1.1Training from unannotated recordings
A reconstruction network learns to recover masked waveform intervals from surrounding samples. A peak detector operates on the reconstructed waveform, and timing rules flag runs of irregular beat intervals.
Generated automatically. Not reviewed by a person. Not legal advice or an opinion.
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.
A reconstruction network learns to recover masked waveform intervals from surrounding samples. A peak detector operates on the reconstructed waveform, and timing rules flag runs of irregular beat intervals.
Training produces a waveform classifier and fixed normalization constants. During analysis, the processor applies those constants to each segment and runs a forward pass with the stored classifier weights.
A detector identifies ventricular peaks and measures successive beat intervals, ventricular complex widths, and waveform amplitudes. A decision tree trained on these measurements assigns a rhythm label to each segment.
A peak detector extracts beat intervals and ventricular complex widths from a labeled calibration recording. A logistic classifier fitted to those measurements predicts beat labels in subsequent segments from the same subject.
Further arrangements of the same subject matter, described here as of the publication date at the foot of this page.
Housing accelerometers record pump vibration, and each training record includes its sensor mounting position and a maintenance inspection label such as normal operation or shaft misalignment. A waveform predictor learns features through an adversarial classifier that predicts mounting position, and a fault model learns to classify inspection states from residual spectra. For a newly instrumented pump, unlabeled vibration windows update the predictor, and per-band residual scaling from those windows adapts the fault model to that pump.
A gradient-boosted decision tree ensemble replaces the generalized state predictor and receives residual features from a waveform predictor trained with an electrode-configuration classifier. For each electrocardiogram window, the feature extractor computes the median absolute residual, its 95th percentile, and residual power in fixed frequency bands. At enrollment, unlabeled windows update the waveform predictor and yield per-feature percentile mappings from the enrollment distribution to the training distribution, which the individualized model applies before evaluating the trees.
A linear dynamical state space model with a Kalman filter serves as the waveform abstraction model, and each electrode configuration has its own observation matrix. Training fits the latent dynamics and observation matrices to predict the next sample while an adversarial classifier penalizes electrode-configuration information in the latent state. A state classifier learns from labeled windows summarized by the magnitude and autocorrelation of the one-step prediction errors. On a target recording, unlabeled windows fit the observation matrix and noise covariance for the recorded configuration, and the fitted covariance normalizes prediction errors before the state classifier evaluates them.
The processor resamples electrocardiogram waveforms to 250 samples per second and segments them into 8-second windows with a 2-second stride. During training, an adversarial classifier predicts the lead configuration from the waveform features, and a rhythm model maps each window's prediction residual to state probabilities. The first 60 seconds of an unlabeled target recording update the waveform predictor and set per-band residual scales for the individualized rhythm model. The model evaluates each subsequent window and averages the latest five probability vectors before assigning a rhythm label.
A bedside telemetry receiver supplies timestamped electrocardiogram samples, lead configuration, and electrode contact status to a waveform analysis service. The service starts with a waveform predictor trained against a lead-configuration classifier and a residual-based state model trained on annotated telemetry records. The service uses unlabeled windows with stable electrode contact to update the waveform predictor and calibrate the state model's residual inputs for the current recording. It sends state probabilities and window timestamps to the telemetry alarm manager, which applies its persistence timer before issuing a notification.
An acquisition controller records active channel identifiers and contact impedance, then starts recalibration when the channel set changes or impedance remains outside its configured range. During initial training, an adversarial classifier predicts electrode configuration from the waveform features, and a state model learns from the prediction residuals. During recalibration, software selects unlabeled windows with stable contact and no clipping, then updates the waveform predictor and computes new per-band residual offsets for the individualized state model. The controller activates the updated predictor and offsets together at a window boundary.
An acquisition module detects ventricular peaks and accumulates counts in 20-millisecond beat interval bins over 30 seconds. The analyzer computes the median interval and interquartile range from the bin counts to flag fast or irregular rhythms.
A training record stores a 10-second electrocardiogram segment sampled at 250 hertz as signed 16-bit integers. The record also stores the electrode positions, amplifier gain, and an annotated rhythm code for the segment.
An input record contains a 30-second differential electrocardiogram sampled at 500 hertz. The record specifies the center-to-center electrode spacing in millimeters and identifies which electrode connects to the positive amplifier input.
A convolutional encoder and decoder reconstruct each waveform while a lead classifier uses a gradient reversal layer to penalize lead information in the encoder features. A rhythm classifier receives the sample-by-sample reconstruction errors and predicts the annotated rhythm code.
A two-minute recording updates the channel means and variances in the trained abstraction model. The generalized classifier uses the same recording to update its input normalization statistics for the resulting reconstruction errors, forming an individualized classifier that predicts rhythm labels on subsequent segments.
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.
A biological-information evaluating device according to an embodiment of the present disclosure includes a signal processing circuit that generates a feature value of an observation target waveform for each of pieces of observation data obtained through observation of a living body for a predetermined period, on a basis of a plurality of pieces of partial observation data contained in each of the pieces of observation data and each having a period shorter than an observationA · US 2021/0145368 A1 · abstract
A biological-information evaluating device according to an embodiment of the present disclosure includes a signal processing circuit that generates a feature value of an observation target waveform for each of pieces of observation data obtained through observation of a living body for a predetermined period, on a basis of a plurality of pieces of partial observation data contained in each of the pieces of observation data and each having a period shorter than an observationA · US 2021/0145368 A1 · abstract
The classifier is proposed to predict a label of a direct exchange domain so as to minimize the loss function of the exchange domain; the domain labels in the loss function are interchanged, and the domain information is implicitly encoded in the prediction probability of the adversarial domain classifier, so the optimization of the loss function based on the interchanged domain makes the learned features independent of the domain.B · US 2022/0222529 A1 · ¶ 33
According to the present application, meta-knowledge based on domain-invariant features is introduced, and common domain features of the same kind of tasks are learnt from different domains corresponding to different data sets to obtain highly transferable domain-invariant knowledge; after meta-knowledge fine-tuning, the model can be fine-tuned for different domains of the same kind of tasks, with a better parameter initialization ability and a higher generalization ability,B · US 2022/0222529 A1 · ¶ 4
An empty cell means only that no passage was reproduced there.
In addition to a human, the target from which the biological information is acquired may include an animal or the like.A · US 2021/0145368 A1 · ¶ 2
The biological-information evaluating device and the method of evaluating biological information according to the embodiment of the present disclosure make full use of biological information in which periodic variation is able to be seen, which makes it possible to know a difference between states with a small number of samples and a short measuring period.A · US 2021/0145368 A1 · ¶ 8
In the fine-tuning stage, the existing compression method of the pre-trained language model for a downstream task is to fine-tune the specific data set of the downstream task, and the effect of the compressed model obtained by training is limited by the specific data set of this kind of tasks.B · US 2022/0222529 A1 · ¶ 3
In view of the shortcomings of the prior art, the purpose of the present application is to provide a method for meta-knowledge fine-tuning and platform based on domain-invariant features.B · US 2022/0222529 A1 · ¶ 4
patent.fyi, US 2026/0354731 A1, claim 1, first published 2026-11-19T02:47:19-05:00, https://patent.fyi/US20260354731/, SHA-256 3db53138e3a8c4ed0ace59715c6f219e77a748b523ec77e882941147019ef8cb