Patent Application
20240298954
The disclosure concerns method and system subject matter dealing with predicting delirium in patients.
About 80% of ICU patients develop delirium, and about 60% of them will still have impaired memory and attention at discharge that prevents them from managing their treatment plan at home. Delirium triples one year mortality risk, and increases risk of need for long term care and hospital readmission costing ˜$164 billion/year. Delirium can occur in any setting or environment and is not limited to the ICU, hospital or healthcare facility. Older adults are at greater risk meaning the prevalence will increase as the population ages and more survive critical illness. Without an accurate method of detection, interventions and prevention strategies cannot accurately be tested. Bedside clinical assessments detect less than 20% of ICU delirium when a validated tool is used, and less than 10% of clinicians report using a validated method despite being a national guideline from more than 10 professional organizations for more than 10 years.
While the lack of an accurate method of detection prevents the use of interventions and prevention strategies, currently delirium is assessed using a bedside clinical screen that requires the patient be able to participate in the assessment. Thus, typically, such bedside clinical assessments are only able to detect less than 20% of ICU delirium even when a validated tool is applied.
Delirium is an acute syndrome manifested by changes in global cognitive function including either disorganized thinking or altered level of consciousness.[1] Delirium occurs in as many as 80% of critically ill older adults and is associated with worse long-term cognitive outcomes.[2, 3] For more than 20 years, at least 10 national and international professional organizations have included routine delirium screening in clinical practice guidelines.[4-6] Despite these recommendations and the availability of more than 40 validated screening tools, less than 10% of clinicians report routine delirium screening.[4, 7] Many critically ill patients are unable to participate in delirium screening, and are therefore untestable, due to coma or deep sedation.
Even with validated screening tools, delirium remains difficult to recognize, therefore, frequently underdiagnosed and undertreated. As duration and severity of delirium increases, it becomes increasingly more difficult to treat. Delirium is associated with a one-year increase in economic burden of more than$44K/patient, making it a global public health crisis.[8]
The electroencephalogram (EEG) is a representative signal describing the condition of the brain.[9, 10] The shape, amplitude, and oscillation speed help describe the condition and assist with diagnostics.[11, 12] Using EEG for delirium detection was first identified in the 1940's. Romano and Engel identified slowing of EEG with increases in sleep and decreases in awake waves when delirium was present.[13, 14] Thus, delirium has been reliably identified by changes in neural activity EEG. Unfortunately, the significant cost associated with technological set up and the need for expert analysis has prevented the use of EEG for delirium detection in the clinical environment.[15, 16]
More recently, user-friendly preprogrammed handheld EEG devices (as represented by
A widely used machine learning technique in research and practice is supervised machine learning where the ground truth is known to the researchers and labeled in the training dataset. Sophisticated deep neural networks are often adopted and optimized to fit the needs of specific learning tasks in real world scenarios. While the deep-learning-based approach often outperforms traditional statistical machine learning algorithms in prediction accuracy, its Blackbox nature often makes the model impractical to use in the medical field. Additional modules are often needed to help explain why or how the decision was made by deep learning algorithms.
Presently disclosed methodology and associated/related systems deal generally with predicting delirium in patients, and more particularly for some embodiments with using limited-lead rapid-response EEG and supervised deep learning methods with vision transformer to predict delirium.
Per presently disclosed subject matter, using a wireless EEG device using fewer leads than traditional EEG and a machine learning algorithm, delirium can be detected prior to patients having symptoms and with significantly greater accuracy and in patients that currently cannot be assessed. Further, such presently disclosed subject matter can advantageously be used in home, hospital, or long term care settings with little to no training.
Per some presently disclosed embodiments, use of a vision transformer with rapid-response EEG is capable of predicting delirium. Such monitoring is feasible in critically ill older adults, for improving delirium detection accuracy, and providing greater opportunity for individualized interventions. Benefits may shorten hospital length of stay, increase discharge to home, decrease mortality, and reduce the financial burden associated with delirium.
In accordance with presently disclosed subject matter, for some embodiments, using machine learning and limited lead EEG delirium can be detected prior to the onset of symptoms with 98% accuracy.
