Method and system for estimating an intracranial pressure value in patients, method for training a model and monitoring panel
The method processes complete intracranial pressure waveforms using machine learning models to estimate intracranial pressure and related parameters directly from the raw waveform, addressing the limitations of existing technologies by enhancing accuracy and real-time monitoring.
Patent Information
- Application Number
- PCT/BR2026/050223
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-09
- Filing Date
- 2026-05-11
- Publication Date
- 2026-11-12
AI Technical Summary
Current non-invasive methods for estimating intracranial pressure and related physiological parameters require pre-processing and filtering of physiological signals, which discards relevant information and fails to provide accurate, real-time measurements.
A method and system that utilize machine learning models to process complete intracranial pressure waveforms directly, without filtering or pre-processing, to estimate absolute intracranial pressure and related parameters from the raw waveform.
Enables accurate, real-time estimation of intracranial pressure and multiple physiological parameters by preserving the complete morphological information of the waveform, improving the reliability and precision of intracranial pressure monitoring.
Smart Images

Figure BR2026050223_12112026_PF_FP_ABST
Abstract
Description
METHOD AND SYSTEM FOR ESTIMATING A PATIENT'S INTRACRANIAL PRESSURE VALUE, MODEL TRAINING METHOD AND MONITORING PANEL Field of Invention
[0001] The present invention describes a method and a system for estimating a patient's intracranial pressure value, as well as a method for training the developed model and a monitoring panel, more specifically in the fields of medicine, neuroscience, electrical engineering, and signal processing. Background of the Invention
[0002] Currently, measuring and monitoring intracranial pressure (ICP) is a fundamental practice in patient care, especially in patients with neurological disorders, since the brain is confined in a closed space where small variations in volume can generate dangerous increases in pressure. Intracranial compliance (ICC), which represents the skull's ability to accommodate changes in volume, is an essential parameter for neurophysiological assessment.
[0003] In this sense, situations such as trauma, hemorrhages, tumors, or edema can increase intracranial pressure (ICP), compromising cerebral blood circulation and reducing the supply of oxygen and nutrients to nerve cells, which can lead to serious and even irreversible damage.
[0004] In this context, ICP monitoring allows professionals to intervene more quickly and accurately in patient treatment, helping the medical team adjust treatments and prevent complications. In addition to aiding in a more precise diagnosis, this measurement is also essential for monitoring the patient's progress and response to therapies, being an important resource that contributes to reducing sequelae and increasing the chances of recovery in critical situations.
[0005] Thus, the search for more precise, safe, and effective methods has driven the development of new technologies. More recently, brain4care developed a mechanical sensor capable of detecting micrometric cranial deformations, which has been incorporated into a set of non-invasive methods for assessing intracranial compliance. This system can record the morphology of the intracranial pressure (ICP) pulse, offering an automated waveform analysis and converting ICP variations into numerical parameters. This is done through indicators and classifiers that dynamically and in real time relate the ICP waveform to numerical scales at the bedside.
[0006] In this sense, the cranial pulse waveform is considered a composite physiological signature, generated by the dynamic interaction of multiple bodily "systems," such as the cerebral arterial pulse drive, intracranial system compliance—which determines how volume changes translate into pressure changes—cerebral vascular autoregulation—which modulates the relationship between systemic arterial pressure and cerebral blood flow—and the resulting cerebral perfusion. Each of these "systems" leaves characteristic marks on the cranial pulse morphology curve.
[0007] Furthermore, decades of clinical literature on intracranial dynamics document that alterations in pulse morphology, such as inversion of the relationship between the percutaneous and tidal components, displacement of the time to peak, pulse widening, and loss of dicrotic components, accompany clinically relevant changes in various neurological conditions.
[0008] In this context, advances in artificial intelligence, particularly in the field of machine learning (ML), have grown exponentially in medicine, including in the field of intracranial pressure measurement. Several studies demonstrate the potential of using blood pressure clusters or waveforms to estimate intracranial pressure (ICP) or even to identify complications during ICP monitoring.
[0009] Furthermore, some existing artificial intelligence models have the ability to interpret certain data windows of pre-processed physiological curves correlated with ICP, and deliver intracranial pressure values and correlated patient parameters, and / or waveforms in graphs displayed on monitors.
[0010] Document US20110201961 A1, for example, describes a model for morphological analysis of ICP pulses, in which a processor acquires the cranial morphology curve and, optionally, the electrocardiogram (ECG) signal, performs pre-processing to refine the curve(s) before input into the model, so that the model can interpret the incoming data. The model automatically detects characteristic P1, P2, and P3 pulses of each pulse and generates numerical metrics linked to the processed waveform, such as amplitudes, time intervals between subpeaks, slopes, curvatures, decay constants, etc. These metrics are recorded over time and compared with stored reference profiles, allowing monitoring of changes in the patient's intracranial and cerebrovascular conditions.
[0011] Document WO2024261499A1 also describes a model for measuring ICP and other brain parameters, which uses iNIRS technology with light to detect a pulsed waveform of blood flow, which is related to ICP. This signal enters a regression model, and generates intracranial parameters as output, such as an intracranial pressure value, a brain reactivity index, and an estimate of the patient's health status.
[0012] In another case, document EP2505137A1 describes a model for estimating intracranial parameters, based on the measurement of cranial electrical bioimpedance over time. The model's input data includes electrical impedance signals and can also receive synchronized auxiliary physiological signals, such as ECG and respiration. A computational module processes the signals and extracts features such as pulsatile variations, amplitudes, phases, and time delays associated with cardiac and respiratory cycles, applying temporal analysis and correlation algorithms to isolate components related to intracranial dynamics. With this, the model generates quantitative estimates of intracranial parameters, including ICP itself and derived metrics, such as indicators of cerebral compliance or perfusion pressure, displayed in numerical and / or graphical format on a monitor.
[0013] In general, these documents reveal the use of artificial intelligence models capable of interpreting pre-processed data windows related to intracranial pressure, and generating a set of values such as absolute ICP and related intracranial parameters.
