An early detection framework for young Chinese learners at risk of reading difficulty using fNIRS and deep learning

Abstract

Reading difficulty (RD), a neurodevelopmental disorder affecting language acquisition in children, necessitates early screening for effective educational interventions. This study proposes the RD-risk Classifier (RDr-C), a novel framework integrating functional near-infrared spectroscopy (fNIRS) with deep learning, specifically combining a dual-layer graph convolutional network (GCN), a bidirectional long short-term memory network (BiLSTM), and multi-head self-attention mechanisms (MSA) for 7-8-year-old children’s literacy assessment. The model was validated using fNIRS signals from 30 participants (16 experimental group, 14 control group) during the visual sign recognition test and phonetic discrimination test, with performance evaluated through 5 runs 5-fold cross-validation experiments. Results show that RDr-C achieved a mean classification accuracy of 99.60% and 99.66% in visual and auditory tests, respectively, significantly outperforming traditional convolutional neural networks (CNN), long short-term memory networks (LSTM), and existing fNIRS classification models (e.g., fNIRS-T, fNIRSNet). Furthermore, leave-one-subject-out cross-validation demonstrates that RDr-C achieves global accuracies of 89.33% and 87.93% on visual and auditory tasks, respectively, with corresponding Kappa coefficients of 0.78 and 0.76, confirming its robustness across individuals. Feature shuffling and wavelet transformation visualizations further confirm robust feature separation, highlighting the model’s ability to capture distinct hemodynamic patterns associated with RD. By integrating the spatial feature extraction of GCN, the temporal modeling of BiLSTM, and the global dependency capture of MSA, this work establishes a non-invasive neuroimaging paradigm for educational neuroscience. The high-precision classification lays a technical foundation for early screening tools, with future applications extending to multimodal brain-computer interfaces and longitudinal intervention monitoring.

Introduction

Reading difficulty (RD), also known as developmental dyslexia, is clinically characterized as a developmental learning disorder that primarily affects reading skills. Individuals with RD experience significant and ongoing challenges in mastering academic skills related to reading, including word recognition accuracy, reading fluency, and comprehension¹. Specifically, dyslexia is defined as learning difficulties characterized by problems with accurate or fluent word recognition, poor decoding, and poor spelling abilities according to the Diagnostic and Statistical Manual of Mental Disorders (DSM-5)². In alphabetic systems, dyslexia is a reflection of the children’s poor spelling ability, affecting about 5-10% of children^3,4. Dyslexia is prevalent among approximately 5%-15% of children globally⁵. However, in morpho-syllabic systems like Chinese, the rate of dyslexia is moderately lower than that in alphabetic writing systems, which is around 4%-7%⁶.

Dyslexia diagnosis and assessment currently rely mainly on clinical questionnaires and assessments of reading and cognitive ability^7–9, and the diagnosis is usually conducted at the upper grades of primary school, as children’s development at lower grades may be highly unstable. However, for children at reading difficulty risk, the later the assessment happens, the more possible missing the best intervention window.

Dyslexia assessment methods

Daniel et al.¹⁰ explored various historical and contemporary methods for identifying dyslexia, highlighting their limitations and evolution. Early theories, such as the visual-perceptual deficit model that attributed dyslexia to magnocellular pathway dysfunction, have been criticized for lacking robust empirical support¹¹. The response to intervention (RTI) model¹², which screens and monitors reading progress while providing tiered interventions, offers advantages such as early intervention and reduced over-identification¹³, but it still faces challenges in reliability in assessing adequate instruction and response^14–19. Hybrid models integrating multiple criteria are now recommended^20–23 , due to the recognition that dyslexia has a multifactorial etiology, in addition to the functional disorders resulting from brain injuries, many children with dyslexia have comorbidities of other diseases, such as autism and Attention-Deficit/Hyperactivity Disorder (ADHD)²⁴.

The Compulsory Education Curriculum Standards for Chinese course (2022 Edition) clearly proposed that children at Grade 1 and Grade 2 need to master 1,600 commonly used words’ recognition and 800 common words’ writing²⁵. However, according to feedback from elementary school Chinese teachers, the proportion of young children who fail to meet the curriculum standard far exceeds 7%, i.e., beyond those diagnosed with dyslexia by medical professionals, there is a substantial number of children with marginal reading and writing difficulties, who are considered at high risk for dyslexia²⁴. Identifying and understanding what hinders this broader at-risk group from developing language skills normally is crucial. Therefore, in regular school education, the most practical and effective detection method remains objective measurement based on curriculum standards²⁶.

With advancements in non-invasive detection devices, various brain function-based assessment methods have been explored²⁷. These tools offer more precise subclassifications for dyslexia compared to subjective measures. For example, Kovelman et al.²⁸ reported that children with reading and writing disorders caused by auditory perception deficits exhibit insufficient activation of the prefrontal cortex (PFC). In addition to auditory perception, deficits regarding speech perception have also been reported, such as reduced brain activation in the dorsal left inferior frontal gyrus (IFG) during an auditory rhyming judgment task²⁹; Fu et al.²⁷ found that both the right IFG and the dorsolateral prefrontal cortex (DLPFC) show less activation in Chinese readers with RD compared with their normal peers. A primary challenge of using functional near-infrared spectroscopy (fNIRS) to reveal and assess reading difficulty is that no consistent conclusion has been reached worldwide regarding the differences in brain function between dyslexics and normal individuals, as the neurological specificity of different languages’ writing systems in language development remains poorly understood.

