An asd child screening method based on fnirs data quality improvement and empathy features

Abstract

The application discloses an ASD child screening method based on fNIRS data quality improvement and empathy characteristics, and belongs to the technical field of brain imaging data processing and medical diagnosis. The method comprises the following steps: reading and verifying a `.nirs` data file; performing optical density conversion, motion artifact correction, band pass filtering and Beer-Lambert law conversion on light intensity data to obtain HbO / HbR / tHb concentration change data; extracting a time sequence based on a preset ROI, calculating ROI activation features and functional connection features; integrating the features into a unified feature package; adopting an IQR-count abnormal sample detection method to improve data quality, training an MLP classification model under the condition of each fold independent standardization, and outputting an ASD screening result. The method improves the accuracy, stability and reproducibility of fNIRS single-mode ASD child screening through fine abnormal value processing and empathy-emotion infection related neural feature extraction.

Description

Technical Field

[0001] This invention relates to the fields of brain imaging data processing and medical diagnostics, specifically to a method for screening children with ASD based on fNIRS data quality improvement and empathy features. More particularly, it relates to a data quality improvement and classification screening method for parent-child interaction scenarios, extracting neural activity and functional connectivity features related to empathy and emotional contagion. Background Technology

[0002] Autism spectrum disorder (ASD) is a neurodevelopmental disorder, and early screening and intervention are crucial for improving prognosis. fNIRS, a non-invasive brain imaging technique, can reflect changes in cortical hemodynamics and has potential applications in pediatric screening.

[0003] However, existing fNIRS screening technologies still have the following shortcomings: 1. The acquired data is susceptible to motion artifacts, physiological noise, and signal drift, resulting in unstable data quality; 2. Outlier handling strategies are crude, often employing a "remove any outlier feature" approach, which can easily lead to over-sampling. 3. The feature extraction dimensions are limited, making it difficult to simultaneously take into account both brain region activation information and network connectivity information; 4. The selection of ROIs is not specific enough, making it difficult to effectively characterize neural processing patterns related to empathy and emotional contagion; 5. Insufficient standardization, traceability, and reproducibility of the processing procedures affect the credibility of clinical applications.

[0004] Therefore, there is a need for a method that can stably extract empathy-related neural features and improve the screening performance of children with ASD while ensuring data quality. Summary of the Invention

[0005] I. Technical problems to be solved The technical problem to be solved by this invention is to provide a screening method for children with ASD based on fNIRS data quality improvement and empathy features. By improving data quality and extracting empathy-related neural features, the accuracy, stability and reproducibility of screening can be improved.

[0006] II. Technical Solution To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A method for screening children with ASD based on fNIRS data quality improvement and empathy features includes the following steps: S1 Data Reading and Verification Read the .nirs file, extract the light intensity data d, timestamp t, measurement list ml, and source detector configuration; estimate the sampling rate and perform data integrity verification, and generate a verification report.

[0007] S2 Data Preprocessing The original light intensity data were sequentially subjected to density conversion, motion artifact correction, bandpass filtering, and Beer-Lambert conversion to obtain HbO / HbR / tHb concentration change data.

[0008] The formula for optical density conversion is:

[0009] Where I is the current light intensity and I0 is the reference intensity.

[0010] The conditions for motion artifact detection can be expressed as:

[0011] The threshold is preferably 3.0.

[0012] The Beer-Lambert conversion can be represented as:

[0013] Where ε_HbO and ε_HbR are extinction coefficients, DPF is the differential path length factor, and L is the source-detector distance.

[0014] S3 ROI Feature Extraction Based on the preset ROI definition, the ROI time series is extracted, and the ROI activation features and functional connectivity features are calculated to form a static brain functional feature vector.

[0015] The ROI activation features include at least: mean, standard deviation, maximum value, minimum value, range, slope, peak time, and baseline-corrected mean.

[0016] The functional connectivity features are obtained by calculating the correlation between ROI pairs.

[0017] S4 Feature Integration The ROI activation features and functional connectivity features are written into the feature package according to a unified key name rule to form a structured input that can be used for modeling.

[0018] S5 Data Quality Improvement and Model Training The IQR-count method is used to identify anomalous samples: the number of anomalous features in each sample is counted, and when the number of anomalous features reaches a threshold, the sample is identified as an anomalous sample and removed; then, independent standardization is performed for each fold and a classification model is trained to output the ASD probability and category results.

[0019] The IQR-count determination process can be represented as:

[0020] when When sample j is identified as an abnormal sample, it is determined to be an abnormal sample.

[0021] The minimum outlier feature is preferably 15.