[0014] However, to date, non-invasive bedside ICP estimation has not yet become a reality, nor has the estimation of other important and essential brain and / or health parameters for assessing cranial compliance. Existing technologies are not capable of directly measuring intracranial pressure and other physiological parameters from a single patient physiological signal. Furthermore, current technologies are unable to display a diversity of physiological parameters from the processing of a single patient physiological signal without the need for filtering / refinement and pre-processing of the input signal in the model.
[0015] Thus, based on the literature reviewed, no documents were found that anticipated or suggested the teachings of the present invention, so the solution proposed here has novelty and inventive activity compared to the state of the art. Summary of the Invention
[0016] Thus, the present invention solves the problems of the prior art by providing a method for estimating an intracranial pressure value from the processing of complete intracranial pressure waveforms, using one or more machine learning models. In this way, the present method provides the generation of an absolute intracranial pressure value, in real time, directly from a physiological signal of intracranial pressure morphology, without the need for pre-processing and / or filtering of the input physiological signal.
[0017] In this sense, the present invention allows intracranial pressure to be measured from the interpretation of a complete waveform of intracranial pressure morphology, that is, the present model is capable of receiving and interpreting the “raw” intracranial pressure morphology curve, without filtering and / or processing prior to the curve entering the model.
[0018] Thus, using the complete waveform of intracranial pressure morphology as input in the present model, instead of using a reduced vector of pre-computed numerical descriptors (state of the art), allows for the complete preservation of the signal's morphological information, including subpeaks, relative temporal phases, amplitudes, rise and fall slopes, durations, and relationships between pulse components.
[0019] In this context, any prior reduction, refinement, and / or filtering of the complete waveform to scalar parameters, as occurs in the state of the art, would discard information that could be relevant for the simultaneous estimation of the absolute value of intracranial pressure and multiple physiological output parameters of the model, especially when the relationship between ICP morphology (input) and physiological parameters (output) involves non-linear or non-obvious combinations of waveform aspects of the morphology.
[0020] In this way, the present invention solves the problems of the prior art by maintaining the waveform of intracranial pressure morphology in its pure form, that is, "the way it is collected," and using it as input to the present model, without any prior reduction, filtering, and / or refinement. This also allows the model to learn these relationships between morphology and output parameters directly from the annotated data, without depending on prior hypotheses about which aspects of the morphology are relevant for each estimated physiological parameter.
[0021] Thus, a primary object of the invention is a method for estimating a patient's intracranial pressure value comprising the steps of: a. interpretation, using a previously trained model, of at least one data window from at least one physiological signal of a patient; and b. return of an absolute value of intracranial pressure, based on the interpretation of the data window, by the model; - in which the patient's physiological signal comprises a complete waveform of intracranial pressure morphology.
[0022] In a second object, the present invention provides a system for estimating a patient's intracranial pressure value, comprising a data display interface communicating with a processor, which reads a data window of at least one physiological signal from a patient, wherein: The processor is equipped with a machine learning model that processes and retrieves, for the data display interface, an absolute value of intracranial pressure from the data window; and The patient's physiological signal comprises a complete waveform of intracranial pressure morphology.
[0023] In a third object, the present invention provides a model training method for estimating a patient's intracranial pressure value, comprising the following steps: a. extraction, by a processor, of one or more data windows from a plurality of distinct intracranial pressure morphology waveforms; and b. comparison of said data windows with absolute values of intracranial pressure; - in which the plurality of waveforms with distinct intracranial pressure morphologies has varied physiological and / or clinical conditions attributed to various sets of physiological parameters and intracranial pressure values.
[0024] In a fifth object, the present invention provides a monitoring panel for intracranial pressure and physiological parameters of a patient, comprising a display showing an absolute value of intracranial pressure and at least one physiological parameter generated by a previously trained model, wherein the model retrieves the absolute value of intracranial pressure and the physiological parameters from a data window received by the model from at least one physiological signal from a patient.
[0025] It is also another object of the invention a method for estimating an intracranial pressure value and a set of physiological parameter values of a patient comprising the steps of: a. interpretation, using a previously trained model, of at least one data window from at least one physiological signal of a patient; and b. return of an absolute value of intracranial pressure and a set of physiological parameters, based on the model's interpretation of the data window; - in which the patient's physiological signal comprises a complete waveform of intracranial pressure morphology.
[0026] 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
[0027] The following figures are presented:
[0028] Figure 1 shows an embodiment of the present invention, illustrating the input of a physiological signal from a patient into the model and the output of an absolute value of intracranial pressure.
[0029] Figure 2 shows an embodiment of the present invention, illustrating the input of a physiological signal from a patient into a model and the output of a set of physiological parameters, including: absolute intracranial pressure value, P2 / P1 ratio, time to peak (TTP), mean arterial pressure (MAP), cerebral perfusion pressure (CPP), cerebrovascular pressure autoregulation, optimum mean arterial pressure (optABP), and pulse amplitude index (PAx).
[0030] Figure 3 shows an embodiment of the present invention, illustrating the input of a complete / integral waveform of a patient's intracranial pressure morphology into a model, and the generation of a set of physiological parameters, including: absolute intracranial pressure value, P2 / P1 ratio, time to peak (TTP), mean arterial pressure (MAP), cerebral perfusion pressure (CPP), cerebrovascular pressure autoregulation, optimum mean arterial pressure (optABP), and pulse amplitude index (PAx).
[0031] Figure 4 shows an embodiment of the monitoring panel (2) of the present invention, where the waveform of the intracranial pressure morphology is collected by a non-invasive intracranial pressure measuring device (3), and is used as input data in a model implemented in a processor (1), which returns and displays on the panel (2) a set of physiological parameters.
[0032] Figure 5 shows a flowchart used for training the machine learning model.
[0033] Figure 6 shows a graph illustrating the correlation between intracranial pressure (ICP / ICP) measured invasively and intracranial pressure (elCP) measured using this model.
[0034] Figure 7 shows a graph illustrating the correlation between measured arterial pressure (ABP) and measured mean arterial pressure (eABP) using this model.
