An acute brain injury prognosis detection system

Abstract

This invention discloses an acute brain injury (TBI) prognosis detection system, comprising: a multi-source data synchronous acquisition and extraction section, a data preprocessing section, a machine learning prediction model, and an acute traumatic brain injury result display system. The system performs multi-source synchronous acquisition and extraction of 19-channel electroencephalogram (EEG) data distributed throughout the brain and data from the left and right middle cerebral arteries (MCA) under a mixed auditory evoked paradigm task. The system preprocesses clinical data, EEG, MCA, and Glasgow Outcome Scale (GOS) scores of TBI patients. The preprocessed data is then input into the machine learning prediction model, which calculates and outputs the probability of poor prognosis in TBI patients. The TBI patient poor prognosis prediction result displays the probability of poor prognosis in TBI patients. This invention enables early, simple, and accurate detection of acute brain injury (TBI) patients.

Description

Technical Field

[0001] This invention relates to the field of trauma prognosis, and more particularly to an acute brain trauma prognosis detection system. Background Technology

[0002] Traumatic brain injury (TBI) refers to injury caused by direct mechanical force applied to the head, with a global annual incidence rate as high as 939 / 100,000. [1] TBI has a high rate of disability and mortality; 10% of patients with mild TBI and 66%–100% of patients with moderate to severe TBI experience long-term cognitive impairment. Early and rapid assessment of the severity of TBI helps clinicians to develop timely treatment plans and improve prognosis. [2] Transcranial Doppler (TCD) has high value in assessing neurological damage. Increased cerebral blood flow velocity can predict the early occurrence of neurological complications such as hypoxic-ischemic encephalopathy and neurocognitive decline. [3] The middle cerebral artery is the largest branch of the internal carotid artery, supplying 80% of the blood to the brain tissue and directly reflecting changes in brain hemodynamics.

[0003] Currently, cerebral hemodynamics and radiomics are being used to assess the condition and prognosis of spontaneous intracerebral hemorrhage. [4,5] However, clinical assessment of TBI severity often relies on cerebral hemodynamics and clinical experience, which can lead to misdiagnosis and missed diagnosis, and also introduces a degree of subjectivity. [6,7] Multimodal detection has become an important trend in medical diagnosis. Among the many methods of detecting neural information, with the development of BCI research, many studies have found that there is a certain correlation between EEG characteristics and disease pathology. EEG is widely used due to its advantages such as simple operation, non-invasiveness, and high temporal resolution.

[0004] References

[0005] [1]Dewan MC, Rattani A, et al. Estimating the global incidence of traumatic brain injury. JNeurosurg. 2018Apr 27;130(4):1080-1097.

[0006] [2] Maas AIR, Menon DK, et al. Traumatic brain injury: integrated approaches to improve prevention, clinical care, and research. Lancet Neurol. 2017 Dec; 16 (12): 987-1048.

[0007] [3]Millet A,Evain JN,et al.Clinical applications of transcranialDoppler in non-trauma critically ill children:a scoping review.Childs NervSyst.2021Sep;37(9):2759-2768.

[0008] [4]Ye H,Gao F,et al.Precise diagnosis of intracranial hemorrhage andsubtypes using a three-dimensional joint convolutional and recurrent neuralnetwork.Eur Radiol.2019Nov;29(11):6191-6201.

[0009] [5]Jain V,de Godoy LL,et al.Cerebral hemodynamic and metabolicdysregulation in the postradiation brain.J Neuroimaging.2022Nov;32(6):1027-1043.

[0010] [6]Koyama T,Uchiyama Y,et al.Comparison of Fractional Anisotropy fromTract-Based Spatial Statistics with and without Lesion Masking in Patientswith Intracerebral Hemorrhage:ATechnical Note.J Stroke CerebrovascDis.2019Nov;28(11):104376.

[0011] [7]Muravskiy A, Polischuk M, et al. Changes in cerebral hemodynamics inboxers with repeated traumatic brain injury. Pol Merkur Lekarski. 2020Oct 23;48(287):312-317. Summary of the Invention

[0012] This invention provides an acute brain injury prognosis detection system, which can perform early, simple, and accurate detection in patients with acute brain injury, as detailed below:

[0013] An acute brain injury prognosis detection system, the system comprising: a multi-source data synchronous acquisition and extraction section, a data preprocessing section, a machine learning prediction model, and an acute traumatic brain injury result display system.

