environmental modulation in response. Such control actions may be delivered through an actuator system in communication with the model, and may be used to enhance sleep quality', modulate neural plasticity, or suppress abnormal arousals. The ability to generate EEG in real time from noncontact or wearable sensors enables such closed-loop systems to operate continuously and non-invasively in naturalistic settings.
[0295] The disclosed systems and methods provide a practical application of machine learning techniques for generating and analyzing electroencephalography (EEG) signals from non-EEG physiological inputs, such as respiratory signals, cardiac signals, or posture information. Unlike abstract mathematical models or mental processes performed in the mind, the claimed invention is directed to a specific technological solution to a technical problem — namely, how to obtain high-quality, analysis-ready EEG signals in contexts where direct acquisition of EEG is infeasible, unreliable, or cost-prohibitive. The invention thus facilitates the deployment of EEGbased clinical decision systems using widely available and easily acquired physiological signals, enabling broader access to neurological and sleep-related diagnostics.
[0296] The techniques disclosed herein are implemented in conjunction with tangible data acquisition devices and specialized signal processing pipelines, including wearable respiratory sensors, accelerometers, and ECG or PPG monitors, that collect raw physiological signals in real-time from a human subject. These physiological signals are then transformed by trained tokenizers and generative models to yield estimated EEG signals that retain spectral and temporal fidelity appropriate for downstream clinical tasks. In some embodiments, the estimated EEG signal is further processed by task-specific EEG processors that generate real-time diagnostic outputs,Attorney Docket No 70009- 127-C- WO
[0297] Client case 26590
[0298] therapeutic recommendations, or device control signals, thereby producing concrete and useful outcomes with real-world applicability.
[0299] In some embodiments, the disclosed system comprises one or more sensors, one or more processors, memory storing instructions that, when executed by the one or more processors, cause the one or more processors to perform one or more of the methods described herein, and at least one output device. The output device may comprise a display device, a communication interface, an alarm generator, or a control interface, and may be configured to output an electroencephalography representation, an indication of pharmacological exposure, an indication of a disease condition, or a corrected electroencephalography representation. In some embodiments, the output device is configured to provide a signal to an external system.
[0300] Moreover, the described approaches enable non-invasive, ambulatory, and continuous health monitoring outside of traditional clinical environments, and supports closed-loop control systems that adaptively respond to detected sleep states or neurological features. For example, estimated EEG signals generated from a breathing belt and motion sensor may be used to trigger auditory stimulation during slow-wave sleep, thereby enhancing sleep quality and memory consolidation. Such applications tie the use of learned models and signal transformation techniques to specific, real-world technological uses, rather than to abstract ideas in the absence of practical implementation.
[0301] In further embodiments, the invention is deployed as part of real-time inference systems configured for low-latency operation on embedded hardware platforms. The disclosed models and components are not merely abstract data transformations; they are implemented as part of systems that interact with physical sensors, respond to environmental noise and artifact conditions, and produce outputs suitable for use in regulated medical workflows. These features demonstrate that the invention is not merely directed to an abstract idea, but is instead integrated into a practical application with significant technological improvements over conventional EEG acquisition and analysis systems.
[0302] Implementations of the approaches described above may be implemented using software comprising instructions that are stored on a non-transitory machine-readable (i.e., computer-readable) medium. Such instructions, when executed on a processor, cause the processor to perform steps described above and / or cause other components (e.g., sensors) to perform such steps (e.g., signal acquisition). The processor may be a general-purpose processor, or a special-purpose processor such as an array processor,Attorney Docket No 70009- 127-C- WO
[0303] Client case 26590
[0304] graphics processing unit tensor processor, digital signal processor, etc. The processor may be a virtual processor implemented that is implemented in software on a physical processor. Aspects of training are not necessarily implemented using the same computing resources as runtime aspects. For example, the various training steps may be implemented on a ser er computer, while the runtime steps may be implemented as integrated in medical equipment. Yet other implementations may use dedicated circuitry', such as application-specific integrated circuits or field-programmable gate arrays to perform some of the steps, such as signal transformation steps which may involve substantial computation or that require low latency in practice.
[0305] 12 RESULTS
[0306] 12.1 Participants and Datasets
[0307] A retrospective analysis was performed on a large and diverse dataset consisting of 28,141 nights from 22,616 individuals aged 3 to 102 years (mean 52.6, standard deviation 25.1), 44.7% of whom are female and 72.8% are white. The dataset was created by pooling several sleep datasets. Overall, data from 14 datasets were combined: 9 were used for training and cross-validation, while 5 were held out as external testing datasets. A table show n in FIG. 6 provides a summary of the dataset characteristics.
