Questions about how individual parts of claim 1 might be changed. A question may describe something different from what the claim language recites, or a version of it; no view is offered on which.
1.1Direct classification of signal features
What ifWhat if the common model computes seizure probabilities directly from signal features without comparing those features with a common seizure template?
Claim language“compares similarity with a common seizure template”
The processor computes band power, line length, and amplitude variation from four-second windows of scalp electrical recordings. A gradient-boosted tree model trained on labeled recordings from multiple users produces separate probabilities for an ongoing seizure and a seizure starting within the next five minutes.
1.2Retain feature summaries and annotations
What ifWhat if the accumulated record contains only derived signal features and independently reviewed seizure annotations?
Claim language“including the bio-signals of the user and seizure detection and prediction results”
The processor converts each recording window into a feature vector containing band power, line length, and interchannel correlation, then releases the waveform from its processing buffer. The system discards its detection and prediction outputs after delivery and pairs the stored feature vectors with seizure onset and end times entered by a reviewer.
1.3Calibrate the sensor input stage
What ifWhat if the accumulated data calibrates the input stage while the common seizure template and common model stay fixed?
Claim language“updating the common seizure template and the common seizure AI model”
An input calibration module estimates each channel’s baseline offset and amplitude scale from artifact-free segments of the user’s recordings. It applies those values to incoming samples before feature extraction so the fixed model receives inputs in the amplitude range used during training.
1.4Configure a personal alert threshold
What ifWhat if the system configures a personal alert threshold while using the common model to compute every seizure score?
Claim language“generating a personal seizure AI model”
A calibration routine tests candidate thresholds against stored scores and reviewed event times. It stores the threshold with the highest event sensitivity among candidates that meet a specified false alert limit, and an alert controller applies that threshold to subsequent common-model scores.
1.5Combine scores from both models
What ifWhat if the common and personal models both contribute to every result without selecting either model for the user?
Claim language“determining a model optimized for the user”
Both models process the same recording window and produce calibrated probabilities for ongoing seizures and for onset within a defined forecast interval. The processor averages the two probabilities for each output type and applies separate thresholds to the resulting detection and forecast scores.
Section 2 · machine-generated
Variations
Further arrangements of the same subject matter, described here as of the publication date at the foot of this page.
Other field of use
Seizure monitoring in veterinary wards
A scalp electrode array and a collar accelerometer record synchronized electrical activity and motion from dogs admitted for seizure observation. A shared model compares incoming windows with a template formed from recordings of multiple dogs and stores the signals, seizure detections, and forecasts for each dog during its observation period. The processor uses the accumulated records to update the shared template and model and train a model for the monitored dog. It selects the model with the lower count of false alerts per hour at a fixed minimum event sensitivity on held-out recordings and uses that model for subsequent monitoring.
Component substitution
Dynamic time warping for template matching
Dynamic time warping replaces pointwise similarity scoring between incoming electroencephalogram feature sequences and the shared template. The template stores representative sequences for background, pre-seizure, and seizure states, and the processor converts the minimum alignment distance for each state into an input for the detection and forecast classifiers. Accumulated signals and outputs contribute candidate sequences to the shared template, and the processor refits the shared classifier and trains a personal classifier from the same collection. It compares both classifiers against event annotations on held-out recordings and runs the classifier with the lower time-weighted classification loss.
Parameter or scale variation
Dual window lengths for seizure monitoring
A recorder samples eight electroencephalogram channels at 256 samples per second per channel and advances both analysis windows once per second. A four-second window supplies features for seizure detection, and a 120-second window supplies trends in band power for forecasting seizure onset within the next five minutes. During a 21-day collection period, the processor stores signals and both outputs, then uses the records to update the shared template and model and train a personal model with the same window lengths. The processor scores both models on held-out days and deploys the model with the lower combined detection and forecast loss.
Combination with existing systems
Seizure monitoring with synchronized video review
A bedside video electroencephalography recorder timestamps the electrode signals and camera frames against the same acquisition clock. The shared model scores the electrode signals against its template, and the recorder stores those scores alongside synchronized video during the collection period. Reviewers mark seizure onset and offset in the playback interface, and the processor uses the marked intervals and accumulated signals to update the shared template and model and fit a personal model. The selected model sends detection and forecast timestamps back to the recorder as bookmarks in the synchronized playback timeline.
Automated or data-driven variant
Automatic quality weighting during model training
An electrode interface measures contact impedance, and a head-mounted accelerometer records motion on the same time base as the electroencephalogram. Software assigns quality weights to each channel window from contact impedance, amplifier saturation, and motion that coincides with abrupt signal changes, then stores the weights with the signals and model outputs. The processor applies the weights when rebuilding the shared template, refitting the shared model, and training the personal model. It selects a model using held-out event annotations and applies the same channel weighting to the selected model during live monitoring.
