Method and system for estimating an intracranial pressure value of a patient, and method for training a model

Abstract

The present invention describes a method for estimating an intracranial pressure value of a patient and a method for training a model to be implemented in said estimation method. The present invention specifically comprises a method for estimating the intracranial pressure values of a patient in real time, which incorporates a trained model, implemented in a processor, which interprets datasets of intracranial pressure morphology signals and estimates the intracranial pressure value of the patient from the dataset of the intracranial pressure signal. The present invention pertains to the fields of medicine, neuroscience, electrical engineering and signal processing.

Description

Descriptive Report of Invention Patent METHOD AND SYSTEM FOR ESTIMATING THE INTRACRANIAL PRESSURE VALUE OF A PATIENT, AND METHOD FOR MODEL TRAINING Field of Invention

[0001] The present invention describes a method for estimating a patient's intracranial pressure values ​​in real time, which incorporates a trained model implemented in a processor that interprets intracranial pressure morphology signal datasets and estimates the patient's intracranial pressure value from the intracranial pressure signal dataset. The present invention lies in the fields of medicine, neuroscience, electrical engineering, and signal processing. Background of the Invention

[0002] In more than two centuries of advances in research on intracranial dynamics, it was understood that the balance between intracranial volumes is fundamental for brain function, since the elevation of intracranial pressure (ICP) leads to the reduction of blood perfusion of brain tissue and interferes with the regulatory mechanisms of cerebral blood flow.

[0003] The intracranial volume / pressure ratio, which represents intracranial compliance (ICC), is a marker of brain health. In addition to monitoring oxygenation, metabolism, and electrical activity of brain tissue, ICC monitoring is essential in neurocritical care. However, the most widely used standard for ICP monitoring remains the insertion of a catheter into the ventricle or brain parenchyma, which has several limitations, including the need for specialized technicians to perform the procedure, the high cost of the procedure, and, most importantly, the technical challenges of evaluating small ventricles and the risks involved with invasive brain procedures.

[0004] In the search for less harmful medical practice, the development of modern, non-invasive techniques, such as Transcranial Doppler (TCD), optic nerve sheath ultrasound, and pupillometry, has allowed physicians to investigate intracranial hypertension (IH) in a variety of clinical settings.

[0005] More recently, a mechanical sensor that detects micrometric cranial deformations, developed by brain4care, has been added to this arsenal of non-invasive methods for investigating intracranial compliance. This system is capable of recording ICP pulse morphology, providing automated waveform analysis and translating ICP variations into numerical parameters through indicators and classifiers that dynamically relate the ICP waveform to numerical scales at the bedside in real time.

[0006] Advances in artificial intelligence, especially machine learning (ML), have been exponential in medicine, with many publications revealing the potential of clustering blood pressure or waveforms to estimate ICP or even identify complications during ICP monitoring. However, noninvasive bedside ICP estimation has not yet become a reality.

[0007] To date, several approaches have been explored for noninvasively estimating ICP values ​​in humans. These approaches employ techniques using transcranial Doppler (TCD) waveforms, arterial blood pressure (ABP), acoustic signals, and / or ultrasonographic optic nerve sheath diameter (ONSD) as model inputs. However, these techniques demonstrate significant variability in the metrics used to determine the accuracy of these noninvasive ICP estimation methods, typically measured as mean absolute difference (MAD), mean absolute error (MAE), and a 95% confidence interval (CI).

[0008] In the search for the state of the art in scientific and patent literature, the following documents were found that deal with the topic:

[0009] WO20231 18556A1 discloses, more generally, a system that uses a trained model to predict an indicative intracranial pressure parameter based on medical data provided from a database, a measuring unit, an intensive care unit, or from a receiver or transceiver for receiving medical data. However, WO2023118556A1 does not mention the use of intracranial pressure morphology curves by the model to generate an intracranial pressure value.

[0010] Document US20130041271 A1 discloses a method and device for monitoring intracranial hemodynamic parameters such as ICP, cerebral blood volume, cerebral blood flow, and cerebral perfusion pressure. Waveforms are extracted from impedance plethysmography signals and used to continuously estimate intracranial hemodynamic parameters. Furthermore, the waveforms are used to construct a signal analysis model to determine an average ICP value over a period of time, for example, a single cardiac cycle. However, document US20130041271 A1 does not mention the use of intracranial pressure morphology curves by the model to generate an intracranial pressure value.

[0011] WO2023049529A1 discloses a method for determining a patient's ICP, which combines ultrasound data obtained from measuring acoustic signals from the patient's brain to determine a measurement of the brain's cerebral blood flow velocity (CBFV), and data from measurements of the brain's arterial blood pressure (ABP), to generate, from both measurements, an input to a model trained to output an ICP measurement and to provide input to the model to obtain a measurement of the patient's ICP. However, WO2023049529A1 does not mention the use of intracranial pressure morphology curves by the model to generate an intracranial pressure value.

[0012] Document CN1 15804582A discloses a non-invasive continuous ICP monitoring system using microwaves. The system has a A processor configured to transmit a first microwave signal and receive a second returned microwave signal, performing filtering. It also has a computer used to analyze the frequency bands of the first microwave signal and the second received and reflected microwave signal to obtain related parameters, as well as a model implemented on the computer to train the related parameters and the first intracranial pressure value to obtain a monitoring model. However, document CN1 15804582A does not disclose the use of intracranial pressure morphology curves by the model to generate an intracranial pressure value.

[0013] In this sense, there is growing recognition in the field that ICP management should be more individualized and differentiated. Several compelling arguments support this perspective, given that, in real clinical practice, medical professionals face a significant challenge in accurately determining true arterial blood pressure across all areas of brain tissue, as measurements taken in the brachial artery tend to be significantly higher than intracranial arterial blood pressure.

