Introduction

Perioperative neurocognitive disorders (PND) are a common neurological complication among elderly patients following surgery, often leading to increased postoperative complications, prolonged hospital stays, and higher hospital costs. PND significantly affects postoperative recovery and quality of life in elderly patients. The primary manifestations of postoperative cognitive dysfunction include impairments across various cognitive domains such as learning, memory, executive function, attention, visuospatial skills, orientation, and visual search ability [1].

Recent research indicates that visual cognitive impairment represents the most prevalent form of dysfunction in PND patients, as evidenced by significant changes in visual-cognition-related assessments [2]. Our previous studies demonstrated that PND patients exhibited marked declines in performance on the Digit Symbol Substitution Test (DSST) and Trail Making Test Part A (TMT-A), indicating notable impairments in visual cognitive functions [3, 4]. Given the structural and functional decline in the visual system of elderly patients, they are particularly vulnerable to visual transmission impairments following surgery, which may further exacerbate cognitive function. Consequently, it is crucial to identify risk factors for postoperative visual cognitive impairment and develop clinically practical, accurate, and easily implementable predictive models for early screening of high-risk patients. Such models could potentially reduce the incidence of postoperative visual cognitive impairment and improve the overall quality of life in elderly patients.

In recent years, resting-state functional magnetic resonance imaging (rs-fMRI) has been widely applied in studies investigating neurodegenerative diseases [5,6,7]. Rs-fMRI facilitates the analysis of brain states and the exploration of neural network mechanisms by characterizing local neural activity and connectivity patterns across the entire brain network. In a previous study, we observed a significant decrease in default mode network connectivity with brain regions involved in visual processing among patients with postoperative cognitive dysfunction [4]. This finding suggests potential impairments in cognitive functions related to visual signaling, including spatial recognition, facial and object recognition, visual word processing, and long-term visual memory. Therefore, the present study aims to further investigate visual cognitive impairment in PND by using DSST and TMT-A assessments. We will explore distinct alterations in brain networks associated with visual cognitive function by constructing resting-state functional connectivity network and employing graph theory analysis. Additionally, we aim to develop a predictive model for postoperative visual cognitive dysfunction using machine learning algorithms based on these brain network features.

Materials and methods

Ethics approval and participants

The study, approved by the Ethics Committee of Huadong Hospital Affiliated to Fudan University (approval number: 20170020), was registered in the Chinese Clinical Trial Registry (ChiCTR-DCD-15006096, 16th/March/2015). A total of 74 patients were enrolled in the study at Huadong Hospital, Fudan University, from September 2017 to February 2019. Written informed consent was obtained from all subjects participating in the trial before enrollment.

Inclusion criteria included (1) Non-cardiac major surgery (expected duration of surgery > 2 h); (2) Age ≥ 60 years; (3) American Society of Anesthesiologists (ASA) classification I-III; and (4) Right-handed. Exclusion criteria included (1) Education duration < 6 years; (2) Baseline mini-mental state examination (MMSE) < 24 points; (3) Existence of mental, nervous system, or liver diseases, kidney dysfunction, and a history of cerebrovascular accidents; (4) Use of psychotropic medications, such as diazepam or anti-anxiety drugs; (5) History of cardiac or cranial-cerebral surgery; (6) Patients with language communication, vision, or hearing impairments; (7) Patients with metal implants in the body unsuitable for MRI.

Cognitive function assessment

TMT-A and DSST assessments were performed 1 to 3 days before surgery and 7 to 14 days after surgery to evaluate cognitive functions related to vision processing. The DSST assesses cognitive functions such as visuomotor coordination, visuospatial scanning, and attention, while the TMT-A evaluates visuospatial abilities and visual search speed.