[0035] Figure 8 shows a graph illustrating the correlation between measured cerebral perfusion pressure and measured cerebral perfusion pressure (eCPP) using this model.
[0036] Figure 9 shows a graph illustrating the correlation between the measured pulse amplitude index (PAx) and the pulse amplitude index (PAx) measured using the present model. Detailed Description of the Invention
[0037] The descriptions that follow are presented by way of example and are not limiting to the scope of the invention, and will provide a clearer understanding of the subject matter of this patent application.
[0038] Intracranial pressure (ICP) is the relationship between the volume of the intracranial space and its components: cerebrospinal fluid (CSF), blood, and brain parenchyma. In each cardiac cycle, blood flow to the brain originates the ICP wave. The intracranial pressure wave has a typically known morphology, possessing three distinct components related to the patient's physiological parameters: peaks P1, P2, and P3. The first peak, P1, is the percussion wave and reflects the systolic blood pressure transmitted from the choroid plexus to the cerebral ventricle. The second peak, P2, is related to brain tissue compliance, which is variable and shows an increase in amplitude as intracranial compliance decreases; when it exceeds the level of wave P1, it suggests a significant drop in intracranial compliance. Finally, wave P3 is associated with aortic valve closure.
[0039] In this context, the cranial pulse waveform can be understood as a composite physiological signature, produced by the dynamic interaction between various "systems" of the organism. These include the pulsatile impulse of the cerebral arteries, intracranial compliance – responsible for defining how volume changes cause pressure alterations – cerebrovascular autoregulation – which adjusts the relationship between systemic arterial pressure and cerebral blood flow – and the resulting cerebral perfusion. Each of these "systems" imprints its own characteristics on the cranial pulse morphology curve.
[0040] Furthermore, the relationship between the pulse subpeaks (percutaneous component, tidal component, dicrotic component to systolic component, tidal component, and diastolic component), their relative amplitudes, their temporal phases, and the overall shape of the pulse reflect the combined state of these systems at a given moment.
[0041] Based on this, the method of the present invention starts from the validated premise that the physiological information composed of the cranial pulse waveform is sufficiently rich that models trained on annotated data are able to estimate an absolute value of intracranial pressure, in addition to a set of correlated physiological parameters.
[0042] Thus, it is possible to use the “raw” waveform of intracranial pressure morphology, that is, the waveform as detected, to estimate an absolute value of intracranial pressure and values of different physiological parameters of a patient, through one or more models implemented in a processor, with such model(s) being capable of interpreting the curves of intracranial pressure morphology and generating numerical values indicative of physiological parameters.
[0043] The invention proposed here involves obtaining, estimating, or calculating an intracranial pressure value and a set of physiological parameters of a patient, based on the reading of intracranial pressure morphology signal curves. This is achieved using one or more artificial intelligence models implemented in a processor, which is capable of at least reading intracranial pressure morphology signal curves and processing the read curve data to return a numerical result representing the absolute value of intracranial pressure and, additionally, a set of physiological parameters of a patient.
[0044] In this sense, the present invention provides, among other advantages, the generation of an absolute value of intracranial pressure and, additionally, a set of values of physiological parameters of a patient, through only one type of morphological curve collected and / or received (intracranial pressure morphology signal curve).
[0045] In a first object, the present invention provides a method for estimating a patient's intracranial pressure value comprising the following steps: a. interpretation, using a previously trained model, of at least one data window from at least one physiological signal of a patient; and b. return of an absolute value of intracranial pressure, based on the interpretation of the data window, by the model; - in which the patient's physiological signal comprises a complete waveform of intracranial pressure morphology.
[0046] For the purposes of the present invention, the complete waveform of intracranial pressure morphology is defined as the "raw" waveform of intracranial pressure morphology, that is, the integral waveform, as collected, without any signal pre-processing or filtering steps.
[0047] In one embodiment, the data window is formed by a representative amount of information relating to a signal of intracranial pressure morphology, such that said data window is composed of all or part of the “raw” curve of the patient's ICP signal. Furthermore, in another embodiment, said data window represents a wave period or the equivalent of an ICP pulse composed of the three components related to the patient's physiological parameters: peaks P1, P2, and P3.
[0048] In one embodiment, the physiological signal comprises selected and / or aggregated pulses of the "raw" waveform of intracranial pressure morphology, in order to obtain a complete representation of the waveform of a patient's intracranial pressure morphology over a given time interval.
[0049] In one embodiment, the physiological signal additionally comprises an electrocardiogram signal from a patient. In this embodiment, the previously trained model is able to interpret the waveform of the intracranial pressure morphology simultaneously and / or together with the patient's electrocardiogram signal, and return an absolute value of intracranial pressure and / or related physiological parameters. In one embodiment, the electrocardiogram signal acts to contribute to greater accuracy and / or measurement performance of the absolute value of ICP and / or related physiological parameters.
[0050] In one embodiment, the previously trained model is able to interpret any other physiological signals and body and / or brain waveforms that may complement the intracranial pressure morphology waveform and contribute positively to the generation and measurement of a more accurate absolute value of intracranial pressure and related physiological parameters.
[0051] In one embodiment, the data window interpreted by the model comprises the filtered and / or "cleaned" "raw" physiological signal of the complete ICP morphology waveform. In another embodiment, the data window comprises at least one static or dynamic image of the ICP morphology waveform, for example, of a patient's ICP pulse, such that this image is input into the model, which is then trained to recognize and process static or dynamic images.
[0052] In one embodiment, the previously trained model is implemented in an algorithm, and the algorithm, in turn, is executed in a processor (1). In one embodiment, the processor (1) receives at least one data window from at least one ICP morphology signal from a patient, processes the data window using the previously trained model, and retrieves an absolute intracranial pressure value and at least one related physiological parameter from the data window.
[0053] In one embodiment, the data window containing the patient's intracranial pressure morphology waveform is acquired in real time at the bedside by means of a non-invasive intracranial pressure measuring device (3), so that the processor (1) is able to receive such data measured by the non-invasive intracranial pressure measuring device (3).
