Artificial intelligence has shown great promise in noninvasive recognition of vulnerable coronary plaques. However, practical data issues in multicenter studies, such as inconsistent data distribution and insufficient or missing data labels, could significantly affect the recognition accuracy. Unsupervised domain adaptation (UDA) can be introduced to address this challenge, but several limits still remain. First, many existing UDA models lack interpretability, hindering healthcare professionals' ability to interpret and trust the model's decision. Second, some methods use pseudolabel to enhance performance, but often overlook the quality assessment of these pseudolabels, potentially leading to negative knowledge transfer. To this end, based on the interpretable Takagi-Sugeno-Kang fuzzy system (TSK-FS), a novel domain adaptive method is proposed to improve model generalizability for vulnerable coronary plaques recognition in multicenter data. First of all, TSK-FS is employed to construct a shared fuzzy feature space for the source domain and the target domain, aiming to better align data distribution. To make full use of the information of unlabeled target domain data and further reduce the negative knowledge transfer, the enhanced pseudolabel learning mechanism is further introduced by combining the graph-based random walking and label filtering. Moreover, Multicenter data of 910 patients with suspected or diagnosed coronary artery disease were collected from three hospitals for experiments. Experimental results demonstrate that the proposed DA-TSK-PLR-FS achieves the promising generalizability across multicenter datasets

This study aims to investigate the predictive value of pre-treatment multi-phasic contrast-enhanced computed tomography (CECT) radiomic features for treatment resistance in patients with rat sarcoma virus (RAS)-mutated colorectal liver metastases (CRLMs) receiving bevacizumab-based chemotherapy. Seventy-three samples with RAS-mutated CRLMs receiving bevacizumab-combined chemotherapy regimens were evaluated. Radiomic features were extracted from arterial phase (AP), portal venous phase (PVP), AP-PVP subtraction image, and Delta phase (DeltaP, calculated as AP-to-PVP ratio) images. Three groups of radiomics features were extracted for each phase, including peritumor, core tumor, and whole-tumor regions. For each of the four phases, a two-sided independent Mann-Whitney U test with the Bonferroni correction and K-means clustering was applied to the remnant features for each phase. Subsequently, the Least Absolute Shrinkage and Selection Operator (LASSO) algorithm was then applied for further feature selection. Six machine learning algorithms were then used for model development and validated on the independent testing cohort. Results showed peritumoral radiomic features and features derived from Laplacian of Gaussian (LoG) filtered images were dominant in all the compared machine learning algorithms; NB models yielded the best-performing prediction (Avg. training AUC: 0.731, Avg. testing AUC: 0.717) when combining all features from different phases of CECT images. This study demonstrates that peritumoral radiomic features and LoG-filtered pre-treatment multi-phasic CECT images were more predictive of treatment response to bevacizumab-based chemotherapy in RAS-mutated CRLMs compared to core tumor features.

To develop a radiomics-based machine learning model for the noninvasive prediction of axillary lymph node metastasis (ALNM) in breast cancer patients, thereby providing an objective basis for clinical diagnosis and treatment. This multicenter retrospective study included 501 patients from four hospitals. All patients underwent preoperative MRI examinations, and their axillary lymph node status was pathologically confirmed. A total of 13,184 radiomics features were extracted from each patient. Feature selection was performed using one-way analysis of variance (ANOVA), Spearman correlation analysis, and the least absolute shrinkage and selection operator (LASSO). Based on the selected features, four machine learning models were constructed, and their predictive performance was assessed using receiver operating characteristic (ROC) analysis in the training, internal validation, and external validation cohorts. Among the four models, the logistic regression (LR) model demonstrated the best predictive performance. The areas under the ROC curve (AUC) for the training, internal validation, and external validation cohorts were 0.753, 0.743, and 0.698, respectively, indicating good predictive accuracy and generalizability. Conclusion: The radiomics-based LR model enables noninvasive prediction of ALNM in breast cancer patients and holds potential clinical value for guiding individualized treatment strategies.

