CN114246551A - System for diagnosing schizophrenia and application thereof - Google Patents

Abstract

The invention belongs to the field of biomedicine, and particularly relates to a system for diagnosing schizophrenia and application thereof. In particular, a computing device for diagnosing schizophrenia based on neurocognitive and/or electrophysiological characteristics is included in the system. The neurocognitive features are measured by RBANS or Stroop testing methods, and the electrophysiological features are analyzed after electroencephalogram detection.

Description

System for diagnosing schizophrenia and application thereof

Technical Field

The invention belongs to the field of biomedicine, and particularly relates to a system for diagnosing schizophrenia and application thereof.

Background

Schizophrenia is one of the most serious mental disorders, and 2000 million people are affected worldwide. Schizophrenia usually develops slowly or subacute in young and old years, and is clinically manifested as syndromes with different symptoms, involving various disorders such as sensory perception, thinking, emotion and behavior, and uncoordinated mental activities. The patient is generally aware of the clear and normal intelligence, but some patients suffer from impairment of cognitive function during the course of the disease. Studies have shown that cognitive deficits are one of the core features of significant neurological dysfunction associated with schizophrenia and are often associated with poor prognosis. Furthermore, prepulse inhibition (PPI), which is based on electrophysiological measurements and exhibits moderate effect magnitude, is considered as an index reflecting the information processing deficiency of schizophrenic patients. In addition, electroencephalography (EEG) is a non-invasive electrophysiological measurement method that is widely used to assess the neural response of the brain to external stimuli. The EEG power spectrum describes the power distribution of each frequency band, typically used for schizophrenia studies. More advanced electroencephalographic analysis methods have appeared in recent years, such as DFA (discrete Electrical analysis) method, LRTC (long-range temporal correlation) method.

The current diagnostic golden standard for schizophrenia is established on the basis of International Classification of diseases, 11 th edition (ICD-11) or "diagnostic and statistical Manual for schizophrenia", 5 th edition (DSM-5). These diagnostic methods rely on descriptive psychopathology, to some extent influenced by the subjective judgment of the neuropsychiatric physician. Therefore, objectively measuring and diagnosing schizophrenia is an urgent need of clinicians. However, due to the heterogeneity of etiology and clinical variation, we still lack markers for the diagnosis of schizophrenia.

Disclosure of Invention

In order to provide a schizophrenia diagnosis system more suitable for clinical use, the invention develops a diagnosis system including neurocognitive features and/or electrophysiological features based on a logistic algorithm, a random forest algorithm, and an XGBoost algorithm.

Feature combination

In one aspect, the present invention provides a combination of features for diagnosing schizophrenia, the combination of features comprising at least two features belonging to neurocognitive features or electrophysiological features;

the neurocognitive features are selected from the group consisting of IMM (immediate memory), LAN (language function), ATT (attention), DEM (delayed memory), INT-C (color interference time), INT-W (word interference time).

The electrophysiological characteristics are selected from PPI, Abs-T (theta frequency band absolute power, absolute theta power), Abs-A (alpha frequency band relative power), Abs-AFp/AO (frontal pillow area alpha frequency band absolute power ratio, absolute power AFp/AO ratio), Abs- (D + T)/(A + B) (absolute power delta + theta/alpha + beta), Rel-D (delta relative power, relative delta power), Rel-T (theta relative power, relative theta power), Rel-A/B (relative power alpha/beta), DFA-A (alpha frequency band DFA), DFA-B (beta frequency band DFA).

Preferably, the neurocognitive characteristics are measured by the RBANS or Stroop test method.

Preferably, the electrophysiological characteristics are obtained by post-electroencephalographic detection analysis.

Preferably, the PPI includes PSC-PPI (perceptually spatially fused PPI), PSS-PPI (perceptually spatially separated PPI).

Preferably, the combination of features is selected from any one of the following:

1)IMM、LAN、ATT、DEM、INT-C、INT-W；

2)PSC-PPI、PSS-PPI、Abs-T、Abs-A、Abs-AFp/AO、Abs-(D+T)/(A+B)、Rel-D、Rel-T、Rel-A/B、DFA-A、DFA-B；

3)1) and 2).

Preferably, said IMM, LAN, ATT, DEM are measured by RBANS.

Preferably, the INT-C, INT-W is measured by the Stroop test.

Preferably, the PSC-PPI, PSS-PPI are obtained by the startle reflex test.

Preferably, the Abs-T, Abs-A, Abs-AFp/AO, Abs- (D + T)/(A + B), Rel-D, Rel-T, Rel-A/B, DFA-A, DFA-B are obtained by electroencephalogram testing.