[0014] Furthermore, the oversimplified concept of cerebral perfusion pressure (CPP) as merely the difference between arterial blood pressure and intracranial pressure in the context of acute brain injury is increasingly being discouraged. This emphasizes the need to treat ICP as a dynamic phenomenon that can vary widely between individuals and even within the same patient over time.

[0015] Indeed, the accuracy of ICP measurements is of paramount importance in clinical practice. According to the standards established by the Association for Advancement of Medical Instrumentation and the Brain Trauma Foundation guidelines, ICP monitoring devices should have a continuous output in the 0-100 mmHg range, with an accuracy of ±2 mmHg in the 0-20 mmHg range, and a maximum prediction error of 10% for ICP >20 mmHg. It is worth noting that Previously proposed noninvasive ICP data-driven models can exhibit varying degrees of mean error, ranging from ±5 mmHg to approximately ±20 mmHg, often accompanied by a high standard deviation. This variability reflects the importance of considering trends and anticipating ICP peaks, rather than relying solely on passive observation until a predefined threshold is exceeded for these prior art estimation models.

[0016] In this context, the state of the art lacks a non-invasive method capable of providing real-time ICP estimates and changes that allow proactive patient management and early intervention, aligning with the broader goal of improving patient outcomes. Summary of the Invention

[0017] Thus, the present invention solves the problems of the prior art by obtaining an ICP value that is generated through a specifically trained artificial intelligence model, which receives the signal from the intracranial pressure morphology curve and estimates the ICP value related to the respective curve. This model is implemented to operate both in real time, that is, collecting real-time data from the intracranial pressure morphology curve, and in cases where the curve is previously collected and stored in memory for later computation / estimation of the ICP value.

[0018] In a first object, the present invention presents a method for estimating the intracranial pressure value of a patient comprising the steps of: a. receiving, by a processor, at least one window of data from a signal of intracranial pressure morphology of a patient; b. processing the data window by means of a model implemented in the processor; and c. generation of an intracranial pressure value, from the data window, by the model, said model being previously trained to receive one or more data windows of the intracranial pressure morphology signal and generate an intracranial pressure value related to the received data window.

[0019] In a second object, the present invention presents a system for obtaining a patient's intracranial pressure value comprising at least one processor that reads at least one data window from at least one intracranial pressure morphology signal from a patient, in which the processor is provided with a model that processes and generates an intracranial pressure value, from the data window read, said model being previously trained to receive one or more data windows from the intracranial pressure morphology signal and generate an intracranial pressure value related to the processed data window received.

[0020] In a third object, the present invention presents a model training method for estimating the intracranial pressure value of a patient, which is performed by an algorithm comprising at least one intracranial pressure morphology signal pre-processing step among: a. extraction, by a processor, of one or more data windows of a plurality of distinct intracranial pressure morphology signals; and b. comparing said data windows with absolute intracranial pressure values ​​of the respective distinct intracranial pressure morphology signals.

[0021] In a fourth object, the present invention presents a computer program comprising instructions executable by the computer, in which the instructions, when executed by the computer, reproduce the method for estimating the intracranial pressure value of a patient.

[0022] Additionally, in another object of the invention, a system is proposed which has a device for measuring intracranial pressure non-invasively, where the system is adapted to implement any of the methods defined above.

[0023] These and other objects of the invention will be immediately appreciated by those skilled in the art and will be described in detail below. Brief Description of the Figures

[0024] The following figures are presented:

[0025] Figure 1 shows a diagram illustrating the input of an intracranial pressure data window into the model and the output of a numerical intracranial pressure value.

[0026] Figure 2 shows a flowchart illustrating the patient and data selection step for the method using Random Forest classification and regression.

[0027] Figure 3 shows a flowchart that illustrates the separation of patients used in testing and training the Random Forest classification and regression method.

[0028] Figure 4 shows a flowchart illustrating the patient and data selection step for the method using gradient histogram boosting with a temporal convolutional network.

[0029] Figure 5 shows a flowchart illustrating the separation of patients used in testing and training the method using gradient histogram boosting with a temporal convolutional network.

[0030] Figure 6 shows a flowchart illustrating the separation of patients used in testing and training the method using the regression tool and gradient histogram boosting method.

[0031] Figure 7 shows a graph with the MAE and MSE results for elCP synchronized with changes in invasively measured intracranial pressure (ICP).

[0032] Figure 8 shows a Bland-Altiman analysis that illustrates the difference average between ICP and elCP.

[0033] Figure 9 shows the Spearman correlation between ICP and elCP.

[0034] Figure 10 shows the error percentages of the calculated differences between the obtained elCP means and real intracranial pressure (ICP). Detailed Description of the Invention

[0035] The following descriptions are presented by way of example and are not limitative of the scope of the invention and will make the subject matter of this patent application more clearly understood.

[0036] Intracranial pressure (ICP) is the relationship between the volume of the intracranial space and its components: cerebrospinal fluid (CSF), blood, and brain parenchyma. During each cardiac cycle, blood flows to the brain, creating the ICP waveform. The intracranial pressure waveform has a typically well-known morphology, with three distinct components related to the patient's physiological parameters: the P1, P2, and P3 peaks. The first P1 peak is the percussion wave and reflects the systolic blood pressure transmitted from the choroid plexus to the cerebral ventricle. The second P2 peak is related to brain tissue compliance, which is variable and increases in amplitude as intracranial compliance decreases; when it exceeds the P1 wave level, it suggests a significant decrease in intracranial compliance. Finally, the P3 wave is associated with aortic valve closure.