Deep learning methods applied on fNIRS signals

Recently, machine learning-based methods have been adopted in the brain-computer interface (BCI) field to enhance the accuracy of fNIRS signal recognition. The characteristics of fNIRS raw data determine that it is not suitable for traditional machine learning models, since its raw data comprises changes in light absorption in a continuous time series or hemoglobin concentration changes (e.g., oxygenated hemoglobin, HbO, and deoxygenated hemoglobin, HbR) after processing. Traditional machine learning models, such as support vector machine (SVM)³⁰, SVM+ linear discrimination analysis (LDA)³¹ and multiple layer perceptron (MLP) based artificial neural networks (ANNs) are widely used to recognize the signals³²; However, considering their limited feature extraction ability, their performance mostly relies on hand-engineered features selected by experts or based on prior studies, such as the mean and slope values of the signal. Consequently, the robustness of the extracted features may be compromised due to factors like individual variability and experimental condition variations. In contrast, deep learning approaches are capable of automatically extracting features directly from raw data, eliminating the need for manual feature engineering.

Deep learning techniques have been increasingly applied to classify fNIRS data. Architectures such as convolutional neural networks (CNN)^33–35, long short-term memory (LSTM)³⁶, Transformer³⁷, and hybrid methods^33,34,38 have demonstrated success in enhancing classification performance. Beyond these direct applications to fNIRS, recent advances in fMRI analysis have introduced sophisticated deep learning frameworks that effectively handle high-dimensional, non-linear, and time-frequency correlated data. For instance, Deep CSAF³⁹ and HB-DFL⁴⁰ leverage non-linear tensor factorization with attention mechanisms to extract interpretable factor features for ADHD and Parkinson’s disease classification. Furthermore, Deep FCAF⁴¹ incorporates frequency-channel attention to emphasize low-frequency fluctuations, while Deep WTFAF⁴² and Deep WSANTF⁴³ embed wavelet transforms into self-attention to capture both high-frequency details and low-frequency structures. More recently, F-GAN-NTD⁴⁴ combines generative adversarial networks with non-negative tensor decomposition to improve interpretability and robustness in depression fMRI analysis. Although developed for fMRI, these methodologies share core challenges with fNIRS: handling non-linearity, high dimensionality, and time-frequency dynamics, and their success suggests promising directions for enhancing fNIRS-based brain-computer interfaces and clinical assessments. Deep learning-based models have demonstrated superior performance on public datasets compared with traditional machine learning models, yet they share some common strategies, such as using a sliding window and data augmentation to expand the sample size, and strengthening the features embedded in the data. Notably, in order to expand the feature space, HbO and HbR were considered as two-dimensional data in most studies to improve the models’ classification accuracy; however, from a biomedical engineering perspective, both signals essentially reflect changes in the participant’s blood oxygen concentration^45,46. Most existing models (e.g., fNIRS-T⁴⁷, fNIRSNet⁴⁸) incorporate both HbO and HbR as different sources of signals. To increase the explainability of the model, this study only employs HbO to assess the models’ usability and efficiency.

In order to recognize and understand children at risk for reading difficulties (RD), capturing their abnormal temporal dependencies in time series data is vitally important. To this end, this study proposes a novel model, the RD-risk Classifier (RDr-C), which integrates a dual-layer graph convolutional network (GCN), a bidirectional long short-term memory network (BiLSTM) with multi-head self-attention mechanisms (MSA) to better recognize the typical patterns of RD at an early age. The dual-layer GCN uses an adjacency matrix constructed from the probe layout corresponding to the fNIRS signal acquisition device to perform spatial feature extraction on the input data. The BiLSTM module processes the data in both forward and backward directions, effectively capturing long-term dependencies. The MSA dynamically weighs the importance of different time steps, allowing the model to focus on the most salient temporal sequences and learn from multiple representational subspaces. This integrated architecture strengthens the capacity to represent spatial features via GCN and temporal dynamics via BiLSTM, while capturing global dependencies via MSA, ultimately improving the performance of early RD identification.

Methods

Participant

The dataset comprised two groups. Group A included 16 children who were reported by their teachers as marginal RD (age 7–8, mean = 7.38, SD = 0.5; 7 female, and 9 male), and Group B (the control group) included 14 children who were identified by their teachers as having normal reading abilities(age 7–8, mean=7.79, SD=0.43; 11 female and 3 male). All participants were Grade 2 students. In order to further verify the reliability of the participants’ labels, we conducted a Chinese character phonetic notation test for 150 Grade 2 students, including the 30 participants, and the suspected RD pupils achieved significantly lower scores (below the mean-1 SD) than their normal peers.

Ethics approval and consent to participate

The study protocol was approved by the Ethics Committee of Sichuan Normal University. Written informed consent was obtained from the parents/legal guardians of all participants. All procedures were conducted in accordance with the Declaration of Helsinki and relevant institutional guidelines.

Chinese character phonetic notation test

The Chinese Character Phonetic Test required participants to label pinyin for characters specified in the Chinese curriculum standards (2022)²⁵. The test included 100 Chinese characters drawn from the second semester of the Grade 1 textbook. More than five Chinese teachers in that school are involved in the word selection work based on their experience and the importance and acquisition difficulty of the words. 150 children, including the participants from the same primary school, were required to participate in the test, in which they needed to note the pinyin on the top of each word within a one-hour time limit. The test sheet is illustrated in Fig. 1.

Fig. 1.

The Chinese character phonetic notation test sheet.

