Method for establishing depression diagnosis model based on dried blood near infrared spectrum
The diagnostic model of depression is established through near-infrared spectroscopy of dry blood combined with machine learning algorithms, which solves the problem that depression diagnosis depends on clinical symptoms, and achieves rapid and accurate identification of depression, simplifies the operation process and improves diagnostic efficiency.
Patent Information
- Application Number
- CN202411275578.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-12
- Publication Date
- 2024-12-31
AI Technical Summary
The diagnosis and treatment of depression mainly relies on clinical symptoms and lack of objective biomarkers, resulting in insufficient early warning and precise diagnosis and treatment methods.
By collecting characteristic signals based on the near-infrared spectroscopy of dry blood, combining machine learning algorithms to establish a depression diagnostic model, use the screened important features to reduce the spectral feature dimension, improve the calculation rate and model accuracy, and achieve rapid identification of results.
No sample preprocessing is required, the operation is simple and fast, reducing water peak interference, showing more characteristic signals, improving the accuracy and efficiency of diagnosis, and achieving rapid identification of depression.
Smart Images
Figure CN119230040A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of depression diagnosis, and in particular to a method for establishing a depression diagnosis model based on dried blood near-infrared spectroscopy. Background Art
[0002] Depression is a mood disorder characterized by significant and persistent low mood. This low mood is disproportionate to one's situation, ranging from melancholy to grief, or even pessimism and world-weariness. In severe cases, psychotic symptoms such as hallucinations and delusions may occur. Depression is highly heterogeneous and has a complex pathogenesis. Currently, its diagnosis and treatment are mainly based on clinical symptoms, which are subjective and empirical. There is an urgent need to develop early warning and precision diagnosis and treatment methods based on objective biomarkers.
[0003] Near infrared spectroscopy is an important spectral analysis technique, with a wavelength range between visible light (Vis) and mid-infrared spectroscopy (MIR), usually defined as 780nm to 2500nm. This spectral region mainly reflects the double frequency and sum frequency vibrations of hydrogen bonds and organic molecules. Its spectral characteristics are highly correlated with the types and contents of organic molecules, so it is particularly suitable for the analysis of organic compounds and biomolecules.
[0004] By obtaining the near-infrared spectral characteristics of blood samples from patients with depression and combining them with machine learning algorithms, a near-infrared spectral feature rapid identification model is established, which is expected to achieve rapid identification of near-infrared spectral spectrum results of patients with depression. Summary of the invention
[0005] The invention provided by the present invention aims to provide a method for establishing a depression diagnosis model based on the near-infrared spectrum of dried blood. The method for establishing a depression diagnosis model based on the near-infrared spectrum of dried blood according to the present invention does not require sample pretreatment, and the spectrum recognition result can be completed only through a trace amount of blood, which is simple and quick to operate. Compared with the near-infrared spectrum of liquid blood collected directly, the collected near-infrared spectrum of dried blood has no strong interference from the water peak and shows more and stronger characteristic signals. The modeling of the selected important features greatly reduces the feature dimension of the near-infrared spectrum, and increases the calculation rate and model accuracy.
[0006] In order to achieve the above effects, the present invention provides the following technical solution: a method for establishing a depression diagnosis model based on dried blood near-infrared spectroscopy, comprising the following steps:
[0007] S1. Collect and store a small amount of blood samples from several clinically depressed patients and healthy controls;
[0008] S2, dropping the blood sample obtained in step S1 onto a transparent quartz glass slide to obtain a dried blood sample, and collecting the near-infrared spectrum of the dried blood to form a spectrum set;
[0009] S3, preprocessing the spectral data of the near-infrared spectrum set obtained in step S2 according to a certain method, and establishing multiple basic models through a machine learning algorithm;
[0010] S4, performing SHAP analysis on the multiple basic models established in step S3, screening and merging important features, establishing models through the machine learning algorithm used in step S3, and comparing these models with the basic models to select the optimal model as the final spectrogram recognition model;
[0011] S5. Collect the near infrared spectrum of the dried blood of the clinical patient to be diagnosed, and substitute it into the spectrum recognition model established in step S4 to obtain a rapid spectrum recognition result.