[0308] In all datasets, EEGs were recorded using scalp electrodes, following standard PSG procedures (FIG. 7A). Multiple electrodes captured electrical activity from different brain regions. The described experiments focus on two EEG channels recorded from electrodes C4 (center of the right side of the head) and Al (left ear), which are commonly used in sleep studies and present in the described datasets. The montage of these channels (i.e., C4-A1) is used to represent each individual's EEG signal. After standard preprocessing steps, such as down-sampling, filtering, and noise removal, the EEG power spectral density (PSD) is calculated using the standard multitaper method, which minimizes spectral leakage and provides high spectral resolution. FIG. 7B shows an example sleep EEG spectrogram, illustrating characteristic patterns such as posterior dominant rhythms (8-12 Hz) associated with a relaxed eyes-closed state23, sleep spindles (12-16 Hz) that help maintain sleep24, and slow w ave activity (0.5- 1.5 Hz) indicative of deep sleep25.
[0309] 12.2 Training and Evaluation Protocol
[0310] A distinction is made between internal datasets, which are included in training and cross-validation, and external datasets, which are used only for testing (see theAttorney Docket No 70009- 127-C- WO
[0311] Client case 26590
[0312] table in FIG. 6). For internal datasets, patients are randomly split into four folds, using three folds for training and the remaining fold for testing. This procedure is repeated four times so that each fold serves as an internal test set once, covering the entire dataset. If a dataset contains multiple visits per patient, all visits for that patient are assigned to either the training set or the test set. External datasets, on the other hand, are excluded from training and used strictly for testing. All performance metrics and statistical analyses are based on pooled predictions across the four test folds. All neural network models in this work are implemented in Python using PyTorch 1.7.1.
[0313] 12.3 EEG Generation Performance
[0314] The described model is evaluated for its ability to accurately translate breathing signals to EEG by comparing generated EEGs to ground-truth counterparts. A qualitative assessment is provided by visualizing representative examples. FIGs. 8a-d show four examples of Al-generated EEG spectrograms alongside their corresponding ground-truth EEG. A side-by-side comparison indicates that the model successfully translates a person’s breathing into their sleep EEG. The generated EEG captures the temporal structure in the ground truth and preserves key characteristics of the original spectrogram, including sleep spindles (FIG. 8a), the posterior dominant rhythm (FIG.
[0315] 8c), and slow wave activities (FIGs. 8a-d). A denoising effect is also observed in the generated EEG. Actual EEG recordings can be susceptible to artifacts — such as ECG contamination if an electrode is placed over a vein, resulting in stripe patterns at multiples of ~1 Hz. As illustrated in FIG. 8d, the described model produces EEG signals that exhibit reduced ECG artifacts compared to the ground truth.
[0316] A quantitative evaluation of how closely the generated EEG matches the groundtruth across all datasets is performed using the mean absolute error (MAE), a standard metric for images and spectrograms, which calculates the average absolute difference between predicted and actual values. In the described experiments, the EEG spectrogram values range from -20 dB to 30 dB and are normalized to [0, 1], The signal-to-noise ratio (SNR) metric is also used to quantify the error relative to the ground truth. The SNR is computed as: SNR=101ogl0S2S-S2, where S is the groundtruth EEG spectrogram and S is the model's output.
[0317] The table shown in FIGs. 9A-B presents the MAE and SNR of the described model across various datasets and EEG frequency bands. The errors are consistently low across most datasets, with the exception of STAGES and P18C, which exhibit higher errors. This increased variability' likely stems from the multi-site nature of the data sources in these datasets, leading to less consistent data distributions.
[0318] Specifically, as different sites may collect data using different devices and protocols,Attorney Docket No 70009- 127-C- WO
[0319] Client case 26590
[0320] the collected data may have higher noise and hence lower quality. In terms of frequency bands, the MAE errors are similar, reflecting balanced performance across bands. The model demonstrates strong generalization, as its performance on external datasets is comparable to that on internal datasets, reflecting model robustness.
[0321] 12.4 EEG Generation Performance across Demographics and Pre-existing Conditions
[0322] Model performance was further assessed across various subpopulations, considering three demographic factors: 1) age (five subgroups: 0-18, 18-40, 40-60, 60-80, 80-100 years), 2) sex, and 3) race (Asian, Black, White, Other). Figure 8a-c shows small differences in mean absolute error (MAE) between groups; overall, however, the model maintains consistently good performance. Slightly higher errors are observed for groups with fewer data samples — such as ages 18-40, as well as Black and Asian individuals. Higher errors are also observed for children (0-18), likely because they tend to have higher EEG power (especially in the delta band) and greater variance, which makes prediction more challenging. These differences may be mitigated by expanding the dataset to achieve more balanced representation across ages and races.
[0323] Model performance was also evaluated across five principal disease categories: cardiovascular, autoimmune, metabolic, neurological, and respiratory. Overall, the model performs well for these conditions (FIG. lOd). Slightly lower errors are observed among individuals with cardiovascular and metabolic conditions.
