A method for identifying patients with conversion to mania in the case of small samples

Abstract

The present invention relates to an identification method, and particularly to an identification method for solving the problem of identifying patients with mania conversion under small samples. It assists clinicians in early differential diagnosis of two diseases, prompts doctors which MDD patients have potential risk of mania conversion, and helps doctors formulate precise diagnosis and treatment plans. It includes the following steps: Step 1, data preprocessing: preprocess the functional magnetic resonance data of brain images; Step 2, feature extraction: based on the preprocessed data, extract functional connectivity features FC based on bilateral amygdala, independent component analysis features, Fractional-ALFF and ReHo features. Step 3, feature selection and model construction, including: Step 3.1, dataset planning; Step 3.2, calculate feature importance; Step 3.3, perform feature selection based on the ADC curve of the model; Step 3.4, verify the results in the test set.

Description

Technical Field

[0001] The present invention relates to an identification method, and in particular to an identification method for solving the problem of identifying patients with potential mania in a small sample size. Background Art

[0002] Bipolar disorder (BD) and major depressive disorder (MDD) are two highly disabling and lethal mental illnesses. Since the two diseases are mainly differentiated by subjective symptomatology and have similar depressive symptoms, the differentiation between the two diseases has always been an urgent problem to be solved in clinical practice. Early diagnosis of the two diseases will affect the choice, outcome and prognosis of treatment plans. Therefore, it is particularly important to differentiate the two diseases early to achieve accurate judgment.

[0003] Patients with potential mania (tBD) are a group of patients who are misdiagnosed as MDD in the early stage, but after long-term follow-up or observation, manic or hypomanic symptoms are found. Such patients often have unreasonable medication conditions, which in turn lead to adverse consequences such as repeated hospitalization and increased suicide risk, causing a heavy burden on patients, families and society. In addition, such patients may have the characteristics of early quality and independence from the disease state of BD, which will help to identify and even predict BD early to achieve early differentiation from MDD.

[0004] In view of the urgent need for individualized and precise diagnosis and treatment of current mood disorders, existing methods classify and study BD and MDD, but there are mainly two limitations:

[0005] First, existing methods have not discovered the quality neuroimaging characteristics shared by patients with potential mania and BD patients, and this characteristic is the key to early identification and prediction of patients with potential mania.

[0006] Second, the establishment and testing of existing models are mostly carried out within the dataset, ignoring the generalization ability of the models for external data, especially for the generalization ability of small-sample patients with potential mania with unclear judgment. Summary of the Invention

[0007] The present invention aims at the defects existing in the prior art and provides an identification method for solving the problem of identifying patients with potential mania in a small sample size. It constructs a model based on the data of patients with clear diagnosis, hoping to establish the most discriminative model and use this model to identify misdiagnosed patients with potential mania, assist clinicians in early differentiation of the two diseases, prompt doctors which MDD patients have potential risk of mania, and assist doctors in formulating precise diagnosis and treatment plans.

[0008] To achieve the above object, the present invention adopts the following technical solutions, including the following steps:

[0009] Step 1, data preprocessing: preprocess the functional magnetic resonance data of brain images.

[0010] Step 2, Feature Extraction: Based on the preprocessed data, functional connectivity features FC, independent component analysis features, Fractional-ALFF, and ReHo features based on bilateral amygdala were extracted.

[0011] Step 3, Feature Selection and Model Construction, including:

[0012] Step 3.1, Dataset Planning.

[0013] Step 3.2, Calculate Feature Importance.

[0014] Step 3.3, Feature Selection Based on the ADC Curve of the Model.

[0015] Step 3.4, Result Verification in the Test Set.

[0016] Furthermore, the preprocessing of the functional magnetic resonance data of the brain images includes:

[0017] Step 1.1, Convert the DICOM format data to NIFTI format.

[0018] Step 1.2, Delete the first 10 time points and retain the data of 190 time points.

[0019] Step 1.3, Perform inter-slice time difference correction, head motion correction, and spatial normalization.