System for controlling a power supply

In another aspect, the present invention provides a system for diagnosing schizophrenia, the system comprising a computing means for diagnosing whether a subject has schizophrenia based on a detection result of any one or any combination of diagnostic features;

the diagnostic features include IMM, LAN, ATT, DEM, INT-W, PPI, DFA-A, DFA-B;

the diagnostic combination of features is selected from NSF, ESF, ASF;

the NSF comprises IMM, LAN, ATT, DEM, INT-C, INT-W;

the ESF comprises PSC-PPI, PSS-PPI, Abs-T, Abs-A, Abs-AFp/AO, Abs- (D + T)/(A + B), Rel-D, Rel-T, Rel-A/B, DFA-A, DFA-B;

the ASF is a collection of NSFs and ESFs.

Preferably, the PPI comprises PSC-PPI, PSS-PPI.

The Cohen's d values of the above diagnostic features are all about 0.5, and the AUC values are all more than 0.7, so that the diagnostic features alone can have better diagnostic effect.

The term "NSF" as used herein is an abbreviation for neurocognitive selected pests, also known as neurocognitive feature subsets.

As used herein, "ESF" is an abbreviation for electrophysiologically selected pests, also known as a subset of electrophysiological characteristics.

The "ASF" is an abbreviation for All selected feeds, i.e., the collection of NSF and ESF, also referred to as the full feature set.

Preferably, the system further comprises a neural feature detection device and/or an electrophysiological detection device:

the neural feature detection device includes:

1) a first neural feature detection device for operating a RBANS;

2) a second neural feature detection device for running a Stroop test;

the electrophysiological detection device includes:

1) a first electrophysiological detection device for operating a detection of the human startle reflex system;

2) a second electrophysiological detection device for operating an electroencephalogram system;

preferably, the system further comprises an information collecting means for inputting subject information;

preferably, the subject information includes detection results, demographic information, and clinical profile information for any one or any combination of diagnostic features.

Preferably, the diagnosis includes suffering from schizophrenia, not suffering from schizophrenia.

Preferably, the system further comprises a modeling device which is modeled by using any one method of a Logitics algorithm, a random forest algorithm and an XGboost algorithm.

The Logitics algorithm, the random forest algorithm and the XGboost algorithm are used for constructing a model by respectively constructing the neurocognitive feature subset, the electrophysiological feature subset or the full feature set through related algorithms.

Preferably, the system further comprises result transmission means for transmitting the diagnosis result, the result transmission means may transmit the diagnosis result to an information communication terminal device to which the patient or the medical staff can refer.

The algorithm comprises a Logitics algorithm, a random forest algorithm and an XGboost algorithm.

Device

In another aspect, the present invention provides an apparatus for diagnosing schizophrenia, the apparatus comprising: a memory and a processor;

the memory is to store program instructions;

the processor is configured to invoke program instructions that, when executed, perform modeling and/or diagnostic operations comprising:

preferably, the operation step of establishing the model is to process the diagnosis feature combination by using any one of a Logitics algorithm, a random forest algorithm and an XGboost algorithm;

preferably, the diagnostic operating steps are:

1) obtaining a test result from any one diagnostic feature or any combination of diagnostic features of the subject;

2) calculating whether the subject suffers from schizophrenia based on any one or any combination of the diagnostic features;

3) and displaying the diagnosis result.

The diagnostic features include IMM, LAN, ATT, DEM, INT-W.

The diagnostic feature combinations include NSF, ESF, ASF;

the NSF comprises IMM, LAN, ATT, DEM, INT-C, INT-W;

the ESF comprises PSC-PPI, PSS-PPI, Abs-T, Abs-A, Abs-AFp/AO, Abs- (D + T)/(A + B), Rel-D, Rel-T, Rel-A/B, DFA-A, DFA-B;

the ASF is a collection of NSFs and ESFs.

Computer readable storage medium

In another aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a method of modeling and/or a diagnostic method:

the method for establishing the model comprises the steps of processing a diagnosis feature combination by using any one of a Logitics algorithm, a random forest algorithm and an XGboost algorithm;

the diagnostic method comprises the following steps:

1) obtaining a test result from any one diagnostic feature or any combination of diagnostic features of the subject;

2) calculating whether the subject suffers from schizophrenia or not according to the detection result of any one diagnosis characteristic or any one diagnosis characteristic combination;

3) and displaying the diagnosis result.