[0324] Conversely, those with neurological conditions exhibit marginally higher errors. Most of these differences can be attributed to data imbalance, and may improve with more data from individuals with neurological conditions.
[0325] 12.5 Performance on Downstream Tasks
[0326] The described experiments validate that the model captures two key aspects of health: the effects of diseases and therapies on human physiology. Some of the studied datasets include information about participants’ conditions and medications, enabling evaluation of model performance on dow nstream tasks related to disease screening and identifying drug effects. These tasks are framed as binary classification tasks, using AUROC as the evaluation metric. For each disease or medication, data are aggregated across relevant datasets and conditions with a sufficiently large number of positive and negative samples are selected.Attorney Docket No 70009- 127-C- WO
[0327] Client case 26590
[0328] 12.6 Screening for diseases and drugs: Comparison of Al-generated and groundtruth EEGs
[0329] These experiments verify that the Al -generated EEGs retain the same disease-and drug-related information as the original EEGs. Al models are trained that take a person’s EEG as input and output whether the person has a specific disease or is taking a particular medication. For each downstream task, a model that uses the AI-generated EEGs is compared against two baselines that use the ground-truth EEG, represented either as a spectrogram image or as a time-series signal. A standard ResNet-1826 (11.7M parameters) is used as the backbone for models processing spectrograms, and a ID convolutional neural network (CNN)27 with residual blocks26 followed by an LSTM layer28 (12.7M parameters) for models working with time-series signals.
[0330] FIG. 11 presents the results. For the four diseases tested, the described model consistently achieves higher AUROC scores than the baselines — Alzheimer’s Disease (81.5% vs. 74.8%), Parkinson’s Disease (87.2% vs. 82.9%), CHF (79.0% vs. 75.8%), and COPD (75.5% vs. 75.1%). For drug effect classification, the model shows comparable or superior performance on three of the four medications (with the exception of benzodiazepines) — anticholinergics (84.7% vs. 79.4%), benzodiazepines (75.4% vs. 79.4%), insulin (77.0% vs. 79.2%), and oral anticoagulants (79.6% vs. 79.3%). This reflects the model’s ability to capture the physiological changes associated with both diseases and drugs. Moreover, the similarity in performance between the baseline that uses the EEG spectrogram and the one that relies on the EEG time series suggests that converting time-series data into a spectrogram does not undermine clinically relevant information. By focusing on power spectrograms, the model retains health-related features from the original EEG signal and supports a range of downstream clinical applications.
[0331] 12.7 Screening for diseases and drugs: Multi-modal foundation model Recent years have demonstrated the value of foundation models that learn a general-purpose representation from a single, simple task, which can then be adapted to multiple downstream tasks. In the text domain, the foundation task is predicting words in context (e.g., next-word prediction or masked-word prediction). In the context of physiological signals, translating one physiological signal into another can serve as a foundation task. Breathing is used as the primary modality. Rather than training separate models for each disease and drug detection task, a foundation model is first trained to translate breathing into EEG, then adapted with per-task supervision. The breathing-to-EEG translation task enables the model to serve as a foundationAttorney Docket No 70009- 127-C- WO
[0332] Client case 26590
[0333] model for breathing, leveraging multi-modal learning to enhance its capabilities and outperform task-specific models.
[0334] FIG. 12 compares the performance of adapting the described breathing-to-EEG model (blue) against training task-specific models (light blue) for disease and drugeffect detection, with all models using the same breathing signals as input. The model’s breathing representation outperforms the raw breathing signal on every task. The largest improvements appear in three disease-screening tasks (Parkinson’s Disease, Alzheimer’s Disease, and Stroke) and two drug-effect assessments (Benzodiazepines and Thiazolidinediones), all linked to neurological functions. This result underscores the benefits of multi-modal learning, allowing the model to extract information from modality B that is implicitly embedded in modality A. By learning to translate breathing into EEG, the model uncovers neurologically relevant features that would otherwise be difficult to detect directly from breathing.
[0335] In fact, the advantages of using breathing-to-EEG translation as a foundation task extend beyond neurological conditions and drug effects. The model also shows notable gains in tasks related to respiratory', cardiovascular, and metabolic conditions. In particular, it surpasses purely per-task training when screening for COPD (by 5.6%), Emphysema (by 4.3%). CHF (by 7.5%), Hypertension (by 3.1%), and more. It likewise demonstrates marked improvement in assessing the effects of Xanthines (by 6.6%), Insulin (by 2.8%), and Diuretics (by 7.4%), each of which is commonly used in treating respiratory’, metabolic, or cardiovascular diseases.
[0336] Overall, the results in Figure 6 show that translating breathing to EEG can provide a foundation model for training many downstream tasks improving the performance of each of them.