[0014] 19-channel EEG data distributed throughout the brain and left and right MCA data were simultaneously collected and extracted from multiple sources under a mixed auditory evoked paradigm task. Clinical data, EEG, MCA, and classification label GOS of brain trauma patients were preprocessed.

[0015] The preprocessed data is input into a machine learning prediction model, which calculates and outputs the probability of a poor prognosis for TBI patients.

[0016] The adverse outcome prediction results for TBI patients show the probability of poor prognosis in TBI patients.

[0017] The paradigm mentioned above is a hybrid auditory evoked paradigm that integrates auditory steady-state evoked response (ASSR) into the auditory P300 system;

[0018] The sound stimulation carrier frequencies were 500Hz and 2000Hz, and the modulation frequency was 40Hz. The auditory stimulation was amplitude modulated (AM) with a modulation depth of 100% to represent the standard stimulus and the deviated stimulus, respectively. In each trial, the standard stimulus and the deviated stimulus appeared randomly in a 4:1 ratio to induce auditory P300. EEG and left and right MCA data under the mixed auditory evoked paradigm were collected simultaneously, and the user's EEG characteristic parameters were calculated and extracted offline.

[0019] The collection and extraction of acute traumatic brain injury data included the following: patient inclusion and extraction of multiple physiological data. Inclusion criteria: age ≥ 18 years; Glasgow Coma Scale (GCS) score ≤ 12. Exclusion criteria: open head trauma; severe multiple injuries; patients transferred from other hospitals for rehabilitation; patients who died upon admission; patients with prior head injury; patients lost to follow-up or with incomplete data; liver and kidney failure; hemorrhagic diseases; coagulation disorders; severe infections; malignant tumors.

[0020] Data collection: ERP and ASSR EEG characteristic parameters, MCA hemodynamic parameters, and patient basic information, as well as cerebral blood flow parameters and EEG data were collected and extracted simultaneously at 1 day, 1 week, and 2 weeks after hospitalization.

[0021] After the data is collected and extracted, it is preprocessed: data cleaning and normalization.

[0022] The machine learning prediction model was constructed using the support vector machine machine learning algorithm. A retrospective collection of 300 TBI patients was conducted, with 210 TBI cases used as the training set and 90 TBI cases used as the test set.

[0023] The preprocessed data is input into the prediction model. The SVM classification model parameters are set as follows: the penalty coefficient for incorrect terms is 200, the kernel type is Gaussian kernel, the kernel coefficient is 5, whether to enable probability estimation is False, whether to use heuristic shrinkage is True, the error precision at which the model stops training is 0.001, the memory required for training is 400, the weights of different penalty parameters for each class are set to None, the maximum number of iterations is -1, the cross-validation parameter is 5, and the model evaluation criteria are ROC and AUC.

[0024] The beneficial effects of the technical solution provided by this invention are:

[0025] 1. This invention uses a brain-computer interface method based on a hybrid auditory evoked paradigm of auditory event-related potential (ERP) and auditory steady-state response (ASSR) to establish a machine learning (ML) model based on EEG parameters and middle cerebral artery hemodynamic parameters to predict the 6-month poor prognosis rate of acute TBI.

[0026] 2. The present invention has been trained and validated on the training set and the test set to ensure that the prediction model established by the present invention can make early, simple and accurate predictions for patients with acute brain injury, thereby assisting in the formulation of clinical decisions.

[0027] 3. The SVM-based machine learning prognostic detection system can predict the early prognosis of TBI patients, helping clinicians to provide personalized adjunctive treatment plans for TBI patients earlier and providing suggestions for the treatment of TBI patients.

[0028] 4. This invention enables the detection of brain consciousness levels and can be extended to fields such as neuroscience and life sciences, yielding considerable social and economic benefits. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the structure of an acute brain injury prognosis detection system.

[0030] Figure 2 A schematic diagram of the timing design for a mixed auditory evoked paradigm. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below.

[0032] This invention provides an acute brain injury prognosis detection system. The system mainly comprises four parts: a multi-source data synchronous acquisition section, a data preprocessing section, a machine learning model prediction section, and a display of acute traumatic brain injury (TBI) prediction results. This invention aims to improve the accuracy of predicting the poor prognosis rate of patients in hospital intensive care units 6 months later, assisting physicians in clinical decision-making and improving patient outcomes. To this end, this invention aims to design a hybrid auditory evoked paradigm, simultaneously acquire EEG and MCA cerebral blood flow parameters, and integrate patient clinical physiological information to build an acute brain injury prediction system.

