Comput Methods Programs Biomed. 2024 Dec 18;260:108568. doi: 10.1016/j.cmpb.2024.108568. Online ahead of print.

ABSTRACT

BACKGROUND: The fusion of multi-modal data has been shown to significantly enhance the performance of deep learning models, particularly on medical data. However, missing modalities are common in medical data due to patient specificity, which poses a substantial challenge to the application of these models.

OBJECTIVE: This study aimed to develop a novel and efficient multi-modal fusion framework for medical datasets that maintains consistent performance, even in the absence of one or more modalities.

METHODS: In this paper, we fused three modalities: chest X-ray radiographs, history of present illness text, and tabular data such as demographics and laboratory tests. A multi-modal fusion module based on pooled bottleneck (PB) attention was proposed in conjunction with knowledge distillation (KD) for enhancing model inference in the case of missing modalities. In addition, we introduced a gradient modulation (GM) method to deal with the unbalanced optimization in multi-modal model training. Finally, we designed comparison and ablation experiments to evaluate the fusion effect, the model robustness to missing modalities, and the contribution of each component (PB, KD, and GM). The evaluation experiments were performed on the MIMIC-IV datasets with the task of predicting in-hospital mortality risk. Model performance was assessed using the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC).

RESULTS: The proposed multi-modal fusion framework achieved an AUROC of 0.886 and AUPRC of 0.459, significantly surpassing the performance of baseline models. Even when one or two modalities were missing, our model consistently outperformed the reference models. Ablation of each of the three components resulted in varying degrees of performance degradation, highlighting their distinct contributions to the model's overall effectiveness.

CONCLUSIONS: This innovative multi-modal fusion architecture has demonstrated robustness to missing modalities, and has shown excellent performance in fusing three medical modalities for patient outcome prediction. This study provides a novel idea for addressing the challenge of missing modalities and has the potential be scaled to additional modalities.

PMID:39709743 | DOI:10.1016/j.cmpb.2024.108568

J Hepatol. 2024 Dec 20:S0168-8278(24)02784-3. doi: 10.1016/j.jhep.2024.12.029. Online ahead of print.

ABSTRACT

BACKGROUND & AIMS: Various hepatocellular carcinoma (HCC) prediction models have been proposed for patients with chronic hepatitis B (CHB) using clinical variables. We aimed to develop an artificial intelligence (AI)-based HCC prediction model by incorporating imaging biomarkers derived from abdominal computed tomography (CT) images along with clinical variables.

METHODS: An AI prediction model employing a gradient-boosting machine algorithm was developed utilizing imaging biomarkers extracted by DeepFore, a deep learning-based CT auto-segmentation software. The derivation cohort (n=5,585) was randomly divided into the training and internal validation sets at a 3:1 ratio. The external validation cohort included 2,883 patients. Six imaging biomarkers (i.e., abdominal visceral fat-total fat volume ratio, total fat-trunk volume ratio, spleen, and liver volume; liver-spleen Hounsfield unit [HU] ratio; and muscle HU) and eight clinical variables were selected as the main variables of our model, PLAN-B-DF.

RESULTS: In the internal validation set (median follow-up duration=7.4 years), PLAN-B-DF demonstrated an excellent predictive performance with a c-index of 0.91 and good calibration function (P=0.78 by the Hosmer-Lemeshow test). In the external validation cohort (median follow-up duration=4.6 years), PLAN-B-DF showed a significantly better discrimination function compared to previous models including PLAN-B, PAGE-B, modified PAGE-B, and CU-HCC (c-index, 0.89 vs. 0.65-0.78; all P<0.001) and maintained a good calibration function (P=0.42 by the Hosmer-Lemeshow test). When patients were classified into four groups according to the risk probability calculated by PLAN-B-DF, the 10-year cumulative HCC incidence was 0.0%, 0.4%, 16.0%, and 46.2% in the minimal-, low-, intermediate-, and high-risk groups, respectively.