[0337] 12.8 Analysis of Model Performance
[0338] The EEG spectrograms generated by the described model sometimes outperform the ground-truth EEG spectrograms on downstream tasks. Two hypotheses are investigated. First, the ground-truth EEG may contain noise introduced during data collection, such as inconsistencies in electrode placement. In contrast, the modelgenerated EEG is inferred solely from respiration signals, thereby avoiding such noise. Second, the ground-truth EEG can occasionally miss certain patterns due to electrode positioning, whereas the model, having learned priors from extensive training data, may reconstruct these patterns. Both hypotheses are explored in the following sections, supported by experimental evidence.Attorney Docket No 70009- 127-C- WO
[0339] Client case 26590
[0340] 12.9 Denoising the recording noise
[0341] EEG signals, which rely on direct scalp measurements, are inherently vulnerable to noise introduced by factors such as poor electrode-to-scalp contact or fluctuations in skin impedance30. These challenges can produce high-impedance artifacts that shift signal amplitudes and alter the power of the EEG. In contrast, Al-generated EEGs, such as produced from the respiration signal, are not subject to these noise sources. Similar denoising benefits of generative Al have been observed in other domains as well.
[0342] Because of this denoising effect, the Al-generated EEG spectrogram can potentially yield better results on downstream tasks. To investigate this hypothesis, downstream tasks are considered where the Al-generated EEGs outperform the ground truth EEGs, and whether some of the improved performance is related to noisy EEG power in the ground truth is examined. The average power of each ground-truth EEG spectrogram is calibrated with that of its corresponding Al-generated EEG (i.e. , the average ground truth power is subtracted and the average power of the generated EEG is added). Performance is then evaluated on three tasks: screening for Alzheimer’s Disease (AD), screening for Parkinson’s Disease (PD), and assessing the impact of anticholinergics. In these tasks, the model’s performance surpasses the ground-truth EEG by substantial margins — AD (4.1%), PD (4.3%), and anticholinergics (5.3%). After calibrating the ground-truth EEG spectrogram to align with the Al-generated power levels, the performance on the original EEG improved in all three tasks. As shown in Figure 10a, this calibration led to notable AUROC gains — AD (2.2%), PD (0.7%), and anticholinergics (3.9%). These findings support the idea that Al-generated EEGs can offer a denoising benefit that enhances performance in downstream applications.
[0343] 12.10 Reconstructing EEG patterns missing in the ground truth
[0344] The described experiments focus on the C4-A1 EEG montage for both training and testing. Although EEG signals across different channels are highly correlated and often exhibit similar patterns, certain channels may fail to capture specific EEG activity. This can occur when an electrode is positioned far from the source of an EEG signal, making it unable to detect the pattern. However, the model-generated EEG may still recover the missing signal.
[0345] The posterior dominant rhythm (PDR) is used as an illustrative example. The C4-A1 EEG channel fails to detect the PDR pattern in approximately 5% of cases in the described dataset. This occurs because C4 is located in the central region of the scalp,Attorney Docket No 70009- 127-C- WO
[0346] Client case 26590
[0347] whereas the PDR primarily originates from the occipital and posterior regions of the brain23, particularly around the 01 and 02 electrodes (Figure 1). As a result the C4 channel may not capture a strong PDR signal.
[0348] Despite being trained exclusively on C4-A1 data, the model can sometimes reconstruct the missing PDR pattern. FIG. 13c provides an example where the C4-A1 EEG channel largely fails to detect the PDR pattern (top row), while the pattern is clearly present in the O1-A2 EEG channel (middle row). The model successfully reconstructs the missing PDR signal with power comparable to that of the O1-A2 channel (bottom row). This reconstruction is possible because, in most training samples, the C4-A1 channel correctly captures the PDR pattern, as verified by the 01-A2 channel. Through the translation task, the model leams to generate PDR features based on breathing signals.
[0349] This capability7may enhance model performance on downstream tasks, such as screening for Parkinson’s disease, a condition known to significantly affect PDR. In a controlled experiment, PDR features were extracted from both ground-truth and AI-generated EEG spectrograms and their effectiveness was assessed in classifying Parkinson’s disease using a linear model. The average power of the 5 Hz EEG band around the peak PDR frequency 34 was used as the PDR feature. FIG. 13b illustrates the performance of this feature in screening for Parkinson's disease. The classification performance is notably higher (AUROC = 67.4%) when using the PDR feature derived from the model’s generated signals compared to the feature extracted from the original C4-A1 EEG channel (AUROC = 55.1%).
