Method and system for predicting post-operative delirium by means of temporal clustering of EEG states

Abstract

The present invention describes a method for monitoring and analysing electroencephalography (EEG) signals to predict the probability of the development of post-operative delirium (POD) in patients undergoing surgery. The method involves receiving a patient's EEG signals, calculating the spectral power in predefined time windows using canonical frequency bands (delta, theta, alpha and beta) and comparing the calculated spectral powers with the centroids of at least two, preferably four, predefined brain state clusters. These clusters include low broadband power (LBP), high broadband power (HBP), alpha prevalent (AP) and delta prevalent (DP). On the basis of the comparison, the brain state is classified and the fractional occupancy of the LBP state, i.e. the proportion of time that the patient spends in this state, is calculated. The method then predicts the probability of developing POD according to the time spent in the LBP state. The invention also includes a system, a computer program product and a computer-readable storage means for implementing the method.

Description

METHOD AND SYSTEM FOR PREDICTING POSTOPERATIVE DELIRIUM BY TEMPORAL GROUPING OF EEG STATES DESCRIPTIVE MEMORANDUM SCOPE

[0001] The present invention relates to a method and system for monitoring and analyzing electroencephalograms (EEGs) that provides a prediction of the probability of postoperative delirium (POD). The method and system are based on the analysis of intraoperative EEG data using a spectral temporal clustering methodology. This analysis allows the classification of a subject's brain state at each moment of surgery into predefined brain states based on their spectral properties. Thus, the subject is characterized based on the relative time spent in each brain state.

[0002] Time spent in low broadband power (LBP) EEG states has been found to be associated with the occurrence of postoperative dysplasia (POD). This allows for the prediction of this condition before it occurs, which in turn enables the development of preventive and personalized interventions to address this adverse clinical outcome. BACKGROUND

[0003] Postoperative delirium (POD) is a transient, adverse cognitive condition characterized by inattention and confusion. It is associated with negative postoperative outcomes such as prolonged hospital stays, increased admissions to intensive care units (ICUs), progression to dementia, and increased mortality. The etiology of POD is multifactorial and complex, including neuroinflammation and blood-brain barrier dysfunction, among other factors.

[0004] Several solutions have been proposed in the patent literature to address the prediction of postoperative delirium (POD). For example, patent application W02023137400A1 describes a method for predicting delirium using electroencephalogram (EEG) analysis and machine learning algorithms. Application WO2022149789A1 also proposes a method for predicting postoperative delirium through EEG analysis, highlighting the use of advanced signal processing techniques. Furthermore, application US20210321928A1 presents devices and methods for providing a prediction of postoperative delirium based on the analysis of physiological and EEG data.

[0005] Despite these advances, there is still a need for more precise and personalized methods to predict postoperative day time (POD), especially in subjects undergoing surgery, taking advantage of techniques currently used during surgery, such as EEG-based anesthesia monitors.

[0006] To date, various EEG-based markers have been used to predict delirium with varying degrees of accuracy. Several of these EEG-based markers are associated with postoperative delirium (POD). For example, associations have been found between decreased alpha power (8–12 Hz), alpha frequency, burst suppression, lower preoperative spectral frequency limit (SEF), and lower EEG power. Most studies reporting associations between intraoperative EEG and POD use overall averages (typically 2 minutes of data) over long surgical periods (hours). However, the brain is a dynamic rather than a stationary system. For example, alpha oscillations are known to appear as spindles or transient increases in amplitude. Delta waves also occur during anesthesia as K-complexes, which are transient and unstable.Therefore, assuming steady-state brain states in a dynamic system seems like an incorrect assumption.

[0007] Additionally, it is known that postoperative depression (POD) is more frequent in older individuals. In this context, the signals usually associated with POD (e.g., alpha waves, SEF, etc.) are also age-related, meaning they do not provide much additional information beyond the subject's age when predicting POD. This increases the need for a different methodology than those currently available, one that not only considers age and age-related factors but also proposes an objective analysis of a subject's brain states at each moment during a surgical procedure, thus predicting the occurrence of POD more accurately.

[0008] The present invention employs an EEG data analysis methodology based on spectral temporal clustering to classify intraoperative brain states according to the spectral power of canonical frequency bands at any given time. Through an analysis of the associations between these clusters and postoperative depression (POD), it was possible to conclude that the probability of developing POD can be predicted, regardless of the subject's age, based on the time a subject spends in low-pressure brain states (LBB).