[0033] The technical process is as follows: by inputting the ERP, ASSR EEG characteristic parameters and MCA hemodynamic parameters of TBI patients at 1 day, 1 week and 2 weeks after hospitalization, based on the fusion analysis of multiple physiological information, the poor prognosis rate of TBI patients can be predicted efficiently, quickly, accurately and timely, providing strong support for clinical decision-making.

[0034] I. Design of an Acute Brain Injury Prognostic Prediction System Based on SVM

[0035] Study Subjects: Clinical data of patients in the craniocerebral trauma ward were used to select eligible patients with acute brain trauma based on the following inclusion and exclusion criteria. Inclusion Criteria: Age ≥ 18 years; Glasgow Coma Scale (GCS) score ≤ 12. Exclusion Criteria: Open craniocerebral trauma; severe multiple injuries; patients transferred from other hospitals for rehabilitation; patients who died upon admission; patients with prior brain injury; patients lost to follow-up or with incomplete data; liver and kidney failure; concomitant hemorrhagic diseases; coagulation disorders; severe infections; concomitant malignancies.

[0036] Data acquisition included ERP and ASSR EEG parameters, middle cerebral artery hemodynamic parameters, and basic patient information (age, gender, Glasgow Coma Scale score at admission) at 1 day, 1 week, and 2 weeks post-hospitalization. Cerebral blood flow parameters and EEG data were acquired simultaneously. This invention constructed a feature set, encompassing all patient information (age, gender, Glasgow Coma Scale score); middle cerebral artery hemodynamic parameters (end-diastolic velocity (EDV), peak systolic velocity (PSV), mean maximum velocity (MV), and pulsatility index (PI); and EEG parameters (amplitude, latency, Fast Fourier Transform (FFT), Event-related spectral perturbation (ERSP), phase-locked value (PLV), signal-to-noise ratio (SNR), and relative power (RP). This multi-source data fusion method comprehensively reflects the patient's physiological state and disease progression, providing richer and more accurate information for prediction.

[0037] The system of this invention embodiment is as follows Figure 1 As shown, its system architecture and technical process include: behavioral data acquisition (Glasgow Coma Scale GCS), multi-source data synchronous acquisition and extraction, neurophysiological index data preprocessing, machine learning model prediction, and display of prediction results for poor prognosis of acute brain injury.

[0038] Regarding the assessment criteria, the embodiments of this invention adopted the following two main dimensions: (1) Assessment of the severity of brain injury: Based on the Glasgow Coma Scale (GCS), a GCS score of 9-12 at hospital indicates moderate brain injury, 6-8 indicates severe, and below 5 indicates extremely severe. The GCS score at hospital is one of the features of the model. (2) Assessment of prognostic outcomes: The Glasgow Outcome Scale (GOS) at 6 months after injury is used for assessment. Patients with a GOS score of 11-15 are considered to have a good prognosis, while those with a score of 10 or below are considered to have a poor prognosis. The Glasgow Outcome Scale (GOS) at 6 months after injury is used as a classification label.

[0039] Subsequently, this embodiment of the invention utilizes a Support Vector Machine (SVM) machine learning algorithm to fuse analysis and construct a 6-month prognostic prediction model for patients. To ensure the accuracy and reliability of the model, patients were randomly divided into a training set and a test set. This embodiment of the invention trains on the test set and validates the model on the test set. This ensures that the predictive model established in this embodiment of the invention can predict the prognosis of traumatic brain injury patients early, simply, and accurately, assisting clinical decision-making.

[0040] Temporal design of mixed auditory evoked paradigms, such as Figure 2 As shown, this embodiment of the invention designs a hybrid auditory evoked paradigm of ERP and ASSR. By integrating ASSR into the auditory P300 system, an EEG synchronous acquisition device is built to offline calculate and extract the user's EEG feature parameters. The auditory stimulation uses amplitude modulation (AM) with a modulation depth of 100%, the sound stimulus carrier frequencies are 500Hz and 2000Hz, and the modulation frequency is 40Hz. In each trial, the standard stimulus and the biased stimulus appear randomly in a ratio of 80%:20%, thereby eliciting auditory P300.