Applications of

In another aspect, the invention provides the use of IMM, LAN, ATT, DEM, INT-W, PSC-PPI, PSS-PPI, DFA-A, DFA-B in the manufacture of a product for diagnosing schizophrenia.

In another aspect, the invention provides the use of the aforementioned combination of features, system, device, computer readable storage medium for the manufacture of a product for the diagnosis of schizophrenia.

Method

In another aspect, the present invention provides a method of establishing a model for diagnosing schizophrenia, the method comprising processing a combination of diagnostic features using any one of a logistic algorithm, a random forest algorithm, and an XGBoost algorithm;

the diagnostic feature set is selected from the group consisting of a neurocognitive feature subset, an electrophysiological feature subset, and a full feature set.

Meanwhile, the invention provides the application of the model established by the method in preparing products for diagnosing schizophrenia.

In another aspect, the present invention provides a method of diagnosing schizophrenia, the method comprising:

1) obtaining a test result from any one or a combination of the aforementioned diagnostic features of the subject;

2) and calculating whether the subject suffers from schizophrenia or not according to the detection result of any one diagnosis characteristic or any one diagnosis characteristic combination.

Drawings

FIG. 1 is a diagnostic ROC curve for a model constructed using Logitics algorithm for NSF, ESF, ASF, with sensitivity on the ordinate and specificity on the abscissa.

FIG. 2 is a diagnostic ROC curve for a model constructed using a random forest algorithm for NSF, ESF, ASF, with sensitivity on the ordinate and specificity on the abscissa.

FIG. 3 is a diagnostic ROC curve for models constructed by NSF, ESF, ASF using the XGboost algorithm, with sensitivity on the ordinate and specificity on the abscissa.

Fig. 4 is a diagnostic scatter plot of a model constructed using the logistic algorithm, with NSF fit values as horizontal coordinates and ESF fit values as vertical coordinates.

FIG. 5 is a diagnostic scatter plot of a model constructed using a random forest algorithm with NSF fit values as horizontal coordinates and ESF fit values as vertical coordinates.

FIG. 6 is a diagnostic scatter plot of a model constructed using the XGboost algorithm with NSF fit values as horizontal coordinates and ESF fit values as vertical coordinates.

Detailed Description

The present invention will be further described with reference to the following examples, which are intended to be illustrative only and not to be limiting of the invention in any way, and any person skilled in the art can modify the present invention by applying the teachings disclosed above and applying them to equivalent embodiments with equivalent modifications. Any simple modification or equivalent changes made to the following embodiments according to the technical essence of the present invention, without departing from the technical spirit of the present invention, fall within the scope of the present invention.

Example 1 influence factor screening and validation

1. Test object

A total of 69 schizophrenic patients and 50 healthy controls were enrolled in the study. The investigators confirmed the diagnosis by the american statistical manual for mental disorder diagnosis fourth edition (DSM-IV) Structured Clinical Interview (SCID) tool. All subjects were right-handed and hearing assessment (pure tone audiometry, 1000Hz) was normal. Inclusion criteria for schizophrenic patients enrollment were as follows:

1) all clinically stable subjects had no history of neurological disease or head trauma,

2) there was no history of shock treatment over the past six months,

3) no alcohol/drug dependence or history of abuse (except tobacco).

Patients were excluded due to unstable conditions or wisdom below 70.

During the study, all patients received antipsychotic medication as usual. Healthy control group (CON) was matched to schizophrenia group (SCZ) gender, age, educational age, and smoking history. Exclusion criteria for the healthy control group included drug abuse, risk of suicide, major head trauma, and neuropsychiatric disorders.

Each subject received a detailed description of the purpose and procedure for participation in the study prior to signing an informed consent. The independent ethics committee of the beijing stability hospital approved the study. Patients were evaluated psychopathologically using the positive and negative symptoms scale (PANSS). The detailed demographic and clinical characteristics are shown in table 1.

TABLE 1 demographic and clinical characteristics of healthy control and schizophrenia groups

Note: a represents the P value of chi-square test; b represents the P value of the independent sample t test.

2. Neurocognitive assessment

We assessed neurocognitive function in subjects using a reproducible set of neuropsychological state measurements (RBANS). RBANS evaluated five cognitive domains: IMM, DEM, VC, ATT, LAN. In addition to RBANS, Stroop testing was also performed. Each subject was asked to complete two perturbation tasks and the duration under color perturbation (INT-C) and duration under word sense perturbation (INT-W) were recorded.

In this part of the experiment, a total of 5 RBANS features (IMM, DEM, VC, ATT, LAN) and 2 Stroop features (INT-C, INT-W) were extracted.