It is to be understood that the presently disclosed subject matter equally relates to both systems and to associated and/or corresponding methodologies. One exemplary such method relates to a method for predicting delirium in patients which integrates a deep learning based model with electroencephalogram (EEG) data, the method comprising training a machine-learned supervised deep learning method model to predict the presence of delirium in a patient based on training data associated with at least a plurality of limited-lead rapid-response EEG training data sets from patients; obtaining EEG test data associated with a target patient to be tested for the presence of delirium; inputting the EEG test data into the machine-learned supervised deep learning method model; and receiving, as output of the model, a positive or negative prediction of whether the target patient is experiencing the presence of delirium.
Another exemplary such method relates to a method for predicting delirium in patients which integrates a deep learning based model with electroencephalogram (EEG) data, the method comprising training a machine-learned supervised deep learning method vision transformer based model to predict the presence of delirium in a patient based on training data associated with at least a plurality of limited-lead rapid-response EEG training data sets from patients; obtaining EEG test data using a limited-lead rapid-response EEG device associated with a target patient to be tested for the presence of delirium; inputting the EEG test data into the machine-learned supervised deep learning method model; and receiving, as output of the model, a positive or negative prediction of whether the target patient is experiencing the presence of delirium.
Other example aspects of the present disclosure are directed to systems, apparatus, tangible, non-transitory computer-readable media, user interfaces, memory devices, and electronic devices for predicting delirium in patients. To implement methodology and technology herewith, one or more processors may be provided, programmed to perform the steps and functions as called for by the presently disclosed subject matter, as will be understood by those of ordinary skill in the art.
Another exemplary embodiment of presently disclosed subject matter relates to a system for predicting delirium in patients which integrates a deep learning based model with electroencephalogram (EEG) data, the system comprising a machine-learned supervised deep learning method model trained to predict the presence of delirium in a patient based on training data associated with at least a plurality of limited-lead rapid-response EEG training data sets from patients; one or more processors; and one or more non-transitory computer-readable media that store instructions that, when executed by the one or more processors, cause the one or more processors to perform operations. Such operations preferably comprise obtaining EEG test data associated with a target patient to be tested for the presence of delirium; inputting the EEG test data into the machine-learned supervised deep learning method model; and receiving, as output of the model, a positive or negative prediction of whether the target patient is experiencing the presence of delirium.
Additional objects and advantages of the presently disclosed subject matter are set forth in, or will be apparent to, those of ordinary skill in the art from the detailed description herein. Also, it should be further appreciated that modifications and variations to the specifically illustrated, referred and discussed features, elements, and steps hereof may be practiced in various embodiments, uses, and practices of the presently disclosed subject matter without departing from the spirit and scope of the subject matter. Variations may include, but are not limited to, substitution of equivalent means, features, or steps for those illustrated, referenced, or discussed, and the functional, operational, or positional reversal of various parts, features, steps, or the like.
Still further, it is to be understood that different embodiments, as well as different presently preferred embodiments, of the presently disclosed subject matter may include various combinations or configurations of presently disclosed features, steps, or elements, or their equivalents (including combinations of features, parts, or steps or configurations thereof not expressly shown in the figures or stated in the detailed description of such figures). Additional embodiments of the presently disclosed subject matter, not necessarily expressed in the summarized section, may include and incorporate various combinations of aspects of features, components, or steps referenced in the summarized objects above, and/or other features, components, or steps as otherwise discussed in this application. Those of ordinary skill in the art will better appreciate the features and aspects of such embodiments, and others, upon review of the remainder of the specification, and will appreciate that the presently disclosed subject matter applies equally to corresponding methodologies as associated with practice of any of the present exemplary devices, and vice versa.
These and other features, aspects and advantages of various embodiments will become better understood with reference to the following description and appended claims. The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments of the present disclosure and, together with the description, serve to explain the related principles.
A full and enabling disclosure of the present subject matter, including the best mode thereof to one of ordinary skill in the art, is set forth more particularly in the remainder of the specification, including reference to the accompanying figures in which:
Repeat use of reference characters in the present specification and drawings is intended to represent the same or analogous features, elements, or steps of the presently disclosed subject matter.
Reference will now be made in detail to various embodiments of the disclosed subject matter, one or more examples of which are set forth below. Each embodiment is provided by way of explanation of the subject matter, not limitation thereof. In fact, it will be apparent to those skilled in the art that various modifications and variations may be made in the present disclosure without departing from the scope or spirit of the subject matter. For instance, features illustrated or described as part of one embodiment, may be used in another embodiment to yield a still further embodiment.