[0054] In another embodiment, the data window containing the patient's intracranial pressure morphology waveform is acquired in real time at the bedside by means of an invasive or minimally invasive intracranial pressure measuring device (3), so that the processor (1) is able to receive such data measured by said device.
[0055] In another embodiment, the processor (1) is able to acquire the data window of the patient's intracranial pressure morphology waveform from a database or file containing the patient's intracranial pressure morphology to be interpreted and processed. This signal can be collected automatically or manually, in the latter case the data is manually entered into the processor (1).
[0056] For the purposes of the present invention, the processor (1) is defined as an electronic component whose central function is to convert raw information into useful results through logical, mathematical or control operations. In other words, the processor (1) is defined as an electronic component responsible for receiving input data, interpreting instructions, performing calculations, analyzing, transforming and / or processing the input data, and generating / returning at least one output data or result, based on the processing of the input data.
[0057] The model implemented in the processor (1) is able to process and mathematically interpret the windowed ICP data. The trained model is able to return a number that indicates the value of the patient's intracranial pressure based on its interpreted ICP morphology waveform.
[0058] In one embodiment, the value of the patient's intracranial pressure generated by the model is given in at least one unit among: mmHg, cmH2O, mmH2O, cmHg, Pascal, Bar, Psi, among others that can be used to represent the respective physical quantity. The resulting value is assigned or related to the collected pulse, that is, through the waveform of the ICP morphology, a numerical value is obtained that represents the ICP in the desired unit.
[0059] In this sense, the model is also capable of operating in real time, receiving ICP morphology waveforms and returning the patient's absolute ICP value and / or at least one related physiological parameter in real time. Therefore, if there are morphological changes in the waveform, the numerical value returned by the model also changes relatively.
[0060] In one embodiment, the present method comprises a step of returning, additionally, at least one physiological parameter simultaneously with the return of an absolute value of intracranial pressure, based on the interpretation, by the previously trained model, of at least one physiological signal from the patient.
[0061] For the purposes of the present invention, the physiological parameter generated / returned by the previously trained model is defined as one or more values and / or range of values and / or measurements and / or curves related to the functioning of a patient's organism, used to assess health status, the performance of biological functions, and possible clinical changes. Furthermore, for the purposes of the present invention, the physiological parameter includes any bodily and / or cerebral parameters, such as, but not limited to: cranial compliance parameters; hemodynamic parameters; autoregulation indices; and / or a combination thereof.
[0062] In one embodiment, the physiological parameter comprises at least one of the following parameters: P2 / P1 ratio; time to peak (TTP); intracranial pressure (ICP); mean arterial pressure (MAP); cerebral perfusion pressure (CPP); cerebrovascular autoregulation (CAR); optimal mean arterial pressure (optABP); pulse amplitude index (PAx); and / or a combination thereof. In one embodiment, such physiological parameters returned by the model are made available and displayed on a cerebral monitoring panel (2) for patient monitoring.
[0063] Specifically, the P2 / P1 ratio indicates the risk of intracranial hypertension. Time to peak (TTP) is used to dynamically assess cerebral compliance. Mean arterial pressure (MAP) is a parameter that represents the blood pressure driven by the heart that needs to overcome resistance within the skull to ensure cerebral blood flow. Cerebral perfusion pressure (CPP) is the value that determines whether the brain is receiving enough blood or whether pressure within the skull is impeding this flow. Cerebrovascular autoregulation (CAR) is the brain's intrinsic ability to maintain stable and adequate cerebral blood flow, even when variations in mean arterial pressure occur.
[0064] In one embodiment, the present method comprises a step of receiving, through a data receiving interface, at least one data window of at least one physiological signal from a patient. In one embodiment, said data receiving interface feeds the previously trained model with the received data window.
[0065] In one embodiment, the data window to be interpreted and processed is acquired by means of an intracranial pressure measuring device placed on the patient. In this embodiment, the data receiving interface is capable of receiving the data window in real time at the bedside, originating from said measuring device. In one embodiment, the intracranial pressure measuring device is a non-invasive measuring device. In another embodiment, the intracranial pressure measuring device is an invasive or minimally invasive measuring device.
[0066] In another embodiment, the data window is acquired by means of a database or a file containing the complete waveform of intracranial pressure morphology of the patient to be interpreted and processed. In this embodiment, the data receiving interface is able to acquire the data window in real time or not, at the bedside, so that the data window can be collected automatically or manually, in the latter case the data window is manually entered into the processor (1).
[0067] In one embodiment, the algorithm comprises one or more models for generating the intracranial pressure value and / or the physiological parameters of a patient. In another embodiment, the algorithm comprises a model for generating each physiological parameter. In yet another embodiment, the algorithm comprises a model for generating all physiological parameters.
[0068] In one embodiment, the model is pre-trained to perform the steps of the method using a regression tool employing Histogram Gradient Boosting (HGB).
[0069] In one embodiment, the model is pre-trained to perform the steps of the method using at least one regression machine learning tool.
[0070] In one embodiment, the machine learning model is a classifier and regressor based on decision trees or used for the formation of decision trees. In another embodiment, a Histogram Gradient Boosting model is used. In yet another embodiment, the model is a Random Forest. Certainly, 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 or resulting intracranial pressure from the collected curve.
[0071] Thus, the analysis guided by the method comprising the model of the present invention allows the monitoring of intracranial pressure and different physiological parameters from the morphological curve of intracranial pressure collected and / or received in real time, providing information derived from wearable technology data and electronic health records that inform evidence-based clinical data, promoting preventive medicine.
[0072] Another object of the invention is a system for estimating a patient's intracranial pressure value comprising at least one data display interface communicating with a processor (1), which reads a data window of at least one physiological signal from a patient, wherein: - the processor (1) is provided with a machine learning model that processes and retrieves, for the data display interface, an absolute intracranial pressure value from the data window; and The patient's physiological signal comprises a complete waveform of intracranial pressure morphology.