Existing Deep Unsupervised Domain Adaptation (UDA) methods primarily rely on aligning black-box feature distributions, which often leads to semantic drift and lacks transparency regarding where the alignment occurs. To address these limitations, we propose an Interpretable Residual TSK UDA Network. First, we introduce a Residual TSK Fuzzy Layer that serves as a structural foundation for alignment. We treat the learnable fuzzy rules as semantic anchors that partition the high-dimensional feature space into interpretable fuzzy sub-regions while preserving feature discriminability via a residual connection. Second, leveraging this structured representation, we propose Rule-Class Conditional Alignment. Unlike global alignment strategies, our Rule-Anchored Semantic Alignment treats fuzzy rules as shared coordinate bases to enforce fine-grained consistency between source and target domains, effectively mitigating negative transfer caused by noisy pseudo-labels. We integrate these components into a dual-classifier mutual learning framework to refine decision boundaries. Extensive experiments on VisDA-2017, Office-Home, and ImageCLEF-DA benchmarks demonstrate that our method achieves state-of-the-art performance and offers intuitive visualizations of the adaptation process through rule activation dynamics.

Background and purpose:The clinical utility of radiomics in head-and-neck (H&N) cancer is hindered by poor reliability caused by delineation uncertainties from the use of binary mask (BinMask). This study introduced a fuzzy mask (FuzzMask) approach to enhance the reliability of computed tomography (CT)-based radiomics for precision prognosis. Materials and methods:This retrospective study included 2,539 H&N cancer patients (855 laryngeal cancer (LC), 1,336 oropharyngeal cancer (OPC), 348 nasopharyngeal carcinoma (NPC)). Delineation uncertainty was simulated via perturbation techniques. Radiomic features (RFs) were extracted using BinMask and FuzzMask, respectively. The evaluation focused on feature reliability and relevance via the intraclass correlation coefficient (ICC) and hierarchical clustering. In addition, the predictive performance and output reliability of penalized Cox's proportional hazard models were assessed using the concordance index (C-index) and ICC, respectively. Results:The FuzzMask improved feature reliability, yielding 21, 29, and 5 additional reliable features for LC, OPC, and NPC cancers, respectively, compared to BinMask. The FuzzMask also reduced feature redundancy, generating up to 70 more clusters in hierarchical clustering, particularly for smaller tumors in complex peritumoral environments although showed marginal improvements in predictive performance (C-index: +0.1% for OPC, +0.4% for NPC, p > 0.05). However, model reliability was enhanced by FuzzMask, with ICC values increasing by 0.024, 0.022, and 0.007 for LC, OPC, and NPC, respectively, compared to BinMask (p ≥ 0.05). Conclusions:The proposed FuzzMask technique significantly improved feature reliability and model robustness against delineation uncertainty, offering greater trustworthiness for clinical translation, although predictive accuracy remains unaffected.

Accurate prostate and prostate cancer MRI segmentation remains challenging because prostate structures and lesions may be small, low-contrast, have blurred boundaries, and be strongly affected by surrounding tissues. To address these issues, this study proposes CSFBNet, a segmentation network that combines cosine-consistency-based feature filtering with foreground-background complementary semantic guidance. Specifically, the cosine-consistency sparse selection block (CSSB) selects more reliable shallow features and reduces redundant background responses. The foreground-background adaptive convolution block (FBAC) further enhances lesion-related semantic cues while suppressing background interference during decoding. Experiments were conducted on the PROMISE12, HY Prostate, and PI-CAI datasets. CSFBNet achieved Dice scores of 0.9017, 0.6539, and 0.4784, respectively, corresponding to improvements of 1.16, 5.68, and 2.13 percentage points over the baseline model. Compared with representative segmentation methods, CSFBNet achieved improved or competitive performance in region-overlap and false-positive suppression metrics. These results suggest that CSFBNet provides a practical automatic segmentation framework for prostate MRI by improving region-overlap performance and false-positive suppression, while small-lesion sensitivity and boundary recovery in challenging PI-CAI cases remain directions for future improvement.

Electroencephalogram (EEG) signals typically contain multiple types of features, including temporal, frequency, and time-frequency domains. Effectively integrating these diverse features has become a key challenge in tasks such as EEG-based disease diagnosis. Nonsparse Multi-Kernel Learning (MKL), which leverages the information from multiple kernels, has been widely and successfully applied in multi-modal feature fusion. However, most existing models rely on the strong assumption that a nonsparse combination of multiple kernels can infinitely approximate a strict binary label matrix, which overly restricts the freedom of label fitting. To address this limitation, this paper proposes a novel nonsparse MKL model for multi-modal feature fusion. Specifically, we introduce a label relaxation strategy to relax the binary label matrix, thereby enhancing the flexibility of label fitting. Meanwhile, a regularization term based on manifold learning is constructed to mitigate the potential risk of overfitting caused by label relaxation. Experimental results on a benchmark dataset containing EEG recordings from 23 patients at Boston Children's Hospital demonstrate the superior performance and promising application prospects of the proposed model.