For each test, we compared the changes (Δx) from baseline scores to follow-up scores in patients. To correct the learning effect, we recruited 30 volunteers, age- and education duration- matched to the surgical patients, who completed the same neurocognitive tests as those performed on the surgical patients. For volunteers, we calculated the changes from baseline scores to follow-up scores to obtain the average learning effects (Δxc) and the standard deviations (SDΔxc) in each test. The Z score was built as follows: (Δx -Δxc)/SD(Δxc). According to the Z-score criterion, a Z-score greater than 1.96 in either the TMT-A or DSST indicates the the presence of postoperative visual cognitive impairment.

MRI data acquisition and preprocessing

On the day prior to surgery, patients underwent head MRI scans using a SIEMENS Skyra 3.0T MRI system. Before the scan, patients’ heads were securely immobilized, and they are instructed to maintain a calm, supine position with their eyes closed. High-resolution 3D T1-weighted structural images were acquired using the MPRAGE sequence with the following scanning parameters: TR = 1900 ms, TE = 3.57 ms, slice thickness = 1.0 mm, slice count = 176, flip angle = 9°, voxel size = 1 mm × 1 mm × 1 mm. The rs-fMRI images were acquired using an echo-planar imaging sequence with the following scanning parameters: TR = 3000 ms, TE = 30 ms, slice count = 35, slice thickness = 4.0 mm, voxel size = 3.4 mm × 3.4 mm × 3.4 mm, flip angle = 90°. During the fMRI imaging, 105 volumes were obtained.

The raw rs-fMRI data were preprocessed using the SPM12 and Restplus software packages [8]. To minimize potential noise interference stemming from the imaging instrument, the first five-time points for each subject were excluded. The remaining data underwent time correction, head movement correction, spatial normalization, and spatial smoothing. Linear trend removal was then applied to the time series, followed by filtering in the range of 0.01 to 0.08 Hz to reduce the effects of low-frequency drift and high-frequency noise. Cerebrospinal fluid and white matter signals were regressed out as confounding covariates. Finally, the spatial volumes were normalized to the Automated Anatomical Labeling (AAL-90) template, and the data were divided into 90 regions of interest (ROIs) [9].

Brain network construction and feature extraction

To construct a brain functional connectivity network, rs-fMRI images were first transformed into brain activity signals. This process involved calculating the average time series signal for each brain region by aggregating the voxel time series signals within that region. The resulting average time series signal reflects the functional activity changes in each specific brain region.

Next, the time series data was partitioned into multiple subsequences along the temporal dimension using a sliding window approach. Let \(\:t\) represent the length of the original signal, \(\:w\) denote the sliding window size, and \(\:s\) represent the sliding step size. The number of resulting subsequences, denoted as \(\:K\) is calculated as \(\:K=\left(\frac{t-w}{s}\right)+1\). Each subsequence \(\:{x}_{i}^{j}\left(n\right)\) corresponds to the \(\:{i}^{th}\) ROI in the \(\:{j}^{th}\) subsequence of case \(\:n\). The set of all these subsequences for the 90 brain regions for case is represented as \(\:{X}^{j}\left(n\right)=\{{x}_{1}^{j}\left(n\right),\:{x}_{2}^{j}\left(n\right),\:\dots\:,\:{x}_{90}^{j}(n\left)\right\}\) for case \(\:n\).

To obtain the functional connectivity (FC) matrix \(\:{A}^{j}\left(n\right)\) corresponding to the \(\:{j}^{th}\) subsequence of case \(\:n\), we calculated the Pearson correlation strength between different ROIs [10]:

$$\:{{a}}_{{p},{q}}^{{j}}\left({n}\right)=\text{{corr}}({\text{x}}_{{p}}^{{j}}\left({n}\right),{\text{x}}_{{q}}^{{j}}\left({n}\right))$$

where, \(\:p\) and \(\:q\) represent two distinct ROIs. Each ROI region was treated as a node, and the correlation coefficient’s strength between regions serves as an edge. This results in the construction of \(\:K\) dynamic brain functional connectivity networks for each individual case. In our experimental setup, the sliding window size and sliding step size were set to 90 and 1, respectively. After obtaining the dynamic functional connectivity (DFC) matrices, we averaged these FC matrices across the temporal dimension to yield a single mean connectivity matrix, which was then used for subsequent feature extraction.