[0041] Sinusoidal amplitude modulation (AM) controls the amplitude of the carrier signal using the modulation frequency to be transmitted, shifting the baseband signal's spectrum to a higher frequency. The carrier frequency exhibits periodic changes, and its envelope shape is similar to that of the modulation signal. The modulation depth is the ratio of the amplitude of the modulation signal to the amplitude of the DC signal, expressed as a percentage. The modulation signal m(t) is superimposed on the DC signal A and multiplied by the carrier signal to form the AM signal, whose time-domain expression is:

[0042] S AM (t)=(A+m(t))*cos(2*pi*f c *t)

[0043]

[0044] Among them, S AM (t) is the modulated signal, A is the DC signal, m(t) is the modulating signal, and f is the DC signal. c Let f be the carrier signal frequency, md be the modulation depth of the signal in %, and peak(m(t)) be the peak value of the modulated signal. For sound stimuli, f c=1000Hz; the modulation signal m(t) is a sinusoidal signal with a modulation frequency of 40Hz. In this embodiment of the invention, the sound is generated by the PsychPortAudio function in the Psychtoolbox of Matlab, in a waveform audio file format (*.wav) with a sampling rate of 44.1kHz, and played through headphones in both ears. To accurately control the presentation of auditory stimuli, the time error of the sound playback is controlled within 5ms to keep the time error within a minimum range. To prevent the sound stimulus from producing obvious popping sounds, the sound is faded in and out for 10ms.

[0045] II. Functions of Each Module in the Brain Trauma Prediction and Prognosis System

[0046] (1) EEG-cerebral blood flow data acquisition module

[0047] EEG acquisition utilized a 19-channel EEG acquisition system and its dedicated software, with a data acquisition sampling rate set to 1000Hz and a 50Hz power frequency notch filter. Different task modes corresponded to different event codes, facilitating subsequent synchronous data extraction and feature analysis. Bilateral middle cerebral artery (MCA) parameters were acquired via bilateral temporal windows using a color transcranial Doppler (TCD) pulse-wave Doppler probe. TCD is a cerebrovascular disease examination method that utilizes the ultrasound Doppler effect to examine the hemodynamics and various hemodynamic parameters of major arteries in the circle of Willis at the base of the brain. TCD offers advantages such as non-invasiveness, portability, radiation-free operation, and repeatability, directly acquiring hemodynamic parameters of the middle cerebral artery: EDV, PSV, MV, and PI. This invention innovatively incorporates EEG features into a multi-source data fusion method for cerebral hemodynamic parameters, comprehensively reflecting the patient's physiological state and disease progression, providing richer and more accurate information for prediction.

[0048] (2) Extraction of hemodynamic parameters of the middle cerebral artery

[0049] Transcranial Doppler ultrasound (TCD) is based on the Doppler effect, emitting ultrasound waves that penetrate the skull and are reflected by red blood cells moving through cerebral blood vessels, thereby acquiring cerebral hemodynamic parameters. A Delikai TCD monitor paired with ICM+ software was used to acquire parameters. A dual-channel transcranial Doppler with a 2MHz probe frequency was used. The patient was placed in a supine position, and the examiner stood at the patient's head and stabilized the skull. The examiner held the probe and probed along the temporal window to a depth of 55-65mm, locating blood flow towards the probe. The probe angle was finely adjusted to make the probe's sound wave direction as parallel as possible to the blood flow direction, and the maximum measured velocity was recorded. The steps for acquiring hemodynamic parameters of the middle cerebral artery included:

[0050] 1) Utilize TCD blood flow imaging technology to detect the distribution and direction of intracranial blood vessels, and locate the MCA position based on depth, velocity, and blood flow direction;

[0051] 2) Use ICM+ software to synchronously acquire hemodynamic parameters of the MCA;

[0052] 3) Define and quantify the blood flow parameters of the MCA, including EDV, PSV, MV, and PI.

[0053] (3) Extraction of EEG neurophysiological features

[0054] Fast Fourier Transform (FFT): ASSR is a frequency-domain neural response. To investigate ASSR evoked responses, it is necessary to convert the EEG signal from the time domain to the frequency domain. This embodiment of the invention uses Fast Fourier Transform for the conversion, as shown in the following formula:

[0055]

[0056] Where e is the base of the natural logarithm, N represents the sampling rate, and x[n] represents the collected discrete EEG signal, which is in complex form and its expression is x^[k]=a+bi.