[0009] Accordingly, the present invention addresses this need by providing a detailed and evidence-based analysis of low-power EEG states during anesthesia, thus providing a valuable tool for clinical practice. DESCRIPTION

[0010] To address the problem at hand, a method and system for monitoring and analyzing electroencephalograms (EEG) is proposed. This system provides a prediction of the probability of postoperative delirium (POD) based on EEG data and its temporal spectral grouping. This would benefit surgical patients, especially the elderly or those undergoing more complex surgeries, as well as healthcare providers, allowing them to improve the management of POD episodes and modify anesthetic or surgical strategies to reduce the likelihood of developing POD. As previously mentioned, POD is a condition that can occur after surgery, particularly in older adults. Its occurrence is associated with increased mortality, morbidity, and cognitive decline; therefore, this solution facilitates the timely allocation of in-hospital resources for prevention treatments and / or specific management, which are not part of the present invention.

[0011] Spectral time clustering is an advanced data analysis technique used to identify patterns and structures in complex datasets that vary in both time and frequency domains. This technique is particularly useful in the context of EEG analysis proposed by the present invention, where the aim is to classify the brain states of subjects during surgical procedures. As its name suggests, spectral time clustering is based on spectral clustering techniques to which the temporal dimension of the analyzed signal is incorporated, which, as detailed below, is crucial to the proposed invention.

[0012] In general terms, spectral clustering is a technique that aims to organize or classify data based on its spectral characteristics, that is, based on how frequencies are distributed within a signal or dataset. In other words, it analyzes the frequency content of a signal to group segments or clusters that have similar patterns in their spectral components. The main components of spectral clustering include: a) Spectral analysis: Before grouping, a spectral analysis is performed, which involves transforming a signal from its time-domain representation (or original domain) to a frequency-domain representation. This is generally done using signal processing tools such as the Fourier Transform, which decomposes the signal into its sinusoidal components of different frequencies. The result is a frequency spectrum that shows which frequencies are present in the signal and with what amplitude or power. b) Definition of spectral characteristics: Spectral characteristics may include aspects such as the power or amplitude of the signal in different frequency bands (e.g., low, medium or high frequencies), the shape of the spectrum or specific frequency patterns. c) Clustering algorithms: Once the spectral features are obtained, unsupervised clustering algorithms (such as k-means, DBSCAN, or spectral algorithms) are used to classify the data into different groups or "clusters." These groups are defined based on the similarities between the spectral features of the analyzed segments. In practice, this means that parts of a signal or datasets with similar spectral characteristics are grouped into the same "cluster".

[0013] In the context of EEG signals, spectral grouping would involve: 1) Obtaining the spectrum of the EEG signal, which implies decomposing the signal into specific frequency bands (such as delta, theta, alpha, and beta); 2) Characterizing the frequencies, where each The signal segment is analyzed in terms of which frequencies predominate at that moment and what power they have; and 3) Apply clustering algorithms, so that with this spectral data, the signal segments that share similar power patterns in the frequency bands are grouped together.

[0014] Next, spectral time-grouping, or spectral estimation grouping over time, takes this concept a step further by considering the temporal dimension, which is crucial for understanding brain dynamics during anesthesia and surgery. Instead of assuming that the brain remains in a stable or stationary state for long periods, this methodology recognizes that the brain is a dynamic system and seeks to classify brain states based on the relative time spent in different brain states, defined by specific canonical frequency bands.

[0015] When the temporal dimension is included in spectral clustering, this analysis is extended to account for how spectral features change over time. This makes it possible to identify not only which spectral features predominate, but also when and for how long these features are present. This approach is useful for dynamic signals such as EEG, where frequency patterns are constantly changing and can be associated with different brain states. By analyzing both frequency and time, transitions between specific states or phases of a signal can be detected as it evolves.

[0016] In this context, the spectral time-grouping methodology is based on several steps that allow for the analysis of the evolution of an EEG signal in terms of its spectral characteristics over time. First, the EEG signal is segmented in time, where it is divided into small time "windows" or segments that may or may not overlap between consecutive windows. These windows, typically of short duration, allow for capturing how the signal changes over time. For example, the time windows can have a resolution of 30 milliseconds, which offers a considerable level of detail.

[0017] Once the EEG signal is segmented, spectral estimation is performed within each window. In this step, spectral analysis is applied, generally using Fourier Transforms, multitapers, or specific filters that separate the signal into different frequency bands—in this case, the delta, theta, alpha, and beta bands, which correspond to specific canonical frequency bands of the EEG signal. This analysis generates a representation of the EEG signal in terms of the different frequencies that compose it in each time window.

[0018] Subsequently, the spectral features obtained from each window are grouped into "clusters." This is done using clustering algorithms, such as k-means, which classifies time segments based on the similarity of their spectral characteristics—that is, the power levels present in the different frequency bands. Each cluster is defined by a specific combination of power levels in these bands, allowing for the identification of different EEG signal states. For example, one cluster might be characterized by high power in the delta band and low power in alpha and beta, while another might show the opposite pattern.