3. Electrophysiological assessment

3.1 startle reflex measurement (prepulse inhibition measurement)

The subject sits comfortably on the couch with the arms fully relaxed in a natural position. The auditory startle reflex was measured by recording the electromyographic signals of the orbicularis oculi muscle using a human startle reflex custom system (manufactured by hoechuan hong Yuan company, Beijing). The electrode impedance is kept at <5k omega. In addition, electromyography was band-pass filtered to 100 and 1000Hz and amplified 10,000 times. Auditory startle stimuli are presented binaural through two headphones. Sound pressure correction was performed using a sound pressure corrector (Larson Davis, AUDit and System 824).

The experimental environment is as follows: the method is carried out in an auditory sound insulation shielding room, and certain brightness and temperature are kept.

Sound materials: the desired sound material was generated using "randn ()" in the MATLAB function library, with a sampling rate of 48 kHz. The generated gaussian noise is passed through a 512-order low-pass digital filter with a cutoff frequency of 10kHz to obtain broadband noise as background noise. In addition, a pre-stimulus sound having a length of 150ms of broadband noise and a startle sound having a length of 40ms of broadband noise are generated. In order to avoid the energy splash phenomenon, each sound stimulus is filtered after the introduction of the pre-pulse stimulus sound and the startle stimulus sound. The sound signal is input to the Sennheiser HD600 headphones by means of a sound card (Audio CODEC'97) and presented to the subject. Sound pressure correction was performed using a sound pressure corrector (Larson Davis, AUDit and System 824). The following are the individual sound specific parameters, a) background sound: white noise, which is divided into a left sound channel or a right sound channel leading 3ms, duration 15s and sound pressure level 60 dB; b) pre-stimulation: white noise, which is divided into a left sound channel or a right sound channel leading 3ms, duration 150ms and sound pressure level 65 dB; c) frightening and stimulating: white noise, duration 40ms, 100 dB.

The test flow comprises the following steps: order to sit on the examination chair, paste 2 Ag/AgCl electrodes under the pupil of the right eye and at the outer side of the pupil by 1.5cm to record the electrical activity of the orbicularis oculi muscle, and paste the electrode on the right mastoid as the ground wire. The eye is 60cm from the eye tracker display. The muscles of the whole body are relaxed, the patient keeps clear-headed as much as possible in the experiment process, and the patient keeps the head fixed as much as possible while watching the screen with two eyes. Firstly, learning each sound material for 3 times, giving the same guide words to all the tested persons, wherein the left channel or the right channel needs to be distinguished by the background sound and the forward stimulus, carrying out monocular 3-point scale eye movement calibration on the tested persons, ordering the tested persons to be annotated to look at the front cross, and starting the test.

The test paradigm is: the pre-stimulation duration is 150ms, divided into left-leading and right-leading. The front stimulus and background sounds (left-leading and right-leading) exhibited 2 states of perceptual spatial separation (front stimulus left-leading, background sound right-leading; front stimulus right-leading, background sound left-leading), each state being repeated 5 times. The pre-and startle stimuli are separated by 120 ms. The number of trials with only startling stimulation was 7. See figure 2-1 for details.

And (3) statistical indexes are as follows: recording 500ms before stimulation as baseline S0; the startle stimulus alone was designated as S1, while the pro-stimulus was designated as S2; calculating the myoelectricity change percentage 1- (S2-S0)/(S1-S0); myoelectric maximum latency; myoelectric maximum response rate; sustained attention level (e.g., sustained time to stare at the cross).

In this experiment, a total of 2 characteristic perceptually spatially separated PPIs and perceptually spatially fused PPIs (PSC-PPI, PSS-PPI) were extracted.

3.2 electroencephalogram recording and processing

3.2.1 EEG data preprocessing

The subject sits comfortably on the couch, closes his eyes, and remains relaxed and quiet for 5 minutes. Electroencephalographic data was acquired using an electroencephalographic system (EGI, Electrical Geodesics, inc., America) at a sampling rate of 1000Hz, with Cz as a reference. EEGLAB (v2019.1) and FieldTrip kits were used for offline pretreatment of EEG data in MATLAB (MATLAB Release 2017b, MathWorks, Inc.). The EEG raw data is first down-sampled to 500Hz and band-pass filtered to 0.5-45 Hz.. For each subject's electroencephalographic data, artifact removal was performed by independent component analysis (ICA, Algorithm: runica) using EEGLAB. ICA components were classified using the EEGLAB plug-in ICLabel tool. Eye movement, blinking, heartbeat, muscle activity, or other artifacts are isolated by the ICA algorithm and corrupted derivatives are interpolated. The EEG data is then manually examined to verify artifact removal, and bad segments are rejected.