Visual sign recognition test

Sixteen suspected RD learners in Group A and 14 normal learners in Group B took the fNIRS-based sign recognition test. The test consisted of two parts: a Chinese character recognition and discrimination test, and an English word recognition and discrimination test. The experimental flow of the sign recognition and discrimination test is shown in the upper panel of Fig. 2. In this test, participants observed the target word displayed at the center of the screen for 20 seconds, then selected it from four candidates. Both the Chinese character candidates and the English word candidates were visually similar ones. All Chinese characters and English words were selected from the Grade 1 textbook, i.e., words that appeared in the test were those that had to be acquired for Grade 2 students.

Fig. 2.

Protocols of the fNIRS-based tests.

Phonetic discrimination test

Learners who participated in the fNIRS-based sign recognition test also completed the fNIRS-based phonetic discrimination test. The experimental procedure is shown in the lower panel of Fig. 2. In this test, participants listened carefully to two sounds (each presented twice), then immediately responded to indicate whether the two sounds were identical. The test included five sets of Pinyin sounds and five sets of English sounds.

fNIRS signal collecting

In this study, fNIRS signals were collected by a multichannel continuous fNIRS measuring system (NirSmart II-3000A, Huichuang, China). The fNIRS device used in this study is a portable one with a sampling rate of 11 Hz, and wavelengths of 730 nm and 850 nm. As shown in Fig. 3a–f⁴⁹, the probe cap includes 22 infrared light sources and 16 detectors following the international 10–20 system (with D3 aligned with the Fpz reference point). The 49 fNIRS channels were grouped into 22 functionally defined ROIs, which covered the frontal, left, and right hemispheres⁴⁹. Figure 4 shows the experimental setup for fNIRS data acquisition.

Fig. 3.

The detectors and channels distribution on the cap⁴⁹.

Fig. 4.

fNIRS data collection.

Data preprocessing

The fNIRS signals obtained from participants during the visual sign recognition test and phonetic discrimination test were raw light intensity data, which were further processed in NirSpark, a software bundled with the NirSmart Device. NirSpark reduced the influence of motion artifacts from the optical density curve by using a cubic spline difference algorithm; then, NirSpark decreased noise caused by physiological fluctuations, such as heart rate and respiration, by applying a band-pass filter with a cut-off frequency of 0.01–0.2 Hz. The optical density data were then converted into relative changes in HbO and HbR according to the modified Beer–Lambert law (with a differential path factor of 6 for wavelengths of 730 nm and 850 nm). The definition of the baseline HbO data, task-related HbO data followed the same research paradigm described by Wang et al.⁵⁰. The change in HbO induced by the test was defined as the difference between the mean values of the task-related and baseline HbO data (HbO concentration when learners in rest).

The segmentation of task-related and baseline state data was executed based on the marks inserted by E-Prime. For each participant, the HbO data were partitioned into ten samples according to the start-end mark intervals of the ten experimental task blocks. Specifically, the baseline state was defined as a segment of 5s within the start and end markers interval of the rest period. Subsequently, the mean HbO value during this baseline period was subtracted from each sample to obtain baseline-corrected HbO data (HbO). In aggregate, a total of 3010 samples were ultimately obtained. After data segmentation, we performed padding and interpolation to align the number of rows in the data to the mean number of rows across all samples. These data were normalized in the range of -1 and 1 using min-max normalization to facilitate model optimization. The normalization was applied as follows:

1

where x is the raw feature values, x’ is the normalized feature values, xmax and xmin are the maximum and minimum values of x, respectively.

RD-risk learners’ recognition model

Structure of RDr-C

To address the prominent characteristics of spatial functional connectivity and temporal variations in fNIRS signals, we propose the RDr-C that integrates a dual-layer GCN and a BiLSTM with an MSA to effectively capture temporal dynamics and global dependencies in fNIRS data, as shown in Fig. 5. The input shape is batch size36349, where 363 denotes the number of time steps, and 49 is the feature dimension per step. The architecture comprises two branches: one branch begins with a GCN, which performs spatial feature extraction on the input data by incorporating an adjacency matrix constructed from the probe layout corresponding to the fNIRS signal acquisition device. The extracted features are then passed to a two-layer BiLSTM module, where each layer contains 32 hidden units, producing concatenated outputs of 64 dimensions. Subsequently, the temporal self-attention module employs 4 parallel attention heads. The input sequences are projected into Query, Key, and Value matrices via independent linear layers, and attention weights are computed using scaled dot-product attention:

2

The multi-head outputs are concatenated and linearly projected back to 128-dimensional features. The second branch focuses on capturing temporal characteristics of the input data through a more powerful BiLSTM with 64 hidden units and outputs of 128 dimensions, and an attention module with 8 heads. The outputs of both branches are dynamically weighted and fused at the final stage, which is mapped to a 2-class output space via a fully connected layer, with Softmax generating the final class probabilities. After the model training, critical features are extracted by shuffling certain features of the input data and observing the corresponding decline in model accuracy, followed by visualization of these features. By combining GCN for spatial feature extraction, BiLSTM for bidirectional temporal modeling, and MSA for long-range dependency extraction, this architecture achieves efficient classification of complex fNIRS signals.

Fig. 5.

The architecture of RDr-C.

Training strategy and parameters

During model training, multiple optimization strategies were employed to enhance convergence speed and generalization. The AdamW optimizer (learning rate =0.001, weight decay =0.0001) was utilized, which adaptively adjusts parameter updates by estimating gradient directions through exponential moving averages of first-order (momentum) and second-order (adaptive scaling) moments, while applying decoupled weight decay regularization directly to the parameters to improve generalization and prevent overfitting.