Based on the adjacency matrix \(\:A\), both local and global features of brain network connectivity were computed. First, five local features were calculated. Specifically:

Assuming that \(\:n\) represents the node of the brain network and \(\:1\) represents the connected edge, the degree of the \(\:{i}^{th}\) brain region can be expressed as follows [11]:

$$\:{k}_{i}={\sum\:}_{j\in\:N}{a}_{ij}$$

where \(\:{a}_{ij}\) represents the connection state between the nodes \(\:i\) and \(\:j\). \(\:{a}_{ij}=1\) means that there is a direct connection between two nodes. Otherwise \(\:{a}_{ij}=0.\)

The feature path length serves as a metric to assess the average connectivity within a network and gauge its overall routing efficiency. It is denoted by \(\:{L}_{i}\) representing the average shortest path length of the \(\:{i}^{th}\) brain region, and its formula is as follows [12]:

$$\:{L}_{i}=\frac{{\sum\:}_{j\in\:N,j\ne\:i}{d}_{ij}}{n-1}$$

where \(\:{d}_{ij}\) represents the shortest path between nodes \(\:i\) and \(\:j\).

The clustering coefficient is used to measure the local density or the degree of collectability of the network. The clustering coefficient of the \(\:{i}^{th}\) brain region can be expressed as:

$$\:{C}_{i}=\frac{2{t}_{i}}{{k}_{i}\left({k}_{i}-1\right)}$$

where \(\:{t}_{i}\) refers to the number of triangle connections around the node \(\:i.\)

The global efficiency is a measure of parallel information transfer in a network. It is defined as the inverse of the harmonic mean of the shortest path length between each pair of nodes. The efficiency of the \(\:{i}^{th}\) brain region is \(\:{E}_{i}\), and the formula is as follows [13, 14]:

$$\:{E}_{i}=\frac{{\sum\:}_{{j}{\in}{N,j\ne\:i}}{{d}_{ij}}^{-1}}{n-1}$$

The local efficiency of the \(\:{i}^{th}\) brain region is defined as follows:

$$\:{E}_{local,i}=\frac{{\sum\:}_{i,h\in\:N,j\ne\:i}{a}_{ij}{a}_{jh}{\left[{d}_{jh}\left({N}_{i}\right)\right]}^{-1}}{{k}_{i}\left({k}_{i}-1\right)}$$

where \(\:{d}_{jh}\left({N}_{i}\right)\) is the shortest path between nodes \(\:j\) and \(\:h\) including node \(\:i\).

Subsequently, seven global features related to brain network connectivity were computed. Five of these global features encompassed the global averages of the previously mentioned five local features, representing the averages across all 90 brain regions. The remaining two global features were modularity and entropy. Modularity in graph theory characterizes the potential communities that can emerge within a network and is defined as:

$$\:Q={{\Sigma\:}}_{u\in\:M}\left[{e}_{uu}-{\left({{\Sigma\:}}_{v\in\:M}{e}_{uv}\right)}^{2}\right]$$

where the brain network is divided into a collection of non-overlapping modules denoted as \(\:M\), and \(\:{e}_{uv}\) represents the ratio of all links connecting nodes within module \(\:u\) to nodes within module \(\:v\). Since various network partitions yield distinct modularity values, we employed the Louvain algorithm to identify the optimal network partition to obtain the highest value. Furthermore, entropy serves as a measure of the complexity of brain network connectivity, where a higher entropy value signifies more intricate brain network connectivity.