[0019] Finally, the temporal dynamics of the clusters are analyzed. Once each time segment within a cluster has been classified, the amount of time the signal spends in each of the defined states and how it transitions between them is analyzed. This step is crucial, as it allows us to study how the signal evolves over time and how the different states (clusters) succeed one another, providing detailed information about the dynamic patterns of the signal, such as the different brain states in the case of EEG analysis.

[0020] For example, in the context of postoperative delirium (POD), this methodology allows for the identification of brain activity patterns associated with an increased risk of POD. By doing so, it improves the prediction and, potentially, the prevention of POD, which would have significant implications for patient care.

[0021] According to one modality, the electroencephalogram (EEG) monitoring and analysis method that provides a prediction of the probability of developing DPO comprises: - Receive EGG data signals from a subject. - Calculate spectral powers of EEG data signals in predefined time windows, based on canonical frequency bands (delta, theta, alpha and beta); - Compare the spectral powers calculated in each time window with centroids of at least two, preferably four, groups or clusters of predefined brain states, where in the case of two groups or clusters we have: general broadband low power state and general broadband high power state, while in the case of four groups or clusters of predefined brain states we have: low broadband power state (LBP), high broadband power state (HBP), prevalent alpha state (PA) or prevalent delta state (PD); - Classify the subject's brain state in each time window into one of these predefined brain states, analyzing the similarity of the subject's spectral powers with the centroids of the predefined clusters, determining the subject's brain state in each time window; - Calculate the fractional occupancy of the general low broadband power state, in the case of two clusters, or of LBP, in the case of four clusters, where the fractional occupancy is the relative amount of time that the subject spends in said state; and - Predict the probability of DPO development based on the fractional occupancy of the general broadband low power state or LBP, as the case may be.

[0022] According to another approach, the EEG monitoring and analysis system that provides a prediction of the probability of developing DPO comprises: - An EEG-based anesthesia monitor, arranged to capture EEG data signals from a subject; - An EEG signal data processing module, configured to: or Receive the captured EGG data signals; or Calculate spectral powers of EEG data signals in predefined time windows, based on canonical frequency bands (delta, theta, alpha and beta); or Compare the spectral powers calculated in each time window with centroids of at least two, preferably four, groups or clusters of predefined brain states, where in the case of two groups or clusters of predefined brain states we have: general broadband low power state and general broadband high power state, while in the case of four groups or clusters of predefined brain states we have: low broadband power state (LBP), high broadband power state (HBP), prevalent alpha state (PA) or prevalent delta state (PD); or Classify the subject's brain state into one of these predefined brain states, analyzing the similarity of the subject's spectral powers with the centroids of the predefined clusters, determining the subject's brain state in each time window; or Calculate the fractional occupancy of the general low broadband power state, in the case of two clusters, or LBP, in the case of four clusters, where the fractional occupancy is the relative amount of time the subject spends in that state; and or Predict the probability of DPO development based on the fractional occupancy of the general broadband low power state or LBP, as the case may be.

[0023] Furthermore, the present invention comprises a computer program product containing a set of instructions that, when executed on a programmable device, enables the implementation of the described method for monitoring and analyzing electroencephalographic (EEG) signals to predict the likelihood of developing postoperative delirium (POD). This computer program product can reside on any programmable device or system capable of receiving and processing EEG signals, including, but not limited to, EEG monitors, data processing servers, or distributed cloud systems, where the program instructions implement the detailed steps of the method.

[0024] The invention also includes a non-transient, computer-readable storage medium containing the aforementioned computer program product. This storage medium can be a physical or virtual device, such as a USB flash drive, hard drive, memory card, or even a remote server, from which the computer program can be accessed for execution on a programmable device. By storing the program instructions, the storage medium allows any compatible device to implement the functions of the described method, ensuring its availability and operability in different usage environments.

[0025] According to one embodiment of the invention, the predefined brain states are obtained from a previously performed temporal spectral clustering process, in which: - the overall broadband low power state corresponds to a cluster with low power across all frequency bands; - the overall broadband high power state corresponds to a cluster with high power across all frequency bands; - the low broadband power (LBP) state corresponds to a cluster with low power in all frequency bands; - the high broadband power (HBP) state corresponds to a cluster with high power across all frequency bands; - the prevalent alpha (PA) state corresponds to a cluster with high power in the alpha band and low power in other bands; and - The prevalent delta state (PD) corresponds to a cluster with high power in the delta band and low power in other bands.

[0026] According to one embodiment of the invention, the spectral temporal grouping process comprises: - segment historical EEG data from multiple subjects into time windows of predefined duration; - perform a time-spectral estimation of the EEG data in each of the time windows, based on the spectral power of canonical frequency bands (delta, theta, alpha and beta), for which spectral analysis techniques can be used, selected from multitaper analysis, Hilbert transform, wavelet transform and / or Fourier transform; and - Apply a clustering algorithm (k-means or other) to the spectral estimation over time to define at least two or preferably four predefined clusters and thereby define each predefined centroid, classifying the spectral estimates into brain states: low and high general broadband power; or LBP, HBP, PA and PD.