[0037] Based on this, some parameters can be extracted from the PIC morphology and become important for its analysis, for example, the ratio between the P2 and P1 peaks, described here as the P2 / P1 ratio, and the time elapsed until the peak (highest amplitude value) of the wave in the given cycle / period, referred to here as TTP (time to peak), among others.

[0038] In this sense, the invention proposed here provides the obtaining, estimation or calculation of a patient's intracranial pressure value from the ICP morphology curve, where an artificial intelligence model is used. implemented in a processor, which is capable of performing, at least, the reading of curves of the morphology signal of a patient's intracranial pressure and performing processing, based on the curve data, to return a numerical result that represents the value of the patient's intracranial pressure.

[0039] Thus, a first object of the invention is a method for estimating the intracranial pressure value of a patient comprising the steps of: a. receiving, by a processor, at least one window of data from an intracranial pressure morphology signal of a patient; b. processing the data window by means of a model implemented in the processor; and c. generating an intracranial pressure value, from the data window, by the model, said model being previously trained to receive one or more data windows of the intracranial pressure morphology signal and generate an intracranial pressure value related to the received data window.

[0040] For the purposes of the present invention, the data window is formed by a quantity of information representative of an intracranial pressure morphology signal, such that said data window is composed of all or part of the patient's ICP signal curve. Further, for purposes of example, said data window represents a period of the wave or the equivalent of an ICP pulse composed of the three components related to the patient's physiological parameters, the P1, P2 and P3 peaks.

[0041] In one embodiment, the data window contains sampled data from a single ICP pulse curve and returns an intracranial pressure value related to the respective curve, such as a vector of discretized numerical data representing the patient's ICP pulse curve. In another embodiment, the processor samples the received signal to extract a set of points from the ICP curve before inputting the information into the model. In one embodiment, the data window The data window received by the processor comprises the filtered and / or "cleaned" "raw signal" of the collected ICP curve. In another embodiment, the data window comprises at least one image of the ICP curve collected from the patient, where said image is sampled by the processor using an image recognition algorithm, where said sampling converts the image into one or more data vectors with numerical values ​​for the composition of the data window. In another embodiment, the data window comprises at least one static or dynamic image of the ICP curve, for example, of a pulse of the patient's ICP, so that such image is input to the model, which is trained to recognize and process static or dynamic images.

[0042] In one embodiment, the processor is capable of receiving real-time ICP morphology data at the bedside, measured using a non-invasive ICP measurement device.

[0043] In another embodiment, the processor is capable of acquiring the PIC morphology signal from a database or file containing the PIC morphology to be interpreted and processed. This signal can be collected automatically or manually, with the latter being the data manually input into the processor.

[0044] In one embodiment, the method proposed here comprises a preliminary preprocessing step of the ICP morphology / curve signal to compose the data window. Alternatively, or in addition, this preprocessing is performed on the data from the data window. Thus, an algorithm is implemented in the processor that performs filtering steps for the intracranial pressure morphology signals, selecting standardized data windows from said signals, and classifying the intracranial pressure morphology signals based on a quality index. Through these steps, it is possible to "clean" the signal to be used in the model, removing unwanted signal artifacts, records with missing or incomplete data, noisy and erroneous values, etc.

[0045] The quality index for classifying PIC morphology signals, For example, this involves a signal quality index (SQI) of 0-1 based on the spectral analysis of one-minute data with reduced sampling and detrended bias. This, combined with the other steps, defines a threshold that allows us to classify whether the collected signal is sufficient to be absorbed by the machine learning model, preventing poor or interfering signals from being used by the model—during its use or training.

[0046] Thus, such steps are performed to adapt the collected ICP curves to a standard that can be absorbed and interpreted by the processor and machine learning model during the method steps, in which the adaptation of the curves includes data cleaning steps (removal of incomplete records, missing data, noisy and erroneous values, etc.), integration of the “cleaned” data and extraction of parameters related to pulse morphology, such as the P2 / P1 ratio, TTP, areas under the curve and results of its dimensionality reduction, so that the numerical value of intracranial pressure estimated at the end of processing is free of inconsistent data.

[0047] Thus, after this preprocessing step, the model implemented in the processor is able to process and mathematically interpret the windowed ICP data. The trained model is able to return a number indicating the patient's intracranial pressure value based on the collected ICP morphology curve. In one embodiment, the patient's intracranial pressure value generated by the model is given in at least one unit among: mmHg, cmH2O, mmH2O, cmHg, Pascals, Bar, Psi, among others that can be used to represent the respective physical quantity. This resulting value is assigned to or related to the collected pulse; that is, through the ICP morphology curve, a numerical value is obtained that represents the ICP in the desired unit.

[0048] In this sense, the model is also capable of operating in real time, receiving ICP curves and, in real time, returning the absolute value of the patient's ICP. Therefore, if there are morphological changes in the curve, the value numeric returned by the model also changes relatively.

[0049] In one embodiment, the machine learning model is a classifier and regressor based on decision trees or used to form decision trees. In one embodiment, a Random Forest model is used. In another embodiment, the model is a Histogram Gradient Boosting model. Of course, other models can be used within the scope of the invention, so that training can be performed using ICP morphology data to return the numerical value of the relative intracranial pressure or the resulting value from the collected curve.

[0050] In this way, the analysis guided by the method comprising the model of the present invention allows the monitoring of patients' intracranial pressure in real time, providing information derived from wearable technology data and electronic health records that inform evidence-based clinical data, promoting preventive medicine.

[0051] The invention also includes a training method for estimating a patient's intracranial pressure value, which is performed by an algorithm that performs pre-processing steps on the ICP morphology signal.

[0052] In one embodiment, the method comprises a preliminary step of training the model by means of a regression tool using a gradient histogram boosting method.