Feature selection and classification model construction

While numerous brain network features were extracted, not all of them contributed effectively to classification prediction models. Some redundant features increased the model’s computational complexity and risk of overfitting, especially when the feature dimension significantly exceeds the number of training samples. Therefore, before constructing the classification model, we implemented a feature screening process to eliminate redundant features. We employed the following sparse representation-based feature screening model [15]:

$$\begin{aligned}\:\widehat{f}=&\:{argmin}_{f}{||y-{f}^{T}W||}_{F}^{2}+{\lambda\:}_{1}tr\left({f}^{T}WL{W}^{T}f\right)\cr &+{\lambda\:}_{2}{||f||}_{\text{2,1}}\end{aligned}$$

where, \(\:y\) represents the gold label, and \(\:W\) signifies the sample feature set. \(\:L\) denotes a Laplacian graph, and \(\:{\lambda\:}_{2}\) corresponds to the regularization term. After calculating the representation coefficient \(\:\widehat{f}\), feature importance was determined by ranking the absolute values of its elements. Specifically, the higher the absolute value, the more significant the feature. Subsequently, a sequential forward selection algorithm was applied to determine the final feature subset for classification. This algorithm systematically adds ranked features one by one to the feature subset, conducts classifier training and validation, and ultimately selects the subset that yields the highest classification performance as the final result.

Using the selected features, a sparse representation classifier (SRC) [16] was constructed to predict postoperative visual cognitive impairment. SRC is a type of K-Nearest Neighbor (KNN) classification algorithm. It uses samples in the training set to sparsely represent the test sample, and determines which of the positive and negative sample sets is closer to the test sample by comparing the residual representation of the test sample between the positive and negative sample sets in the training set (i.e., the residual representation is smaller). Finally, the category of the test sample is judged as the closer category. The sparse representation model can be expressed as:

$$\:\widehat{\beta\:}=\underset{\beta\:}{argmin}{||f-F\beta\:||}_{2}^{2}+\gamma\:{||\beta\:||}_{p}$$
(1)

where \(\:f\) represents the test sample feature,\(\:F=[{F}_{+},{F}_{-}]\),\(\:{F}_{+}=[{f}_{1},{f}_{2},\cdots\:,{f}_{m}]\) is the positive sample feature set in the training sample, \(\:m\) is the number of positive samples. \(\:{F}_{-}=[{f}_{1},{f}_{2},\cdots\:,{f}_{n}]\) is the negative sample feature set in the training sample, \(\:n\) is the number of negative samples, \(\:\gamma\:\) is the sparse representation control parameter. After obtaining the sparse representation coefficient \(\:\beta\:\), calculate the residual:

$$\:{r}_{c}\left(f\right)=f-F{\delta\:}_{c}\left(\beta\:\right),\:c=\text{1,2}$$
(2)

\(\:{\delta\:}_{c}(\cdot\:)\)indicates the coefficient corresponding to the selected type of feature. The final category of the sample to be tested is \(\:Id\left(f\right)=\)\(\underset{c}{argmin}{r}_{c}\left(f\right)\). Unlike conventional classifiers trained using parameters, such as support vector machines (SVM), sparse representation is a non-parametric classifier. During the classification process, it does not require training a large set of model parameters, thereby reducing the model’s reliance on sample size. This approach was compared to three other classifiers, namely, SVM, random forest (RF) and XGBoost, to evaluate model performance.

Our method contains the following important hyperparameters: sliding window size \(\:w\) and sliding step size \(\:s\) in the construction of brain network adjacency matrix, \(\:{\lambda\:}_{1}\) and \(\:{\lambda\:}_{2}\) in the feature selection model (Eq. 1), and γ in the sparse representation classifier. For \(\:w\) and \(\:s\), we set them to 90 points and 1 point respectively based on the experience of the existing reference [17]. For \(\:{\lambda\:}_{1}\) and \(\:{\lambda\:}_{2}\), we determine their values through experiments. Specifically, we set \(\:{\lambda\:}_{1}\) and \(\:{\lambda\:}_{2}\) to any one of the three sets of parameters: 0.1, 0.5 and 1, and then compare the performances of the models under 9 combinations of \(\:{\lambda\:}_{1}\) and \(\:{\lambda\:}_{2}\). Finally, we determined \(\:{\lambda\:}_{1}\) and \(\:{\lambda\:}_{2}\) to be 0.1 and 1 respectively. For \(\:\gamma\:\), we also selected the best one from three parameters of 0.005, 0.01 and 0.05 through experimental comparison, and finally determined \(\:\gamma\:\) to be 0.01. We have added the explanation of hyperparameter selection in the last paragraph of the method section.