[0057] Event-Related Spectrum Perturbation EEG Time-Frequency Analysis (ERSP): Sound stimuli modulated at a frequency of 40 Hz induce synchronous oscillations in neurons of the corresponding auditory cortex. ERSP was used to explore the effects of auditory stimuli from different tasks on auditory cortex activation. Multiple trials (n trials) of ERSP were defined as follows:

[0058]

[0059] Where n represents the total number of trials, F k (f,t) 2 This represents the energy spectrum estimate of the k-th trial at frequency f and time t. ERSP calculation with baseline is used, selecting a data segment from 200 ms before the stimulus (standard stimulus and deviated stimulus) to 1000 ms after the stimulus, and using the mean spectral value 200 ms before the stimulus as the baseline to calculate ERSP values ​​in the frequency range of 20-50 Hz.

[0060] Phase-locked value (PLV): The PLV is represented by a non-random distribution of phase or phase difference. It can be used to analyze nonlinear and non-stationary systems. The brain can be considered a nonlinear dynamic system, and PLV can be used to analyze the synchronicity of neural activity. PLV is defined as the average vector length calculated based on the phase angle difference.

[0061]

[0062] Where n represents the number of trials, e represents the natural logarithm, i represents the imaginary unit, and p represents the radian phase difference for each trial t. The value of PLV ranges from 0 to 1, representing the phase difference between two signals. When PLV is 0, it means that the phase difference is evenly distributed over time on the unit circle of the complex plane, indicating no synchronization. The closer PLV is to 1, the stronger the synchronization between the two signals. When PLV is 1, it means that the two signals are completely synchronized.

[0063] Signal-to-noise ratio (SNR): SNR is a commonly used metric for evaluating brain-computer interface systems. SNR is determined by the ratio of the target frequency amplitude to the noise level. The noise level is defined as the average of the n adjacent amplitudes on either side of the target frequency in the frequency spectrum.

[0064]

[0065] Where y(f) represents the amplitude of the target frequency, Δf represents the frequency resolution, and n represents the number of points near the signal. In this embodiment of the invention, the target frequency is 40Hz, the frequency resolution is 1.25, and n is 5. That is, in this embodiment of the invention, the signal-to-noise ratio is determined by the ratio of the amplitude at 40Hz to the average value of the five adjacent amplitudes on both sides of 40Hz.

[0066] Relative Power (RP): Relative power can be used to assess the degree of ASSR induction. Studies have found a significant positive correlation between relative power values ​​and CRS-R scores. The formula for calculating RP for a specific frequency band is:

[0067]

[0068] Among them, P relative P(f) represents the relative power value of a certain frequency band, P(f) represents the signal power spectrum of the entire frequency band, and [f1,f2] represents the upper and lower limits of a certain frequency band. L ,f H [] represents the upper and lower limits of the entire frequency band. Relative power is a frequency domain characteristic. To calculate the RP at 40Hz, the frequency band value range is 40Hz. To reduce the interference of low-frequency ERP and power frequency interference of equipment on the RP, the full frequency band value range is 35-45Hz. That is, the RP in this embodiment of the invention is calculated by dividing the power at 40Hz by the total power in the frequency range near 40Hz (35-45Hz).

[0069] III. Data Preprocessing:

[0070] Includes the following steps:

[0071] 1) Retain data from the first hospitalization for brain injury for the same patient;

[0072] 2) Delete patient name, examination date, patient ID, and other information, and only save cerebral blood flow data;

[0073] 3) Age data should be formatted as integers.

[0074] 4) For the digital processing of gender, "1" represents male and "0" represents female;

[0075] 5) For patients' GCS scores, "-1" indicates moderate brain injury, "0" indicates severe injury, and "1" indicates extremely severe injury;

[0076] 6) Patients with poor prognosis based on the GOS scale score are labeled as "1", and patients with good prognosis are labeled as "0".

[0077] 7) Patients with a poor prognosis rate due to brain trauma were labeled as 1, and patients without a poor prognosis rate due to brain trauma were labeled as 0;

[0078] 8) Missing data points are filled using the KNN (K Nearest Neighbors) algorithm;

[0079] 9) For handling outliers, the upper and lower quartiles are used first for determination (the normal range is between [Q2-1.5×Q3] and [Q1+1.5×Q2], where Q1 is the upper quartile, Q2 is the lower quartile, and Q3 is the median difference Q3=Q1-Q2). The points determined to be outliers are replaced by the median corresponding to the index.

[0080] 10) Normalize the data, i.e. (original data - minimum value) / (maximum value - minimum value).