In general, the present disclosure is directed to methodology and associated/related systems dealing generally with predicting delirium in patients. More particularly, some embodiments are directed to using limited-lead rapid-response EEG and supervised deep learning methods with vision transformer to predict delirium.
Per presently disclosed embodiments, to evaluate rrEEG waveforms, signal parameters may be extracted and analyzed using computer based statistical algorithms capable of accurate predictive detection, so that such presently disclosed rrEEG device embodiments may provide a feasible physiologic method to support delirium detection.
As otherwise noted herein, one known general machine learning technique in research and practice is referred to as supervised machine learning, where the ground truth is known to the researchers and labeled in the training dataset. For some exemplary embodiments disclosed herewith, a supervised deep learning model used is Vision Transformer (ViT).[19-21]
As represented in
A proof-of-concept pilot approach for Methods of Detecting Neurological Delirium (MIND) uses a prospective design to evaluate use of supervised deep learning, ViT, and a rrEEG device for predicting delirium in critically ill older adults requiring mechanical ventilation. Such initial prospective proof-of-concept approach uses a rrEEG device providing data from all cerebral lobes, and the presently disclosed ViT method to evaluate rrEEG for the presence of delirium in critically ill patients. For such initial approach, written informed consent was obtained from the participant's legally authorized representative prior to any research activities, and all research procedures were conducted in accordance with the ethical standards set by the UMCIRB IRB and the Helsinki Declaration of 1975.
The protocol for review of such approach was previously published.[22] In brief, seventeen patients meeting inclusion and exclusion criteria were recruited from three intensive care units (cardiac, medical, and surgical ICU) in a large rural academic medical center in North Carolina between March 2019 and March 2020. All participants 50 years old or older, required mechanical ventilation for greater than 12 hours, and were English speaking. Exclusion criteria included acute brain injury, seizures, or condition preventing participation in the delirium screening.
Each day, participants were assessed for the ability to participate in a delirium screening determined using the Richmond Agitation Sedation Scale (RASS).[23, 24] The RASS is a 10 level scale (+4 “combative” to −5 “unarousable”) with excellent inter-rater reliability (r=0.956, lower 90% confidence limit=0.948; k=0.73, 95% confidence interval=0.71, 0.75).[23, 24] A RASS score of −2 or higher (able to open eyes for >10 seconds to voice) met eligibility. Demographic and clinical characteristics were obtained from the electronic medical record.
Bedside Behavioral Assessment for Delirium, The Confusion Assessment Method for the ICU (CAM-ICU), is a modified version of the CAM developed to assess mechanically ventilated and non-verbal patients.[25, 26] The CAM-ICU is based on the gold standard for delirium identification, the Diagnostic and Statistical Manual for Mental Disorder IV (DSM-IV) and one of two delirium screening tools recommended for use by the Society of Critical Care Medicine (SCCM).[1, 4, 27] The CAM-ICU has also been validated for use with older adults.[6, 28] Therefore, the CAM-ICU was considered the gold standard for determining delirium status. The CAM-ICU requires patient participation to identify four key delirium features including an a) acute onset or fluctuation in mental status within the previous 24 hours, (b) inattention, (c) altered level of consciousness [Richmond Agitation Sedation Scale (RASS) not=0], and (d) disorganized thinking.[25, 26] When used in research, the CAM-ICU has high sensitivity and specificity, 93 and 98% respectively and high inter-rater reliability at k.0.79.
Rapid-response EEG (rrEEG) headbands that circumscribe the head (as represented in
rrEEG Processing
rrEEG data were processed to remove artifacts such as muscle movement in the face and interference from nearby devices such as ventilators and cardiac monitors using high and low frequencies filters. Data are then re-referenced to estimate physiological noise and divided into multiple discrete time periods called epochs. rrEEG data is further “cleaned” using individual component analysis (IDA) to remove noises and generate features needed for machine learning algorithms. Component analysis is a widely accepted method to separate artifact from cortical processes data.[18, 29] The benefit of IDA with higher order statistics is the ability to simply subtract artifacts by directly examining the data's independent components.
EEG analytic techniques across studies have varied. Therefore, both traditional and three machine learning methods were used to analyze these data, specifically random forest (series of decision trees), stepwise linear discriminant analysis (removes variables that do not help classify the data, in this case delirium−/delirium+), and support vector machine (computer builds a model to provide the greatest difference between categories, in this case delirium−/delirium+). Due to challenges with feature selection, a supervised deep learning method, ViT, was primarily used.