[0053] The model training step includes a step of extracting, by the processor, one or more parameters from a plurality of distinct intracranial pressure morphology signals, such as, for example, areas, dimensionality reductions, P2 / P1 ratio and TTP, for training the model, as well as comparing said data windows with absolute intracranial pressure values ​​of the respective distinct intracranial pressure morphology signals.

[0054] For purposes of understanding the present invention, said distinct intracranial pressure morphology signals comprise morphology signals of the ICP of different patients, with different pathological conditions, with different skull conditions, with different diagnoses, etc.

[0055] In one embodiment, said extracted parameter comprises at least one of: segmented area under the ICP curve, where one unit represents 20% of the waveform length; waveform amplitude; total area under the ICP curve; correlation coefficient; wave slope derivative; dimensionality reduction elements (ISOMAP); intracranial compliance scale; P2 / P1, being the ratio between the P1 and P2 peaks of the ICP waveform; TTP, being the time to peak - the highest amplitude value of the wave; or a combination thereof.

[0056] In one embodiment, such extracted data and / or parameters are refined for dimensionality reduction using the regression tool using the gradient histogram boosting method, based on distinct absolute intracranial pressure values. In one embodiment, such absolute intracranial pressure values ​​come from a database.

[0057] Furthermore, in one embodiment, to optimize model performance, data and / or parameters are selected using a grid search approach to obtain the mean absolute error (MAE) and mean square error (MSE).

[0058] Thus, based on this method, the model is then trained to identify the ICP morphology signal pattern and extract numbers corresponding to the ICP morphology, generating a numerical value of the patient's ICP in real time.

[0059] Therefore, it is an objective of the present invention that the said method for estimating the intracranial pressure value of a patient is used as a guide for the accurate interpretation of ICP curves, allowing the ICP curve measured non-invasively to be converted into a numerical value for subsequent accurate and efficient medical diagnosis.

[0060] In this context, without limitation to the scope of the invention, only signs of PIC morphology are used for model training, not requiring the use of any other type of curve for this purpose.

[0061] Furthermore, an object of the invention is a system for estimating the intracranial pressure value of a patient comprising a processor that reads at least one window of data originating from at least one intracranial pressure morphology signal of a patient, in which the processor is provided with a model that processes and generates an intracranial pressure value, from the window of data read, said model being previously trained to receive one or more windows of data from the intracranial pressure morphology signal and generate an intracranial pressure value related to the window of processed data received.

[0062] The system comprises a connection interface that receives PIC morphology data and transmits it to the processor. Within the scope of the invention, said connection interface is a physical or electronic medium that allows the reception and forwarding / transmission of data to the processor. For example, the connection interface is a physical data bus that connects an electronic device to the processor. In another example, the connection interface is an instruction or routine implemented in software that receives data and forwards it to the processor. In another alternative example, the connection interface is a bus connectable to a peripheral through which a user can input PIC curve data. Such examples, of course, should not limit the scope of the invention.

[0063] In one embodiment, the connection interface is communicative with a non-invasive intracranial pressure measuring device, which collects real-time data on the patient's ICP morphology and transmits it to the processor.

[0064] In another embodiment, the connection interface communicates with a database to acquire intracranial pressure morphology signals, which are transmitted to the processor.

[0065] In one embodiment, the system comprises a value display of intracranial pressure communicating with the processor, which displays the intracranial pressure value obtained through processing. This allows for quick and convenient visualization of the numerical intracranial pressure value obtained.

[0066] Furthermore, an object of the invention is a system and method for estimating the intracranial pressure value in mmHg of a patient, comprising generating an intracranial pressure value, in mmHg, from a window of data originating from an intracranial pressure morphology signal of the patient, the intracranial pressure value in mmHg being generated by means of processing through a model implemented in a processor, in which the model is previously trained to receive one or more data windows of the intracranial pressure morphology signal and generate an intracranial pressure value, in mmHg, related to the received data window.

[0067] In this way, the system and methods of the present invention can be used as a guide for interpreting intracranial pressure morphology curves, allowing the non-invasively measured ICP curve to be converted into a numerical value in mmHg for subsequent accurate and efficient medical use.

[0068] The examples shown here are intended only to exemplify one of the numerous ways of carrying out the invention, without, however, limiting its scope. Example 1 - Method for estimating intracranial pressure in mmHg

[0069] In this example, a system and method were developed to estimate a numerical value of intracranial pressure, in mmHg, from a patient's ICP curve. In this sense, the present method uses only ICP morphology curves / waves to estimate an intracranial pressure value in mmHg in real time. The patient's ICP curves are measured in real time at the bedside using a measuring device. non-invasive PIC.

[0070] Figure 1 shows a simplified schematic diagram representing a pulse from a patient's intracranial pressure curve being input to the machine learning model. In this example, a decision tree model trained with a Histogram Gradient Boosting (HGB) technique was used. With the trained model, when inputting the pulse from the ICP curve, the model returns a numerical ICP value in mmHg, which is relative to the curve read by the model.

[0071] In this example, the data window received by the previously trained model is a vector of amplitudes related to the collected PIC curve itself, where the amplitudes are obtained by temporal sampling of the collected signal. The number of points, i.e., the vector length, is previously defined to be sufficient and compatible with the model parameters. In this example, without any limitation to the scope of the invention, a set of approximately 260 points from the PIC curve was used, each with amplitude values ​​in micrometers assigned to each point, these values ​​being interpreted by the model. The sensor shown in WO2013041973, the applicant's own, was used to acquire this signal.

[0072] The sensor collects the "raw signal" of the ICP curve in real time at the bedside. The processor then performs a filtering process to "clean" the signal and extract the values ​​from each signal point to compose the data window. This filtering process includes selecting standardized data windows and classifying the ICP signals according to a quality index, to adapt the ICP curves to a standard that can be captured and interpreted by the model.