A Cosine Annealing Warm Restarts scheduler was implemented to refine parameter tuning, periodically restarting the learning rate using cosine annealing (). This scheduler helps escape local minima and improves convergence by providing multiple warm restarts during training, with the learning rate following a cosine decay pattern within each restart cycle:

3

Gradient clipping with L2 norm (maximum norm = 1.0) was implemented to constrain gradient magnitudes and prevent exploding gradients, ensuring stable training dynamics throughout the optimization process:

4

Label smoothing regularization () was applied to the cross-entropy loss function to prevent overconfident predictions and improve model generalization. This technique replaces hard target labels with a weighted combination of the true label and a uniform distribution over all classes:

5

Hierarchical Dropout layers with different rates were strategically placed throughout the architecture to suppress overfitting: Dropout with rate p=0.2 was applied within the BiLSTM layers, while Dropout with rate p=0.3 was implemented after GCN layers and in the feature fusion modules, randomly deactivating neuron outputs during training and scaling features appropriately during inference to improve model robustness. The dropout operation can be expressed as:

6

The models were trained on Intel Core i9-14900K and NVIDIA GeForce RTX 4090 using PyTorch v2.2.1.

Leave-one-subject-out cross-validation

To evaluate the generalization ability of the proposed model across different individuals, we conducted leave-one-subject-out cross-validation (LOSO-CV). In this strategy, data from a single subject are used as the test set, while the remaining subjects serve as the training set. This process is repeated until every subject has been tested once. Since all samples within each subject share the same label, we compute overall performance metrics (accuracy, precision, recall, F1-score, and Kappa coefficient) from the aggregated predictions across all subjects, rather than reporting the mean and standard deviation of accuracy. The LOSO-CV approach provides a rigorous assessment of the model’s ability to generalize to unseen individuals, which is crucial for real-world applications where individual variability is significant.

Employing two open-sourced models to make a comparison

We deployed two open-source models from existing research as comparison models: fNIRS-T⁴⁷ and fNIRSNet⁴⁸. The detailed architectures of these models are described as follows: The fNIRS-T model employs a dual-branch Transformer architecture for fNIRS classification. Its spatial branch uses 2D convolutions with 530 kernels and 14 strides to extract multi-channel spatiotemporal patterns, while the channel branch applies 130 kernel convolutions with identical strides for single-channel temporal analysis. Both branches integrate positional embeddings and [CLS] tokens, feeding sequences into Transformer encoders equipped with multi-head self-attention, GELU-activated MLPs, and layer normalization for global dependency modeling. The fNIRSNet processes fNIRS signals through two convolutional stages. The DHR module extracts features via 140 convolutions and batch normalization, while the global module applies 401 depthwise separable convolutions for spatial-temporal compression. Features are flattened and projected to class outputs through a fully connected layer, integrating channel-specific filtering and hierarchical representation learning.

Previous studies have demonstrated suboptimal performance of traditional machine learning models in fNIRS data classification tasks. In this study, we established a comparative framework by selecting two representative deep neural network architectures (CNN and LSTM) alongside the widely adopted fNIRS-T and fNIRSNet, against which our novel RDr-C is rigorously evaluated. The CNN model consists of two 2D convolutional layers, two fully connected layers, and a softmax layer. The first convolutional layer has 16 filters, and the second convolutional layer has 32 filters, both with a kernel size of 33 and padding of 1. Each convolutional layer is followed by a 22 max-pooling layer to reduce spatial dimensions. The fully connected layers include one with 128 units and another with 2 units for classification. ReLU activation is applied after each convolutional and fully connected layer, except for the output layer. The LSTM model is composed of four stacked LSTM layers and three fully connected layers. Each LSTM layer has 64 hidden units, and the fully connected layers contain 512, 128, and the number of classes of hidden nodes, respectively. ReLU activation is applied after the first two fully connected layers, and the final layer outputs logits for classification.

Feature reconstruction model for RD-risk learners

To better analyze the features utilized by our RDr-C model for fNIRS data classification, we conducted feature shuffling on the data from each channel corresponding to brain regions in the Brodmann Area (BA) parcellation. By randomly permuting specific features, we assessed their impact on the drop in the model’s recognition accuracy to quantify the importance of each brain region’s features for classification, as shown in Fig. 6. Subsequently, a wavelet transformation was applied to the data of K1 to demonstrate why it is the most significant feature for identifying RD-risk children from their normal peers.

Fig. 6.

Procedure of the key feature extraction.

In fNIRS signal analysis, to uncover the time-frequency characteristics of infra-slow neural oscillations, this study employed continuous wavelet transform methodology. Initially, the mapping relationship between scale and frequency was established based on the central frequency characteristics of the Morlet wavelet: given the mother wavelet’s central frequency and sampling interval , the conversion between scale s and actual frequency was determined by the formula . To cover the physiologically significant infra-slow frequency range of 0.01-0.2 Hz, we computed the corresponding scale boundaries and , then generated 100 equally spaced scale values in logarithmic space using

7

For each brain region channel’s time series x(t), we performed a continuous wavelet transformation using

8

where represents the Morlet wavelet basis function. To obtain the overall activity pattern of brain regions, the wavelet coefficients were spatially averaged across channels within the same brain region:

9

Finally, the average time-frequency characteristics for the experimental and control groups were computed separately as

10

and

11

We revealed functional abnormalities by comparing the between-group differences in time-frequency patterns within the critical 0.01–0.2 Hz frequency band.