[0020] Step 1.4, Smooth the image using a Gaussian kernel of 6mm×6mm×6mm.

[0021] Step 1.5, Perform detrending and filtering.

[0022] Furthermore, the dataset division includes: Since the number of MDD in the dataset is large, it is divided into a training set, a validation set, and a test set according to a ratio of 6:2:2; BD data is used as the training set; and tBD is divided into a validation set and a test set according to a ratio of 1:1.

[0023] Furthermore, the calculation of feature importance includes: Using data preprocessing and feature extraction methods for the training set data, and applying random forest to calculate the feature importance score;

[0024] That is, the parameter setting of the random forest is performed using the grid search method:

[0025] First, pass in other parameters except for the parameters that need to determine the best (other parameters include: the default value of oob_score is False; whether to use out-of-bag samples to estimate the generalization accuracy. criterion measures the splitting criterion. n_jobs: the number of parallel jobs in the program. random_state: random number setting.)

[0026] Then, input the values of the parameters to be optimized in the form of a list or dictionary, set the model evaluation criterion scoring to 'roc_auc'; finally, set refit to True, and the model uses the optimal parameters obtained from the cross-validation training set to fit the entire dataset again with the optimal parameter results; after obtaining the model, extract the feature importance scores in the model.

[0027] Furthermore, the feature selection based on the ADC curve of the model includes: using an SVM classifier, selecting the top N features with high importance scores according to the grid search method, where the range of N is 100 - 500; then testing on the tBD and MDD validation sets, plotting the ADC curve, and selecting the N value corresponding to the optimal ADC curve as the key feature.

[0028] Furthermore, the result verification in the test set includes: applying the SVM classification model corresponding to the selected N value to the test sets of tBD and MDD for result verification.

[0029] The beneficial effects of the present invention compared with the prior art.

[0030] The present invention can give the probability that a patient with depressive disorder may turn into bipolar disorder (mania conversion) at the early stage of the disease, and the accuracy rate can reach 80%. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The following further describes the present invention in conjunction with the drawings and specific embodiments. The protection scope of the present invention is not limited only to the description of the following content.

[0032] Figure 1 It is a schematic diagram of FC feature visualization.

[0033] Figure 2 It is a schematic diagram of ALFF feature visualization.

[0034] Figure 3 It is a schematic diagram of REHO feature visualization.

[0035] Figure 4 It is a ROC curve diagram of classification.

[0036] Figure 5 It is a flowchart of the method for identifying mania-converted patients under small samples. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0037] Such as Figures 1-5As shown below, specific embodiments: In the case where there are few patients with switching mania and it is impossible to construct an accurate recognition model, based on the assumption that there are common qualitative neuroimaging features between BD patients and patients with switching mania, the data of MDD and BD that have been clearly diagnosed after follow-up are used to construct a model and applied to identify a small number of patients with switching mania in order to achieve early recognition and prediction of patients with switching mania. This method mainly includes four modules: data preprocessing, feature extraction, feature selection, and model construction.

[0038] 1. [Data Preprocessing] The brain images of the present invention use functional magnetic resonance data, and the preprocessing uses the DPARSFA software package. The specific processing steps are as follows: Convert the DICOM format data into NIFTI format; Delete the first 10 time points and retain the data of 190 time points; Inter-slice time difference correction; Head motion correction; Spatial normalization; Smooth the image using a Gaussian kernel of 6mm×6mm×6mm; Remove linear drift; Filtering process.

[0039] 2. [Feature Extraction] Based on the processed data, the present invention extracts functional connectivity features FC based on bilateral amygdala, independent component analysis (ICA) features, Fractional-ALFF and ReHo features.

[0040] 3. [Feature Selection and Model Construction] This step is divided into dataset planning; Calculating feature importance; Based on the ADC curve of the model, perform feature selection; Test in patients with switching mania. The specific steps are as follows:

[0041] 1) Dataset division: Since the number of MDD in the dataset is large, it is used as the training set, validation set, and test set in a ratio of 6:2:2. The BD data is used as the training set. And tBD is divided into a validation set and a test set at a ratio of 1:1.