[0073] With the filtered signal, the data window is input to the model, which interprets the sampled signal from the real-time ICP curve and returns a numerical value of the patient's intracranial pressure in mmHg. In this sense, the ability of this method to provide real-time ICP estimates and changes becomes increasingly valuable for management. proactive patient management and early intervention, aligning with the broader goal of improving patient outcomes, as it obtains the ICP morphology curve non-invasively and also generates an absolute ICP value related to the collected curve.

[0074] In this way, the present system and method provides simplicity of handling and immediate acquisition of information, in addition to the absence of any type of energy emission and suitability for locations with fewer financial resources.

[0075] Furthermore, the model's use of the measured ICP curve itself represents a true physiological signal of ICP dynamics on a beat-by-beat basis, which is an advantage compared to prior-art noninvasive ICP estimation models, which use only secondary parameters related to ICP dynamics, such as cerebral blood velocity and mean arterial blood pressure. In this context, the present method is more advantageous because it is based on a single physiological signal, avoiding potential discrepancies arising from a multiparameter architecture.

[0076] Thus, the present method enables the inclusion of artificial intelligence mechanisms capable of providing the value of the patient's ICP estimated at the point of care, as well as continuous non-invasive monitoring so that ICP becomes a more universally accessible vital sign. Example 2 - Machine Learning Framework Using Random Forest Classification and Regression Models

[0077] The method developed in this example aims to use artificial intelligence to read a data window from an intracranial pressure curve to generate an estimated numerical value of a patient's intracranial pressure based on the interpretation of the processed data window.

[0078] In the present method, the patient's ICP curves are obtained by means of of a non-invasive ICP measurement device attached around the patient's head, which measures ICP and provides curves of the patient's ICP morphology in real time at the bedside.

[0079] In this way, the method can be used as a guide for the accurate interpretation of ICP curves, allowing the non-invasively measured ICP curve to be converted into a numerical value in mmHg for subsequent accurate and efficient medical use.

[0080] For this purpose, two machine learning (ML) tools were used, a combination of classification and regression models using Random Forest.

[0081] Thus, all data underwent initial preprocessing using a Random Forest classifier. This tool allowed for the identification and segregation of data based on ICP variation ranges without relying on ICP information from invasive sensors. Subsequently, the regressor was used to train within these distinct ICP ranges, resulting in unique models for different ICP bands. The patients included in this study were randomly divided into different groups to facilitate model development and evaluation.

[0082] Specifically, 80% of the patients were allocated to the training group, while the remaining 20% ​​were reserved for testing model performance. Furthermore, 10 patients were reserved for final replication and model validation. Furthermore, parameters derived from morphological segments, such as the P2 / P1 ratio, time-to-peak, area under the curve, and results from dimensionality reduction techniques (e.g., ISOMAP), were used in the development of this model.

[0083] To estimate ICP values ​​using this model, a subgroup of patients, comprising 20% ​​of those involved in model development, was used as the test group.

[0084] The model creation process involved cross-validation, meaning that the patient data used for testing and training were were mixed at each stage of the cross-validation process, comprising five stages in total. This methodology ensured robust validation by avoiding potential biases that could arise from having the same patient data in the training and testing datasets.

[0085] Finally, the model was saved in a file structure suitable for machine learning, allowing for easy replication and application to larger datasets in the future. Example 3 - Validation of the model that uses classification and regression by Random Forest to estimate non-invasive intracranial pressure from its morphology curves

[0086] Data analysis began with a series of 73 patients, collectively representing 64,113 pulses. These data were segmented into 10-second windows, generating 5,594 distinct windows.

[0087] Nine patients were excluded from the initial set due to low-quality pooled morphology data, resulting in a revised set of 64 patients and 4,422 10-second windows. Subsequently, an additional 14 patients who had undergone decompressive craniectomy procedures were removed from the analysis.

[0088] Thus, data from 50 patients were preprocessed. Ten patients were randomly selected from the sample to ensure they were not used as data sources for the machine learning (ML) training and testing phases, but exclusively for model replication and validation. Figure 2 shows a flowchart illustrating the aforementioned data selection step.

[0089] Consequently, 40 patients were used to develop the model. Of these, 80% were allocated to the training process, while the remaining 20% ​​were used to test the generated model. The model developed for ICP estimation therefore uses a synergistic approach that combines classification and regression models. using Random Forest.

[0090] The estimate of the intracranial pressure value produced by the model was tested in 20% of the patients in the group involved in the development of the model, and also in the 10 patients previously excluded from the sample.

[0091] The model was developed using cross-validation, ensuring that the patient data used in the training and testing groups were intermingled. This means that no single patient's data is present in either the training or testing groups at each stage of the cross-validation process. This ensures the robustness of the evaluation, avoiding potential biases arising from the overrepresentation of any individual patient in the training and testing datasets, as illustrated in Figure 3.

[0092] The model was statistically evaluated using the Mean Absolute Error (MAE) and Mean Squared Error (MSE) metrics, whose equations (1) and (2), respectively, are described below:

[0093] Means and medians were used to evaluate the outcome in patients excluded from the sample, while the area under the receiver operator characteristic curve (ROC) was calculated to assess the reliability of the model for ICP above 20 mmHg, indicating intracranial hypertension (IH). The model and statistical calculations were performed using Python 3.7 in a Jupyter Notebook environment hosted on a high-performance server.

[0094] Table 1 below shows the distribution of patients used in this test, the majority of whom were women with an average age of 43 years, presenting traumatic brain injury in 77% of cases.

[0095] Table 1 - Characteristics of patients participating in the test.