Finally, all electrodes were re-referenced to full brain average.

3.2.2 Power spectral features

The Power Spectral Density (PSD) of each electrode was evaluated using a fast fourier transform (FFT, Welch method, 2 second sliding window, 50% overlap, 0.5Hz frequency step) yielding a range of 0.5-45Hz. The frequency bands are selected as follows: delta (1.0-4.0Hz), theta (4.0-8.09/19 Hz), alpha (8.0-14.0Hz), beta (14.0-30.0 Hz). D. T, A, B denote the delta, theta, alpha and beta bands, respectively. AL and AR were calculated from the average power of the alpha bands of the left hemisphere (Fp1, F3, C3, P3, O1, F7, T3, T5) and the right hemisphere (Fp2, F4, C4, P4, O2, F8, T4, T6). The average of (D + T) L and (D + T) R is the sum of the left and right hemispheres δ and θ. AFp and AO are calculated by averaging the alpha band power in the Fp channel (Fp1, Fp2) and O channel (O1, O2). For each electrode, the absolute power (Abs) and the relative power (Rel) for each frequency band are calculated.

In the PSD, a total of 20 features were extracted: Abs-D, Abs-T, Abs-A, Abs-B, Abs-A/T, Abs-A/B, Abs- (D + T)/(A + B), Abs- (D + T) L/(D + T) R, Abs-AL/AR, Abs-AFp/AO, Rel-D, Rel-T, Rel-A, Rel-B, Rel-A/T, Rel-A/B, Rel- (D + T)/(A + B), Rel- (D + T) L/(D + T) R, Rel-AL/AR, Rel-AFp/AO, see in particular Table 2.

3.2.3 Detrending undulation (DFA) features

DFA is an analytical method based on the scale-free theory for estimating the long-range time correlation (LRTC) in power-law form. That is, if the time series data has a non-random temporal structure of slowly decaying autocorrelations, the DFA can quantify the rate at which these correlations decay, as shown by the DFA power law exponent. Some evidence suggests that DFA reflects brain maturation and may prove to be a potential biomarker for the pathophysiology of neurodevelopmental disorders.

The calculation of DFA used the neurophysiological biomarker kit (NBT: http:// www.nbtwiki.net /). First, all electrodes were filtered with delta, theta, alpha and beta oscillations, respectively. An amplitude envelope is then generated from each frequency band.

Finally, each participant estimates the DFA value of each electrode and stores the DFA value of each frequency band, respectively. DFA-D, DFA-T, DFA-A, DFA-B was calculated by averaging all electrodes in the delta, theta, alpha and beta bands.

3.2.4 Fractal Dimension (FD) features

Brain complexity can be described as the highly structured temporal structure observed in EEG signals in an intermediate case between pure randomness (e.g. white noise) and no variability (constancy or pure periodicity). We used the EEGLAB plug-in, myFractal (https:// github. com/polyamides/myFractal) to calculate FD for each electrode.

Finally, the FD feature is extracted by averaging the FD values of all the electrodes.

4. Statistical analysis

Statistics were performed in RStudio (version 1.2.5033, RStudio, Inc.) using R software (version 3.6.3). Demographic and basic clinical data including gender, age of education, smoking history, course, age of onset, chlorpromazine equivalent dose, as well as PANSS total score, PANSS positive score, PANSS negative score and PANSS general psychopathological score. All demographic and clinical variables, except gender and smoking history, were expressed as mean ± SDs. Independent t-test and chi-square test were used to assess potential differences in demographics, clinical variables, neurocognition and electrophysiology for CON and SCZ. According to the Benjamini-Hochberg method, the False Discovery Rate (FDR) was calculated to adjust the P value of multiple tests. P <0.05 was considered statistically significant.

5. Classification

All analyses were performed using R3.6.3. To select the best features for SCZ and CON classification, we first evaluated the classification ability of each feature. The Cohen's d value for each feature was calculated using a self-compiling function. Receiver operating characteristic curve (ROC) analysis for each feature was created using the prom software package in R. ROC values for each feature include accuracy, sensitivity, specificity, and area under the ROC curve (AUC).

All features are classified into two categories: neurocognitive features and electrophysiological features. The rfe function (R caret software package) was used for feature selection by a multivariate recursive feature elimination method. rfe first fit the model to all features using a bag tree algorithm. The models are ranked according to their importance to each feature. In each iteration of feature selection, the ranked features are retained, the model is reassembled, and performance is evaluated. The final selection of each set of features was based on 10-fold cross-validation.