[0350] 12.11 Discussion
[0351] The disclosed framework for translating respiratory' signals into EEG is grounded in the concept that physiological signals can be modeled as a language composed of discrete, interpretable patterns. A multi-modal foundation model maps breathing sequences to EEG spectrograms during sleep. The model leams to generate high-fidelity EEG representations that preserve clinically meaningful features. Beyond accurate reconstruction, the generated EEG supports a wide range of downstream clinical tasks, including the screening of neurodegenerative. cardiovascular, respiratory, and metabolic diseases, as well as the assessment of medication effects. In some cases, the generated signals outperform ground-truth EEG, reflecting the model’s ability to denoise inputs, recover patterns missed in standard recordings, and generalize across diverse populations. These results represent a foundational step toward multi-modal generative models that reason over complex human physiology'Attorney Docket No 70009- 127-C- WO
[0352] Client case 26590
[0353] and produce clinically useful representations, opening opportunities for scalable, data-driven diagnosis, monitoring, and personalized care.
[0354] Several limitations merit consideration. The described experiments focus on sleep EEG data with a specific frequency range and temporal resolution, which may not generalize to other EEG contexts or wakeful states. The availability of high-quality, large-scale physiological datasets is limited, due to the complexity and cost of acquisition. These constraints impact the diversity of training data and contribute to imbalances across demographic subgroups, which can lead to uneven model performance. Moreover, the inherently low interpretability' of EEG, compounded by the complexity7of deep generative models, poses challenges for explainability'.
[0355] Although the model sen es as a foundation model of physiological signals, it may not exhibit the full zero-shot generalization seen in language or vision foundation models; some task-specific data and supervision may remain necessary for optimal performance.
[0356] The described results demonstrate that EEG can be clinically informative beyond its traditional uses. Historically, EEG has been regarded as a specialized diagnostic tool, primarily employed in sleep medicine, epilepsy evaluation, or intensive care monitoring. Its reliance on electrodes, technical setup, and expert interpretation has limited its accessibility, scalability, and integration into broader clinical workflows. The described work reveals that EEG signals contain physiological signatures indicative of a wide spectrum of conditions, spanning from neurodegenerative diseases such as Alzheimer’s and Parkinson’s, to cardiovascular conditions like stroke and congestive heart failure, and further to pharmacological effects, establishing EEG as a multi-functional clinical modality.
[0357] More broadly, the described results show the feasibility' of generating EEG signals from other physiological signals. The empirical results described herein focus on generating EEG from respiratory' signals; however, given the feasibility7of generating breathing from cardiovascular signals such as ECG36 or PPG37, the approach may7be extended to generate EEG from other physiological signals either directly, or indirectly by first generating breathing then translating it to EEG. Furthermore, combining multiple input signals (e.g., breathing and PPG) may improve
[0358] performance.
Claims
Attorney Docket No 70009- 127-C- WOClient case 26590WHAT IS CLAIMED IS:
1. A method for generating an electroencephalography representation of neural activity of a subject, the method comprising:obtaining a non-electroencephalography physiological signal generated using at least one sensor and representing a physical measurement of the subject; processing the physiological signal using a first trained neural network to generate a sequence of codes representing neural activity of the subject; decoding the sequence of codes using a second trained neural network to generate an electroencephalography representation; andoutputting the electroencephalography representation.
2. The method of claim 1, further comprising acquiring the nonelectroencephalography physiological signal using one or more sensors in contact with or proximate to the subject.
3. The method of claim 1. wherein the first trained neural network was trained using pairs of physiological signals and corresponding electroencephalography signals, and wherein training comprises minimizing a loss function dependent on at least one of similarity between sequences of codes generated from the physiological signals and sequences of codes generated from the corresponding electroencephalography signals, similarity between electroencephalography representations reconstructed from those sequences of codes, or performance of a downstream task.
4. The method of claim 1. wherein the second trained neural network comprises a decoder, and wherein the decoder was trained jointly with a neural network encoder configured to process one or more electroencephalography signals and encode the one or more electroencephalography signals to produce a sequence of codes representing neural activity, the decoder being configured to decode the sequence of codes to produce approximate electroencephalography signals approximating the processed electroencephalography signals.
5. The method of any one of claims 1-4, wherein the sequence of codes comprises discrete codes selected from a finite set.Attorney Docket No 70009- 127-C- WOClient case 265906. The method of any one of claims 1-4, wherein the sequence of codes comprises numerical values.
7. The method of any one of claims 1-4, wherein the sequence of codes comprises vectors in a multi-dimensional numerical space.
8. The method of any one of claims 1-4, wherein the codes encode spectral -temporal features of neural activity.
9. The method of claim 1, wherein the physiological signal comprises at least one of a respiratory signal, electrocardiogram (ECG), photoplethysmogram (PPG), electromyogram (EMG), or motion signal, wherein the at least one physiological signal is obtained by direct contact measurement, by proximate non-contact measurement, or by measurement using radio-frequency signals reflected off the subject.
10. The method of claim 1. wherein the physiological signal comprises a plurality of physiological signals.
11. The method of claim 1, wherein the electroencephalography representation is output for use in detection, screening, monitoring, or assessment.