[0027] According to one embodiment of the invention, the processing of brain states over time may include calculating the fractional occupancy of each brain state (LBP, HBP, DP, and AP, or others). It may also include calculating the transition probabilities between each pair of states; that is, for each state, the probability of changing to each of the other possible states. Likewise, it may include calculating the average time spent in each state, either in general or conditionally after having been in a specific previous state. Furthermore, it may include calculating the stability of the aforementioned measures (fractional occupancy, duration, transition probabilities) over time; that is, how consistent they are over a period of time estimated with the variance, entropy, or another similar measure.

[0028] By implementing the method of the invention, it was possible to identify at least two, preferably four, clusters of brain states: high delta, high alpha / beta, Low power and high power, analyze the fractional occupancy of each cluster and its relationship with clinical variables such as age, preoperative cognitive function and the occurrence of DPO and, thereby, determine that time spent in the low power state is significantly correlated with the occurrence of DPO, independent of the idea and other known clinical variables.

[0029] By implementing the clustering process of the invention, it was possible to identify at least two, preferably four, clusters of brain states: high delta, high alpha / beta, low power, and high power; to analyze the fractional occupation of each cluster and its relationship with clinical variables such as age, preoperational cognitive function, and the occurrence of DPO; and thereby determine that the time spent in the low power state is significantly correlated with the occurrence of DPO, independent of other known clinical variables.

[0030] According to one embodiment, the method of the invention comprises a preprocessing step to obtain clean EEG data. In this embodiment, the system's processing module is configured to preprocess the EEG data signal, obtaining clean EEG data. The EEG data preprocessing to obtain clean data is predefined based on thresholds for amplitude, total spectral power, variance, and other parameters. Several mechanisms are currently in use in the industry. Essentially, all EEG segments with an amplitude, spectral power, and / or variability higher than a threshold—which depends on the EEG-based anesthesia monitor and the context in which the recording was made—are eliminated.

[0031] According to one modality, the method of the invention, in addition to determining and classifying brain states, allows the identification of the transition between states during the grouping stage.

[0032] According to one embodiment, the method and system of the invention receive EEG data signals during an intraoperative procedure or surgery.

[0033] According to one embodiment, the method and system of the invention comprise that the classification of each brain state of the subject based on the predefined clusters is carried out by comparing the calculated spectral powers of the subject in each time segment with the centroids of the clusters obtained previously, and the brain state of the subject is defined as the cluster closest to the spectral power vector based on the smallest Euclidean distance between the subject's spectral power vector and the centroids of the predefined clusters.

[0034] In one configuration, the processing module is an external unit that communicates data with the anesthesia monitor or is integrated into the anesthesia monitor. Alternatively, the anesthesia monitor, or a communications unit connected to it, is configured to send EEG data to a processing module hosted on a user device or in the cloud.

[0035] According to one modality, the processing module is configured to issue alerts about the probability of DPO occurrence, where such alerts can be displayed on one or more of: an anesthesia monitor screen or indicator, a user device, or an external unit screen.

[0036] The described method and system allow for more accurate prediction of postoperative dysplasia (POD) through the analysis of low-power EEG states during anesthesia. This invention has the potential to improve the prevention and management of POD in patients undergoing cardiac surgery, providing a valuable tool for clinical practice.

[0037] The invention presents a number of advantages, including: It allows predicting DPO over the subject's age, sex, and previous cognitive function; It is non-invasive and fast (delivering results from 2 to 5 minutes after the start of the intraoperative period); It does not require more time from the user; It can be coupled to EEG-based anesthetic depth monitors that are currently in use. BRIEF DESCRIPTION OF THE FIGURES

[0038] As part of the present invention, the following representative figures are presented, which show preferred modalities and, therefore, should not be considered as limiting to the definition of the disclosed technical subject matter. Figure 1: Represents the EEG analysis workflow for predicting postoperative delirium (POD). It includes the EEG signal, spectral estimation over time, cluster segmentation, and the POD predictive model. Fig. 2: Shows the power distribution in four frequency bands (delta, theta, alpha and beta) and the segmentation into four clusters: high delta, high alpha / beta, low power and high power. Fig. 3: Illustrates the characterization of the centroids of the clusters in the example. Fig. 4: Shows the temporal evolution of the EEG signal amplitude and the identification of clusters over time. Fig. 6: Illustrates that the occupancy of the clusters is correlated with important variables in the example. DESCRIPTION OF PREFERRED OPTIONS

[0039] Figure 1 illustrates the complete EEG analysis process used to predict postoperative delirium (POD). The workflow begins with EEG signal acquisition during surgery. A time-spectral estimation is then performed using techniques such as Fourier Transform (FFT), wavelets, or Hilbert Transform. The power in four frequency bands (delta, theta, alpha, and beta) is segmented into four predefined clusters using the k-means algorithm. Finally, a logistic regression predictive model is applied to predict the occurrence of POD.