Two optimal subsets (neurocognitive feature set NSF and electrophysiological feature set ESF) and one combined feature set ASF (containing the aforementioned neurocognitive and electrophysiological feature sets) are then obtained.

After the above three feature sets are determined, a classification model is estimated from the three feature sets using a Logistic algorithm (R statistical data package), a random forest algorithm (R random forest package), and an XGBoost algorithm (R XGBoost package).

Finally, the performance of these models was verified using a 10-fold cross-validation method. The results of the verification are then averaged. Classification performance was assessed by accuracy, sensitivity and specificity. In addition, the performance of each model was also evaluated using the ROC curve.

Experimental result 1: statistical comparison of all features between SCZ and CON

TABLE 2 statistical comparison of all features extracted from neurocognitive and electrophysiological assessments

Note: p < 0.05; p < 0.01; p < 0.001.

The results of the statistical analysis of all features are shown in table 2. Overall, all neurocognitive characteristics of SCZ and CON are statistically different. It can also be observed that all PPI characteristics differ significantly between the two groups. In the characteristics of the electroencephalogram power spectrum, only the Abs-T, Abs-AFp/AO and the Rel-D, Rel-T have the difference of statistical significance. Furthermore, the θ band of CON is DFA (DFA-T) lower than SCZ, while the α and β bands are significantly higher than SCZ (DFA-A, DFA-B). There was no significant difference in FD between SCZ and CON groups. In all functions, the P values of IMM, LAN, ATT, DEM, PSS-PPI and DFA-B were less than 0.001.

Experimental results 2: single feature diagnostic performance

To select the best feature to distinguish between SCZ and CON, we first evaluated the Cohen's d and ROC values for each feature. ROC values include accuracy (%), sensitivity (%), specificity (%), and AUC (%).

As shown in Table 3, Cohen's d values greater than 0.5 for a total of 14 features achieved levels above the moderate effect, including IMM, VC, LAN, ATT, DEM, INT-C, INT-W, PSC-PPI, PSS-PPI, Abs-T, Abs AFp/AO, Rel-T, DFA-A, DFA-B. Wherein Cohen's d for LAN, TT, DEM and PSS-PPI is greater than 0.8, indicating a level of greater efficacy; cohen's d for IMM reached 1.42, showing a very large level of effect. Neurocognitive features exhibit better performance than electrophysiological features in terms of ROC values. IMM is the most neurocognitive characteristic classification method, the accuracy is 84.03%, and the AUC reaches 91.87%.

TABLE 3 Cohen's d and Single feature Classification Performance

Note: one column for Cohen's d values, grey labeled table representing Cohen's d values >0.5, above moderate effect;

AUC values in one column, gray labeled table represents AUC > 70%, reaching the standard of accurate diagnosis.

There were 5 AUC's greater than 70% from neurocognitive features (IMM, LAN, ATT, DEM, INT-W), showing that these features have potential for use in the development of tools to aid in the diagnosis of schizophrenia. PSS-PPI is the best potential electrophysiological feature with an accuracy of 80.67% and an AUC of 84.32%.

AUC of 4 electrophysiological characteristics (PSC-PPI, PSS-PPI, DFA-A, DFA-B) was greater than 70%, indicating that these characteristics have potential for use in the development of tools to aid in the diagnosis of schizophrenia.

Experimental result 3: classification Performance of feature combinations

And selecting the neurocognitive characteristic set and the electrophysiological characteristic set by using a multivariate recursive characteristic elimination method to distinguish the CON and the SCZ to the maximum extent. In addition, model validation was performed using 10-fold cross-validation to prevent overfitting.

The resulting neurocognitive feature subset (NSF subset) comprises IMM, LAN, TT, DEM, INT-C, INT-W. The electrophysiological signature subset (ESF subset) comprises PSC-PPI, PSS-PPI, Abs-T, Abs-A, Abs-AFp/AO, Abs- (D + T)/(A + B), Rel-D, Rel-T, Rel-A/B, DFA-A, DFA-B. The ASF set is a collection of two feature subsets as above.

Then, a classification model is built from the NSF subset, the ESF subset and the ASF set using a logistic algorithm, a random forest algorithm, an XGBoost algorithm. On this basis, the models were evaluated using a 10-fold cross-validation method. Finally, the fit value for the probability for schizophrenic patients is between 0 and 1. By setting the cutoff value to 0.5, the accuracy, sensitivity, specificity were calculated, and then the ROC curve was evaluated from the fitted values and the AUC was calculated.