[0096] The model was statistically evaluated using the Mean Absolute Error (MAE) and Mean Squared Error (MSE) metrics. As illustrated in Table 2 below, the MAE yielded favorable results. Both the training and test populations demonstrated close margins, with a margin of 3.7, a value within the clinically beneficial range. In both tests, there is a remarkable correspondence between the values ​​derived from the test and the training populations, indicating the absence of overfitting. This supports the robustness of the model and its readiness for clinical deployment, assuming that the described margins of error are acceptable in the specific clinical context.

[0097] Table 2 - MAE and MSE results for the test and training groups.

[0098] The 10 patients who were excluded before machine learning training were processed using the generated model. Table 3 below presents the mean values ​​along with their corresponding standard deviations for the invasive ICP measurement sensor (ICPm) and the ICP value estimated by the model developed in this example (elCP). The last column presents the discrepancy between these numbers, outlining the average model error per patient, measured in millimeters of mercury (mmHg).

[0099] Table 3 - Average values ​​for the invasive ICP measurement sensor (ICPm) and for the developed model (elCP).

[0100] Table 4 below shows the median values ​​of these parameters, along with their standard deviations. It also highlights discrepancies between them, representing the error identified when analyzing each patient's medians, in mmHg.

[0101] Table 4 - Median values ​​for the invasive ICP measurement sensor (ICPm) and for the developed model (elCP).

[0102] The analyses show that 70% of the patients submitted to the generated models presented a difference between the invasive and non-invasive sensor of less than 2 mmHg, demonstrating the clinical applicability of the developed model. Example 4 - Machine Learning Framework Using Gradient Histogram Boosting with a Temporal Convolutional Network

[0103] The method developed in this example aims to use artificial intelligence to read a data window from an intracranial pressure curve to generate an estimated numerical value of a patient's intracranial pressure based on the interpretation of the processed data window.

[0104] In the present method, the patient's ICP curves are obtained by means of a non-invasive ICP measuring device attached around the patient's head, which measures the ICP and provides the patient's ICP morphology curves in real time at the bedside.

[0105] In this way, the method can be used as a guide for the accurate interpretation of ICP curves, allowing the non-invasively measured ICP curve to be converted into a numerical value in mmHg for subsequent accurate and efficient medical use.

[0106] For this purpose, two machine learning (ML) tools were used, a combination of gradient histogram boosting with a temporal convolutional network.

[0107] Thus, all data were subjected to initial preprocessing via Gradient Histogram Boosting. This tool allowed the Data identification and segregation based on ICP variation ranges without relying on ICP information from invasive sensors. Subsequently, a temporal convolutional network was used to train within these distinct ICP ranges, resulting in unique models for different ICP bands. The patients included in this study were randomly divided into different groups to facilitate model development and evaluation.

[0108] Specifically, 80% of the patients were allocated to the training group, while the remaining 20% ​​were reserved for testing model performance. Furthermore, 10% of the sample was reserved for final replication and model validation. Furthermore, parameters derived from morphological segments, such as P2 / P1 ratio, time-to-peak, area under the curve, and results from dimensionality reduction techniques (e.g., ISOMAP) were used in the development of this model.

[0109] To estimate ICP values ​​using this model, a subgroup of patients, comprising 20% ​​of those involved in model development, was used as the test group.

[0110] The model creation process involved cross-validation, meaning that the patient data used for testing and training were mixed at each stage of the cross-validation process, comprising five stages in total. This methodology ensured robust validation by avoiding potential biases that could arise from having the same patient data in the training and testing datasets.

[0111] Finally, the model was saved in a file structure suitable for machine learning, allowing for easy replication and application to larger datasets in the future. Example 4 - Model validation using gradient histogram boosting with a temporal convolutional network for non-invasive intracranial pressure estimation from its morphology curves

[0112] Data analysis began with a series of 67 patients, collectively representing 64.1 13 pulses. These data were segmented into 10-second windows, generating 5,594 distinct windows.

[0113] Nine patients were excluded from the initial set due to low-quality pooled morphology data, resulting in a revised set of 58 patients and 4,422 10-second windows. Subsequently, an additional 14 patients who had undergone decompressive craniectomy procedures were removed from the analysis.

[0114] Thus, data from 44 patients were preprocessed. Four patients were randomly selected from the sample to ensure they were not used as data sources for the machine learning (ML) training and testing phases, but exclusively for model replication and validation. Figure 4 shows a flowchart illustrating the aforementioned data selection step.

[0115] Consequently, 40 patients were used to develop the model. Of these, 80% were allocated to the training process, while the remaining 20% ​​were used to test the generated model. The model developed for ICP estimation uses a synergistic approach that combines Histogram Gradient Boosting (HGB) with a Temporal Convolutional Network (TCN).

[0116] The estimate of the intracranial pressure value produced by the model was tested in 20% of the patients in the group involved in the development of the model, and also in the 4 patients previously excluded from the sample.

[0117] The model was developed using cross-validation, ensuring that the patient data used in the training and testing groups were intermingled. This means that no single patient's data is present in either the training or testing groups at each stage of the cross-validation process. This ensures the robustness of the evaluation, avoiding potential biases arising from the overrepresentation of any individual patient in the training and testing datasets, as illustrated in Figure 5.

[0118] The model was statistically evaluated using the Mean Absolute Error (MAE) and Mean Squared Error (MSE) metrics, whose equations (3) and (4), respectively, are described below:

[0119] Means and medians were used to evaluate the outcome in patients excluded from the sample, while the area under the receiver operator characteristic curve (AUROC) was calculated to assess the reliability of the model for ICP above 20 mmHg, indicating intracranial hypertension (IH). The model and statistical calculations were performed using Python 3.7 in a Jupyter Notebook environment hosted on a high-performance server.