[0040] Figure 2 shows the power distribution across the four frequency bands (delta, theta, alpha, and beta) and the segmentation into four clusters: high delta, high alpha / beta, low power, and high power. Each cluster represents a distinct brain state based on the spectral power. The figure highlights how the EEG data are grouped into these clusters for further analysis.

[0041] Figure 3 presents graphs of the spectral power of the four clusters as a function of frequency. Each graph shows the power in decibels (dB) for the delta, theta, alpha, and beta frequency bands. The clusters are identified as high delta, high alpha / beta, low power, and high power. These graphs allow visualization of the differences in spectral power between the different clusters.

[0042] Figure 4 shows the temporal evolution of the EEG signal amplitude and the identification of clusters over time. The EEG signal amplitude is represented in microvolts (pV), and its variation over time is observed. Furthermore, the corresponding clusters are identified at different times or time windows, allowing for the analysis of the temporal dynamics of brain states during surgery.

[0043] Time clustering is a technique used to analyze and classify electroencephalography (EEG) data in different brain states during anesthesia. This process is essential for predicting postoperative delirium (POD) in patients undergoing cardiac surgery. The steps and details of this method are described below: a) Acquisition of EEG data: - During surgery, the subjects' EEG signals are recorded using electrodes placed on the scalp. - EEG signals are captured continuously and stored for later analysis. b) Data preprocessing: - EEG signals are preprocessed to remove artifacts and noise. This includes filtering out unwanted frequencies and correcting for motion. - The signals are segmented into short time windows (e.g., 2 seconds) for more detailed analysis. c) Spectral estimation: - A spectral estimation of the EEG signals is performed in each time window using techniques such as the Fourier Transform (FFT) or the Hilbert Transform. - Spectral estimation allows obtaining the power of signals in different frequency bands (delta, theta, alpha and beta). d) Clustering of brain states: - The k-means algorithm is applied to group the time windows into different clusters based on the spectral power of the frequency bands. - The clusters represent different brain states, such as high delta, high alpha / beta, low power, and high power. - Each time window is assigned to one of the clusters, allowing brain states to be identified over time. e) State transition analysis: - The transition between different brain states is analyzed over time. This includes identifying patterns of change between clusters. - The fractional occupancy of each cluster is calculated, that is, the relative time that the subject spends in each brain state during surgery. f) DPO predictive model: - A logistic regression model is used to predict the occurrence of DPO based on the fractional occupancy of the clusters. - The model considers the relationship between time spent in low power states and the probability of developing DPO. - The model is adjusted to control for other known clinical variables, such as age and preoperative cognitive function. Example 1: EEG monitoring and analysis method to predict DPO: EEG signal reception

[0044] A subject undergoing surgery is monitored with an EEG system, which records their brain activity throughout the procedure.

[0045] The EEG signals, which contain information about brain activity over time, are received and processed by the system. Spectral power calculation

[0046] The subject's EEG signals are divided into 3-second time windows with 99% overlap.

[0047] In each window, the spectral powers corresponding to the canonical frequency bands are calculated: delta (0.5-4 Hz), theta (4-8 Hz), alpha (8-12 Hz), and beta (12-30 Hz). For example, in a 3-second window, the signal's energy in each of these bands is analyzed. Comparison of spectral powers with the centroids of predefined groups or clusters

[0048] The spectral powers calculated in each window are compared with the centroids of four groups or clusters of predefined brain states, obtained from previous studies.

[0049] The four predefined clusters are: - LBP (Low Broadband Power): Low power state across all frequency bands. - HBP (High Broadband Power): High power state across all frequency bands. - PA (Prevalent Alpha): High power state in the alpha band. - PD (Prevalent Delta): State with high power in the delta band.

[0050] The system measures the Euclidean distance between the subject's spectral power vector and the cluster centroids for each time window. The closest brain state is identified based on the smallest Euclidean distance. Classification of brain state in each window:

[0051] For each 3-second window, the subject's brain state is classified according to the nearest cluster, for example: - If the spectral powers in that window are closer to the LBP centroid, the subject is classified in the low broadband power (LBP) state. - If in another window the smallest distance is with the centroid of PA, then the subject is classified in the prevalent alpha state (PA).

[0052] This process is repeated for each time window throughout the entire duration of the monitoring. Calculation of the fractional occupancy of the LBP state

[0053] At the end of the analysis of all time windows, the fractional occupancy of the LBP state is calculated, which is the relative amount of time the subject spends in the low bandwidth power (LBP) state compared to the total monitoring time.