[0073] For the purposes of the present invention, a data display interface is defined as a component responsible for presenting processed data in a visual and / or textual and / or audio and / or interactive manner, aiming to facilitate the interpretation and decision-making of medical professionals. Furthermore, for the purposes of the present invention, the data display interface may present information in different formats, such as text, numbers, images, graphs, time waves, visual indicators, animations, and / or audible alerts. In one embodiment, the data display interface is, for example and without limitation, a monitoring panel, graphic monitor / screen, interactive dashboard, smartphone / tablet / notebook display, digital or printed reports, smartwatches and wearables, headphones / speakers, web interfaces, voice command interfaces, etc.
[0074] In one embodiment, the physiological signal entering the model comprises selected and / or aggregated pulses of the complete (raw) waveform of intracranial pressure morphology over a time interval, in order to obtain a complete representation of a patient's intracranial pressure morphology waveform. In one embodiment, the ICP morphology waveform, processed by the model, is representative of a monitoring time interval. In one embodiment, the selection and / or aggregation of pulses from the complete waveform is decided by the processor / model, where such a decision is motivated by computational efficiency given the low variability between adjacent pulses in continuous monitoring. In one embodiment, at certain times, there is no need to interpret the raw data for the entire minute.In this sense, since neighboring pulses in the PIC morphology waveform tend to be similar, pulses can be selected and / or aggregated in order to have a complete representation of the pulse morphology.
[0075] In one embodiment, the present system comprises a data receiving interface communicating with the processor (1) and / or the data display interface, wherein the data receiving interface is configured to receive the patient's physiological signal, i.e., the complete waveform of the patient's intracranial pressure morphology, which is interpreted by the model implemented in the processor (1).
[0076] In one embodiment, the data receiving interface receives the physiological signal from an intracranial pressure measuring device and / or from a database or file containing the complete ICP morphology waveform of a patient. In another embodiment, the data receiving interface communicates with an intracranial pressure measuring device and / or a database.
[0077] In one embodiment, the data receiving interface is configured to additionally receive an electrical activity signal from a patient's heart, originating from an electrocardiogram device, which is interpreted by the model implemented in the processor (1). In one embodiment, the data receiving interface communicates with an electrocardiogram device. In one embodiment, the data receiving interface communicates with any other devices for measuring physiological signals / body and / or brain parameters and databases, being able to receive different body and / or brain waveforms and physiological signals, which can be interpreted by the model implemented in the processor (1).
[0078] In one embodiment, the model is able to return, additionally, at least one physiological parameter simultaneously with an absolute value of the patient's intracranial pressure, based on the model's interpretation of at least one physiological signal from the patient. In one embodiment, the physiological parameters comprise: parameters associated with cranial compliance, correlated hemodynamic parameters, indices derived from autoregulation, and / or a combination thereof, forming a set of physiological parameters.
[0079] In a concrete application, the physiological parameter comprises at least one of: a. ratio P2 / P1; b. time to reach peak (TTP); c. mean arterial pressure (MAP); d. cerebral perfusion pressure (CPP); e. cerebrovascular pressure autoregulation (CVA); f. optimal mean arterial pressure (optABP); g. pulse amplitude index (PAx); and / or h. a combination of the same.
[0080] It is, yet another object of the invention, a model training method for estimating a patient's intracranial pressure value, which is performed by an algorithm executed by a processor (1), which performs preprocessing steps of the ICP morphology signal.
[0081] The aforementioned training method comprises the following steps: a. extraction, by a processor (1), of one or more data windows from a plurality of distinct intracranial pressure morphology waveforms; and b. comparison of said data windows with absolute values of intracranial pressure; wherein: The plurality of waveforms with distinct intracranial pressure morphologies is associated with various physiological and / or clinical conditions, attributed to different sets of physiological parameters and intracranial pressure values.
[0082] For the purposes of understanding the present invention, said distinct intracranial pressure morphology waveform signals comprise ICP morphology signals from distinct patients, with distinct pathological conditions, with different cranial conditions, with distinct diagnoses, etc.
[0083] The morphology of the cranial pulse exhibits significant variability between individuals and between different clinical states of the same person throughout monitoring. For example, there is sensitivity to physiological noise (movement, respiratory cycle, vasomotor activity), to ICP recording artifacts, and to variations inherent to the patient's clinical condition (acute intracranial hypertension, hydrocephalus, cerebral atrophy, postoperative, etc.).
[0084] This variability makes it non-trivial to ensure that the waveform processed by one or more models is representative of the physiological state of a given interval. In view of this, the present model training method is performed with a broad and diverse database of patients and clinical conditions, and comprises an internal architecture capable of generalizing to atypical morphologies.
[0085] From this, in a concrete application, the model is trained with a plurality of intracranial pressure morphology waveforms, possessing varied physiological and / or clinical conditions attributed to varied sets of physiological parameters and intracranial pressure values.
[0086] In one embodiment, the model training method is performed using a regression tool employing the Histogram Gradient Boosting (HGB) method.
[0087] In one embodiment, the model training method is performed using at least one regression machine learning tool.
[0088] In one embodiment, the model training method uses a classifier and regressor based on decision trees or used for the formation of decision trees. In another embodiment, a Histogram Gradient Boosting is used. In yet another embodiment, a Random Forest is used. Certainly, other tools can be used within the scope of the invention, so that it is possible to perform the training from ICP morphology data to return the absolute numerical value of the relative intracranial pressure or the resultant of the collected curve and / or at least one physiological parameter.
[0089] In this way, based on this method, the model is then trained to identify the complete morphology signal pattern of the ICP and extract numbers corresponding to the ICP morphology, generating a numerical value of the ICP and at least one related physiological parameter of the patient in real time.
[0090] In one embodiment, the present model is trained to be able to generate and estimate at least one physiological parameter simultaneously with the absolute value of intracranial pressure.
[0091] In one embodiment, the estimated intracranial pressure and physiological parameter values are trained by a specific model. In another embodiment, the estimated intracranial pressure and physiological parameter values are trained by one or more models implemented in the processor.
[0092] The preliminary steps in processing the ICP morphology signal include the extraction, by a processor (1), of one or more data windows from a plurality of distinct intracranial pressure morphology signals for training one or more models, as well as the comparison of said data windows with absolute intracranial pressure values of the respective distinct intracranial pressure morphology signals.
