Search

EP-4740867-A1 - SYSTEM FOR GENERATING PHYSIOLOGICAL TIME-SERIES VIA DEEP LEARNING MODEL DETECTION OR PREDICTION

EP4740867A1EP 4740867 A1EP4740867 A1EP 4740867A1EP-4740867-A1

Abstract

Methods and systems are provided for detecting and predicting uterine activity (UA) from an ultrasound receive signal and/or electrohysterogram (EHG) data, using a neural network model (420) trained using ultrasound time-series data (402) and/or EHG data as input data (404) and time-series tocodynamometer (toco) data as ground truth data (405). After training, the trained UA detection model (430) may be used to detect a timing, duration, and intensity of uterine contractions in new patients with more accuracy than the toco. An output of the trained UA detection model (430) may also be used to train a UA prediction model (470) to predict a next uterine contraction (1210) based on a series of preceding uterine contractions. As a result, the uterine contractions may be detected and/or predicted during electronic fetal monitoring (EFM) without the use of an additional toco device, reducing a cost of the EFM and a number of sensing devices relied on for the EFM.

Inventors

  • KHAIR, MOHAMMAD
  • MANICKAM, Kalaivani

Assignees

  • GE Precision Healthcare LLC

Dates

Publication Date
20260513
Application Date
20251015

Claims (15)

  1. A method, comprising: acquiring ultrasound time-series data of an abdomen and/or a fetus of a pregnant mother (602); inputting the ultrasound time-series data into a first trained neural network model (606); generating a plot of a uterine activity (UA) of the mother or the fetus from an output of the first trained neural network model (610); and displaying the plot on a display device (610).
  2. The method of claim 1, wherein the ultrasound time-series data is acquired from a receive signal of an ultrasound probe prior to processing the ultrasound time-series data to generate ultrasound physiological signals.
  3. The method of claim 2, wherein the ultrasound physiological signals are displayed on the display device concurrently with the plot.
  4. The method of claim 1, wherein the first trained neural network model is trained using stored ultrasound time-series data acquired from a plurality of subjects as input data, and stored time-series data of the mother's abdomen acquired from the same plurality of subjects at a same time as the stored ultrasound time-series data using a tocodynamometer as ground truth data.
  5. The method of claim 1, wherein the first trained neural network model is trained using stored time-series data of the mother's abdomen and/or fetus acquired from a plurality of subjects using an electrohysterogram (EHG) device as input data, and stored time-series data of the mother's abdomen and/or fetus acquired from the same plurality of subjects at a same time as the stored EHG time-series data using a tocodynamometer as ground truth data.
  6. The method of claim 1, wherein inputting the ultrasound time-series data into the first trained neural network model further comprises: generating a plurality of overlapping ultrasound time-series windows of the ultrasound time-series data at predefined increments of time (604); and inputting an overlapping ultrasound time-series window of the plurality of overlapping ultrasound time-series windows into the first trained neural network model at each increment of time (606).
  7. The method of claim 6, wherein generating a plot of the UA of the mother from the output of the first trained neural network model further comprises: receiving as an output of the first trained neural network model a plurality of overlapping windows of UA time-series data, each overlapping window of UA time-series data generated at the predefined increment of time; aggregating the plurality of overlapping windows of the UA time-series data by averaging values associated with a same time in each overlapping window, to create a continuous stream of UA time-series data (608); generating the plot of the UA from the continuous stream of UA time-series data (610).
  8. The method of claim 7, further comprising: generating training data for a second trained neural network model from the plurality of overlapping windows of UA time-series data, the training data including training pairs comprising an overlapping window of UA time-series data as input data, and a portion of a subsequent overlapping window of UA time-series data as ground truth data (508); and using the training data, training the second trained neural network to predict a subsequent uterine contraction from a series of preceding uterine contractions (512).
  9. The method of claim 1, wherein the first trained neural network model comprises one of a convolutional neural network (CNN) and a recurrent neural network (RNN).
  10. The method of claim 9, wherein the first trained neural network model is a CNN, and a first layer of the CNN is a wavelet decomposition transform layer that decomposes an input signal of the CNN into primary frequency components.
  11. A uterine activity (UA) detection system (302), comprising: a display device (334); and a processor (304) communicably coupled to the display device (334), and a non-transitory memory (306) including instructions that when executed cause the processor (304) to: during a first, training stage of a UA detection model (420) of the UA detection system (302): generate a first plurality of overlapping windows (408) of ultrasound time-series data (402) acquired from a plurality of ultrasound scans performed on a plurality of fetuses and/or abdomens of pregnant mothers; generate a second plurality of overlapping windows (408) of UA time-series data (402) acquired via a tocodynamometer (toco) from the pregnant mothers at a same time as the ultrasound time-series data; train the UA detection model (420) on training pairs (412) including a time-series window (408) of the first plurality of overlapping windows as input data (404) and a time-series window of the second plurality of overlapping windows as ground truth data (405); and during a second, inference stage of the UA detection model (420): receive ultrasound time-series data (456) from an ultrasound probe placed over a fetus or a pregnant mother's abdomen (452); generate a third plurality of overlapping windows (458) of the ultrasound time-series data; at predefined increments of time, input an overlapping window (458) of the third plurality of overlapping windows (458) into the trained UA detection model (430); receive as outputs of the trained UA detection model (430), at each predefined increment of time, a fourth plurality of overlapping windows of UA time-series data (460); aggregate the fourth plurality of overlapping windows of UA time-series data (460) by averaging values associated with a same time in each overlapping window, to create a continuous stream of UA time-series data (464), the continuous stream of UA time-series data (464) showing uterine contractions of the pregnant mother (452); and display a plot of the continuous stream of UA time-series data on the display device.
  12. The UA detection system (302) of claim 11, wherein the ultrasound time-series data (456) is acquired from a receive signal of an ultrasound probe prior to processing the ultrasound time-series data (456) to generate an ultrasound physiological signals.
  13. The UA detection system (302) of claim 11, wherein further instructions are stored in the memory (306) that when executed, cause the processor (304) to process the ultrasound time-series data (456) to generate an ultrasound physiological signals (720), and display the ultrasound physiological signals (720) on the display device (334) concurrently with the plot (722) of the continuous stream of UA time-series data (464).
  14. The UA detection system (302) of claim 11, wherein the UA detection model (420) is trained using UA time-series data (402) acquired from the pregnant mothers during the plurality of ultrasound scans via an electrohysterogram (EHG) device as ground truth data.
  15. The UA detection system (302) of claim 11, wherein further instructions are stored in the memory (306) that when executed, cause the processor (304) to: generate training data for a UA prediction model (420) from the plurality of overlapping windows (408) of UA time-series data (402), the training data including training pairs (412) comprising an overlapping window (408) of UA time-series data (402) as input data, and a portion of a subsequent overlapping window (408) of UA time-series data (402) as ground truth data (405); and using the training data, train the UA prediction model (420) to predict a subsequent uterine contraction (1210) from a series of preceding uterine contractions (1202, 1204, 1206, 1208).