Many deep learning applications, having sophisticated data processing techniques and feature engineering, are often not needed because the deep neural networks can learn those subtle features directly from the input data. Therefore, two types of data were studied. The first type of input data was preprocessed (muscle movement and device interference removal) and IDA cleaned, while the second type of input received only preprocessing with no IDA cleaning.
rrEEG devices extracted data every 4 ms, and each data sample contained readings from 10 sensors. A continuous number of rows of data are organized into a data slice, which is an 8×n array, where n represents the number of rows. These arrays are resized to 224×224 using bilinear interpolate and treated as images to feed into the ViT model. Note the diagrammatical representations of
To understand how the results are related to the size of the slices, five different lengths were chosen: 25 rows (0.1 s), 125 rows (0.5 s), 250 rows (1 s), 400 rows (1.6 s), and 1250 rows (5 s). Due to the small sample size, data are augmented using an overlapping window scheme, where the starting row of the next data slice is located somewhere in the current data slice, rather than after the last row of the current data slice. An example of 30% overlapping data slice of 1 second (250 rows) is graphically represented in
The following default hyper-parameters of the model were used in conjunction with present initial embodiments of the presently disclosed subject matter: batch size=64, learning rate=0.001, depth=12, and heads=8. Overlapping ratio of 75%, 90%, and 95% were reviewed. Since no major differences were discovered among these different overlapping ratios, 90% windows overlapping ratio was used to report the results. rrEEG data converges very quickly when using the ViT model. In most situations, training accuracies achieved more than 99% in as little as 3 epochs. To avoid overfitting, models trained after 5 epochs were used to evaluate the testing datasets.
To better understand the value of the ViT model in rrEEG data analyses, a public dataset Goldberger, Amaral[30] were also used to perform a binary classification task. The data slice of 1,250 rows was adopted and the overlapping ratio was set to 90%. The model achieved testing accuracy of 86.33%, which is better than state-of-the-art algorithms SleepEEGNet at the accuracy of 80.03%[31]. The initial review results show that ViT is a better fit to analyze rrEEG data than existing algorithms.
Fifteen different models (5 data slice sizes×3 overlapping rates) were used to evaluate the performance of the ViT models. Since overlapping rates did not impact the accuracy results, only findings of the overlapping rate=90% are reported, see
When a data slice contains at least a full cycle of 2 Hz waves (0.5 second, or 125 rows), the accuracy increased to 72.82%. When a data slice contains at least a full cycle of 1 Hz waves (1 second, or 250 rows), the accuracy is further increased to around 95%. Better results are associated with lower frequency waves. Further review and testing would be needed to fully understand if long wave signals, such as delta waves, are truly the predictors for delirium.
As for the training speed, under the setting of 1250 rows and 90% overlapping rate, about 2 hours were required. The training time is proportional to the size of training set. Recall that the data slice will be resized to 224×224, model with 250 s row, and 90% overlapping rate would take 10 hours to train.
Since ViT worked well on data cleaned using IDA, it begs the question: would ViT provide good prediction results on datasets not cleaned by IDA? To answer this question, uncleaned data were also fed into the ViT model to evaluate the results. Data slices of 1,250 were chosen with an overlapping rate of 90%. Both training and testing accuracies reached greater than 99.99%. The model converged within 3 epochs, as shown in
The datasets were limited to only 12 subjects and used only on delirium predictions. One question could be what about other rrEEG datasets. We did another experiment to understand if ViT is still applicable by applying ViT to a public EEG data set[30] using a binary classification task. Under a window size of 1250 points, the model reached 86.33% test accuracy.
The datasets generated and/or analyzed in conjunction with present initial embodiments of the presently disclosed subject matter are not publicly available but are available from the corresponding author on reasonable request.
This is the first prospective application of presently disclosed technology using a 10-electrode rrEEG device providing data from all lobes in the cerebrum, supervised deep learning, and ViT to evaluate rrEEG for the presence of delirium. This initial approach established that such monitoring is feasible in critically ill older adults across medical, surgical, and cardiac ICUs. One significant determination regarding the presently disclosed technology is that, using supervised deep learning and a ViT platform, patients were accurately classified as delirium positive or delirium negative based on identifiable characteristics in rrEEG, thus predicting the presence of delirium. These results were replicated using three methods of machine learning, including stepwise linear discriminant analysis, support vector machines, and supervised deep learning using ViT. Compared to prior studies using various machine learning and preprocessing methods, the ViT model using the hyperparameters mentioned above has provided greater accuracy.