[0022] The classification model preferably uses a multilayer perceptron (MLP), and the output probability is:

[0023] The training loss function is the binary classification cross-entropy:

[0024] Compared with the prior art, the present invention has at least the following beneficial effects: 1. Provide a complete, automated, traceable, and reproducible data processing workflow; 2. By using IQR-count for fine-grained outlier handling, we can reduce over-desampled data and improve data utilization. 3. Simultaneously extract ROI activation and functional connectivity features to enhance the representation ability of neural patterns related to empathy and emotion contagion; 4. Reduce the risk of data leakage and improve the model's generalization stability by standardizing each fold independently; 5. It achieved good accuracy and stability in ASD screening tasks for children, and has practical application value.

[0025] To illustrate the technical effects of the present invention, experimental results are provided as an example below.

[0026] Table 1. Performance Comparison of Different fNIRS Model Configurations (50% Discount) Model Configuration Accuracy of each fold Average accuracy Standard deviation illustrate F1-DataQuality-03 (Quality Improvement Baseline) [0.7895, 0.9474, 0.8421, 0.7222, 0.7778] 0.8158 0.0760 IQR outlier removal F1-DataQuality-04 (Z-score outlier removal) [0.7297, 0.8378, 0.7778, 0.8056, 0.8611] 0.8024 0.0460 Good stability F1-Base-02 (seed=2025) [0.6739, 0.8000, 0.7778, 0.7556, 0.8222] 0.7659 0.0511 Multiple substituent stability verification F1-Base-01 (Early Baseline) [0.6522, 0.8444, 0.7778, 0.7778, 0.8000] 0.7704 0.0639 Standard BCE, 132-dimensional features F1-DataQuality-Improved-02 (main configuration, seed=42) [0.7941, 0.8824, 0.7879, 0.8788, 0.8182] 0.8323 0.0407 iqr_count, min_features=15 As shown in Table 1, F1-DataQuality-Improved-02 performs better in both mean accuracy and standard deviation, making it a more reasonable choice as the current master model configuration.

[0027] Table 2. Representative ablation / improvement experimental results (relative to F1-DataQuality-Improved-02) Experimental setup Average accuracy Standard deviation relative change in conclusion F1-DataQuality-03 (Quality Improvement Baseline) 0.8158 0.0760 -0.0165 Both accuracy and stability are lower than optimal. F1-DataQuality-Improved-01 (iqr_count, min10) 0.8069 0.0516 -0.0254 A low threshold leads to decreased accuracy. F1-DataQuality-Improved-03 (iqr_count, min20) 0.8066 0.0766 -0.0257 A high threshold leads to performance degradation F1-DataQuality-Improved-04 (iqr_count, min25) 0.7849 0.0710 -0.0474 Over-filtered samples F1-DataQuality-Improved-06 (iqr_count, min18) 0.8197 0.0271 -0.0126 Better stability but lower accuracy than optimal F1-DataQuality-Improved-02 (iqr_count, min15) 0.8323 0.0407 0 Optimal configuration As shown in Table 2, the IQR-count strategy with `min_features=15` achieves the best balance between accuracy and stability. Attached Figure Description

[0028] Figure 1 Here is a flowchart of the fNIRS data processing procedure; Figure 2 Here is a diagram of the fNIRS MLP model architecture; Figure 3 This is a comparison chart of model performance; Figure 4 A diagram illustrating the effects of data quality improvements; Figure 5 Here is a comparison chart of the performance of each section; Figure 6 Feature importance analysis diagram; Figure 7 Activate the heatmap for the ROI; Figure 8 This is a graph showing the results of the ablation experiment; Figure 9 This is a schematic diagram of the IQR-count outlier detection method. Detailed Implementation

[0029] The present invention will be further described below with reference to embodiments. It should be understood that the following embodiments are for illustrative purposes only and not for limiting the scope of protection of the present invention.

[0030] Example 1: Data Reading and Preprocessing Read the `.nirs` file and extract fields such as d, t, and ml; Perform optical density conversion, motion artifact correction, bandpass filtering, and Beer-Lambert conversion on the data to output hbo, hbr, and thb data.

[0031] Example 2: ROI Feature Extraction Extract ROI time series according to the preset ROI channel mapping; Calculate ROI activation features (8-dimensional statistical features) and ROI pair connectivity features (Pearson correlation) to generate static brain function features.

[0032] Example 3: Feature Integration and Outlier Handling Write the ROI and connectivity features into a unified feature package; IQR-count is used to count the number of abnormal features for each sample and remove abnormal samples, thereby reducing the problem of accidentally deleting valid samples in high-dimensional scenarios.

[0033] Example 4: Modeling and Evaluation The MLP classification model was trained using independently standardized features at each fold, and evaluated using 5-fold cross-validation. Output metrics such as accuracy, AUC, ASD recall, and confusion matrix to evaluate screening performance.