These model assessment indicators (accuracy, sensitivity, specificity and AUC) are shown in table 4. The ROC curves for each model are shown in fig. 1-3. As can be seen from table 4 and fig. 1-3, the model developed by including all the selected feature sets showed better classification performance regardless of the algorithm used. The NSF subset model and the ESF subset model showed roughly comparable classification accuracy, but within the AUC value range, the NSF subset model showed better performance than the ESF subset.

In addition, among all compared algorithms, the XGBoost algorithm performs more stably and accurately in terms of accuracy and AUC. The ASF set model adopting the XGboost algorithm achieves the optimal diagnosis effect.

To better demonstrate the difference in classification ability among the three algorithms, we created scatter plots using NSF subset fit values as horizontal coordinates and ESF subset fit values as vertical coordinates (fig. 4-6). The fit values for random forest and XGboost algorithms are more densely distributed in the figure than the Logistic algorithm. This also indicates that the random forest and XGBoost algorithms provide better performance in differentiating schizophrenic patients from healthy people.

TABLE 4 Classification Performance of feature combinations

Claims (10)

1. A combination of features for diagnosing schizophrenia, said combination of features comprising at least two features belonging to neurocognitive features or electrophysiological features;

the neurocognitive features are selected from immediate memory, language ability, attention, delayed memory, duration under color interference, duration under word sense interference;

the electrophysiological characteristics are selected from PPI, theta frequency band absolute power, alpha frequency band relative power, a frontal occipital area alpha frequency band absolute power ratio, absolute power delta + theta/alpha + beta, delta relative power, theta relative power, relative power alpha/beta, alpha frequency band DFA and beta frequency band DFA;

preferably, the PPI comprises a perceptually spatially fused PPI, a perceptually spatially separated PPI.

2. The combination of features of claim 1, wherein the combination of features is selected from any one of the group consisting of:

1) instant memory, language ability, attention, delayed memory, duration under color interference, duration under word meaning interference;

2) PPI, theta frequency range absolute power, alpha frequency range relative power, a frontal occipital area alpha frequency range absolute power ratio, absolute power delta + theta/alpha + beta, delta relative power, theta relative power, relative power alpha/beta, alpha frequency range DFA and beta frequency range DFA;

3)1) and 2);

preferably, the PPI comprises a perceptually spatially fused PPI, a perceptually spatially separated PPI.

3. The combination of features according to claim 1 or 2, wherein the immediate memory, language ability, attention, delayed memory are measured by RBANS,

the duration under the color interference and the duration under the word sense interference are measured by a Stroop test method,

the perception space fusion PPI and the perception space separation PPI are obtained by a shock reflex system test,

the theta frequency band absolute power, the alpha frequency band relative power, the alpha frequency band absolute power ratio of the frontal occipital area, the absolute power delta + theta/alpha + beta, the delta relative power, the theta relative power, the relative power alpha/beta, the alpha frequency band DFA and the beta frequency band DFA are obtained through electroencephalogram tests.

4. A system for diagnosing schizophrenia, the system comprising a computing device for diagnosing whether a subject has schizophrenia based on a result of detection of any one or any combination of diagnostic features;

the diagnosis features comprise immediate memory, language ability, attention, delayed memory, duration under word sense interference, PPI, alpha frequency range DFA and beta frequency range DFA;

the diagnostic feature set is selected from a neurocognitive feature subset, an electrophysiological feature subset, or a full feature set;

the neurocognitive feature subset comprises immediate memory, language ability, attention, delayed memory, duration under color interference and duration under word sense interference;

the electrophysiological characteristic subset comprises a perception space fusion PPI, a perception space separation PPI, theta frequency range absolute power, alpha frequency range relative power, a frontal occipital area alpha frequency range absolute power ratio, absolute power delta + theta/alpha + beta, delta relative power, theta relative power, relative power alpha/beta, alpha frequency range DFA and beta frequency range DFA;

the full feature set is a collection of neurocognitive and electrophysiological feature subsets;

preferably, the PPI comprises a perceptually spatially fused PPI, a perceptually spatially separated PPI.

5. The system of claim 4, further comprising a neural feature detection device and/or an electrophysiological detection device:

the neural feature detection device includes at least one of:

1) a first neural feature detection device for operating a RBANS;

2) a second neural feature detection device for running a Stroop test;

the electrophysiological detection device includes at least one of:

1) first electrophysiological detection device for operating a startle reflex system

2) A second electrophysiological detection device for operating an electroencephalogram system.