[0012] Further, the following steps are included: according to the operating steps in S1, the volume of the small amount of blood sample is 5-10 microliters, the storage temperature is -80°C, and the test is completed within 24 hours after thawing at room temperature.
[0013] Further, the method comprises the following steps: according to the operation steps in S2, the transparent quartz glass slide needs to be inspected for flatness and cleanliness before use.
[0014] Further, the method comprises the following steps: according to the operation steps in S2, the near-infrared spectrum set is collected by a calibrated and verified near-infrared spectrometer to ensure the accuracy of the data.
[0015] Further, the following steps are included: According to the operation steps in S2, the specific method of collecting near infrared spectrum in step S2 is:
[0016] S201, dropping the blood sample onto a transparent quartz glass slide and letting it dry naturally for 30 minutes;
[0017] S202, dripping blood again onto the dried blood and letting it dry naturally for 30 minutes;
[0018] S203, after overlapping and dropping for 1-3 times, collecting near infrared spectrum in transmission mode.
[0019] Further, the following steps are included: according to the operation steps in S2, the collection range of the near infrared spectrum is 4000-12000cm -1 , resolution range 1-16cm -1 , the scan times range is 2-64 times.
[0020] Further, the method includes the following steps: according to the operation steps in S3, the near-infrared spectrum set performs preprocessing on the spectrum data including at least one of SG smoothing (Savitzky-Golay, SG), wavelet transform (Wavelet Transform, WT), moving window average (Moving Window Average, MWA), standard normal variable transformation (Standard Normal Variate Transformation, SNV), multiple scattering correction (Multiple Scattering Correction, MSC), normalization (Normalization, Norm), mean centering (Mean Centering, MC), standardization (Standardization, S), first-order derivative, and second-order derivative.
[0021] Further, the steps include: according to the operation steps in S3, the multiple machine learning algorithms include principal component-linear discriminant analysis (Principal Component Analysis-Linear Dscriminant Analysis, PCA-LDA), principal component-quadratic discriminant analysis (Principal Component Analysis-Quadratic Discriminant Analysis, PCA-QDA), partial least squares discriminant analysis (Partial Least Squares Discriminant Analysis, PLSDA), K nearest neighbor (K-Nearest Neighbor, KNN) classification, logistic regression (Logistic Regression) model, support vector machine (Support Vector Machine, SVM), random forests (Random Forests, RF), extreme gradient boosting tree (Extreme Gradient Boosting, XGBoost), multilayer perceptron (Multilayer Perceptron, MLP) one or more.
[0022] Further, the method comprises the following steps: according to the operation steps in S4, the method for screening important features is to sort the characteristic wavelengths according to the size of the SHAP absolute value mean, and then take 0.5%-60% of the total number of characteristic values.
[0023] Further, the following steps are included: according to the operation steps in S5, the performance evaluation parameters of the spectrum recognition model include accuracy, area under the receiver operating characteristic curve (AUC), sensitivity and specificity.
[0024] The present invention provides a method for establishing a depression diagnosis model based on dried blood near-infrared spectroscopy, which has the following beneficial effects:
[0025] (1) The present invention does not require sample pretreatment and can complete the spectrum recognition result with only a small amount of blood, which is simple and quick to operate.
[0026] (2) Compared with the near-infrared spectrum of liquid blood collected directly, the near-infrared spectrum of dried blood collected by the present invention has no strong interference from the water peak and shows more and stronger characteristic signals.
[0027] (3) The present invention uses the selected important features for modeling, which greatly reduces the feature dimension of the near-infrared spectrum and increases the calculation rate and model accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] Figure 1 The figure is a flow chart of the method for establishing a depression diagnosis model based on dried blood near-infrared spectroscopy according to the present invention.
[0029] Figure 2 This is the receiver operating characteristic curve of the final near-infrared spectroscopy recognition model in Example 1 on the test set. DETAILED DESCRIPTION
[0030] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments.