[0081] IV. Model Building:

[0082] This model employs a Support Vector Machine (SVM) machine learning algorithm to construct a predictive model. A retrospective study of 300 patients with total brain injury (TBI) was conducted, with 210 TBI cases used as the training set and 90 TBI cases as the test set. Cerebral hemodynamic parameters of the cerebral cerebrovascular autologous (MCA), EEG neurophysiological characteristics, and basic patient information (age, sex, GCS score) were fused. Based on the GOS score at discharge, patients were categorized into those with good and poor prognoses. The effectiveness of this invention has been fully demonstrated through the training process on the training set and the validation process on the test set.

[0083] The SVM classification model parameters are set as follows: the penalty coefficient for incorrect terms is 200, the kernel type is Gaussian kernel, the kernel coefficient is 5, whether to enable probability estimation is False, whether to use heuristic shrinkage is True, the error precision at which the model stops training is 0.001, the memory required for training is 400, the weights of different penalty parameters for each class are None, the maximum number of iterations is -1, the cross-validation parameter is 5, and the model evaluation criteria are ROC and AUC (using AUC for model evaluation).

[0084] This invention presents a hybrid auditory evoked response paradigm combining auditory ERP and ASSR, innovatively integrating non-invasive EEG features, clinically relevant MCA hemodynamic parameters, and basic patient physiological information to predict the prognosis of patients with acute brain injury. This invention has enormous potential and far-reaching impact in acute brain injury and emergency medicine. Further research could lead to a more comprehensive human brain injury prediction system, potentially yielding substantial social and economic benefits.

[0085] Unless otherwise specified, the model numbers of the various devices in this embodiment of the invention are not limited, and any device that can perform the above functions is acceptable.

[0086] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0087] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A system for detecting the prognosis of acute brain trauma, characterized in that, The system comprises a multi-source data synchronous acquisition and extraction part, a data preprocessing part, a machine learning prediction model and an acute traumatic brain injury result display. The 19-lead EEG data and the left and right middle cerebral artery data distributed throughout the brain are collected and extracted synchronously under the mixed auditory evoked paradigm task, and the clinical data, EEG, MCA and classification label GOS of the brain trauma patients are preprocessed. The preprocessed data is input into the machine learning prediction model, and the prediction model outputs the probability of poor prognosis of the TBI patient after calculation. The TBI patient poor rate prediction result displays the probability of poor prognosis of the TBI patient. The mixed auditory evoked paradigm is a mixed auditory evoked paradigm in which ASSR is integrated into the auditory P300 system. The sound stimulus carrier frequency is 500 Hz and 2000 Hz, the modulation frequency is 40 Hz, the auditory stimulus is subjected to amplitude modulation AM with a modulation depth of 100%, and the standard stimulus and the deviation stimulus are respectively standard stimulus and deviation stimulus, in each trial, the standard stimulus and the deviation stimulus appear randomly in a ratio of 4:1, the standard stimulus appears four times, and the deviation stimulus appears once, thereby inducing auditory P300; the EEG and left and right MCA data under the mixed auditory evoked paradigm are synchronously collected and offline calculated to extract the brain electrical feature parameters of the user. The acute traumatic brain injury data is collected and extracted as follows: the patient is admitted and the multi-physiological data is extracted, the inclusion criteria are as follows: age ≥ 18 years old; the patient's Glasgow coma score is ≤ 12; the exclusion criteria are as follows: open craniocerebral trauma; severe multiple injuries; patients transferred from other hospitals for rehabilitation; patients who died on admission; patients with brain trauma before this disease; loss to follow-up, incomplete data; liver and kidney failure; hemorrhagic disease; coagulation dysfunction; severe infection; malignant tumor; Data collection: ERP and ASSR brain electrical feature parameters, MCA hemodynamic parameters and patient basic information at 1 day, 1 week and 2 weeks after hospitalization, brain blood flow parameters and EEG data are synchronously collected and extracted; After the data collection and extraction, the data is preprocessed: data cleaning and normalization processing; The machine learning prediction model is constructed by using a support vector machine machine learning algorithm, and a total of 300 TBI patients are retrospectively collected, 210 TBI patients are used as the training set, and 90 TBI patients are used as the test set; The preprocessed data is input into the prediction model, and the SVM classification model parameters are set as follows: the penalty coefficient of the error term is 200, the kernel function type is Gaussian kernel, the kernel function coefficient is 5, whether to enable probability estimation is False, whether to use heuristic shrinkage is True, the error precision of the model stopping training is 0.001, the specified memory required for training is 400, the different penalty parameter weights for each class are set as None, the maximum number of iterations is -1, the cross-validation parameter is 5, and the model evaluation standard is ROC, AUC.

Images