[0120] Table 5 below shows the distribution of patients used in this test, the majority of whom were women with an average age of 43 years, presenting traumatic brain injury in 77% of cases.

[0121] Table 5 - Characteristics of patients participating in the test.

[0122] The model was statistically evaluated using the Mean Absolute Error (MAE) and Mean Squared Error (MSE) metrics. As illustrated in Table 6 below, the MAE yielded favorable results. Both the training and test populations demonstrated close margins, with a margin of 2.60, a value within the clinically beneficial range. In both tests, there is a remarkable correspondence between the values ​​derived from the test and the training populations, indicating the absence of overfitting. This supports the robustness of the model and its readiness for clinical deployment, assuming that the described margins of error are acceptable in the specific clinical context.

[0123] Table 6 - MAE and MSE results for the test and training groups.

[0124] The four patients excluded before machine learning training were processed using the generated model. Table 7 below presents the mean values ​​along with their corresponding standard deviations for the invasive ICP measurement sensor (ICPm) and the ICP value estimated by the model developed in this example (elCP). The last column presents the discrepancy between these numbers, outlining the average model error per patient, measured in millimeters of mercury (mmHg).

[0125] Table 7 - Average values ​​for the invasive ICP measurement sensor (ICPm) and for the developed model (elCP).

[0126] Table 8 below shows the median values ​​of these parameters, along with their standard deviations. It also highlights discrepancies between them, representing the error identified when analyzing each patient's medians, in mmHg.

[0127] Table 8 - Median values ​​for the invasive ICP measurement sensor (ICPm) and for the developed model (elCP).

[0128] The analyses show that 70% of the patients submitted to the generated models presented a difference between the invasive and non-invasive sensor of less than 2 mmHg, demonstrating the clinical applicability of the developed model. Example 5 - Machine Learning Framework using Regression Tool and Gradient Histogram Boosting Method

[0129] The method developed in this example aims to use artificial intelligence to read a data window from an intracranial pressure curve to generate an estimated numerical value of a patient's intracranial pressure based on the interpretation of the processed data window.

[0130] In the present method, the patient's ICP curves are obtained by means of a non-invasive ICP measuring device attached around the patient's head, which measures the ICP and provides the patient's ICP morphology curves in real time at the bedside.

[0131] In this way, the method can be used as a guide for the accurate interpretation of ICP curves, allowing the non-invasively measured ICP curve to be converted into a numerical value in mmHg for later use. precise and efficient doctor.

[0132] To this end, the model used to estimate intracranial pressure employed a gradient histogram boosting regressor, a learning technique that sequentially trains decision trees to correct errors made by previous trees. This approach effectively captures complex relationships and nonlinear patterns in the data by constructing trees based on histograms of parameter values.

[0133] The process involved extracting various parameters from the waveform collected by the noninvasive ICP measurement device. This process resulted in the extraction of approximately 1,180 parameters, including: segmented area under the ICP curve, where one unit represents 20% of the waveform length; waveform amplitude; total area under the ICP curve; correlation coefficient; wave slope derivative; ISOMAP dimensionality reduction elements; intracranial compliance scale; P2 / P1, the ratio of the P1 and P2 peaks of the ICP waveform; TTP, the time to peak; or a combination thereof. Among the various parameters, 15 parameters were selected according to their degree of importance in the calculation. Spearman's correlation was used to select the key parameters (mentioned in Example 6).

[0134] To optimize model performance, parameters were selected through a grid search approach, with a primary focus on achieving the optimal mean absolute error (MAE) and mean squared error (MSE) for the test set, while avoiding overfitting and underfitting through 10-fold cross-validation and parameter analysis.

[0135] To address imbalances in the dataset, such as the limited representation of intracranial hypertension data (ICP >20 mmHg), a class-weighting scheme was implemented that assigns weights inversely proportional to the frequency of each class. This method ensures that cases with lower representation, such as hypertension, intracranial, receive greater importance during the training process, thus improving the model's performance in these critical cases.

[0136] Thus, by calculating the frequency of normal and elevated ICP instances and mapping the frequencies to their corresponding weights, the model's ability to learn from imbalanced data was improved, which ultimately led to more accurate predictions across all ICP ranges.

[0137] Thus, parameters derived from the morphological segments of the intracranial pressure waveform collected by the non-invasive ICP measuring device, such as the P2 / P1 ratio, time to peak (TTP), and area under the curve, and results from dimensionality reduction techniques, such as ISOMAP, were used in the development of the model for this example. Furthermore, it is important to emphasize that no parameters derived from any other physiological signal (e.g., ABP, cerebral blood velocity) were included in the model.

[0138] The entire workflow adopted a robust 10-fold cross-validation strategy to ensure model reliability. The strategy involved separating data from the same patient into training and testing in each fold, with 80% allocated to model training and 20% to testing. Additionally, data from 20 patients were reserved for replication and final validation, with their data kept separate and not used in the model development phase, as illustrated in Figure 5.

[0139] After model training, the saved models for each fold were used to estimate PIC values ​​in the validation dataset (n=20). Furthermore, the model produced 10 calibrations, and the resulting estimated PIC value was determined as the median of these calibration models. Example 6 - Model validation using regression tool and gradient histogram boosting method for estimation of non-invasive intracranial pressure from its morphology curves

[0140] As shown in Figure 6, a total of 136 patients were considered for evaluation. Ten patients were excluded from the initial set due to the poor quality of the derived pooled intracranial pressure morphology curves or the short duration of monitoring, resulting in a revised set of 126 patients. Subsequently, 14 additional patients who had undergone decompressive craniectomy procedures were removed from the analysis, resulting in 112 patients whose data were segmented into 10-second windows, resulting in a total of 11,604 windows, equivalent to approximately 150,000 pulses.