Multi-modal image fusion (MMIF) significantly enhances visual quality and improves the accuracy of medical diagnoses by combining complementary information from multiple modalities. By integrating the strengths of various modalities, the unique characteristics of each can be leveraged to enhance image detail, contrast ratio, and overall visual quality in complex environments. However, current methodologies face substantial challenges in preserving image fidelity and achieving cross-modal coherence. Traditional spatial-domain approaches struggle with reconciling local structures and global frequency information, while cross-scale misalignment often leads to detail loss and inconsistencies across modalities. To address these challenges, we propose an innovative fusion framework, MFS-Fusion. The core of our design is two novel modules: a Multi-Scale Fourier Enhancement Module (MS-FEM) that innovatively models long-range interdependencies in the frequency domain to achieve global coherence, and a Cross-Scale Spatial Refinement Module (CS-SRM) that ensures precise feature alignment and edge preservation. Extensive experiments on several fusion tasks, including infrared-visible image fusion (IVIF) and medical image fusion (MIF), demonstrate that our method outperforms existing approaches in terms of detail retention, structural integrity, and information fidelity, effectively overcoming the challenges of cross-modal misalignment and detail loss. Furthermore, the fused images generated by MFS-Fusion significantly enhance the performance of downstream tasks, such as object detection and semantic segmentation. The code is publicly available at https://github.com/CrisT777-JN/MFS-Fusion.

Thoracic Cone-beam computed tomography (CBCT) is routinely collected during image-guided radiation therapy (IGRT) to provide updated patient anatomy information for lung cancer treatments. However, CBCT images often suffer from streaking artifacts and noise caused by under-rate sampling projections and low-dose exposure, resulting in loss of lung anatomy which contains crucial pulmonary tumorous and functional information. While recent deep learning-based CBCT enhancement methods have shown promising results in suppressing artifacts, they have limited performance on preserving anatomical details containing crucial tumorous information due to lack of targeted guidance. To address this issue, we propose a novel feature-targeted deep learning framework which generates ultra-quality pulmonary imaging from CBCT of lung cancer patients via a multi-task customized feature-to-feature perceptual loss function and a feature-guided CycleGAN. The framework comprises two main components: a multi-task learning feature-selection network (MTFS-Net) for building up a customized feature-to-feature perceptual loss function (CFP-loss); and a feature-guided CycleGan network. Our experiments showed that the proposed framework can generate synthesized CT (sCT) images for the lung that achieved a high similarity to CT images, with an average SSIM index of 0.9747 and an average PSNR index of 38.5995 globally, and an average Pearman’s coefficient of 0.8929 within the tumor region on multi-institutional datasets. The sCT images also achieved visually pleasing performance with effective artifacts suppression, noise reduction, and distinctive anatomical details preservation. Functional imaging tests further demonstrated the pulmonary texture correction performance of the sCT images, and the similarity of the functional imaging generated from sCT and CT images has reached an average DSC value of 0.9147, SCC value of 0.9615 and R value of 0.9661. Comparison experiments with pixel-to-pixel loss also showed that the proposed perceptual loss significantly enhances the performance of involved generative models. Our experiment results indicate that the proposed framework outperforms the state-of-the-art models for pulmonary CBCT enhancement. This framework holds great promise for generating high-quality pulmonary imaging from CBCT that is suitable for supporting further analysis of lung cancer treatment.

PurposeTo guide the preselection of highly repeatable radiomic features (RFs) in downstream analysis without further analysis its repeatability, a detailed radiomic feature robustness databank (RF-RobustDB) was established via image perturbation.MethodsData on 1,274 oropharyngeal carcinoma (OPC) patients who had undergone pretreatment computed tomography (CT) imaging, collected from a public dataset. The original images and corresponding masks underwent systematic perturbations to simulate potential variations encountered during CT image rescanning, including translational shifts, rotational changes, random noise additions, and contour modifications. For each radiomic feature (RF), including unfiltered, wavelet-filtered, and Laplacian-of-Gaussian (LoG)-filtered features, we systematically quantified robustness against these perturbations by intraclass correlation coefficients (ICCs).ResultsOut of 1395 first- and high-order RFs, 470 demonstrated excellent repeatability, i.e., a mean ICC of greater than 0.9. The use of these preselected highly repeatable RFs in model development improved the mean concordance (C) index in two external validation cohorts and reduced the mean C index gap between the training and external validation cohorts. These results demonstrate that the preselected high repeatable RFs from RF-RobustDB can effectively enhance radiomic model generalizability.ConclusionsThe methodology employed to establish the RF-RobustDB is highly transferable to other tumor sites and different imaging modalities, which will facilitate the creation of RF-RobustDBs to guide the development of universally applicable radiomic models.