Results

RDr-C’s classification performance

To demonstrate the classification performance of the model, we conducted 5 runs 5-fold cross-validation to compare the performance of RDr-C with that of other state-of-the-art models. We selected accuracy, precision, recall, F1-score, and Kappa coefficient as the indices to illustrate the classification performance; the results are summarized in Table 1 (mean pm std %). Our RDr-C achieved the best performance across all evaluation metrics. Furthermore, its superior performance compared to all baseline models demonstrated statistically significant differences (p < 0.001 in all cases, as determined by the Wilcoxon signed-rank test). Notably, unlike prior studies, conventional deep learning baselines such as CNN and LSTM exhibited performance comparable to the widely recognized fNIRS-T model in our dataset. This result underscores the inherent discriminative features of the fNIRS data, further validating the effectiveness of fNIRS technology in RD identification.

Table 1.

Average classification result.

Dataset	Method	Accuracy	Precision	Recall	F1-score	Kappa
Visual Data	CNN	90.44 pm 3.77	90.48 pm 3.95	90.40 pm 3.69	90.32 pm 3.73	0.81 pm 0.08
	LSTM	89.52 pm 3.76	89.80 pm 3.95	89.12 pm 3.92	89.32 pm 3.79	0.78 pm 0.08
	fNIRS-T	91.84 pm 4.02	92.36 pm 3.60	91.42 pm 4.31	91.60 pm 4.18	0.83 pm 0.08
	fNIRSNet	97.24 pm 3.52	97.24 pm 3.52	97.16 pm 3.59	97.16 pm 3.56	0.94 pm 0.07
	RDr-C	99.60 pm 0.71	99.61 pm 0.69	99.60 pm 0.71	99.60 pm 0.71	0.99 pm 0.01
Auditory Data	CNN	91.86 pm 3.13	91.33 pm 4.96	93.32 pm 4.63	92.17 pm 3.05	0.84 pm 0.06
	LSTM	77.38 pm 6.90	76.90 pm 8.12	80.96 pm 14.40	77.96 pm 8.64	0.54 pm 0.14
	fNIRS-T	92.14 pm 8.40	92.87 pm 9.44	93.23 pm 5.28	92.75 pm 6.50	0.84 pm 0.17
	fNIRSNet	97.59 pm 2.33	97.76 pm 3.51	97.57 pm 3.01	97.61 pm 2.36	0.95 pm 0.05
	RDr-C	99.66 pm 0.86	99.41 pm 1.78	99.88 pm 0.59	99.63 pm 0.95	0.99 pm 0.02

LOSO-CV results

To further assess the model’s generalization, we performed LOSO-CV on both visual and auditory datasets. Figure 7 shows the per-subject accuracy of RDr-C compared with baseline models. The overall performance metrics obtained from LOSO-CV are summarized in Table 2. RDr-C achieved an overall accuracy of 89.33% on the visual task and 87.93% on the auditory task, with corresponding Kappa coefficients of 0.78 and 0.76. In contrast, baseline models such as CNN, LSTM, fNIRS-T, and fNIRSNet exhibited significantly lower global accuracies, particularly on the auditory dataset, where fNIRS-T achieved only 78.28% and CNN 63.79%. These results demonstrate that RDr-C maintains strong discriminative ability even when tested on unseen individuals, underscoring its potential for practical early screening.

Fig. 7.

Per-subject classification accuracy of RDr-C and baseline models under LOSO-CV on (a) visual and (b) auditory datasets.

Table 2.

LOSO-CV result.

Dataset	Model	Accuracy	Precision	Recall	F1-score	Kappa
Visual Data	CNN	66.00	64.95	78.75	71.19	0.31
	LSTM	61.33	63.58	64.38	63.98	0.22
	fNIRS-T	75.33	72.63	86.25	78.86	0.50
	fNIRSNet	77.00	76.00	83.12	79.40	0.53
	RDr-C	89.33	85.16	96.88	90.64	0.78
Auditory Data	CNN	63.79	63.47	70.67	66.88	0.27
	LSTM	64.14	64.38	68.67	66.45	0.28
	fNIRS-T	78.28	77.02	82.67	79.74	0.56
	fNIRSNet	77.59	74.01	87.33	80.12	0.55
	RDr-C	87.93	87.58	89.33	88.45	0.76

Training and test curves

The training and test curves of CNN, LSTM, fNIRS-T, fNIRSNet, and RDr-C are shown in Fig. 8 (a) and (b) for visual and auditory data processing, respectively. Compared with all the comparison models, the RDr-C demonstrates the highest and most stable test accuracy. Additionally, it exhibits the fastest convergence speed among all the evaluated models.

Fig. 8.

Training and test curves.

The training curve of the original LSTM shows significant fluctuations in training and validation losses and accuracies, and the validation accuracy is relatively low, indicating that the model has difficulty in stably learning data features and has insufficient generalization ability. This proves the effectiveness of using the attention mechanism to help the LSTM extract key features. Meanwhile, the fluctuations shown by fNIRS-T also demonstrate that using the attention mechanism alone and dividing samples with a sliding window does not achieve the best results in this study. The cortical activity of children with RD is a relatively long-term phased process, which requires the LSTM to capture the dependencies in long-term sequences.

The training and validation losses of CNN decrease relatively fast, but the accuracy never reaches a higher level, reflecting the characteristic of local connections of the CNN model, which is deficient in representing the overall features of sequences with long-term dependencies. The relatively high validation accuracy of fNIRSNet confirms the effectiveness of considering delayed hemodynamic responses in the field of fNIRS data classification. However, as a lightweight model, its limited representational ability leads to relatively high losses and unstable accuracies.