[0093] Thus, it is an objective of the present invention, the aforementioned method for estimating intracranial pressure values and physiological parameters of a patient, which allows the measured ICP morphology curve to be converted into numerical values of intracranial pressure and physiological parameters for subsequent accurate and efficient medical diagnosis.
[0094] Yet another object of the invention is a monitoring panel for intracranial pressure and physiological parameters of a patient, comprising a display showing an absolute value of intracranial pressure and at least one physiological parameter generated by a previously trained model, wherein the model retrieves the absolute value of intracranial pressure and the physiological parameters from a data window received by the model from at least one physiological signal from a patient.
[0095] In one embodiment, the monitoring panel (2) comprises a display that can present the information generated by the model in different formats, such as text, numbers, images, graphs, time waves, visual indicators, animations and / or sound alerts.
[0096] In a specific embodiment, the monitoring panel (2) is an interactive dashboard with a graphical interface that allows visualization and quick access to intracranial pressure values and physiological parameters generated by the model, based on the received data window, in addition to allowing medical professionals to interact with the values generated and displayed on the panel (2).
[0097] Furthermore, another object of the invention is a computer program comprising instructions executable by the computer, wherein the instructions, when executed by the computer, reproduce the method for estimating a patient's intracranial pressure value. Examples:
[0098] The examples shown here are intended only to illustrate one of the numerous ways of carrying out the invention, without, however, limiting its scope. Example 1 - Use of the complete waveform of intracranial pressure morphology for the generation and measurement of a set of physiological parameters of a patient.
[0099] In this example, a method was developed for estimating a set of patient physiological parameters, based on an artificial intelligence model implemented in an algorithm executed by a processor (1).
[0100] As shown in Figure 1, one of the objectives of this example is to use a physiological signal from a patient as input to a previously trained model, and to return an absolute value of intracranial pressure.
[0101] As shown in Figure 2, one of the objectives of this example is to use a physiological signal from a patient as input into a previously trained model, and to return a set of physiological parameters, including: P2 / P1 ratio; time to peak (TTP), intracranial pressure (ICP / elCP) in mmHg, mean arterial pressure (MAP) in mmHg, cerebral perfusion pressure (CPP) in mmHg, cerebrovascular pressure autoregulation (CAR), and two indices generated in this developed model, namely the optimal mean arterial pressure (optABP) in mmHg and the pulse amplitude index (cerebral regulation, PAx).
[0102] As shown in Figure 3, one of the objectives of this example is to use the complete waveform of intracranial pressure morphology from a patient as input to a previously trained model, and to retrieve the set of physiological parameters.
[0103] In this example, the developed artificial intelligence model receives as input a complete waveform of the cranial morphological pulse, captured by means of a non-invasive sensor from brain4care, and produces as output a set of physiological parameters (P2 / P1, TTP, elCP, MAP, CPP, CAR, optABP, PAx, or a subset thereof, depending on the usage configuration).
[0104] The waveform processed by the models is representative of a monitoring time interval. The strategy for selecting or aggregating this waveform within the time interval is an internal choice of the system's operation, motivated by computational efficiency given the low variability between adjacent pulses in continuous monitoring. That is, at certain times, there is no need to read the raw data for the entire minute. Since neighboring pulses tend to be similar, the present model can select pulses or aggregate pulses in order to have a complete representation of the pulse.
[0105] Furthermore, the entire transformation of the input ICP morphological waveform into output physiological parameters occurs within the algorithm where the trained models are included, without external steps of extraction or reduction of the signal to pre-computed morphological parameters provided as input. After all, each output parameter may be more sensitive to parts of the wave than others, and these do not necessarily share the same characteristics.
[0106] In this sense, the morphology of the cranial pulse presents significant variability between subjects and between different clinical states of the same subject throughout the monitoring. There is sensitivity to physiological noise (movement, respiratory cycle, vasomotor activity), to capture artifacts in the measurement sensor, and to variations inherent to the patient's clinical condition (acute intracranial hypertension, hydrocephalus, cerebral atrophy, postoperative, etc.).
[0107] This variability makes it non-trivial to ensure that the waveform processed by the models is representative of the physiological state of a given interval. In view of this, there was a demand for training the model on a broad and diverse base of patients and clinical conditions, and for developing an internal architecture capable of generalizing to atypical morphologies.
[0108] State-of-the-art approaches, based on pre-computed morphological descriptors (specific amplitudes, time intervals between subpeaks, externally calculated P2 / P1 ratios), tend to fail precisely in these atypical morphologies, where manual subpeak detection rules lose reliability.
[0109] Thus, the choice to use the complete waveform of the cranial morphological pulse, instead of a reduced vector of pre-computed numerical descriptors (state of the art), fully preserves the morphological information of the signal: subpeaks, relative temporal phases, amplitudes, rise and fall slopes, durations, and relationships between pulse components.
[0110] As is typical in the state of the art, any prior reduction to scalar parameters discards information that may be relevant for the simultaneous estimation of multiple output parameters, especially when the relationship between morphology and parameters involves non-linear or non-obvious combinations of waveform aspects.
[0111] Thus, in the method of the present example, the waveform is maintained in its entirety at the model input, which allows the models to learn these relationships directly from the annotated data, without depending on prior assumptions about which aspects of the morphology are relevant for each estimated parameter.
[0112] In this context, the cranial pulse waveform is a composite physiological signature, generated by the dynamic interaction of multiple “systems”: the cerebral arterial pulse drive, the compliance of the intracranial system (which determines how volume changes translate into pressure changes), cerebrovascular autoregulation (which modulates the relationship between systemic arterial pressure and cerebral flow), and the resulting cerebral perfusion. Each of these “systems” leaves characteristic marks on the pulse morphology.
[0113] The relationship between the subpeaks of the pulse (percutaneous component, tidal component, dicrotic component to systolic component, tidal component, and diastolic component), their relative amplitudes, their temporal phases, and the overall shape of the pulse reflect the combined state of these "systems" at a given moment.