To assess the model’s performance, we randomly partitioned the dataset into a cross-validation set and an independent test set, using a 2:1 ratio. The K parameter in cross-validation was set to 10. The evaluation metrics included accuracy (ACC), sensitivity (SEN), specificity (SPE), receiver operating characteristic (ROC) curve analysis, and the area under the curve (AUC). Statistical analyses were conducted using R software (version 3.5.2; www.r-project.org) and Python (version 3.7, www.python.org). Univariate analyses were conducted to compare the clinical characteristics of the two cohorts, using either the Mann-Whitney U test or the chi-square test, depending on appropriateness.

Results

Subject characteristics

A total of 74 patients were enrolled in this study. Among them, 16 patients experienced postoperative visual cognitive impairment, resulting in an incidence rate of 21.6%. There were no statistically significant differences observed in terms of age, gender, education, height, weight, body mass index (BMI), smoking history, or surgical history between the two groups (Table 1).

Table 1 Subject characteristics
Full size table

Functional connectivity network construction

Figure 1 provides a visual comparison of the functional connectivity networks between the postoperative visual cognitive impairment group and the control group. The functional connectivity displayed here represents the result obtained by averaging the dynamic functional connectivity (DFC) matrices across the temporal dimension. To facilitate visual comparison of brain network activity between the two groups, Fig. 1A and B present the brain network connectivity matrices of patients with and without postoperative visual cognitive impairment, respectively. To further enhance the comparison, Fig. 1C displays the difference value in average connectivity matrices between the two groups. Additionally, Fig. 1D highlights 15 pairs of brain regions where the absolute difference value exceeds 0.18. Detailed information on the specific pairs of brain regions and their corresponding differences can be found in Table 2. Notably, the largest difference in functional connectivity was observed between the right middle frontal gyrus (MFG. R) and the right middle orbital frontal gyrus (ORBmid. R).

Fig. 1
figure 1

Whole-Brain functional connectivity networks (mean connectivity of dynamic functional connectivity). (A) Average whole-brain functional connectivity network matrix for patients experiencing postoperative visual cognitive impairment; (B) Average whole-brain functional connectivity network matrix for patients without postoperative visual cognitive impairment; (C) Mean difference value in functional connectivity between the two groups; (D) Pairs of brain regions with an absolute difference value in functional connectivity exceeding 0.18

Full size image
Table 2 The absolute value of functional connectivity difference is greater than 0.18 in 15 pairs of brain regions
Full size table

Graph theory analysis

A total of 457 features were extracted for each patient, including 450 local features (derived from the analysis of 90 brain regions, each yielding five local features) and seven global features. Among these features, 16 features showed statistically significant differences between the two groups (P < 0.05). Figure 2 provides a visual representation of the data distribution for these 16 features, illustrating their association with 10 specific brain regions: Right Inferior Occipital Gyrus (IOG.R), Left Medial Superior Frontal Gyrus (SFGmed.L), Left Cuneus (CUN.L), Left Amygdala (AMYG.L), Left Triangular Part Inferior Frontal Gyrus (IFGtriang.L), Left Precentral Gyrus (PreCG.L), Right Posterior Cingulate Gyrus (PCG.R), Right Temporal Pole Middle Temporal Gyrus (TPOmid.R), Left Inferior Occipital Gyrus (IOG.L), Left Angular Gyrus (ANG.L).