Model selection by daily activity state
A wearable accelerometer and a clock supply activity measurements and time of day to a classifier that labels recording windows as sleep, quiet wakefulness, or active wakefulness. The processor stores those labels with the signals and monitoring outputs during collection and uses them as features when updating the shared template and model and training the personal model. It compares the two fitted models within each state using held-out annotated recordings and stores the selection for each state. During live monitoring, the processor uses the model assigned to the current state and changes models after the new state persists for 30 seconds.
Specific version of a part
Clustered feature templates
The processor extracts band power and line length from labeled windows collected from multiple users and clusters seizure and pre-onset windows separately. The common model compares each incoming feature vector with the stored cluster centroids using cosine similarity and emits separate detection and forecast scores.
Time-aligned recording and result storage
The recorder stores four channels of scalp electrical samples at 256 samples per second in one-minute blocks over a 14-day collection interval. A separate table records the start time of each analysis window, its detection score, its forecast score, and the forecast interval.
Refresh centroids and score weights
The processor adds reviewed user windows to a retained set of recordings from multiple users and recomputes the seizure and pre-onset feature centroids. It fits new logistic output weights from the updated centroid similarity scores and installs the centroids and weights together.
Compare models on later recordings
The processor evaluates the updated common model and the personal model on a later block of reviewed user recordings held out from both model updates. It compares event sensitivity and false alerts per day for detection and forecasting, then selects the model with the higher mean sensitivity among models that meet both false alert limits.
Section 3 · quoted from USPTO records
Other work
Passages reproduced word for word from two earlier-published U.S. patent documents, placed beside the language of claim 1. No characterization is offered. Each passage is marked A or B for the document it comes from; ¶ numbers are the paragraph numbers in the USPTO’s published full text of that document.
1.1detecting seizure detection and prediction results from bio-signals of the user by using a common seizure AI model that compares similarity with a common seizure template composed of bio-signals collected from a plurality of users;
Based on novel computational intelligence algorithms, a realistic posterior probability function P(ST|x) representing the probability of one or more seizures starting within the next T minutes, given observations x derived from intracranial EEG (IEEG) or other signals, is periodically synthesized for a plurality of prediction time horizons (scales T, or resolutions 1/T), e.g., a second, a minute, ten minutes, an hour, etc.
A · US 2008/0021342 A1 · ¶ 8
The shape (i.e., all the geometrical information that is invariant to position (including rotation) and scale) of these curves may be used for detection of changes in brain state such as epileptic seizures and their properties may be characterized through use of statistical shape analysis (e.g., Procrustes analysis), of the different embodiments of “matched filtering” or of other geometrical (Euclidian and non-Euclidian) methods.
B · US 2022/0304616 A1 · ¶ 289
1.2accumulating user data including the bio-signals of the user and seizure detection and prediction results during a first period of time;
Step 2: An expert neurologist inspects the IEEG that was digitally archived in the external storage device over the last period, and marks the times of unequivocal electrographic onset (UEO) in all recorded seizures (logic block 404).
A · US 2008/0021342 A1 · ¶ 39
1.3updating the common seizure template and the common seizure AI model using the accumulated user data;
1.4generating a personal seizure AI model using the accumulated user data;
1.5determining a model optimized for the user among the updated common seizure AI model or personal seizure AI model; and
1.6detecting seizure detection and prediction results from the bio-signals of the user using the model optimized for the user.
The invention unifies seizure onset detection as a degenerate form of prediction at the finest time resolutions.
A · US 2008/0021342 A1 · ¶ 13
An empty cell means only that no passage was reproduced there.
Statements in these documents
Drugs are the most common form of treatment, but their efficacy is limited.
A · US 2008/0021342 A1 · ¶ 4
When coupled with optimally determined thresholds for alarm or therapy activation, probabilities defined in this fashion provide anticipatory time-localization of events in a synergistic logarithmic-like array of time resolutions, thus effectively circumventing the performance vs. prediction-horizon tradeoff of single-resolution systems.
A · US 2008/0021342 A1 · ¶ 8
There is a need for improved epilepsy treatments that address cardiac impairments associated with seizures.
B · US 2022/0304616 A1 · ¶ 5
Restoration of the sympathetic-parasympathetic tone to normal levels, a therapeutic objective that may be accomplished by enhancing parasympathetic activity through among others, electrical stimulation of the vagus nerve, may decrease the rate and severity of cardiac and autonomic co-morbidities in these patients.
B · US 2022/0304616 A1 · ¶ 4
Cite aspatent.fyi, US 2026/0354740 A1, claim 1, first published 2026-11-19T02:47:19-05:00, https://patent.fyi/US20260354740/, SHA-256 bc5cb926619c2e4296e0534c2ce56cd6f5a52a2ff78379c180014c27b4965253