[0034] In one embodiment, the 50% accuracy of the primary configuration F1-DataQuality-Improved-02 is [0.7941, 0.8824, 0.7879, 0.8788, 0.8182], the mean accuracy is 0.8323, and the standard deviation is 0.0407.

[0035] This configuration offers higher accuracy and less variability, resulting in better engineering usability.

[0036] Fusion and classification can be represented as:

[0037]

[0038] Example 5: Ablation and Stability Verification By ablation analysis of outlier handling strategies, feature combinations, and training configurations, we verify the contributions of IQR-count and multi-dimensional feature schemes to performance improvement and stability enhancement.

[0039] In one embodiment, the IQR-count was compared and verified with the training configuration, and the results are as follows.

[0040] Table 3 Results of Data Quality Improvement and Stability Verification Comparison items Configuration Average accuracy Standard deviation illustrate Comparison of outlier handling thresholds F1-DataQuality-Improved-01 (min_features=10) 0.8069 0.0516 The threshold is too low, increasing the risk of accidental deletion. Comparison of outlier handling thresholds F1-DataQuality-Improved-03 (min_features=20) 0.8066 0.0766 The threshold is too high, resulting in more retained noise. Threshold neighborhood stability F1-DataQuality-Improved-06 (min_features=18) 0.8197 0.0271 Higher stability but lower accuracy than optimal Quality Improvement Baseline Comparison F1-DataQuality-03 (IQR) 0.8158 0.0760 Overall performance is below the optimal threshold configuration Optimal Model Comparison F1-DataQuality-Improved-02 (iqr_count,min_features=15) 0.8323 0.0407 Current optimal model The above results show that the IQR-count threshold optimization scheme represented by F1-DataQuality-Improved-02 can achieve the best screening performance (0.8323) while taking into account stability, thus supporting the effectiveness of the scheme of this invention.

Claims

1. A method for screening children with ASD based on fNIRS data quality improvement and empathy features, characterized in that, Includes the following steps: S1. Read and verify the `.nirs` data file, and extract information such as light intensity data, timestamps, and measurement lists; S2. Perform optical density conversion, motion artifact correction, bandpass filtering, and Beer-Lambert law conversion on the light intensity data in sequence to obtain hemoglobin concentration change data; S3. Extract ROI time series based on preset brain region definitions, and calculate ROI activation features and functional connectivity features; S4. Integrate the ROI activation features and functional connectivity features into a unified feature package; S5. Perform data quality improvement and model training, including using the IQR-count rule based on the number of abnormal features to detect and remove abnormal samples, performing independent feature standardization by cross-validation fold, and outputting ASD screening results based on the multilayer perceptron classification model; wherein the abnormal sample detection adopts the method of "judging by the number of abnormal features of the sample" rather than "removing any abnormal feature".

2. The ASD screening method for children based on fNIRS data quality improvement and empathy features according to claim 1, characterized in that, The optical density conversion in step S2 uses the formula... Where I is the current light intensity, and I0 is the reference intensity, the reference intensity being the average value of the corresponding channel over time; the motion artifact correction in step S2 is achieved through threshold detection and linear interpolation, and the anomaly detection condition is... The threshold is 3.

0.

3. The ASD screening method for children based on fNIRS data quality improvement and empathy features according to claim 1, characterized in that, The bandpass filtering in step S2 uses a 4th-order Butterworth filter with a passband of 0.01Hz to 0.1Hz and zero-phase filtering. In step S3, seven ROIs are preset, and for each ROI, activation features including at least mean, standard deviation, maximum, minimum, range, slope, peak time, and baseline-corrected mean are calculated, and functional connectivity features between ROI pairs are calculated.

4. The ASD screening method for children based on fNIRS data quality improvement and empathy features according to claim 1, characterized in that, The abnormal sample detection in step S5 adopts the IQR-count method, which determines the abnormal sample by counting the number of abnormal features. When the number of abnormal features is greater than or equal to the threshold, it is determined to be an abnormal sample. The threshold min_outlier_features is 15 and the IQR multiplier factor is 1.

5.

5. The ASD screening method for children based on fNIRS data quality improvement and empathy features according to claim 1, characterized in that, The feature standardization in step S5 adopts a fold-independent standardization strategy, that is, in each fold of cross-validation, the standardized parameters are fitted only using the training set and used for the validation set transformation.

6. The method for screening children with ASD based on fNIRS data quality improvement and empathy features according to claim 1, characterized in that, The classification model in step S5 is a multilayer perceptron, with an input layer, two hidden layers, and an output layer. The dimensions of the hidden layers are 256 and 128. Dropout regularization is used, the loss function is binary cross-entropy loss, and the optimizer is Adam.

Images