Epilepsy is a neurological disorder characterized by transient cerebral dysfunction arising from abnormal cortical neuronal discharges, which can precipitate sudden muscle contractions, loss of consciousness, and convulsions. As one of the most prevalent brain disorders, it affects a substantial share of the global population. Electroencephalography (EEG) - the most widely used and extensively studied noninvasive modality for monitoring cerebral electrical activity - underpins clinical and computational approaches to epilepsy detection. In this work, we propose a Consistency-Regularized Multi-View Transfer Learning (CR-MVTL) algorithm for epileptic EEG analysis and AI-assisted detection. Compared with prevailing methods, CR-MVTL offers: (i) strong cross-view complementarity that promotes information sharing and improves recognition; (ii) high generalizability that enables rapid parameter adaptation when target-domain data are scarce; and (iii) dynamic view-weighting that attenuates weak views while amplifying informative ones. We evaluate CR-MVTL across 12 simulated multi-view transfer scenarios. Extensive experiments demonstrate absolute improvements in average accuracy of 7.25 and 6.41 percentage points in the two-view and three-view settings, respectively, over competitive baselines.

In recent years, the application of machine learning techniques for identifying epileptic electroencephalogram (EEG) signals has become increasingly prevalent. In this study, we propose a method called Multi-View Learning and Regularized Label Relaxation Linear Regression (MVRLR) for the recognition of epileptic EEG signals. To be specific, first, we generate multi-view data by employing three different feature extraction methods to capture the EEG features from different perspectives. Then, in the linear regression framework, we introduce a nonnegative label relaxation matrix to transform the strict binary label matrix into a relaxed variable matrix, allowing for more flexible label fitting and expanding the boundaries between different classes. Additionally, we utilize manifold learning to construct a class tightness graph, which serves as a regularization term to prevent overfitting. Finally, we employ Shannon entropy to quantify the overall uncertainty within the probability distribution. To validate the effectiveness of the proposed method, we conduct experiments using a multi-view epileptic EEG dataset and compare it with several other classification algorithms. The extensive experimental results demonstrate that the proposed MVRLR method achieves higher classification accuracy. By applying the proposed method, we can more accurately identify epileptic EEG signals, providing robust support for early diagnosis and personalized treatment of epilepsy.

The first-order Takagi-Sugeno-Kang (TSK) fuzzy classifier with a fully combined fuzzy rule base (FuCo-FRB) is a potent and interpretable classifier for multiple input and multiple output (MIMO) tasks. However, FuCo-FRB poses an exponential increase in the number of fuzzy rules, which creates challenges for efficient identification of the parameter matrix for MIMO tasks. To address these challenges, we propose a lightweight TSK fuzzy classifier (LW-TSK-FC) to achieve the balance of training efficiency and predictive performance, particularly for MIMO tasks. The proposed LW-TSK-FC has three key advantages: 1) an adaptive weighting method based on directly connected FRB enables the efficient generation of fuzzy rules while retaining the distribution characteristics of FuCo-FRB; 2) the consequent network of the first-order TSK is optimized to a novel series structure by using matrix factorization. This new structure increases the depth of the consequent network, enabling the implementation of kernel functions and least learning machine (LLM), which enhance both calculation efficiency and predictive performance; and 3) the consequent network of LW-TSK-FC contains only weight parameters, which can be quantitatively identified using LLM. This results in significant improvements in training efficiency. Comparison experiments demonstrate that LW-TSK-FC achieved superior predictive performance and training efficiency, particularly for tasks involving higher-order statistical features. Furthermore, experiments on a real-world clinical application show that LW-TSK-FC efficiently handled 1000-dimensional data and achieved the best testing AUC. This lightweight structure holds tremendous potential for application in deep or stacked fuzzy neural networks.