Evaluating RDr-C’s generality via using an open-sourced dataset

To demonstrate the generalizability of our model, we evaluated the RDr-C’s performance on an open-sourced dataset⁵¹ that contains fNIRS signals from eight participants during a mental arithmetic (MA) task, and compared our model with two open-source models (fNIRS-T and fNIRSNet). Table 3 presents the classification results:

Table 3.

Average classification result.

Method	Accuracy	Precision	Recall	F1-score	Kappa
fNIRS-T	88.68 pm 3.76	89.04 pm 3.87	88.64 pm 3.75	88.60 pm 3.79	0.77 pm 0.08
fNIRS-NET	72.60 pm 8.69	78.47 pm 4.40	72.60 pm 8.67	70.05 pm 11.18	0.45 pm 0.17
RDr-C	81.97 pm 5.67	81.98 pm 5.67	82.08 pm 5.70	81.74 pm 5.70	0.64 pm 0.11

Although our model did not achieve the top performance on this dataset, it still delivered satisfactory results (81.97% accuracy), demonstrating reasonable cross-task adaptability. The performance gap between RDr-C and fNIRS-T (88.68%) can be primarily attributed to the substantial domain shift between the source task (audiovisual cognition in children) and the target task (mental arithmetic in adults). It should be noted that there is currently a lack of publicly available fNIRS datasets targeting audiovisual cognition in young children, and no established benchmark data closely aligns with the objective task of this study. The public dataset used primarily involves adult subjects, and its cognitive paradigm, task difficulty, and neural activation patterns differ substantially from the reading cognition task in children that is the focus of our work. This discrepancy between task domains is the main reason for the fluctuation in model performance. Nevertheless, RDr-C’s ability to exceed 81% accuracy under such cross-task conditions indicates a degree of robustness in feature extraction. However, this result also reveals a limitation of the current model in cross-task generalization when the target domain diverges significantly from the training domain. Addressing this limitation will be a key focus of our future work.

Feature extraction, visualization, and explanation

The top 5 most significant features identified are presented in Table 4.

Table 4.

Feature importance.

Dataset	Brain Region	Channels	Importance (%)
Visual Data	K1	41,49	5.20
	A1	35,36,42,43	4.00
	J2	6,10,11	3.40
	J1	39,40,47	1.93
	C1	37,44	1.67
Auditory Data	K1	41,49	7.86
	H2	18,19,24,26	7.31
	C1	37,44	4.21
	A2	12,13,15,16	3.59
	B2	21,22	2.07

A wavelet transformation was applied to the effect of K1 to demonstrate why it is the most significant feature for identifying RD-risk children from their normal peers. The visualizations, as shown in Fig. 9a,b illustrate different patterns that K1 plays in RD-risk and normal children.

Fig. 9.

Heatmap comparison of region K1.

Feature extraction ability comparison among different models

We used t-Distributed Stochastic Neighbor Embedding (t-SNE), a non-linear dimensionality reduction technique, to project the high-dimensional features into a two-dimensional space for visualization. Figure 10a illustrates the feature visualization of the RDr-C, while Fig. 10b–e depict the visualizations of the four baseline models for comparison on the visual dataset (Fig. 11 on the auditory dataset). The features learned by the RDr-C are clearly separated into two compact clusters, each corresponding clearly to one experimental class. This pronounced separability and tight intra-cluster grouping indicate that our model has acquired highly discriminative features. In contrast, the t-SNE visualizations of LSTM and fNIRS-T, though showing some discernible separation, exhibit significant overlap between classes. This observation confirms that incorporating an attention mechanism is crucial for the LSTM to extract salient features and identify important long-term dependencies, a capability that is fully realized in our RDr-C. Similarly, while fNIRSNet demonstrates a certain degree of separability, its clusters are considerably more scattered than those of RDr-C. This comparison substantiates that the RDr-C proposed in this study achieves the best classification performance by learning the most distinct and robust feature representations among all models evaluated.

Fig. 10.

The t-SNE visualization of the visual dataset.

Fig. 11.

The t-SNE visualization of the auditory dataset.

Ablation study

To investigate the critical roles of the GCN module, bidirectional LSTM module, and the MSA module in the performance of the RDr-C, this study conducted ablation studies to reveal the contributions of each component by comparing core metrics across different model configurations. The experimental results shown in Table 5 indicated that removing either the GCN, the bidirectional LSTM, or the MSA module from the complete model led to a decline in model performance, with varying degrees of reduction in all metrics. This confirms that each component plays a necessary and non-redundant role in the model’s ability to classify fNIRS signals related to reading difficulty. Notably, the model without the MSA module still exhibited significantly superior classification performance compared to the traditional LSTM model(as shown in Table 1), further confirming that the fNIRS data of RD-risk children exhibit global perspective feature patterns. Conversely, the complete RDr-C, which integrates GCN, BiLSTM, and MSA, achieved optimal performance. This demonstrates a clear synergistic effect: the GCN captures spatial connectivity, the BiLSTM models bidirectional temporal dynamics, and the MSA highlights the most salient temporal dependencies across the sequence. The ablation results validate that this integrated architecture is fundamentally responsible for the model’s high classification accuracy, highlighting the value of combining spatial, temporal, and attention-based mechanisms for fNIRS-based RD-risk identification.

Table 5.

Ablation study results.