[0054] For example, if the subject was in the LBP state for 15 minutes in a 60-minute surgery, the fractional occupancy would be: 15 minutes Fractional occupancy of LBP = — — - = 0.25 or 25% 62 minutes Prediction of the probability of developing DPO

[0055] Based on previous studies, the system knows that a high fractional occupancy in the LBP state (e.g., greater than 20%) is associated with a higher risk of developing Postoperative Delirium (POD).

[0056] In this example, given that the subject spent 25% of the time in the LBP state, the system predicts a high probability of developing DPO, providing this information to physicians so they can make appropriate clinical decisions.

[0057] In this regard, it is important to highlight that the developed system and method do not require the surgery to be completed to perform either the fractional occupancy calculation or the prediction of the probability of postoperative delirium. In fact, it has been certified that with only a few minutes of EEG signal recording (for example, 5 minutes), not the entire surgery, the system and method of the invention provides accurate calculations and predictions regarding the occurrence of postoperative delirium. This is highly useful because the system and method of the invention can consistently provide information throughout the entire surgery, allowing for: 1) greater accuracy in the prediction as the surgery progresses, and 2) valuable information to be provided to healthcare professionals to enable them to make better clinical decisions both during and after surgery. Example 2: Prediction results

[0058] In this example, data from 380 subjects were analyzed, 265 of whom underwent a complete assessment of postoperative delirium (POD) during three postoperative days. Table 1 shows the main characteristics of the subjects separated by the occurrence of POD. Subjects who developed POD were, on average, older, had lower preoperative cognitive function, underwent longer surgeries, and were more likely to be female, which is consistent with previous studies. Table 1 no DPO DPO p-value N [%] 236 [89.1] 29 [10.9] Average (years) [ STD] 68.5 [6.1] 72.4 [6.7] 0.0029 a Preop t-MoCA [STD] 18.1 [2.1] 17.6 [3.0] 0.030 a BMI [STD] 28.1 [5.2] 28.0 [5.0] 0.92 a Surgery Duration (hours) [STD] 5.8 [1.2] 6.4 [1.4] 0.021 a a Mann-Whitney test; b Binomial test Grouping of bandwidth powers into temporary clusters

[0059] To better understand the dynamic changes occurring in a subject's EEG activity during anesthesia, the spectral power of four canonical frequency bands (delta, theta, alpha, and beta) was calculated in overlapping windows. This resulted in band-specific spectral estimates with a temporal resolution of 30 ms. The spectral power of each frequency band was normalized (z-score) to the spectral powers of the 380 study subjects. Following this, each time point of these 4-dimensional data (4 frequency bands) was separated into one of four cluster groups using the k-means clustering algorithm. Each group was defined by the position of its centroid in the 4-dimensional band power space, which in turn describes the contribution of each frequency band to each cluster. Figure 3A shows the contribution of each frequency band to the centroid of each cluster using a bar chart.Cluster 1 (blue) has a strong positive contribution from the delta band and a negative contribution from the alpha and beta bands. Conversely, Cluster 2 (orange) exhibits strong alpha-beta components and a negative delta component. Clusters 3 and 4 have similar components for each frequency band but with opposite signs; Cluster 3 (green) has negative contributions from all frequency bands, while Cluster 4 (red) has high and positive contributions from all bands.

[0060] Two of the cluster centroids obtained showed marked differences between frequency bands. Cluster 1 showed a positive contribution from the delta band and negative components for the alpha and delta frequency bands, hence its designation as a high-delta cluster. The centroid of Cluster 2 showed the opposite, with positive contributions from alpha and beta while exhibiting a negative delta component (high-alpha cluster). The two remaining clusters showed more undifferentiated behavior with respect to particular frequency bands: Cluster 3 showed low spectral power in all frequency bands (low-power cluster; low broadband power), and Cluster 4 showed high components for all frequency bands (high-power cluster). The spectral differences of the four clusters are illustrated in Fig. 3B, which shows representative power spectral density (PSD) graphs for each cluster.

[0061] To analyze the amount of time spent in each of these four EEG states, the fractional occupancy of each cluster was calculated for each subject studied. Table 2 shows that, on average, the percentage of time spent in each cluster is similar across all clusters, between 20% and 30% (25% expected by chance). However, all cluster occupancies showed very high variability (standard deviation greater than 30%). Because of this, The median occupation among the subjects was calculated (Table 2). The large difference between some average and median occupations revealed that the distribution of the % of time spent in each cluster was strongly asymmetric, with many subjects spending minimal time in some clusters. Table 2: Occupancy Occupancy STD Min Max [%] Average [%] Mean [%] [%] High Delta Cluster 30.23 20.08 30.28 0 99.50 High Alpha Cluster 29.90 13.58 33.04 0 100 Beta Low Power Cluster 18.77 0.55 31.52 0 100 High Power Cluster 21.10 1.65 33.03 0 100 Low-power cluster occupancy predicts DPO