For example, van Sleuwen et al. (2022) used a 3-channel limited lead device to measure physiologically based methods using the CAM-Severity Score[32]. To obtain three channels or waveforms, they used 6-second EEG strips obtained from a 4-electrode frontal montage. Using this montage on 252 delirious and 121 non-delirious patients, they obtained accuracies of 0.63-0.73 on the ROC curve, meaning that the model accurately predicted 63-70% of true positives for every possible decision threshold of the model. Similarly, Yamanashi, Malicoat[33] were able to obtain AUCs of 63-76% using a bispectral EEG with two channels. As a result, they recommended further studies may benefit from deep learning models such as one of the aspects used in present evaluations.
The merit of this handheld rrEEG is its objectiveness compared to currently available bedside screening methods. Additionally, this method does not require the use of a large EEG machine and specialized technicians for electrode placement that frequently limit large volume mass screening such as that needed to screen for delirium in ICUs. rrEEG is easy to use by busy hospital staff with minimal training. Use of pre-programmed algorithms, such as the one described here, limits the need for skilled interpretation required when using traditional EEG. Therefore, this method of detection is not limited by subjectivity and rationalizing of results associated with bedside screening.
While the ViT model used above has good performance, there are still opportunities for further improved prediction. For example, using a window of 10-60 seconds instead of the 1-5 second windows typically used in this type of analysis may have provided more data points. During the pre-processing phase, the presently disclosed technology used sequence extraction, meaning samples are obtained in a chronological sequential order and at a finite length. It is possible that constant correction or interpolation (removing data obtained from lead electrode that is bad) may provide a further cleaned data set for analysis.
Limitations of the present analysis of presently disclosed technology include the small sample of 13 participants with 7 experiencing delirium during the monitoring period as determined by the CAM-ICU. EEG changes occur prior to symptom onset and therefore, some of the participants may have had subsyndromal delirium detected using rrEEG that were not detected using bedside screening (CAM-ICU). Delirium assessments were conducted by the researcher rather than using clinician assessments providing stronger reliability of delirium status. Heterogeneity of the sample with varying etiologies and medication exposures could have resulted in some rrEEG changes being reflected more than others in a subset of patients, thus minimizing generalizability. Despite limitations, consistent results have been obtained across methods of analysis (frequency ratios, supervised learning, and deep learning) providing strength to the findings.
In this analysis, we trained a ViT model to analyze rrEEG data under the constraint of a small sample and hence a limited amount of data. Despite the use of a sequence extraction method for preprocessing, 97% accuracy is significantly better than the 40% accuracy of clinician derived CAM-ICU assessments. Therefore, this method has strong potential for improving accuracy of delirium detection, providing greater opportunity to implement and evaluate individualized interventions when delirium is more amenable to treatment. Additionally, rrEEG is significantly cheaper than traditional EEG without losing the benefit of being able to charge for a billable service.
This written description uses examples to disclose the presently disclosed subject matter, including the best mode, and also to enable any person skilled in the art to practice the presently disclosed subject matter, including making and using any devices or systems and performing any incorporated methods. The patentable scope of the presently disclosed subject matter is defined by the claims, and may include other examples that occur to those skilled in the art. Such other examples are intended to be within the scope of the claims if they include structural and/or step elements that do not differ from the literal language of the claims, or if they include equivalent structural and/or elements with insubstantial differences from the literal languages of the claims. In any event, while certain embodiments of the disclosed subject matter have been described using specific terms, such description is for illustrative purposes only, and it is to be understood that changes and variations may be made without departing from the spirit or scope of the subject matter. Also, for purposes of the present disclosure, the terms “a” or “an” entity or object refers to one or more of such entity or object. Accordingly, the terms “a”, “an”, “one or more,” and “at least one” can be used interchangeably herein.
The present application claims the benefit of priority of U.S. Provisional Patent Application No. 63/489,032, titled Using EEG To Predict Delirium Using Limited Lead Device, filed Mar. 8, 2023, and which is fully incorporated herein by reference for all purposes.
| Number | Date | Country | |
|---|---|---|---|
| 63489032 | Mar 2023 | US |