Dataset	Method	Accuracy	Precision	Recall	F1-score	Kappa
Visual Data	Without GCN	98.40 pm 1.63	98.39 pm 1.64	98.45 pm 1.60	98.40 pm 1.63	0.97 pm 0.03
	Without LSTM	97.99 pm 2.04	98.04 pm 2.00	97.97 pm 2.05	97.98 pm 2.05	0.96 pm 0.04
	Without MSA	98.40 pm 1.83	98.41 pm 1.81	98.45 pm 1.77	98.40 pm 1.83	0.97 pm 0.04
	RDr-C	99.60 pm 0.71	99.61 pm 0.69	99.60 pm 0.71	99.60 pm 0.71	0.99 pm 0.01
Auditory Data	Without GCN	98.48 pm 1.64	98.52 pm 1.62	98.47 pm 1.64	98.48 pm 1.64	0.97 pm 0.03
	Without LSTM	97.03 pm 2.04	97.17 pm 1.97	97.01 pm 2.04	97.03 pm 2.05	0.94 pm 0.04
	Without MSA	98.48 pm 1.64	98.53 pm 1.60	98.49 pm 1.63	98.48 pm 1.64	0.97 pm 0.03
	RDr-C	99.66 pm 0.86	99.41 pm 1.78	99.88 pm 0.59	99.63 pm 0.95	0.99 pm 0.02

Discussion

This paper proposes a deep learning classification framework named “RDr-C” based on a two-layer GCN, a BiLSTM and an MSA. On the fNIRS dataset, i.e., the HbO changes collected from the RD risk children when they were performing the basic visual words and auditory sounds recognition and discrimination tests, the framework achieves an accuracy of approximately 99.60% on both datasets, substantially outperforming other existing models. The LOSO-CV results further confirm that RDr-C captures hemodynamic patterns that generalize across individuals. While the global accuracies (89.33% and 87.93%) are lower than those obtained in the 55 cross-validation, they remain higher than all baseline models. This suggests that although individual variability exists, the features learned by RDr-C are sufficiently robust to distinguish at-risk children from normal peers even when tested on unseen subjects. The t-SNE visualizations in Figs. 10 and 11 also support this observation, showing clear separation between classes in the feature space. Existing RD identification frameworks are designed exclusively for diagnosed RD students, not for those at early risk of RD. In fact, the participants’ accuracy performance in our study showed no significant differences either in auditory cognition (mean RDrisk = 7.80, mean normal = 8.35; two-tailed p = 0.19) or visual cognition (mean RDrisk = 8.37, mean normal = 9.07; two-tailed p = 0.27). This indicates that these students exhibit neither significant auditory developmental delays nor notable deficits in visual cognitive development. This further underscores the importance of our framework in the early identification of students at risk of RD. Our feature extraction and reshaping model reveals that the reading difficulties observed in these borderline RD students stem primarily from asynchronization of the multiple cognitive functions, specifically manifested as poor control of fine finger movements in task-related tasks (i.e., performed poorly in the visual recognition test, but not in the auditory recognition test).

Habib⁵² proposed three triggering mechanisms of dyslexia: attentiveness disturbance (visual–attentional dyslexia), language disturbance (phonological dyslexia), and motor disturbance (a dyspraxic form of dyslexia). Based on the RDr-C’s superior recognition ability, key features that dominate this deep-learning architecture to recognize the children at RD risk from normal peers were selected. A wavelet transformation was applied to demonstrate why K1 is the most significant feature for identifying RD-risk children. As visualized in Fig. 9a,b, K1 exhibits markedly distinct time-frequency activation patterns between RD-risk and normal children.

The visualization of the role that K1 plays in normal children during visual tests suggests their fine motor control ability of the right fingers. This is evidenced by the thermographic red allocated on the middle frequency band (it is meaningful as the data has bypassed the high-frequency noises) regularly appearing alongside the timeline within a single subtask. In contrast, their RD-risk counterparts showed poor motor-sensory control ability. However, this distinct pattern was not observed in the data from the auditory tests.

Although it has been well known that children with dyslexia generally suffer from motor coordination problems^53,54, the visualization of the kinetic patterns that RD-risk children showed in this study displayed another role that K1 plays in identifying RD-risk children from normal children. Most importantly, the atypical kinetic patterns were not directly discernible from behavioral results, but were found through a deep-learning framework that reshaped the input features to extract them.

The abnormal kinesthetic brain function revealed in this study may characterize a context-dependent neuro-coordination dysfunction rather than a general sensorimotor disorder. Specifically, it may reflect a difficulty in cognitive resource allocation or real-time coordination within the sensorimotor feedback system used to maintain fine motor stability when individuals perform high-cognitive-load visual or auditory processing tasks. This difficulty is fundamentally different in nature from the pervasive, top-down sustained inhibitory control deficits commonly found in ADHD^55,56. Therefore, when they performed a less demanding task, i.e., in the auditory test, this motor dysfunction disappeared. This finding also aligns with the previous conclusions that sensorimotor difficulties are not stably associated with dyslexia^57,58, as they represent a dynamic dysregulation under specific cognitive states rather than a stable, cross-context ability deficit.

Although it has been widely acknowledged that the assessments targeting cognitive functions such as working memory and reasoning have limited value for identifying dyslexia and guiding intervention¹⁴, our results provide further empirical support for this position: none of the 150 children in our sample showed significant cognitive dysfunction even though they had displayed obvious dyslexic features. Fletcher and Miciak explicitly argued that comprehensive cognitive assessments are not necessary for the identification and treatment of learning disabilities. Empirical studies have repeatedly demonstrated the poor reliability of identification methods based on “patterns of cognitive strengths and weaknesses”^16,18, and interventions targeting cognitive abilities have negligible effects on improving reading outcomes⁵⁹. A meta-analysis by Burns et al.⁶⁰ further showed that after controlling for prior reading scores, cognitive measures accounted for only 1–2% of the variance in reading growth, suggesting its weak far-transfer effect.