12. The method of claim 1, wherein the electroencephalography representation is output for use in sleep staging, arousal detection, sleep uality assessment, or sleep disorder screening.
13. The method of claim 1, wherein the physiological signal is obtained while the subject is in a sleep state.
14. The method of claim 1, wherein the electroencephalography representation comprises a plurality' of channels corresponding to scalp electrode locations.
15. The method of claim 1. further comprising:receiving an electroencephalography signal including noise, artifacts, or missing intervals;Attorney Docket No 70009- 127-C- WOClient case 26590wherein generating the sequence of codes comprises processing both the electroencephalography signal and the physiological signal; and wherein decoding the sequence of codes comprises generating a corrected electroencephalography representation having reduced artifacts or reconstructed signal components.
16. A method for determining pharmacological exposure of a subject, the method comprising:obtaining a first physiological signal generated using at least one sensor and representing a physical measurement of a first physiological subsystem of the subject;processing the first physiological signal using atained neural network to generate a sequence of codes representing activity associated with a second physiological modality of a second physiological subsystem of the subject, the second physiological subsystem being different from the first physiological subsystem;determining, based at least in part on the sequence of codes or another representation generated using the trained neural network, an indication of pharmacological exposure of the subject; andoutputting the indication for use in detection, screening, monitoring, or assessment of pharmacological exposure.
17. The method of claim 16, further comprising acquiring the first physiological signal using one or more sensors in contact with or proximate to the subject.
18. The method of claim 16, wherein the trained neural network was trained using pairs of physiological signals and corresponding electroencephalography signals, and wherein training comprises minimizing a loss function dependent on at least one of similarity between sequences of codes generated from the physiological signals and sequences of codes generated from the corresponding electroencephalography signals, similarity between electroencephalography representations reconstructed from those sequences of codes, or performance of a downstream task.
19. The method of claim 16, wherein the sequence of codes comprises discrete codes selected from a finite set.Attomey Docket No 70009- 127-C- WOClient case 2659020. The method of claim 16, wherein the sequence of codes comprises numerical values.
21. The method of claim 16, wherein the sequence of codes comprises vectors in a multi-dimensional numerical space.
22. The method of any one of claims 19-21, wherein the codes encode latent physiological features associated with pharmacological effects.
23. The method of claim 16, wherein determining the indication is performed without reconstructing an electroencephalography representation.
24. The method of claim 16, wherein the first physiological signal comprises at least one of a respiratory signal, electrocardiogram (ECG), photoplethysmogram (PPG), electromyogram (EMG), or motion signal, wherein the at least one physiological signal is obtained by direct contact measurement, by proximate non-contact measurement, or by measurement using radio-frequency signals reflected off the subject.
25. The method of claim 16, wherein the indication of pharmacological exposure comprises a probability, score, class label, or binary output identifying presence or absence of exposure to a pharmacological agent.
26. The method of claim 16, wherein the pharmacological exposure is associated with one or more agents belonging to a pharmacologic class selected from the group consisting of neuropsychiatric agents, autonomic nervous system agents, bronchodilators, cardiovascular agents, anticoagulants, diuretics, antidiabetic agents, analgesics, opioids, anesthetics, sedatives, hypnotics, anticonvulsants, antipsychotics, beta-blockers, antiarrhythmics, antihypertensives, antihistamines, respiratory tract agents, corticosteroids, and sex hormones.
27. The method of claim 16, wherein the first physiological signal is obtained while the subject is in a sleep state.
28. The method of claim 16, wherein determining the indication comprises processing physiological signals obtained over a plurality of time points.Attorney Docket No 70009- 127-C- WOClient case 2659029. The method of claim 16, wherein determining the indication requires applying one or more additional processing operations comprising processing by one or more additional trained neural networks, one or more linear projections, one or more nonlinear projections, one or more classifiers, one or more regressors, one or more statistical models, one or more rule-based operations, one or more transformations, or any combination thereof.
30. The method of claim 29, wherein at least one of the one or more additional trained neural network is trained using labels derived from medication records, prescription data, clinician annotations, or treatment schedules.
31. A method for assessing a risk or condition related to health of a subject, the method comprising:obtaining a signal generated using at least one sensor and representing activity of a first physiological subsystem of the subject;processing the signal using a trained neural network, the trained neural network having been trained to generate, from signals representing the first physiological subsystem, a sequence of codes representative of activity' associated with a second physiological subsystem of the subject, the second physiological subsystem being different from the first physiological subsystem;determining, based at least in part on at least one representation generated using the trained neural network, an indication of the risk or condition related to the health of the subject; andoutputting the indication for use in detection, screening, monitoring, or assessment of the risk or condition related to the health of the subject.