[0114] Decades of clinical literature on intracranial dynamics have documented that alterations in pulse morphology, such as inversion of the relationship between the percutaneous and tidal components, displacement of the time to peak, pulse widening, and loss of dicrotic components, accompany clinically relevant changes in various neurological conditions.
[0115] Thus, the central hypothesis, empirically confirmed during the development and validation of the models in this example, is that this composite information present in the signal morphology is sufficiently rich for regression models trained on annotated data to simultaneously estimate parameters associated with compliance (P2 / P1, TTP, ICP / elCP), correlated hemodynamic parameters (MAP, CPP), and autoregulation-derived indices (CAR, optABP, PAx).
[0116] In this sense, confirmation was obtained from two complementary sources: paired databases of invasive and non-invasive monitoring in diverse clinical populations, used to train and validate the models against invasive ground truth, and published clinical studies that demonstrated clinically relevant predictive performance of the estimated parameters against clinical outcomes.
[0117] Advantages of the approach and main gains observed:
[0118] A single set of models is capable of generating multiple physiological output parameters simultaneously, while maintaining physiological consistency between them. Parameters that describe related aspects, such as compliance, autoregulation, and perfusion, share the same learned internal representation of morphology, which tends to produce mutually consistent estimates.
[0119] Robustness to inter-patient and intra-patient variations, where full-waveform-based models, trained on a diverse dataset, generalize better to cases not seen during training than approaches based on manually engineered descriptors.
[0120] - Independence from manual subpeak detection rules, where, in atypical morphologies, frequent in neurological ICU settings, automatic detection of P1, P2, and P3 may fail. The present system does not depend on this external detection as a prerequisite for generating its outputs.
[0121] - Ability to capture non-linear relationships between different morphological aspects, which might be invisible to approaches based on isolated parameters (state of the art).
[0122] - Internal architecture agnostic to future evolutions: changes in how models learn to internally represent morphology, through retraining with new data, architectural adjustments, or incorporation of more advanced algorithms, do not require alteration of the system's input-output interface. Example 2 - Machine learning structure using regression models
[0123] The method developed in this example aims to use an artificial intelligence model to read a data window of a complete intracranial pressure waveform to generate a set of numerical values of a patient's physiological parameters, based on the interpretation of the processed data window.
[0124] The set of physiological parameters generated by the model includes: P2 / P1 ratio; time to peak (TTP), intracranial pressure (ICP / elCP) in mmHg, mean arterial pressure (MAP) in mmHg, cerebral perfusion pressure (CPP) in mmHg, cerebrovascular autoregulation (CAR), and two indices generated in this developed model, namely the optimal mean arterial pressure (optABP) in mmHg and the pulse amplitude index (cerebral regulation, PAx). In this sense, the set of physiological output parameters is displayed on a patient physiological parameter monitoring panel (2), as shown in Figure 2, facilitating the observation of the parameters by the medical professional.
[0125] As illustrated in Figure 3, the patient's ICP morphology curves / waveforms are obtained using a non-invasive ICP measurement device (3) coupled around the patient's head, which detects the patient's ICP morphology curves in real time at the bedside. The ICP morphology curves are read by a processor (1) equipped with an algorithm, which analyzes said curves using a previously trained model and generates the set of physiological parameters based on the curves read, so that the set of physiological parameters can be visualized on a brain monitoring panel (2).
[0126] For model training, a machine learning (ML) tool was used, with the regression tool employing the Histogram Gradient Boosting (HGB) method, which works through successive decisions by evaluating probability histograms.
[0127] For model training, the hyperparameters consist of 4,000 individual decision trees with a maximum depth of 15 successive splitting steps before reaching a final decision. Furthermore, the model optimization strategy used includes adjusting the number of trees and the depth using the grid-search method. The generalization technique used for model training was Bootstrap. Additionally, the generalization hyperparameters have a maximum of 50 terminal nodes (leaves), meaning each tree in the model can have a maximum of 50 terminal nodes. The learning rate used is 0.005, which defines the size of the adjustment made by the model at each training step.
[0128] To estimate the numerical values of the patient's physiological parameters using this model, training was performed with a database of 113 patients. 20 patients were selected for model validation, and the remaining 92 patients were used for model training, with one patient discarded because it had only a 10-second window of the ICP waveform. Thus, the data were divided into 17% for testing and 83% for training in each of the 10 cross-validation folds, as illustrated in Figure 5.
[0129] Patient data were obtained from a database containing physiological parameters for each patient, who were monitored by medical equipment collecting various physiological parameters such as P2 / P1 ratio, TTP, MAP, CPP, CAR, optABP, PAx, and the absolute value of ICP. Additionally, an intracranial pressure waveform detection sensor was used to collect the signal from each patient. The data were then structured to correlate the intracranial pressure waveform with the physiological parameters P2 / P1 ratio, TTP, MAP, CPP, CAR, optABP, PAx, and ICP for each patient, and used for training and validating the model. Example 3 - Validation of the model for estimating physiological parameter values of a patient from intracranial pressure (ICP) morphology waveforms.
[0130] The data analysis began with a series of 112 patients divided into two groups, with group 1 being the first.o group with ICP <20 mmHg and the 2 o group with ICP >=20 mmHg. These data were segmented into 10-second windows, generating 10982 (94.64%) distinct windows in 1 o group and 622 (5.36%) distinct windows in 2 o group. The generated data is shown in Table 1 below.