[0062] To assess the significance of this novel approach to analyzing intraoperative EEG data—namely, the temporal clustering of band powers—an initial analysis was performed by calculating Spearman correlations between three key variables (age, preoperative cognitive function measured with t-MoCA, and DPO) and seven intraoperative EEG characteristics: mean cluster occupancy and spectral powers. Figure 5 shows that the amount of time spent in the High Power cluster, as well as alpha power and total spectral power, are inversely correlated with subject age, meaning that older subjects spend less time in the High Power cluster and have lower alpha and total power overall. Older subjects also show a significant inverse correlation with time spent in the Low Power cluster (all p < 0.001).Preoperative cognitive function showed the same significant correlations but in opposite directions: an inverse relationship with LP occupation and direct relationships with HP occupation, alpha power, and total power (all p < 0.005). The onset of postoperative delirium showed only two significant correlations: a direct correlation with the amount of time subjects spent in the Low Power cluster (rho = 0.147, p = 0.0167), and an inverse relationship with alpha activity power (rho = 0.141, p = 0.0217).

[0063] Figure 5 shows a heat map displaying Spearman's rank correlation coefficients for each pair of important variables (left) and EEG features (bottom). * p < 0.05, ** p < 0.01, *** p < 0.001, uncorrected. LP stands for Low Power and HP for High Power.

[0064] Alpha power has been previously described as a predictor of postoperative dysplasia (POD); however, it has been suggested that alpha is a marker of aging, a known risk factor for POD, rather than a completely independent predictor of POD (REF). To investigate whether the time a subject's EEG spends in the long limb cluster could be an independent predictor of POD, a logistic regression was first performed to predict POD using age and long limb occupation as predictors. This model was found to have a significant predictive value for overall POD (AUC = 0.711, p = 0.00122). Importantly, both age (OR = 1.092, 95% CI = [1.026–1.162], p = 0.0054) and occupation The LP cluster (OR = 1.011, 95% CI = [1.0001 - 1.022], p = 0.048) showed significant contributions to predicting DPO. They also showed low collinearity (VIF = 1.013). This indicates that LP occupation could be an independent predictor of DPO.

[0065] To further assess whether LP occupation is an independent predictor of DPO, a second logistic regression model was constructed that included two other variables known to be predictors of DPO: preoperative cognitive function and sex. This 4-predictor model for DPO showed a better ROC curve AUC than the previous one (AUC = 0.743). Each included variable significantly contributed to predicting postoperative depression (POD) (age OR = 1.070 [1.006–1.139]; occupation LP OR = 1.0137 [1.0016–1.0259]; sex (M) OR = 0.252 [0.102–0.622]; preoperative cognitive function OR = 0.805 [0.681–0.952]). This model showed an R² of 0.150 (Me Fadden), an accuracy of 0.909, a specificity of 0.996, and a sensitivity of 0.207. Furthermore, all variables exhibited low collinearity (all VIF values ​​< 1.14). This demonstrates that occupation in the LP cluster is a strong predictor of POD, independent of other clinical variables known to be associated with POD.

Claims

CLAIMS 1. A method for monitoring and analyzing electroencephalograms (EEG) to predict the probability of occurrence of postoperative delirium (POD), CHARACTERIZED in that it comprises the following steps: - receive EEG data signals from a subject; - calculate spectral powers of EEG data signals in predefined time windows, based on canonical frequency bands (delta, theta, alpha and beta); - compare the spectral powers calculated in each time window with centroids of at least two groups or clusters of predefined brain states, where the two clusters of predefined brain states comprise: general broadband low power state and general broadband high power state; - classify the subject's brain state in each time window into one of these predefined brain states, analyzing the similarity of the subject's spectral powers with the centroids of the predefined clusters, determining the subject's brain state in each time window; - calculate the fractional occupancy of the general broadband low power state; and - predict the probability of DPO development based on the fractional occupancy of the general broadband low power state.

2. The method of claim 1, CHARACTERIZED in that the predefined brain states comprise four clusters of brain states, wherein said clusters correspond to: - Low Bandwidth Power (LBP) state; - High Bandwidth Power (UBP) state; - prevalent alpha state (PA); and - Prevalent delta state (PD).