32. The method of claim 31, further comprising acquiring the signal using one or more sensors in contact with or proximate to the subject.Attorney Docket No 70009- 127-C- WOClient case 2659033. The method of claim 31, wherein the trained neural network was trained using pairs of signals representing the first physiological subsystem and corresponding signals representing the second physiological subsystem, and wherein training comprises minimizing a loss function dependent on at least one of similarity between sequences of codes generated from signals representing the first physiological subsystem and sequences of codes generated from the corresponding signals representing the second physiological subsystem, similarity’ between representations of the second physiological subsystem reconstructed from those sequences of codes, or performance of a downstream task.
34. The method of claim 31, wherein the sequence of codes comprises discrete codes selected from a finite set.
35. The method of claim 31, wherein the sequence of codes comprises numerical values.
36. The method of claim 31, wherein the sequence of codes comprises vectors in a multi-dimensional numerical space.
37. The method of claim 31, wherein the sequence of codes representative of activity- associated with the second physiological subsystem is obtained by encoding electroencephalography signals.
38. The method of claim 31, wherein the at least one representation generated using the trained neural network comprises all or a subset of the sequence of codes representative of activity associated with the second physiological subsystem.
39. The method of claim 31, wherein the first physiological subsystem is a respiratory subsystem, a cardiovascular subsystem, a neuromuscular subsystem, a nervous system, or a musculoskeletal subsystem, and wherein the signal representing the first physiological subsystem comprises at least one of a respiratory signal, electrocardiogram (ECG), photoplethysmogram (PPG), electromyogram (EMG), electroencephalography (EEG) signal, or motion signal, wherein the at least one signal is obtained by direct contact measurement, by proximate non-contact measurement, or by measurement using radio-frequency signals reflected off the subject.Attorney Docket No 70009- 127-C- WOClient case 2659040. The method of claim 31, wherein the second physiological subsystem is a respiratory subsystem, a cardiovascular subsystem, a neuromuscular subsystem, a nervous system, or a musculoskeletal subsystem, and wherein the sequence of codes representative of activity associated with the second physiological subsystem corresponds to a respiratory signal, electrocardiogram (ECG), photopl ethy smogram (PPG), electromyogram (EMG), electroencephalography (EEG) signal, or motion signal.
41. The method of claim 31, wherein determining the indication comprises applying one or more additional processing operations to the at least one representation generated using the trained neural network.
42. The method of claim 41, wherein the one or more additional processing operations comprise processing by one or more additional neural networks, one or more linear projections, one or more nonlinear projections, one or more classifiers, one or more regressors, one or more statistical models, one or more rule-based operations, one or more transformations, or any combination thereof.
43. The method of claim 31, wherein the risk or condition related to the health of the subject comprises a disease risk or condition.
44. The method of claim 43, wherein the disease risk or condition comprises a neurological, metabolic, cardiovascular, pulmonary, or immune-related condition.
45. The method of claim 31, wherein the indication comprises a probability, severity score, risk score, class label, or binary output.
46. The method of claim 31, wherein determining the indication comprises processing sequences of codes obtained from physiological signals acquired at a plurality of time points.
47. The method of claim 31, wherein determining the indication comprises comparing the sequence of codes with a reference distribution derived from healthy subjects or subjects having a known condition.
48. The method of claim 31, wherein determining the indication is based further on age, sex, medication status, sleep state, activity state, or subject history.Attomey Docket No 70009- 127-C- WOClient case 2659049. The method of claim 31, wherein the signal representing the first physiological subsystem is obtained partially or fully dunng sleep of the subject.
50. A method for training a neural network system to generate a neural-activity representation from physiological signals, the method comprising:receiving electroencephalography signals;training an encoder model to map the electroencephalography signals to sequences of codes;training a decoder model to reconstruct electroencephalography representations from the sequences of codes;receiving physiological signals and corresponding electroencephalography signals;training a mapping model to generate sequences of codes from the physiological signals based on a loss function dependent on at least one of:(i) similarity between codes,(ii) similarity between reconstructed electroencephalography representations, or(iii) performance of a downstream task; andconfiguring the mapping model to generate neural-activity representations from physiological signals.
51. The method of claim 50, wherein the sequences of codes comprise discrete codes selected from a finite set.
52. The method of claim 50, wherein the sequences of codes comprise numerical values.
53. The method of claim 50, wherein the sequences of codes comprise vectors in a multi-dimensional numerical space.
54. The method of claim 50, wherein the reconstruction loss comprises a time-domain loss, spectral -domain loss, perceptual loss, adversarial loss, or combination thereof.Attomey Docket No 70009- 127-C- WOClient case 2659055. The method of claim 50, wherein the physiological signals and corresponding electroencephalography signals are temporally synchronized, and wherein training the mapping model comprises compensating for delay between the physiological signals and the electroencephalography signals.