The rapid advancement of medical imaging technologies has led to an exponential increase in medical image data, making efficient retrieval from large-scale datasets critical for improving diagnostic accuracy and speed. However, two key challenges hinder this process: first, the presence of uncertain and subtle lesions in medical images that are often difficult to discern, and second, class imbalance across different case types within medical image databases. These inherent challenges significantly degrade the performance of existing hashing algorithms. In recent years, methods based on the Takagi-Sugeno-Kang fuzzy system (TSK-FS) have shown promising performance in medical image modeling. Inspired by these advances, this article proposes a novel fuzzy hashing network (FHN) based on TSK-FS to enhance retrieval performance by effectively handling both uncertainty and data imbalance in medical imaging. The FHN first introduces a novel fuzzification mechanism that incorporates the concept of a self-attention mechanism to effectively capture the complex underlying features in medical images, thereby enhancing the data discriminability in fuzzy spaces. Meanwhile, a new consequent parameter learning mechanism is developed for defuzzification by introducing the Transformer network, which aims to improve the inference efficiency and generalization capability of the FHN. Based on these two mechanisms, FHN's capability of analyzing and handling uncertain data is significantly enhanced. Furthermore, a novel hash center loss is designed to capture global relationships while emphasizing local structural information, thereby improving the handling of imbalanced data and significantly enhancing retrieval performance.

PURPOSE:Existing prognostic staging systems depend on expensive manual extraction by pathologists, potentially overlooking latent patterns critical for prognosis, or use black-box deep learning models, limiting clinical acceptance. This study introduces a novel deep learning-assisted paradigm that complements existing approaches by generating interpretable, multi-view risk scores to stratify prognostic risk in hepatocellular carcinoma (HCC) patients. METHODS:510 HCC patients were enrolled in an internal dataset (SYSUCC) as training and validation cohorts to develop the Hybrid Deep Score (HDS). The Attention Activator (ATAT) was designed to heuristically identify tissues with high prognostic risk, and a multi-view risk-scoring system based on ATAT established HDS from microscopic to macroscopic levels. HDS was also validated on an external testing cohort (TCGA-LIHC) with 341 HCC patients. We assessed prognostic significance using Cox regression and the concordance index (c-index). RESULTS:The ATAT first heuristically identified regions where necrosis, lymphocytes, and tumor tissues converge, particularly focusing on their junctions in high-risk patients. From this, this study developed three independent risk factors: microscopic morphological, co-localization, and deep global indicators, which were concatenated and then input into a neural network to generate the final HDS for each patient. The HDS demonstrated competitive results with hazard ratios (HR) (HR 3.24, 95% confidence interval (CI) 1.91-5.43 in SYSUCC; HR 2.34, 95% CI 1.58-3.47 in TCGA-LIHC) and c-index values (0.751 in SYSUCC; 0.729 in TCGA-LIHC) for Disease-Free Survival (DFS). Furthermore, integrating HDS into existing clinical staging systems allows for more refined stratification, which enables the identification of potential high-risk patients within low-risk groups. CONCLUSION:This novel paradigm, from identifying high-risk tissues to constructing prognostic risk scores, offers fresh insights into HCC research. Additionally, the integration of HDS complements the existing clinical staging system by facilitating more detailed stratification in DFS and Overall Survival (OS).