3. The method of claim 2, CHARACTERIZED in that the four predefined brain states are obtained from a previously performed spectral temporal grouping process, in which: - the overall broadband low power state corresponds to a cluster with low power across all frequency bands; - the overall broadband high power state corresponds to a cluster with high power across all frequency bands; - the prevalent alpha (PA) state corresponds to a cluster with high power in the alpha band and low power in other bands; and - The prevalent delta state (PD) corresponds to a cluster with high power in the delta band and low power in other bands 4. The method of claim 2, CHARACTERIZED in that the spectral temporal grouping process comprises: - segment EEG data into time windows of predefined duration; - perform a time-spectral estimation of the EEG data in each of the time windows, based on the spectral power of canonical frequency bands (delta, theta, alpha and beta), using spectral analysis techniques selected from multitaper analysis, Hilbert transform, wavelet transform and / or Fourier transform; and - Apply a clustering algorithm (k-means or other) to the spectral estimation over time to define the four predefined clusters and thereby define the predefined centroids, classifying the spectral estimates into brain states.

5. The method of claim 4, CHARACTERIZED in that the classification of the subject's brain states is performed by comparing the spectral powers calculated in each time segment with the centroids of the clusters obtained previously, and the subject's brain state is defined as the cluster closest to the spectral power vector based on the smallest Euclidean distance between the subject's spectral power vector and the centroids of the predefined clusters.

6. The method of claim 1, CHARACTERIZED in that the calculation of the fractional occupancy includes the identification of the transition between brain states over time, and the temporal dynamics of the subject are analyzed, evaluating the time spent in each brain state.

7. The method of claim 6, CHARACTERIZED in that the probability of transition between brain states is also calculated, wherein each probability reflects the possibility that the subject will change from one brain state to another in a given period of time.

8. The method of claim 6, CHARACTERIZED in that the average time spent in each brain state before the transition to another state is calculated.

9. The method of claim 1, CHARACTERIZED in that the process of calculating the spectral powers includes a preprocessing stage of the EEG data to obtain clean signals, eliminating segments with artifacts or noise based on predefined thresholds of amplitude, total spectral power and variability.

10. The method of claim 1, CHARACTERIZED in that the probability of DPO development can be predicted based on the fractional occupancy of the LBP state with a monitoring time of less than 10 minutes.

11. An electroencephalogram (EEG) monitoring and analysis system for predicting the probability of occurrence of postoperative delirium (POD), CHARACTERIZED in that it comprises: - an EEG monitor configured to capture EEG data signals from a subject; - an EEG signal data processing module configured to: or receive EEG data signals captured during a procedure; or calculate spectral powers of EEG data signals in predefined time windows, based on canonical frequency bands (delta, theta, alpha and beta); or compare the spectral powers calculated in each time window with centroids of at least two groups or clusters of predefined brain states, wherein the two clusters of predefined brain states comprise: general broadband low power state and general broadband high power state; or classify the subject's brain state into one of these predefined brain states, analyzing the similarity of the subject's spectral powers with the centroids of the predefined clusters; or calculate the fractional occupancy of the general low broadband power state; and or predict the probability of DPO development based on the fractional occupancy of the general broadband low power state.

12. The system of claim 11, CHARACTERIZED in that the predefined brain states comprise four clusters of brain states, wherein said clusters correspond to: - Low Bandwidth Power (LBP) state; - High Bandwidth Power (HBP) state; - prevalent alpha state (PA); and - Prevalent delta state (PD).

13. The system of claim 12, CHARACTERIZED in that the four predefined brain states are obtained from a prior analysis based on the spectral temporal grouping of historical EEG data.

14. The system of claim 12, CHARACTERIZED in that the centroids of the predefined clusters are determined by applying the k-means algorithm to historical EEG data.

15. The system of claim 11, CHARACTERIZED in that the EEG monitor is configured to preprocess the EEG data signal, eliminating artifacts and noise based on predefined thresholds of amplitude and variability of spectral power.

16. The system of claim 11, CHARACTERIZED in that the processing module is configured to calculate the transition probabilities between the different classified brain states.

17. The system of claim 11, CHARACTERIZED in that the processing module is configured to issue alerts on the probability of occurrence of DPO, based on the results obtained from the EEG analysis.

18. The system of claim 11, CHARACTERIZED in that the processing module is configured to display the results on an EEG monitor screen or on a connected device.

19. The system of claim 11, CHARACTERIZED in that the processing module is integrated within the EEG monitor or as an external unit that communicates with the monitor.

20. The system of claim 11, CHARACTERIZED in that the prediction of DPO can be performed during the surgical procedure with only 5 minutes of captured EEG signals.

21. The system of claim 12, CHARACTERIZED in that the fractional occupancy of the different brain states is calculated in real time during the procedure.

22. A computer program product, CHARACTERIZED in that it comprises instructions that, when executed by a processor, cause a computer to execute the steps of the electroencephalogram (EEG) monitoring and analysis method to predict the probability of occurrence of postoperative delirium (POD) according to any one of claims 1 to 11.

23. A non-transient computer-readable medium, CHARACTERIZED in that it stores the computer program product of claim 22.

Images