Finn et al.⁶¹ found that typical abnormalities in dyslexia manifest as impaired coordination between brain networks (e.g., weakened visual-prefrontal connectivity, delayed left-lateralization of language networks), rather than isolated activation deficits in specific brain regions. Collectively, this evidence suggests that the core difficulties of dyslexia may stem less from the higher-order cognitive dysfunction of the prefrontal system. The non-prefrontal kinesthetic anomaly discovered during the visual tests, not during the auditory tests, not only supports this hypothesis, but also echoes the view from neuroimaging studies that dyslexia is characterized by task-dependent neural abnormality patterns.

Early neuroimaging studies^62,63 found that during visual word processing, the activation in the left IFG and left middle temporal gyrus (MTG) of children with RD showed weaker correlations with semantic association strength, and the bottom-up effective connectivity from the fusiform gyrus (FG) to the MTG was significantly reduced; whereas during auditory word processing tests, their neural activity abnormalities might be less important than during visual tests. This study broadens this view by demonstrating that the kinesthetic anomaly in children with RD or RD-risk is not a single, stable sensorimotor deficit, but more likely a context-dependent, real-time coordination dysfunction in the neuro-motor feedback system when performing reading-related tasks of different difficulty levels. This dysfunction may be related to the differential demands for attention, perception, and motor coordination resources imposed by task complexity. Therefore, the kinesthetic anomaly revealed in this study is likely not a general sensorimotor disorder indicator for RD recognition, but rather a state-dependent, accompanying feature embedded within the patterns of RD-risk children when processing reading-related tasks with different brain functions’ cooperation and coordination.

In summary, the kinesthetic anomaly discovered in those RD-risk children is distinct in nature from higher-order cognitive control deficits, does not primarily involve PFC in its neural basis, and exhibits different task-dependent patterns. This suggests that children displaying reading difficulty features may not show basic and independent cognitive dysfunction; instead, they may show a specific, context-dependent inefficiency in the different brain functions’ cooperation and coordination. This dissociation among brain functions may result in a deficit in the fine-tuning of sensorimotor feedback during the execution of cognitive tasks. This interpretation resonates strongly with findings from previous studies on Parkinson’s disease and ADHD⁴⁰ as well as depression⁴⁴, which emphasize that neurodevelopmental disorders typically manifest as context-dependent coordination failures among brain networks rather than isolated regional deficits. This perspective further motivates our adoption of deep learning approaches to capture dynamic, task-specific coordination patterns in fNIRS signals, enabling a more nuanced understanding of the nature of reading difficulties.

Conclusion and limitations

In summary, this study establishes a robust framework for early RD identification using fNIRS and the RDr-C, demonstrating its potential as a non-invasive, cost-effective tool for early screening. Beyond classification performance, the model’s ability to implicitly capture spatiotemporal brain activity patterns provides data-driven insights that complement and enrich the neuroscience theory of dyslexia.

Although the RDr-C has achieved promising results in identifying RD-risk children from fNIRS data collected in this study, the research still has several limitations. First, RDr-C’s 81.97% accuracy on an adult mental arithmetic dataset, which is lower than fNIRS-T’s 88.68%, exposes its limited cross-task generalization from children’s audiovisual reading to adults’ mental arithmetic. A more direct evaluation was constrained by the absence of publicly available fNIRS datasets targeting audiovisual cognition in young children. Second, concerning the common challenges in fNIRS data processing, only fNIRS and behavioral data were used to predict RD-risk children; more data from different modalities should be incorporated to make a feature augmentation and complementation. Moreover, given the relatively small sample size, the current findings should be regarded as preliminary, and larger samples are required to validate the generalizability of RDr-C. Third, while the GCN, LSTM model with multi-head attention architecture delivers superior classification performance, it introduces substantial computational complexity. Fourth, the black-box nature of deep learning models leads to limited interpretability of results, which has not yet been unified with data analysis outcomes such as brain region activation from a statistical perspective.

Now, we have started the following works: first, replicating the same experiment on more children to expand the participant sample and augment the dataset with additional modalities, such as handwriting images and eye-tracking data; and second, initiating longitudinal studies to validate the model’s generalizability in predicting RD children on a real timeline. These efforts aim to bridge the gap between laboratory research and clinical practice, ultimately enhancing early detection and facilitating personalized support for children with reading difficulties.

Author contributions

P.Y. conceptualised the study and wrote the manuscript. P.Y. and Y.D. conducted the experiments, collected the data, and processed part of the raw data. Y.G. and Y.Z. provided professional guidance and participated in designing the evaluation and intervention methods and hold the teaching reform and research projects of Sichuan Normal University [JWC20240116, JWC20240107]. X.Z. designed, monitored and assessed the Chinese Character Phonetic Notation Test for all participants. L.W. and Z.L. analysed the data, and Z.L. holds the funding “Research project of Ministry of Education of China [23YJC880062]”. D.W. supervised the fNIRS-related processes and provided financial support through the funding “Major Science and Technology Special Program of Jiangsu Province [BG2024025]”. J.Y. was the project administrator and provided overall supervision for the entire research.

Funding

This work is supported by the Major Science and Technology Special Program of Jiangsu Province [BG2024025]; Research project of Ministry of Education of China [23YJC880062]; the teaching reform and research project of Sichuan Normal University [JWC20240107; JWC20240116].

Data availability

The data is only available on request due to the individual privacy agreement made with parents.

Declarations

Competing interests

The authors declare no competing interests.