Fig. 2
figure 2

Data distribution of 16 graph features showing statistically significant differences between the two groups. The postoperative visual cognitive impairment group is represented in green, while the non-postoperative visual cognitive impairment group is depicted in blue

Full size image

Figure 3 further delineates the associations, revealing that three local features correspond to IOG.R, two to IOG.L, two to SFGmed.L, two to CUN.L, and two to AMYG.L. The remaining brain regions each have only one local feature corresponding to them.

Fig. 3
figure 3

Ten brain regions associated with 16 graph-theoretic features. AMYG, amygdala; ANG, angular gyrus; CUN, cuneus; IFGtriang, triangular part inferior frontal gyrus; IOG, inferior occipital gyrus; PCG, posterior cingulate gyrus; PreCG, precentral gyrus; SFGmed, medial superior frontal gyrus; TPOmid, temporal pole middle temporal gyrus; L, left; R, right

Full size image

Feature selection and model construction

Following the sparse representation-based feature selection method, a set of 29 features was chosen for the final machine learning model. These 29 features correspond to 27 distinct brain regions: Left dorsolateral superior frontal gyrus (SFGdor.L); Right Paracentral Gyrus (PCL.R); Left Insular Gyrus (INS.L); Left Hippocampus (HIP.L); Left Transverse Temporal Gyrus (HES.L); Right Inferior Opercular Frontal Gyrus (IFGoperc.R); Right Rectus Gyrus (REC.R); Left Parahippocampal Gyrus (PHG.L); Left postcentral gyrus (PoCG.L); Right Supramarginal Gyrus (SMG.R); Left Precuneus (PCUN.L); Left Thalamus (THA.L); Right Middle Temporal Gyrus (MTG.R); Left Precentral Gyrus (PreCG.L); Left Anterior Cingulate Gyrus (ACG.L); Left Amygdala (AMYG.L); Left Cuneus (CUN.L); Right Dorsolateral Superior Frontal Gyrus (SFGdor.R); Left Superior Occipital Gyrus (SOG.L); Left Inferior Occipital Gyrus (IOG.L); Left Supramarginal Gyrus (SMG.L); Right Precentral Gyrus (PreCG.R); Right Triangular Part Inferior Frontal Gyrus (IFGtriang.R); Left Medial Superior Frontal Gyrus (SFGmed.L); Right Anterior Cingulate Gyrus (ACG.R); Left Superior Temporal Gyrus (TPOsup.L); Right Middle Temporal Gyrus (TPOmid.R).

As shown in Table 3, the proposed method yielded promising results, achieving an AUC of 0.918, ACC of 0.898, SEN of 0.900, and SPE of 0.897 on the cross-validation set. Moreover, on the independent test set, our approach achieved an AUC of 0.877, ACC of 0.840, SEN of 0.833, and SPE of 0.842. Furthermore, when compared to the RF, SVM and XGBoost classifiers, the SRC classifier not only attained the highest AUC and ACC but also demonstrated more balanced performance in terms of SEN and SPE. Figure 4A shows the ROC curves of our proposed model on both the cross-validation set and the independent test set. Figure 4B presents a comparative analysis of the four classifiers—SRC, RF, SVM and XGBoost —on the independent test set. Finally, Fig. 4C and D depict the relationship between the model’s prediction scores and the actual categories of the samples on the cross-validation set and the independent test set, respectively.

Table 3 Comparison of prediction results of different models
Full size table
Fig. 4
figure 4

Performance evaluation of the prediction models. (A) ROC curves for the cross-validation set and independent test set of the SRC model. (B) ROC curves of the RF, SVM, XGBoost and SRC models on the independent test set. (C) Prediction score distribution of the SRC model for the cross-validation set. (D) Prediction score distribution of the SRC model for the independent test set. Yellow and blue bars representing the two true categories of the samples. ACC, accuracy; AUC, the area under the receiver operating characteristic curve; RF, random forest; SEN, sensitivity; SPE, specificity; SRC, sparse representation classifier; SVM, support vector machine