6. The system according to claim 4 or 5, wherein the system further comprises a modeling device for modeling by using any one of a Logitics algorithm, a random forest algorithm and an XGboost algorithm;

preferably, the system comprises the following means:

1) a modeling device which is modeled by any one of a Logitics algorithm, a random forest algorithm and an XGboost algorithm;

2) a neurological feature detection device and/or an electrophysiological detection device;

3) computing means for diagnosing whether the subject has schizophrenia based on the detection of any one or any combination of the diagnostic features;

4) and a display device for displaying the diagnosis result.

7. A method of creating a model for diagnosing schizophrenia, the method comprising processing a combination of diagnostic features using any one of an algorithm;

the diagnostic feature set is selected from a neurocognitive feature subset, an electrophysiological feature subset, or a full feature set;

the neurocognitive feature subset comprises immediate memory, language ability, attention, delayed memory, duration under color interference and duration under word sense interference;

the electrophysiological characteristic subset comprises a perception space fusion PPI, a perception space separation PPI, theta frequency range absolute power, alpha frequency range relative power, a frontal occipital area alpha frequency range absolute power ratio, absolute power delta + theta/alpha + beta, delta relative power, theta relative power, relative power alpha/beta, alpha frequency range DFA and beta frequency range DFA;

the full feature set is a collection of neurocognitive and electrophysiological feature subsets;

preferably, the algorithm comprises a Logitics algorithm, a random forest algorithm and an XGboost algorithm;

preferably, the algorithm is a random forest algorithm or an XGBoost algorithm.

8. An apparatus for diagnosing schizophrenia, the apparatus comprising: a memory and a processor;

the memory is to store program instructions;

the processor is configured to invoke program instructions that, when executed, perform modeling and/or diagnostic operations comprising:

the operation step of establishing the model is to process a diagnosis feature combination by using any one method of a Logitics algorithm, a random forest algorithm and an XGboost algorithm;

the diagnostic operating steps are:

1) obtaining a test result from any one diagnostic feature or any combination of diagnostic features of the subject;

2) calculating whether the subject suffers from schizophrenia or not according to the detection result;

3) displaying the diagnosis result;

the diagnostic features include immediate memory, language ability, attention, delayed memory, duration under semantic interference, perceptual space fusion PPI, perceptual space separation PPI, ALPHA-band DFA, BETA-band DFA;

the diagnostic feature set is selected from a neurocognitive feature subset, an electrophysiological feature subset, or a full feature set;

the neurocognitive feature subset comprises immediate memory, language ability, attention, delayed memory, duration under color interference and duration under word sense interference;

the electrophysiological characteristic subset comprises a perception space fusion PPI, a perception space separation PPI, theta frequency range absolute power, alpha frequency range relative power, a frontal occipital area alpha frequency range absolute power ratio, absolute power delta + theta/alpha + beta, delta relative power, theta relative power, relative power alpha/beta, alpha frequency range DFA and beta frequency range DFA;

the full feature set is a collection of neurocognitive and electrophysiological feature subsets.

9. A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out a method of modeling and/or a diagnostic method of:

the method for establishing the model comprises the steps of processing a diagnosis feature combination by using any one of a Logitics algorithm, a random forest algorithm and an XGboost algorithm;

the diagnostic method comprises the following steps:

1) obtaining a test result from any one diagnostic feature or any combination of diagnostic features of the subject;

2) calculating whether the subject suffers from schizophrenia or not according to the detection result;

3) displaying the diagnosis result;

the diagnostic feature set is selected from a neurocognitive feature subset, an electrophysiological feature subset, or a full feature set;

the neurocognitive feature subset comprises immediate memory, language ability, attention, delayed memory, duration under color interference and duration under word sense interference;

the electrophysiological characteristic subset comprises a perception space fusion PPI, a perception space separation PPI, theta frequency range absolute power, alpha frequency range relative power, a frontal occipital area alpha frequency range absolute power ratio, absolute power delta + theta/alpha + beta, delta relative power, theta relative power, relative power alpha/beta, alpha frequency range DFA and beta frequency range DFA;

the full feature set is a collection of neurocognitive and electrophysiological feature subsets.

10. Use of immediate memory, language skills, attention, delayed memory, duration under semantic interference, PPI, Α -band DFA, b-band DFA, the combination of features described in claim 1, the system of claim 4, the model constructed by the method of claim 7, the apparatus of claim 8, the computer-readable storage medium of claim 9 in the manufacture of a product for diagnosing schizophrenia;

preferably, the PPI comprises a perceptually spatially fused PPI, a perceptually spatially separated PPI.

Images