[0031] Embodiment 1
[0032] See also Figure 1 The method for establishing a depression diagnosis model based on dried blood near-infrared spectroscopy includes the following steps: collect a small amount of serum samples from 400 patients with clinical depression and 400 healthy people, each sample is 50 microliters, and store them in a -80℃ refrigerator for later use. After the frozen serum sample is thawed at room temperature, 10 microliters is taken with a pipette and dripped onto a transparent quartz glass slide. After 30 minutes of natural drying, serum is dripped onto the dried serum again and dried naturally for 30 minutes. After overlapping dripping 1-3 times, the near-infrared spectrum is collected in the transmission mode. The spectrum collection range is 4000-12000cm -1 , resolution range 4cm -1, the number of scans was 32 times, and 400 spectra were collected from all 400 samples to form a spectrum set for spectrum recognition model. The original infrared spectrum set was subjected to 12 preprocessing methods, including baseline correction, MWA, SG smoothing, MSC, SNV, S, first-order derivative, second-order derivative, WT, SG smoothing + first-order derivative, MWA + second-order derivative, and MSC + first-order derivative, to obtain 12 preprocessed spectrum sets. For each spectrum set, the Kennard-Stone algorithm was used to divide it into a training set and a test set, where the training set was modeled using 9 machine learning algorithms, including PCA-LDA, PCA-QDA, PLSDA, KNN classification, logistic regression, SVM, RF, XGBoost, and MLP. According to the accuracy of each model on the test set (number of correctly predicted samples / total number of samples), the models were sorted from high to low, and the 5 best basic models were selected. Furthermore, SHAP analysis was performed on each basic model, and after sorting by the SHAP absolute value mean, 20% of the total number of eigenvalues, i.e. 400 feature bands, were taken. The 400 feature bands obtained from each basic model were merged, and modeled using the PLSDA and XGBoost algorithms. The two models obtained after merging features were compared with the five best basic models, and the best model was selected as the final spectrogram recognition model based on the accuracy. Figure 2 The receiver operating characteristic curve shows that the prediction accuracy of the final spectrogram recognition model on the test set is 87%, the AUC is 0.90, and the sensitivity and specificity are 93% and 75%, respectively. Finally, the near-infrared spectra of the serum samples of the clinical patients to be diagnosed are collected and substituted into the spectrogram recognition model to obtain the diagnosis results of whether the clinical patients suffer from depression.
[0033] Embodiment 2
[0034] See also Figure 1 The method for establishing a depression diagnosis model based on dried blood near-infrared spectroscopy includes the following steps: a small amount of plasma samples are collected from 200 patients with clinical depression and 200 healthy people, each sample is 60 microliters, and stored in a -80°C refrigerator for later use. After the frozen plasma sample is thawed at room temperature, 5 microliters are taken with a pipette and dripped onto a transparent quartz glass slide. After it is naturally dried for 20 minutes, the plasma is dripped onto the dried plasma again and dried naturally for 20 minutes. After overlapping and dripping 4 times, the near-infrared spectrum is collected in the transmission mode. The spectrum collection range is 4000-12000cm -1 , resolution range 2cm -1, the number of scans was 64 times, and 800 spectra were collected from all 800 samples to form a spectrum set for the spectrum recognition model. The original infrared spectrum set was subjected to 12 preprocessing methods, including baseline correction, MWA, SG smoothing, MS C, SNV, S, first-order derivative, second-order derivative, WT, SG smoothing + first-order derivative, MWA + second-order derivative, and MSC + first-order derivative, to obtain 12 preprocessed spectrum sets. For each spectrum set, the Kennard-Stone algorithm was used to divide it into a training set and a test set, where the training set was modeled using 9 machine learning algorithms, including PCA-LDA, PCA-QDA, PLSDA, KNN classification, logistic regression, SVM, RF, XGBoost, and MLP. According to the accuracy of each model on the test set (number of correctly predicted samples / total number of samples), the models were sorted from high to low, and the 8 best basic models were selected. Furthermore, SHAP analysis was performed on each basic model, and after sorting according to the SHAP absolute value mean, 10% of the total number of eigenvalues, i.e., 200 characteristic bands, were taken. The 200 characteristic bands obtained from each basic model were merged, and modeled using the PLSDA and MLP algorithms. The two models after the merged features were compared with the 8 best basic models, and the best model was selected as the final spectrogram recognition model based on the accuracy. At this time, the prediction accuracy of the final spectrogram recognition model on the test set was 90%, the AUC was 0.87, and the sensitivity and specificity were 93% and 71%, respectively. Finally, the near-infrared spectra of the plasma samples of clinical patients to be diagnosed were collected and substituted into the spectrogram recognition model to obtain the diagnosis results of whether the clinical patients suffered from depression.