Full size image

Figure 5 presents the decision curve for the proposed model [18]. From this figure, it is clear that the model can achieve favorable Net Benefit across a broad range of Threshold Probability values. Figure 6 provides a categorical analysis of the 29 features that were ultimately employed in the model. Notably, only local features derived from the adjacency matrix were incorporated into the final feature set for modeling. Among these five local features, the local clustering coefficient features constitute the majority, accounting for 10 out of the 29 selected features. Then we selected the 10 most important features from these features and drew the SHAP (Shapley Additive Explanations). It can be seen from the SHAP that the points corresponding to these features are generally distributed on both sides of the midline, and the difference in the feature values on both sides of the midline is significant, indicating that these features have a greater impact on the model output. In addition, for the output of the model prediction results, the two features SFGdor.R and SFGdor.L show a monotonically decreasing trend with the model output results, while other features show a monotonically increasing trend.

Fig. 5
figure 5

Decision curve for the proposed model

Full size image
Fig. 6
figure 6

Categories of features used for final modeling. (A) feature importance visualization. (B) SHAP of the 10 most important features

Full size image

Discussion

The results demonstrate that preoperative brain FC network features can serve as effective predictors of postoperative visual impairment using a well-constructed machine learning model. The key findings are as follows:

  1. 1)

    By computing the adjacency matrix for both the visual impairment and non-visual impairment groups, distinctive variations in brain region connectivity between the two groups became evident. These discrepancies hold significance in predicting the occurrence of postoperative visual impairment.

  2. 2)

    Six overlapping brain regions were identified as statistically significant through independent sample t-tests and were also selected by the machine learning method. Visual-Cognition Network Alterations Underlie Impairment Risk. IOG resides within the ventral visual processing stream (occipito-temporal pathway), critical for object recognition and visual feature integration [19]. SFGmed anchors the default mode network (DMN) and executive control systems [20]. CUN is a hub of the dorsal attention network (DAN) [21] governing spatial orientation and visuomotor coordination. Preoperative disconnectivity in this tripartite system may predispose patients to postoperative impairment. This underscores the effectiveness of the feature selection technique grounded in sparse representation for predicting visual cognitive impairment.

  3. 3)

    The machine learning model developed in this study achieves high-performance metrics, including a notable area under the ROC curve and high prediction accuracy. These findings hold promise for guiding clinical decisions in anesthesia for elderly patients undergoing surgery.

Several prior studies have underscored the pivotal role of whole-brain FC analysis in predicting postoperative cognitive impairment [22, 23]. These studies revealed that individuals with cognitive impairment tend to exhibit altered brain structural connectivity or functional connectivity patterns prior to surgery. However, previous research has predominantly focused on correlations between specific brain regions’ activity and cognitive impairment, often employing rs-fMRI to investigate thalamic functional connectivity alterations in the early and late stages of amnestic mild cognitive impairment. Regrettably, these studies do not encompass an examination of the comprehensive whole-brain network activity [24]. Since most cognitive processes engage multiple brain regions, exclusively assessing the functional connectivity of one or a select few brain regions may inadvertently overlook crucial information that influences diagnostic accuracy. Furthermore, the utilization of machine learning alone can be insufficient for exploring factors potentially associated with visual spatial perception. Machine learning predominantly screens variables without incorporating comprehensive omics features, thereby neglecting the potential synergistic effects of multiple features [25].

In comparison to previous studies, low-order network features gauge the time-varying correlation between the activities of any two brain regions but cannot adequately capture correlations among multiple brain regions. To explore the connection between preoperative whole-brain network activity and postoperative visual cognitive impairment, we have devised a feature extraction method that amalgamates low-order and high-order brain networks. Notably, this study reveals that high-order FC network features played a pivotal role in achieving favorable results. A study has indicated that assessing correlations among various sets of brain regions significantly contributes to the diagnosis of brain disorders based on FC networks [26]. Consequently, we initially craft a cluster-based high-order FC network to quantify interactions among distinct brain regions, delving deeper into their interactive relationships. Besides, we adopt sparse representation-based feature selection methods, diverging from statistical estimation techniques such as independent sample T-test, to sift through variables. This approach effectively unearths high-resolution features linked to postoperative visual cognitive impairment. The feature selection process considers not only the correlation between features and labels but also addresses feature redundancy, ensuring the selection of a feature subset that most effectively characterizes the sample categories. Finally, this study culminates in the development of a comprehensive prediction model for postoperative visual cognitive impairment, yielding highly satisfactory predictive results.