CONCLUSION: This AI prediction model, integrating deep learning-based auto-segmentation of CT images, offers improved performance in predicting HCC risk among patients with CHB compared to previous models.

IMPACT AND IMPLICATIONS: The AI-driven HCC prediction model (PLAN-B-DF), employing an automated CT segmentation algorithm, demonstrates a significant improvement in predictive accuracy and risk stratification among patients with CHB. Using a gradient-boosting algorithm and CT metrics such as visceral fat volume and myosteatosis, PLAN-B-DF outperforms previous models based solely on clinical and demographic data. This model not only shows a higher c-index compared to previous models, but also effectively classifies CHB patients into different risk groups. This model uses machine learning to analyze the complex relationships among various risk factors contributing to HCC occurrence, thereby offering more personalized surveillance for CHB patients.

PMID:39710148 | DOI:10.1016/j.jhep.2024.12.029

Magn Reson Imaging. 2024 Dec 20:110310. doi: 10.1016/j.mri.2024.110310. Online ahead of print.

ABSTRACT

Dynamic contrast-enhanced (DCE) MRI is an important imaging tool for evaluating tumor vascularity that can lead to improved characterization of tumor extent and heterogeneity, and for early assessment of treatment response. However, clinical adoption of quantitative DCE-MRI remains limited due to challenges in acquisition and quantification performance, and lack of automated tools. This study presents an end-to-end deep learning pipeline that exploits a novel deep reconstruction network called DCE-Movienet with a previously developed deep quantification network called DCE-Qnet for fast and quantitative DCE-MRI. DCE-Movienet offers rapid reconstruction of high spatiotemporal resolution 4D MRI data, reducing reconstruction time of the full acquisition to only 0.66 s, which is significantly shorter than compressed sensing's order of 10 min-long reconstructions, without affecting image quality. DCE-Qnet can then perform comprehensive quantification of perfusion parameter maps (Ktrans, vp, ve), and other parameters affecting quantification (T1, B1, and BAT) from a single contrast-enhanced acquisition. The end-to-end deep learning pipeline was implemented to process data acquired with a golden-angle stack-of-stars k-space trajectory and validated on healthy volunteers and a cervical cancer patient against a compressed sensing reconstruction. The end-to-end deep learning DCE-MRI technique addresses key limitations in DCE-MRI in terms of speed and quantification robustness, which is expected to improve the performance of DCE-MRI in a clinical setting.

PMID:39710009 | DOI:10.1016/j.mri.2024.110310

Phys Med. 2024 Dec 21;129:104873. doi: 10.1016/j.ejmp.2024.104873. Online ahead of print.

ABSTRACT

PURPOSE: Automated treatment plan generation is essential for magnetic resonance imaging (MRI)-guided adaptive radiotherapy (MRIgART) to ensure standardized treatment-plan quality. We proposed a novel cross-technique transfer learning (CTTL)-based strategy for online MRIgART autoplanning.

METHOD: We retrospectively analyzed the data from 210 rectal cancer patients. A source dose prediction model was initially trained using a large volume of volumetric-modulated arc therapy data. Subsequently, a single patient's pretreatment data was employed to construct a CTTL-based dose prediction model (CTTL_M) for each new patient undergoing MRIgART. The CTTL_M predicted dose distributions for subsequent treatment fractions. We optimized an auto plan using the parameters based on dose prediction. Performance of our CTTL_M was assessed using dose-volume histogram and mean absolute error (MAE). Our auto plans were compared with clinical plans regarding plan quality, efficiency, and complexity.

RESULTS: CTTL_M significantly improved the dose prediction accuracy, particularly in planning target volumes (median MAE: 1.27 % vs. 7.06 %). The auto plans reduced high-dose exposure to the bladder (D0.1cc: 2,601.93 vs. 2,635.43 cGy, P < 0.001) and colon (D0.1cc: 2,593.22 vs. 2,624.89 cGy, P < 0.001). The mean colon dose decreased from 1,865.08 to 1,808.16 cGy (P = 0.035). The auto plans maintained similar planning time, monitor units, and plan complexity as clinical plans.