[0035] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A method for establishing a depression diagnosis model based on dried blood near-infrared spectroscopy, characterized in that: The following steps are involved: S1. Collect and store a small amount of blood samples from several clinically depressed patients and healthy controls; S2, dropping the blood sample obtained in step S1 onto a transparent quartz glass slide to obtain a dried blood sample, and collecting the near-infrared spectrum of the dried blood to form a spectrum set; S3, preprocessing the spectral data of the near-infrared spectrum set obtained in step S2 according to a certain method, and establishing multiple basic models through a machine learning algorithm; S4, performing SHAP analysis on the multiple basic models established in step S3, screening and merging important features, establishing models through the machine learning algorithm used in step S3, and comparing these models with the basic models to select the optimal model as the final spectrogram recognition model; S5. Collect the near infrared spectrum of the dried blood of the clinical patient to be diagnosed, and substitute it into the spectrum recognition model established in step S4 to obtain a rapid spectrum recognition result.
2. The method for establishing a depression diagnosis model based on dried blood near infrared spectroscopy according to claim 1, characterized in that: The method comprises the following steps: according to the operation steps in S1, the volume of the small amount of blood sample is 5-10 μl, the storage temperature is -80°C, and the test is completed within 24 hours after thawing at room temperature.
3. The method for establishing a depression diagnosis model based on dried blood near infrared spectroscopy according to claim 1, characterized in that: The method comprises the following steps: according to the operation steps in S2, the transparent quartz glass slide needs to be inspected for flatness and cleanliness before use.
4. The method for establishing a depression diagnosis model based on dried blood near infrared spectroscopy according to claim 1, characterized in that: The following steps are involved: According to the operation steps in S2, the near-infrared spectrum set is collected by a calibrated and verified near-infrared spectrometer to ensure the accuracy of the data.
5. The method for establishing a depression diagnosis model based on dried blood near infrared spectroscopy according to claim 1, characterized in that: The method comprises the following steps: According to the operation steps in S2, the specific method of collecting near infrared spectrum is: S201, dropping the blood sample onto a transparent quartz glass slide and letting it dry naturally for 30 minutes; S202, dripping blood again onto the dried blood and letting it dry naturally for 30 minutes; S203, after overlapping and dropping for 1-3 times, collecting near infrared spectrum in transmission mode.
6. The method for establishing a depression diagnosis model based on dried blood near infrared spectroscopy according to claim 1, characterized in that: The method comprises the following steps: according to the operation steps in S2, the near infrared spectrum is collected in the range of 4000-12000cm -1 , resolution range 1-16cm -1 , the scan times range is 2-64 times.
7. The method for establishing a depression diagnosis model based on dried blood near infrared spectroscopy according to claim 1, characterized in that: The following steps are involved: According to the operation steps in S3, the near-infrared spectrum set performs preprocessing on the spectrum data by at least one of SG smoothing, wavelet transform, moving window averaging, standard normal variable transformation, multivariate scattering correction, normalization, mean centering, standardization, first-order derivative, and second-order derivative.
8. The method for establishing a depression diagnosis model based on dried blood near infrared spectroscopy according to claim 1, characterized in that: The following steps are involved: According to the operation steps in S3, the multiple machine learning algorithms include principal component-linear discriminant analysis, principal component-quadratic discriminant analysis, partial least squares discriminant analysis, K nearest neighbor classification, logistic regression model, support vector machine, random forest, extreme gradient boosting tree, and one or more of multilayer perceptron.
9. The method for establishing a depression diagnosis model based on dried blood near infrared spectroscopy according to claim 1, characterized in that: The following steps are involved: According to the operation steps in S4, the method for screening important features is to sort the characteristic wavelengths according to the SHAP absolute value mean size and then take 0.5%-60% of the total characteristic value quantity.
10. The method for establishing a depression diagnosis model based on dried blood near infrared spectroscopy according to claim 1, characterized in that: The method comprises the following steps: according to the operation steps in S5, the performance evaluation parameters of the spectrum recognition model include accuracy, area under the receiver operating characteristic curve, sensitivity and specificity.