[0042] 2) Calculate feature importance: For the training set data, use the above data preprocessing and feature extraction methods, and apply random forest to calculate the feature importance score. Specifically, use the grid search method to set the parameters of the random forest: First, pass in other parameters except the parameters that need to determine the best. Then input the values of the parameters to be optimized in the form of a list or dictionary, and set the model evaluation criterion scoring to 'roc_auc'. Finally, set refit to True, and the model uses the best parameters obtained from cross-validating the training set, and then fits the entire dataset again with the best parameter results. After obtaining the model, extract the feature importance scores in the model.

[0043] 3) Feature selection based on the ADC curve of the model: The present invention uses an SVM classifier to select the top N features with high importance scores according to the grid search method, where the range of N is 100 to 500. Then, it is tested on the validation set (tBD and MDD), the ADC curve is plotted, and the N value corresponding to the optimal ADC curve is selected, which is the key feature.

[0044] 4) Finally, the SVM classification model corresponding to the selected N value is applied to the test sets of tBD and MDD for result verification.

[0045] It can be understood that the above specific description of the present invention is only for explaining the present invention and is not limited to the technical solutions described in the embodiments of the present invention. Those of ordinary skill in the art should understand that the present invention can still be modified or equivalently replaced to achieve the same technical effects; as long as the usage requirements are met, they are all within the protection scope of the present invention.

Claims

1. A method for identifying patients with mania conversion under small samples, characterized in that: It includes the following steps: Step 1, data preprocessing: Preprocess the functional magnetic resonance data of brain images; Step 2, feature extraction: Based on the preprocessed data, extract functional connectivity features FC, independent component analysis features, Fractional-ALFF, and ReHo features based on the bilateral amygdala; Step 3, feature selection and model construction, including: Step 3.1, dataset planning; Step 3.2, calculate feature importance; Step 3.3, perform feature selection based on the ADC curve of the model; Step 3.4, verify the results in the test set; The division of the dataset includes: Since the number of MDD in the dataset is large, it is used as the training set, validation set, and test set at a ratio of 6:2:2; BD data is used as the training set; and tBD is divided into the validation set and test set at a ratio of 1:1; The feature selection based on the ADC curve of the model includes: Using an SVM classifier, select the top N features with high importance scores according to the grid search method, where the range of N is 100-500; then test on the tBD and MDD validation sets, draw the ADC curve, and select the N value corresponding to the optimal ADC curve, which is the key feature.

2. The identification method for patients with mania conversion under small samples according to claim 1, wherein: The preprocessing of the functional magnetic resonance data of brain images includes: Step 1.1, convert the DICOM format data to NIFTI format; Step 1.2, delete the first 10 time points and retain the data of 190 time points; Step 1.3, perform interlayer time difference correction, head motion correction, and spatial normalization; Step 1.4, smooth the image using a Gaussian kernel of 6 mm×6 mm×6 mm; Step 1.5, perform detrending and filtering processing.

3. A method for identifying patients with mania conversion in small samples according to claim 1, characterized in that: The calculation of feature importance includes: Using data preprocessing and feature extraction methods for the training set data, and applying random forest to calculate the feature importance score; That is, use the grid search method to set the parameters of the random forest: First, input other parameters except the parameters that need to determine the best; Then input the values of the parameters to be optimized in the form of a list or dictionary, set the model evaluation criterion scoring to 'roc_auc'; finally, set refit to True, the model uses the best parameters obtained from cross-validating the training set, and then fits the entire dataset again with the best parameter results; after obtaining the model, extract the feature importance scores in the model.

4. A method for identifying patients with conversion to mania in a small sample according to claim 1, characterized in that: The verification of the results in the test set includes: Applying the SVM classification model corresponding to the selected N value to the tBD and MDD test sets for result verification.

Images