BACKGROUND:Breast cancer (BC) is the most common type of cancer among women. Axillary lymph node metastasis (ALNM) is strongly correlated with distant metastasis, recurrence, and overall survival rates in BC. Therefore, accurate detection of ALNM holds valuable implications for patient prognosis and treatment plan selection. PURPOSE:The objective of this study is to develop a novel domain adaptative radiomics pipeline based on domain adaptation (DA) to predict ALNM based on multi-parametric magnetic resonance imaging (MRI) for multicenter studies. METHODS:396 BC lesions collected from 391 patients at the first three centers were used as source domain data for model training and internal validation. 105 BC lesions collected from 105 patients at the fourth center were used as target domain data for model external validation. Each BC lesion was scanned with eight MRI sequences, including T2-weighted, non-fat-saturated T1-weighted and dynamic contrast-enhanced sequences (phases 0-5). From each MRI sequence, 1648 radiomics features were extracted, resulting in a total of 13184 features extracted from each lesion. Variance threshold, the max-relevance and min-redundancy (mRMR), and the least absolute shrinkage and selection operator (LASSO) were employed for feature selection. Then a classifier based on balanced distribution adaptation (BDA) was developed for ALNM prediction. Unlike traditional radiomics models, the BDA classifier was designed to reduce data distribution differences across the three centers in the source domain by minimizing maximum mean discrepancy (MMD). The design of the BDA classifier is based on the assumption that the average distribution difference between the three centers in the source domain is similar to the distribution difference between the target domain data and the source domain data. Thus, when the BDA classifier is externally validated in the target domain, we can expect to achieve good predictive performance. The area under the receiver operating characteristic curve (AUC) with 95% confidence intervals (CI) was used to evaluate the predictive performance of the domain adaptative radiomics pipeline in the external validation cohort. To highlight the effectiveness of our proposed pipeline, the traditional radiomics pipeline with six machine learning models (support vector machine (SVM), k-nearest neighbor (KNN), random forest (RF), decision tree (DT), extreme gradient boosting (XGBoost), linear discriminant analysis (LDA)) was also trained. The traditional radiomics pipeline with six models was compared with the domain adaptative radiomics pipeline in terms of AUC. RESULTS:After feature selection, 12 radiomics features were selected for the following modeling tasks. The external validation of the domain adaptative radiomics pipeline outperformed the other models, with an AUC of 0.781 (95%CI: 0.692-0.870). The best-performing RF model among the six traditional radiomics models demonstrated an AUC of only 0.700 (95%CI: 0.605-0.795). The p-value for the comparison between BDA and RF model was 0.048. CONCLUSION:Compared with the traditional radiomics pipeline, the proposed domain adaptative radiomics pipeline based on multi-parametric MRI achieved better performance for ALNM prediction in this multicenter study.

The expeditious and precise diagnosis of radiation dermatitis (RD) is expected to mitigate the afflictions endured by patients. Dealing with the intricacies inherent in RD data, feature volatilities, inter-feature redundancies, and elusive quantification of their salience, this study propounds a pioneering hierarchical Takagi-Sugeno-Kang fuzzy classification model that can quantitatively eliminate the redundancy between training features, named RDE-T-S. With the introduction of a mechanism for feature optimization alongside a partially rules-stochastic linking strategy, this study endeavors to surmount the challenges posed by feature importance delineation and rule superfluity, thereby fortifying the model's architectural robustness and streamlining the training paradigm. Expounding the original input space, this study establishes a pioneering methodology for optimizing the input space of base building units, thereby facilitating the rapid unfolding of the original manifold structure with the aim of expediting model convergence. In pursuit of maximal classification efficacy, an integrated optimization output mechanism is advanced, harnessing a multiplication- weighted methodology predicated upon the fusion of subjective and objective weight permutations; thus augmenting the adaptability of model parameterization. Of particular note is the emulation prowess of RDE- T-S that adeptly mirrors human cognitive processes, capitalizing on ubiquitous experiential knowledge (rules) to navigate analogous tasks with finesse. The experimental results show that the average training accuracy of RDE-T-S on five datasets reached 95%, which is indeed superior to current popular classifiers. Therefore, it is a promising option for assessing the severity of RD.

To investigate the optimal training dataset size (TDS) for respiration prediction accuracy using a long short-term memory (LSTM) model. The respiratory signals of 151 patients acquired with the real-time position management system were retrospectively included in this study. Among the dataset, 101 respiratory signals were utilized to evaluate the impact of the TDS on prediction accuracy, while the remaining 50 signals were employed for setting the default hyperparameters. The prediction accuracy of the LSTM model using eight different TDSs (10 s, 20 s, 30 s, 60 s, 90 s, 110 s, 130 s, and 150 s) was examined and evaluated by the root mean square error (RMSE) between the real and predicted respiratory signals. The interplay effects of the main hyperparameters, the ahead time and the different testing data lengths using different TDSs were also measured. For the 520 ms ahead time, the root mean square error values of the LSTM model using the eight different training data sizes listed above were 0.146 cm, 0.137 cm, 0.134 cm, 0.125 cm, 0.120 cm, 0.121 cm, 0.121 cm, and 0.119 cm, respectively. The LSTM model achieved the highest prediction accuracy when the TDS was 150 s. The prediction accuracy was stable when the TDS exceeded 90 s. TDS selection could influence the respiration signal prediction accuracy of the LSTM model. The relationship between TDS and the prediction accuracy of the LSTM model was not linear. The 90 s seemed to be an optimal TDS for the respiration signal prediction tasks using the LSTM model, as it was the shortest time at which a favorable prediction accuracy was maintained in this study.