56. The method of claim 50, wherein the encoder model and decoder model are trained before training the mapping model.
57. The method of claim 50, wherein training the mapping model and training the decoder model are performed jointly.
58. The method of claim 50, wherein training the mapping model comprises a masked prediction task.
59. The method of claim 50, wherein training the mapping model comprises contrastive learning between physiological signals and corresponding electroencephalography signals.
60. The method of claim 50, wherein the mapping model is fine-tuned for a target sensor type or target subject population.
61. The method of claim 50, wherein training comprises domain adaptation across different sensor hardware or acquisition environments.
62. The method of claim 50, wherein configuring the mapping model comprises calibrating the mapping model using subject-specific physiological data.
63. A method for improving an electroencephalography signal of a subject, the method comprising:obtaining an electroencephalography signal generated using at least one sensor and including noise, artifacts, or missing intervals;obtaining a non-electroencephalography physiological signal generated using at least one sensor and representing a physical measurement of the subject; processing the electroencephalography signal and the physiological signal using one or more trained neural networks to generate a sequence of codes;Attorney Docket No 70009- 127-C- WOClient case 26590generating, from the sequence of codes, a corrected electroencephalography representation; andoutputting the corrected electroencephalography representation having reduced artifacts or reconstructed signal components.
64. The method of claim 63, further comprising acquiring the electroencephalography signal and the non-electroencephalography physiological signal using one or more sensors in contact with or proximate to the subject.
65. The method of claim 63, wherein generating the sequence of codes comprises conditioning on the physiological signal.
66. The method of claim 63, wherein the corrected electroencephalography representation has improved signal-to-noise ratio or spectral fidelity.
67. The method of claim 63, wherein the noise or artifacts comprise motion artifacts, muscle artifacts, electrode pop artifacts, line noise, baseline drift, or missing channels.
68. The method of claim 63, wherein the corrected electroencephalography representation is generated in real time.
69. A non-transitory computer-readable medium storing instructions that, when executed by one or more processors, cause the one or more processors to perform the method of any one of claims 1, 16, 31, 50, or 63.
70. A system for processing physiological signals of a subject, the system comprising:one or more sensors configured to acquire at least one physiological signal or electroencephalography signal representing a physical measurement of the subject;one or more processors; andmemory storing instructions that, when executed by the one or more processors, cause the one or more processors to:process the signal using one or more trained neural networks to generate a sequence of codes representing physiological or neural activity;Attorney Docket No 70009- 127-C- WOClient case 26590generate, from the sequence of codes, a result comprising at least one of (i) an electroencephalography representation of neural activity,(ii) an indication of pharmacological exposure,(iii) an indication of a risk or condition related to health of a subject, or (iv) a corrected electroencephalography representation; andoutput the result via at least one output device.
71. The system of claim 70, wherein the sequence of codes comprises discrete codes selected from a finite set.
72. The system of claim 70, wherein the sequence of codes comprises vectors in a multi-dimensional numerical space.
73. The system of claim 70, wherein the one or more sensors comprise at least one of a respiratory sensor, electrocardiogram (ECG) sensor, photoplethysmogram (PPG) sensor, electromyogram (EMG) sensor, motion sensor, or radio-frequency sensor configured to measure physiological signals reflected off the subject.
74. The system of claim 70, wherein the one or more sensors are configured to acquire a plurality of physiological signals, and the one or more processors are configured to process the plurality of physiological signals jointly.
75. The system of claim 70, wherein the output device is configured to provide a signal to an external system configured to process electroencephalography signals.
76. The system of claim 70, wherein the instructions, when executed by the one or more processors, cause the one or more processors to perform all steps of any one of claims 1, 16, 31, 50, or 63.
77. A system for generating an electroencephalography representation of a subject, the system comprising:one or more sensors configured to acquire a non-electroencephalography physiological signal representing a physical measurement of the subject; one or more processors; andAttorney Docket No 70009- 127-C- WOClient case 26590memory storing instructions that, when executed by the one or more processors, cause the one or more processors to:process the physiological signal using a first trained neural network to generate a sequence of codes representing neural activity7;decode the sequence of codes using a second trained neural network to generate an electroencephalography representation; andoutput the electroencephalography representation.
78. The system of claim 77, wherein the sequence of codes comprises discrete codes selected from a finite set.
79. The system of claim 77, wherein the sequence of codes comprises vectors in a multi-dimensional numerical space.
80. The system of claim 77, wherein the one or more processors are configured to process the physiological signal using a sequence model.
81. The system of claim 77, wherein the system is configured to provide the electroencephalography representation to a downstream system configured to process electroencephalography signals.
82. The system of claim 77, wherein the instructions, when executed by the one or more processors, cause the one or more processors to perform all steps of claim 1.
83. A system comprising:one or more processors; andmemory storing instructions that, when executed by the one or more processors, cause the one or more processors to perform all step of any one of claims 1, 16, 31, 50, or 63.