Systems and methods for diagnosing a health condition based on patient time series data
Inventors
Wagner, Tyler • Aravamudan, Murali • BABU, Melwin • Barve, Rakesh • Soundararajan, Venkataramanan • Prasad, Ashim • CARPENTER, Corinne • Carlson, Katherine
Assignees
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Publication Number
US-12327638-B2
Publication Date
2025-06-10
Expiration Date
Abstract
Disclosed systems, methods, and computer readable media can diagnose a health condition based on patient time series data. For example, a method for diagnosing a health condition based on patient time series data includes identifying a training set of health records comprising a first set of patient time series data, training a neural network using the training set of health records, and executing the trained neural network model to diagnose a health condition based on a second set of patient time series data. In further examples, the first set of patient time series data and the second set of patient time series data can each comprise electrocardiogram data and the health condition can comprise pulmonary hypertension.
Core Innovation
The disclosure describes digital diagnostic techniques that use patient time-series data captured earlier than a predetermined amount of time before a date of a positive diagnosis for a health condition. A training set of health records includes a first set of patient time series data forming a preemptive set, and a neural network model is trained using the training set of health records and then executed to diagnose the health condition based on a second set of patient time series data.
The neural network model is trained by receiving initial parameters that are transfer learned from an independently learned self-supervised network. The training set of health records is divided into a training set, a validation set and a test set, and the initial parameters are modified as a function of the training set, the validation set and the test set. The approach includes selecting representations of the patient time series data such as ECG waveform-based inputs and, in some cases, ECG spectrogram representations including derived discrete metrics like a QT interval.
The disclosure further covers diagnosing by processing the second set of patient time-series data, including segmenting the second set into multiple time windows of a predetermined duration and aggregating per-window outputs. Comparative and example configurations described include a neural network model such as a single-branch 1D convolutional network with residual connections and use of overlapping windows. A pulmonary hypertension example is described using patient time-series health records and diagnostic capability relative to a positive diagnosis date, including preliminary results for AL amyloidosis using augmented curation and ECG-based classification.
Claims Coverage
The independent claims cover a method, a system, and a medical instrument for pre-diagnosis digital diagnosis using a neural network with transfer-learned initialization from an independently learned self-supervised network, trained with training/validation/test splits, and executed to diagnose a health condition based on patient time-series data. Across the independent claims, the inventive features are focused on preemptive time-series health records, transfer learning from self-supervised learning, training set splitting and parameter modification, and diagnosis execution on new time-series data, with further narrowing in dependent claims to ECG and pulmonary hypertension and additional representations or discrete metrics.
Preemptive time-series health-record training data for diagnosis
Identifying a training set of health records comprising a first set of patient time series data, wherein the first set of patient time series data comprises a preemptive set of patient time series data captured earlier than a predetermined amount of time before a date of a positive diagnosis for a health condition.
Transfer-learned initialization from an independently learned self-supervised network
Receiving initial parameters of the neural network model, wherein the initial parameters of the neural network model are transfer learned from an independently learned self-supervised network.
Training/validation/test splitting with parameter modification
Dividing the training set of health records into a training set, a validation set and a test set; and modifying the initial parameters of the neural network model as a function of the training set, the validation set and the test set.
Diagnosing a health condition from new patient time-series data using the trained neural network model
Executing the trained neural network model to diagnose the health condition based on a second set of patient time series data.
Neural-network diagnosis using preemptive health records implemented as a system
Identifying a training set of health records comprising a first set of patient time series data with a preemptive set earlier than a predetermined amount of time before a date of a positive diagnosis; training a neural network model using initial parameters transfer learned from an independently learned self-supervised network; dividing the training set into a training set, a validation set and a test set; modifying the initial parameters as a function of the training set, the validation set and the test set; and executing the trained neural network model to diagnose the health condition based on a second set of patient time series data.
ECG-based medical instrument for pulmonary hypertension diagnosis using transfer-learned neural network
An electrocardiogram monitor for capturing electrocardiogram data of a patient; and at least one hardware processor connected to the electrocardiogram monitor, wherein the at least one hardware processor receives the electrocardiogram data from the electrocardiogram monitor and executes a trained neural network model to diagnose the patient for pulmonary hypertension based on the electrocardiogram data, wherein training the trained neural network model comprises identifying a training set of health records comprising a first set of patient time series data with a preemptive set earlier than a predetermined amount of time before a date of a positive diagnosis and training the neural network model using initial parameters transfer learned from an independently learned self-supervised network with division into a training set, a validation set and a test set and modifying the initial parameters as a function of the training set, the validation set and the test set.
Across the independent claims, the core claim coverage centers on using preemptive patient time-series health records earlier than a predetermined amount of time before a positive diagnosis date, training a neural network with initial parameters transfer learned from an independently learned self-supervised network, dividing health records into training/validation/test sets and modifying parameters based on those sets, and executing the trained neural network to diagnose a health condition from a second set of patient time-series data. Dependent claim refinements further ground the inputs and targets to ECG-based pulmonary hypertension and can add representations and discrete metrics such as a QT interval and segmenting time series into multiple time windows.
Stated Advantages
Not explicitly described in patent.
Documented Applications
Not explicitly described in patent.
Interested in licensing this patent?