[0131] Table 1 - Characteristics of the patients participating in the test and data generated. ICP <20 mmHg ICP >20 mmHg Window (10 seconds) 10982 (94.64%) 622 (5.36%) Age (years) 51.60±16.70* 43.30±16.40* ICP (mmHg) 9.81 ±4.74* 25.10±6.46* Sex: Male 76% 81% Female 24% 19% Pathology (%): Traumatic brain injury 38 52 Subarachnoid hemorrhage 29 19 Intracerebral hemorrhage 2 23 Ischemic stroke 1 0 Mass 2 1 Subdural hematoma 0 0 Stroke 0 6 Other 0 0 Distribution of centers (% of patients) (included in the data analysis): University Hospital of 57 São Paulo Federal University of São Paulo 4 Hospital João XXIII 3.5 State Emergency Hospital and 3.5 Senator Humberto Lucena's Trauma University of Porto, São 6 Hospital John Stanford University 20 Johns Hopkins University 6
[0132] The model was statistically evaluated using the Mean Absolute Error (MAE) and Mean Squared Error (MSE) metrics, showing favorable results. For the ICP value in mmHg, the MAE obtained was 2.99637. These ICP values were synchronized with changes in intracranial pressure measured invasively (ICP), as illustrated in Figure 6. For the mean arterial pressure (MAP) value in mmHg, the MAE obtained was 1.47301. These MAP values were synchronized with changes in measured mean arterial pressure, as illustrated in Figure 7. For the cerebral perfusion pressure (CPP) value in mmHg, the MAE obtained was 3.64153. These CPP values were synchronized with changes in measured cerebral perfusion pressure, as illustrated in Figure 8.As illustrated in Figures 6, 7, and 8, there is a direct correlation between the measured values of ICP, MAP, and CPP and the values generated by the model of the present invention, demonstrating the accuracy of the developed model. Furthermore, Figure 9 shows the relationship between invasive pulse amplitude index (PAx) measurements and the PAx measured using the model of the present invention (nPAx), also illustrating the correlation between the data and the accuracy of the developed model.
[0133] Thus, the proposed model was able to function within the acceptable limits recommended by the Association for the Advancement of Medical Instrumentation, which stipulates a maximum error of ±2 mmHg for ICP measurements below 20 mmHg.
[0134] Those skilled in the art will appreciate the knowledge presented here and will be able to reproduce the invention in the forms presented and in other variants and alternatives, covered by the scope of the following claims.
Claims
Claims 1. Method for estimating a patient's intracranial pressure value characterized by comprising the following steps: a. interpretation, using a previously trained model, of at least one data window from at least one physiological signal of a patient; and b. return of an absolute value of intracranial pressure, from the interpretation of the data window, by the model; - in which the patient's physiological signal comprises a complete waveform of intracranial pressure morphology.
2. A method according to claim 1, characterized in that the physiological signal comprises selected and / or aggregated pulses of the physiological signal, in order to obtain a complete representation of the waveform of the intracranial pressure morphology of a patient, over a period of time.
3. A method, according to claim 1, characterized by comprising a step of simultaneously returning an absolute value of intracranial pressure and at least one physiological parameter, based on the interpretation, by the previously trained model, of at least one physiological signal from a patient.
4. Method, according to claim 1, characterized by comprising a step of returning an intracranial pressure value simultaneously with at least one physiological parameter among: a. ratio P2 / P1; b. time to reach peak (TTP); c. mean arterial pressure (MAP); d. cerebral perfusion pressure (CPP); e. cerebrovascular pressure autoregulation (CVA); f. optimal mean arterial pressure (optABP); g. pulse amplitude index (PAx); and / or h. a combination of the same; based on the interpretation, by the previously trained model, of at least one physiological signal from a patient.
5. A method according to claim 1, characterized by comprising a receiving step, via a data receiving interface, of at least one data window of at least one physiological signal from a patient, the data window being acquired by means of an intracranial pressure measuring device placed on the patient, and / or by means of a database containing the complete waveform of the patient's intracranial pressure morphology.
6. Method, according to claim 1, characterized in that the model is trained with a plurality of intracranial pressure morphology waveforms, possessing varied physiological and / or clinical conditions attributed to varied sets of physiological parameters and intracranial pressure values.
7. System for estimating a patient's intracranial pressure value characterized by comprising at least one data display interface communicating with a processor (1), which reads a data window of at least one physiological signal from a patient, wherein: - the processor (1) is provided with a machine learning model that processes and retrieves, for the data display interface, an absolute value of intracranial pressure from the data window; and The patient's physiological signal comprises a complete waveform of intracranial pressure morphology.
8. System, according to claim 7, characterized in that the physiological signal comprises selected and / or aggregated pulses of the complete intracranial pressure morphology waveform over a time interval, so as to obtain a complete representation of the intracranial pressure morphology waveform of a patient.
9. System, according to claim 7, characterized in that the machine learning model additionally returns at least one physiological parameter simultaneously with an absolute value of intracranial pressure, based on the model's interpretation of at least one physiological signal from a patient.
10. System according to claim 9, characterized in that the physiological parameter comprises at least one of: a. ratio P2 / P1; b. time to reach peak (TTP); c. mean arterial pressure (MAP); d. cerebral perfusion pressure (CPP); e. cerebrovascular pressure autoregulation (CVA); f. optimal mean arterial pressure (optABP); g. pulse amplitude index (PAx); and / or h. a combination of the same.
11. System, according to claim 7, characterized by comprising a data receiving interface communicating with the processor, wherein said interface receives a data window of at least one physiological signal from a patient.
12. System, according to claim 7, characterized in that the machine learning model is trained with a plurality of intracranial pressure morphology waveforms, possessing varied physiological and / or clinical conditions attributed to varied sets of physiological parameters and intracranial pressure values.
13. Model training method for estimating a patient's intracranial pressure value characterized by comprising the following steps: a. extraction, by a processor (1), of one or more data windows from a plurality of distinct intracranial pressure morphology waveforms; and b. comparison of said data windows with absolute values of intracranial pressure; wherein: - The plurality of waveforms with distinct intracranial pressure morphologies has varied physiological and / or clinical conditions attributed to different sets of physiological parameters and intracranial pressure values.
14. Training method, according to claim 13, characterized in that the model is trained using a regression tool employing a gradient histogram reinforcement method.
15. A monitoring panel for intracranial pressure and physiological parameters of a patient, characterized by comprising a display showing an absolute value of intracranial pressure and at least one physiological parameter generated by a previously trained model, wherein the model returns the absolute value of intracranial pressure and the physiological parameters from a data window received by the model from at least one physiological signal from a patient.
