Discriminating features based sepsis prediction

Patent metadata

Field	Value
Publication number	US-11817217-B2
Application number	US-202017117375-A
Country	US
Kind code	B2
Filing date	Dec 10, 2020
Priority date	Mar 21, 2020
Publication date	Nov 14, 2023
Grant date	Nov 14, 2023

Abstract

Sepsis is one of the most prevalent causes of mortality in Intensive Care Units (ICUs) and delayed treatment is associated with increase in death and financial burden. There is no single laboratory test or clinical sign that by itself can be considered diagnostic of sepsis. The present disclosure provides discriminating domain specific continuous and categorical features that can reliably classify a subject being monitored into a sepsis class or a normal class. A combination of physiological parameters, laboratory parameters and demographic details are used to extract the discriminating features. Even though the parameters may be sporadic in nature, the systems and methods of the present disclosure make use of a sliding time window to generate continuous features that capture the trend in the sporadic data; and a binning approach to generate categorical features to discriminate deviation from the normal class and facilitate timely treatment.

First claim

What is claimed is:

1. A processor implemented method comprising the steps of:
receiving, via one or more hardware processors, test signals corresponding to one or more parameters, wherein the one or more parameters include physiological parameters, laboratory parameters and demographic parameters, and wherein the test signals are obtained periodically from a subject being monitored, wherein the subject is housed in an intensive care unit (ICU);
pre-processing, via one or more hardware processors, the received test signals by imputing missing data using at least one of Last Observation Carried Forward (LOCF), Next Observation Carried Backward (NOCB), mean imputation, and interpolation techniques, wherein the imputation of the missing data is performed subject-wise in a training phase of at least one classifier and time window-wise for the received test signals associated with the subject being monitored;
extracting, via the one or more hardware processors, a plurality of discriminating features from the received test signals, wherein the plurality of discriminating features is a combination of continuous features and categorical features extracted during the training phase of the at least one classifier, wherein the plurality of discriminating features comprises:
(i) the continuous features associated with the one or more parameters being the physiological parameters and the laboratory parameters over an empirically determined sliding time window (i hours);
(ii) the categorical features by using a binning approach on the one or more parameters being the physiological parameters and the laboratory parameters;
(iii) the categorical features corresponding to the demographic parameters; and
(iv) the categorical features corresponding to a sepsis domain;
normalizing, via the one or more hardware processors, the extracted plurality of discriminating features using stored values of standard deviation and mean associated with the one or more parameters from the training phase, wherein the normalizing is performed across sliding time windows;
ranking, via the one or more hardware processors, the normalized extracted plurality of discriminating features using a Minimum Redundancy Maximum Relevance (MRMR) method to identify a subset of the plurality of discriminating features such that the subset is representative of a sepsis condition in the subject being monitored; and
predicting, via the one or more hardware processors, a classification of the subject being monitored into one of a sepsis class and a normal class based on the identified subset using the at least one classifier trained during the training phase.

2. The processor implemented method of claim 1 , wherein the training phase of the at least one classifier comprises:
receiving training signals corresponding to the one or more parameters, and wherein the training signals are obtained periodically from a plurality of subjects;
extracting the plurality of discriminating features from the received training signals;
normalizing the extracted plurality of discriminating features using values of standard deviation and mean associated with the one or more parameters, wherein the normalizing is performed across the plurality of subjects;
ranking the normalized extracted plurality of discriminating features using the MRMR method to identify the subset of the plurality of discriminating features;
under-sampling a majority class to balance the sepsis class and the normal class in the received training signals; and
training the at least one classifier using the identified subset corresponding to the balanced sepsis class and the normal class.

3. The processor implemented method of claim 2 , wherein the step of extracting the plurality of discriminating features from the received training signals is preceded by preprocessing the received training signals by imputing missing data therein.

4. The processor implemented method of claim 2 , wherein the at least one classifier includes at least one of a Random Forest (RF) model and an Adaptive Logistic Regression of Ensemble learning (LB) model.

5. The processor implemented method of claim 1 , wherein the plurality of discriminating features comprises:
(i) the continuous features associated with the one or more parameters being the physiological parameters and the laboratory parameters over the empirically determined sliding time window (i hours) including:
a) count of valid data records associated with the one or more parameters;
b) count of out of range data records associated with the one or more parameters;
c) difference between a value associated with the one or more parameters at an i th hour and an (i−1) th hour;
d) difference between a value associated with the one or more parameters at the i th hour and a mean of past (i−1) hours; and
e) difference between a value associated with the one or more parameters at the i th hour and a variance of the past (i−1) hours;

(ii) the categorical features by using the binning approach on the one or more parameters being the physiological parameters and the laboratory parameters including:
a) Category;
b) Upper limit (UL);
c) Lower limit (LL); and
d) One hot dummy encoding;

(iii) the categorical features corresponding to the demographic parameters including:
a) Current hospital admission duration;
b) Intensive Care Unit (ICU) type;
c) Gender;
d) Age binned in empirically determined categories of ranges of the age;
e) Current ICU stay duration binned in steps of empirically determined hours;
f) The LL of current ICU stay duration;
g) The UL of current ICU stay duration;
h) The LL of age; and
i) The UL of age; and

(iv) the categorical features corresponding to the sepsis domain including:
a) National Early Warning Score (NEWS);
b) Modified Early Warning Score (MEWS); and
c) Acute Physiology and Chronic Health Evaluation II (APACHE II).

6. The processor implemented method of 4 , further comprising obtaining a final class of the subject being monitored by fusing predictions by the RF model and the LB model using one or more logical operations.

7. The processor implemented method of 6 , further comprising administering antibiotics to the subject being predicted as belonging to the class sepsis.

8. A system comprising:
one or more data storage devices operatively coupled to one or more hardware processors and configured to store instructions configured for execution via the one or more hardware processors to:
receive test signals corresponding to one or more parameters, wherein the one or more parameters include physiological parameters, laboratory parameters and demographic parameters, and wherein the test signals are obtained periodically from a subject being monitored, wherein the subject is housed in an intensive care unit (ICU);
pre process the received test signals by imputing missing data using at least one of Last Observation Carried Forward (LOCF), Next Observation Carried Backward (NOCB), mean imputation, and interpolation techniques, wherein the imputation of the missing data is performed subject-wise in a training phase of at least one classifier and time window-wise for the received test signals associated with the subject being monitored;
extract a plurality of discriminating features from the received test signals, wherein the plurality of discriminating features is a combination of continuous features and categorical features extracted during the training phase of the at least one classifier, wherein the plurality of discriminating features comprises:
(i) the continuous features associated with the one or more parameters being the physiological parameters and the laboratory parameters over an empirically determined sliding time window (i hours);
(ii) the

Assignees

• Tata Consultancy Services Ltd

Inventors

• Sharma Varsha
• Bhattacharyya Chirayata
• Bhattacharjee Tanuka
• Poduval Murali
• Khandelwal Sundeep
• Dutta Choudhury Anirban

Classifications

• G16H50/30 Primary

  for calculating health indices; for individual health risk assessment · CPC title

• A61B5/412

  Detecting or monitoring sepsis · CPC title

• A61B5/7267

  involving training the classification device · CPC title

• G06F18/2113

  by ranking or filtering the set of features, e.g. using a measure of variance or of feature cross-correlation · CPC title

• G06F18/2132

  based on discrimination criteria, e.g. discriminant analysis · CPC title