In this study, the reasons why the sparse representation classifier outperforms SVM and RF can be summarized into three points [16, 27, 28]. (1) Sparsity reduces the model’s dependence on training data. Based on the sparsity constraint, the sparse representation classifier effectively suppresses the interference of irrelevant categories by reconstructing using only a small number of training samples that are most relevant to the test sample, thereby enhancing the discrimination ability, which is particularly suitable for the small sample learning problem in this study. (2) The low model complexity reduces the risk of overfitting. Compared with SVM, which relies on the selection of kernel functions and parameter adjustment, SRC is a simple non-parametric training classifier that only needs to build a training sample dictionary. The low complexity of the model reduces the risk of overfitting. (3) SRC is suitable for high-dimensional and low-sample problems SRC is particularly suitable for tasks where the feature dimension is greater than the number of samples, such as medical images, genetic data and other scenarios. In such tasks, SVM may overfit, while RF may not be able to fully utilize the feature structure.

Some improvement strategies can be implemented to enhance the performance metrics of classification models: Feature Engineering Optimization: Conduct in-depth analysis to extract highly relevant features, reducing redundancy and improving the discriminative power of the model. Data Augmentation and Class Balancing: Employ techniques such as oversampling, undersampling, or generating synthetic data to balance class distributions and mitigate model bias towards majority classes. Ensemble Learning Methods: Combine the strengths of multiple models to construct ensemble models, aiming for more robust and accurate classification results.

In our machine learning-based classification prediction model, 29 features were selected to make the final prediction regarding postoperative visual impairment. Among the brain regions selected by these features and those identified through independent sample T-test, six of them are matched, i.e., IOG. R, SFGmed. L, CUN. L, AMYG. L, PreCG. L, and TPOmid.R. Additionally, MTG.R, THA. L, and PoCG. L corresponds with important brain regions highlighted in previous studies. These connections have been linked to direct enhancements in individual memory [29,30,31]. Consequently, we hypothesize that the features associated with these connections may hold a pivotal role in predicting postoperative visual impairment.

This study has some limitations. First, the sample size is relatively small. While our dataset of 74 participants suffices to detect brain regions associated with preoperative and postoperative visual cognitive impairment, we aim to expand our patient cohort in the future to further validate and optimize our approach. Second, although the machine learning-based FC analysis method in this study has achieved good results, in recent years, there have been continuous studies that have found that deep learning-based FC analysis methods can achieve better results than machine learning-based methods in many cases (such as sufficient experimental data). Thapaliya et al. proposed a graph convolutional networks-based FC matrice analysis method, BrainRGIN, which outperformed traditional machine learning methods in all intelligence prediction tasks [32]. Kan et al. proposed a Transformer-based FC matrice analysis method that achieved significant improvements over traditional methods on both public and private datasets [33]. Furthermore, Thapaliya et al. directly analyzed raw time-series data to establish a goal-specific functional connectivity matrix, eliminating the need for a predefined FC matrix. The results of large-scale independent test experiments showed that the adaptively learned adjacency matrix is better than the inherent adjacency matrix [34]. Inspired by these works, in the future we will focus on rs-fMRI analysis based on adaptive FC matrix construction and graph networks.

Conclusion

In this study, we construct a predictive model for postoperative functional impairments related to visual cognition by employing graph theory analysis on resting-state functional connectivity networks. The proposed method may provide some guidance for the treatment decisions of some elderly patients.