[0141] Furthermore, data from 92 patients were preprocessed (9,591 10-second windows) and used to generate and train the machine learning model. Overall, the entire population sample (n=112) exhibited only 5.36% of the data with values ​​above 20 mmHg.

[0142] The model was statistically evaluated using the Mean Absolute Error (MAE) and Mean Squared Error (MSE) metrics. As illustrated in Table 9 below, the MAE and MSE yielded favorable results. The MAE for the validation dataset was 3.00 mmHg and the MSE was 13.56 mmHg (N=20, 2013 10-second segments). These values ​​(elCP for this example), including those in Table 9, were synchronized with changes in invasively measured intracranial pressure (ICP), as illustrated in Figure 7.

[0143] Table 9 - MAE and MSE results for the test and training groups.

[0144] Table 10 below presents the elCP values ​​for the validation dataset individually, which displays the mean ICP and elCP values ​​along with the observed difference for each patient.

[0145] Table 10 - Average values ​​for the invasive ICP measurement sensor (ICPm) and for the developed model (elCP).

[0146] Bland-Altiman analysis for this population revealed a mean difference between ICP and elCP of -0.21 mmHg and a standard deviation of ±3.68 mmHg, as illustrated in Figure 8. From these values, a 95% confidence limit for predicting ICP of less than ±7.5 mmHg was obtained. Furthermore, there was a moderate relationship between ICP and elCP (Spearman's correlation coefficient r=0.43, p<0.05), as illustrated in Figure 9.

[0147] When calculating the elCP averages for each 10-second window, differences in relation to the actual intracranial pressure (ICP) of less than 2 were obtained mmHg in 36.61% of the validation windows, as illustrated in Figure 10. In 71.90% of the windows, the difference was less than 4 mmHg, and in 91.48%, the error was less than 6 mmHg. It is important to note that less than 9% of the windows presented a difference between ICP and elCP greater than 6 mmHg. Furthermore, an analysis of the mean difference between the 10-second windows with ICP <20 mmHg and >20 mmHg was performed. It was observed that, for the first group, the MAE was 2.97 mmHg, while for the second group it was 6.41 mmHg.

[0148] Thus, in almost 40% of the validation sample (10-second segments), the proposed model was able to operate within the acceptable limits recommended by the Association for the Advancement of Medical Instrumentation, which stipulates a maximum error of ±2 mmHg for ICP measurement below 20 mmHg.

[0149] Those skilled in the art will appreciate the knowledge presented herein and will be able to reproduce the invention in the presented embodiments and in other variants and alternatives, covered by the scope of the following claims.

Claims

Claims 1. Method for estimating the intracranial pressure value of a patient, characterized by comprising the steps of: a. receiving, by a processor, at least one window of data from an intracranial pressure morphology signal of a patient; b. processing the data window by means of a machine learning model implemented in the processor; and c. generating an intracranial pressure value, from the data window, by the model, said model being previously trained to receive one or more data windows of the intracranial pressure morphology signal and generate an intracranial pressure value related to the received data window.

2. Method for estimating the intracranial pressure value of a patient, according to claim 1, characterized by comprising a preliminary step of pre-processing the data window, by means of the processor, comprising the steps of: filtering the intracranial pressure morphology signals; selecting standardized data windows of said signals; and classifying the intracranial pressure morphology signals based on a quality index.

3. Method for estimating the intracranial pressure value of a patient, according to claim 1, characterized by comprising a preliminary step of training the model by means of a regression tool using a gradient histogram boosting method.

4. Method for estimating the intracranial pressure value of a patient, according to any one of claims 1 to 3, characterized in that the model training step comprises at least one of the steps of: a. extracting, by a processor, one or more data windows from a plurality of distinct intracranial pressure morphology signals; and b. comparison of said data windows with absolute intracranial pressure values ​​of the respective distinct intracranial pressure morphology signals.

5. Method for estimating a patient's intracranial pressure value, according to claim 1, characterized in that the processor receives at least one data window, in which the data window comprises data from the intracranial pressure curve, and returns an intracranial pressure value.

6. System for obtaining a patient's intracranial pressure value, characterized by comprising a. an interface that receives data from a patient's intracranial pressure morphology signal; and b. a processor communicating with the interface that reads the patient's intracranial pressure morphology signal data, wherein the processor is provided with a machine learning model that processes and generates an intracranial pressure value from the read data, said model being previously trained to receive one or more data windows from the intracranial pressure morphology signal and generate an intracranial pressure value related to the received data window.

7. System for obtaining a patient's intracranial pressure value, according to claim 6, characterized in that the connection interface is communicative with: a non-invasive intracranial pressure measuring device, said device acquiring the intracranial pressure morphology signal; and / or a database for acquiring intracranial pressure morphology signals.

8. System for obtaining a patient's intracranial pressure value, according to claim 6, characterized by comprising at least one display communicating with the processor, which displays the intracranial pressure value generated by the processor.

9. Model training method for estimating the intracranial pressure value of a patient characterized by being performed by an algorithm comprising at least one intracranial pressure morphology signal preprocessing step among: a. filtering the distinct intracranial pressure morphology signals; b. classifying the intracranial pressure morphology signals based on a quality index; c. extracting, by a processor, one or more data windows from a plurality of distinct intracranial pressure morphology signals; and d. comparing said data windows with absolute intracranial pressure values ​​of the respective distinct intracranial pressure morphology signals.

10. Training method for estimating the intracranial pressure value of a patient, according to claim 9, characterized in that the algorithm performs the steps of the method by means of a regression tool using a gradient histogram boosting method.

11. Computer program comprising instructions executable by the computer, characterized in that the instructions, when executed by the computer, reproduce the method for estimating the intracranial pressure value of a patient of claim 1.

Images