Description

TECHNICAL FIELD Embodiments of the subject matter disclosed herein relate to detecting or predicting physiological signals of a patient, including fetal heart rate activity and uterine activity in a pregnant mother. BACKGROUND Uterine Activity (UA) corresponding to a mother's uterine contractions may be measured using a tocodynamometer, or toco, which measures a displacement of the mother's abdomen via a belt mounted mechanical pressure transducer. The UA may be represented as time-series data, which may be used to determine a frequency, an amplitude, a duration, and/or a period of the uterine contractions. However, the data generated by the toco may not be as precise as desired, due to a dependence on abdominal position and variable tension on the belt, as well as elevation ambient pressure dynamics that may affect the data. Clinical ultrasound is an imaging modality that employs ultrasound waves to probe the internal structures of a body of a patient and produce a corresponding physiological signal. For example, an ultrasound probe comprising a plurality of transducer elements emits ultrasonic pulses which reflect or echo, refract, or are absorbed by structures in the body. The ultrasound probe then receives reflected echoes, which are processed into a signal. Transabdominal ultrasound performed on a pregnant mother also captures contractions of the uterine muscle, which is positioned in the path of an abdominal ultrasound beam. However, time-series data generated by a toco may not indicate the frequency and duration of the uterine contractions as accurately as the transabdominal ultrasound. The ultrasound has more sensitivity to tissue compression and is also less positionally dependent when transducer is placed on mother's abdomen, as compared to the Toco. SUMMARY The current disclosure at least partially addresses one or more of the above identified issues via a method, comprising acquiring ultrasound time-series data of an abdomen and/or fetus of a pregnant mother; inputting the ultrasound time-series data into a first trained neural network model; generating a plot of a uterine activity (UA) of the mother from an output of the first trained neural network model; and displaying the plot on a display device. The above advantages and other advantages, and features of the present description will be readily apparent from the Detailed Description below when taken alone or in connection with the accompanying drawings. It should be understood that the summary above is provided to introduce in simplified form a selection of concepts that are further described in the detailed description. It is not meant to identify key or essential features of the claimed subject matter, the scope of which is defined uniquely by the claims that follow the detailed description. Furthermore, the claimed subject matter is not limited to implementations that solve any disadvantages noted above or in any part of this disclosure. BRIEF DESCRIPTION OF THE DRAWINGS Various aspects of this disclosure may be better understood upon reading the following detailed description and upon reference to the drawings in which: FIG. 1 shows a block diagram of an exemplary ultrasound system;FIG. 2 shows an exemplary tocodynamometer;FIG. 3 is a schematic diagram of an exemplary UA detection system used to predict uterine contractions in pregnant mothers;FIG. 4 is an information flow diagram showing an exemplary system for training and deploying a UA detection model used by the UA detection system;FIG. 5 is a flowchart illustrating an exemplary method for training the UA detection model;FIG. 6 is a flowchart illustrating a method for using a UA detection model to predict uterine contractions;FIG. 7 is a first graph of time-series data showing UA in a pregnant mother, as prior art;FIG. 8 is a second graph of time-series data showing UA in a pregnant mother, as prior art;FIG. 9 shows an exemplary array of UA data values extracted from UA time-series data;FIG. 10 shows an example of UA time-series windows extracted from the array of UA data values;FIG. 11 is a third graph of time-series data showing UA in a pregnant mother, as prior art;FIG. 12 is a graph showing a plot of an exemplary output of the UA detection model; andFIG. 13 is a schematic diagram showing an exemplary neural network architecture of the UA detection model. DETAILED DESCRIPTION Accurately evaluating the well-being of a fetus during labor and delivery may lead to a healthy newborn and mother. Electronic Fetal Monitoring (EFM) is a common obstetrical procedure in the US, and most women undergo EFM during pregnancy, labor, and delivery. The rationale for use of EFM assumes that Fetal Heart Rate (FHR) abnormalities accurately reflect risk of hypoxia (inadequate low oxygenation of the fetus), and that early recognition of hypoxia could induce intervention to improve outcomes for the fetus. EFM may also be used to detect post-partum hemorrhage due to lack of contraction of uterus. EFM is