CONCLUSIONS: We proposed an online ART autoplanning method for generating high-quality plans with improved organ sparing. Its high degree of automation can standardize planning quality across varying expertise levels, mitigating subjective assessment and errors.

PMID:39709892 | DOI:10.1016/j.ejmp.2024.104873

Comput Biol Med. 2024 Dec 21;185:109593. doi: 10.1016/j.compbiomed.2024.109593. Online ahead of print.

ABSTRACT

Over the past few decades, machine learning and deep learning (DL) have incredibly influenced a broader range of scientific disciplines. DL-based strategies have displayed superior performance in image processing compared to conventional standard methods, especially in healthcare settings. Among the biggest threats to global public health is the fast spread of malaria. The plasmodium falciparum infection, the disease origin causes the intestinal illness. Fortunately, advances in artificial intelligence techniques have made it possible to use visual data sets to quickly and effectively diagnose malaria which has also proven to be cost and time effective. In literature, several DL approaches have previously been used with good precision but suffer from computational inefficiency and interpretability. Therefore, this research proposes a generalized fractional order-based explainable lightweight convolutional neural network model to overcome these limitations. The fractional order optimization algorithms have proven worth in terms of estimation accuracy and convergence speed for different applications. The proposed fractional order optimizer-based model offers an improved solution to malaria disease diagnosis with a percentage accuracy of 95 % using the standard NIH dataset and outperforms the existing complex models concerning speed and effectiveness. The proposed fractionally optimized lightweight CNN model has shown substantial performance on the external MP-IDB dataset and M5 test set as well by achieving a generalized test accuracy of 92 % and 90.4 % which verifies the robustness and generalizability of the proposed solution under available circumstances. Moreover, the efficacy of the proposed lightweight architecture is endorsed through evaluation metrics of precision, recall, and F1-score.

PMID:39709870 | DOI:10.1016/j.compbiomed.2024.109593

Comput Biol Med. 2024 Dec 21;185:109576. doi: 10.1016/j.compbiomed.2024.109576. Online ahead of print.

ABSTRACT

BACKGROUND AND OBJECTIVES: There is an increasing and renewed interest in Electronic Health Records (EHRs) as a substantial information source for clinical decision making. Consequently, automatic de-identification of EHRs is an indispensable task, since their dissociation from personal data is a necessary prerequisite for their dissemination. Nevertheless, the bulk of prior research in this domain has been conducted using English EHRs, given the limited availability of annotated corpora in other languages, including Spanish.

METHODS: In this study, the automatic de-identification of medical documents in Spanish was explored. A private corpus comprising 599 genuine clinical cases was annotated with eight different categories of protected health information. The prediction problem was approached as a named entity recognition task and two deep learning-based methodologies were developed. The first strategy was based on recurrent neural networks (RNN) and the second, an end-to-end approach, was based on Transformers. In addition, we have implemented a procedure to expand the amount of texts employed for model training.

RESULTS: Our findings demonstrate that Transformers surpass RNNs in the de-identification of clinical data in Spanish. Particularly noteworthy is the excellent performance of the XLM-RoBERTa large Transformer, achieving a rigorous strict-match micro-average of 0.946 for precision, 0.954 for recall, and an F1 score of 0.95 when applied to the amplified version of the corpus. Furthermore, a web-based application has been created to assist specialized clinicians in de-identifying EHRs through the aid of the implemented models.

CONCLUSION: The study's conclusions showcase the practical applicability of the state-of-the-art Transformers models for precise de-identification of clinical notes in real-world medical settings in Spanish, with the potential to improve performance if continual pre-training strategies are implemented.

PMID:39709869 | DOI:10.1016/j.compbiomed.2024.109576

Talanta. 2024 Dec 18;285:127427. doi: 10.1016/j.talanta.2024.127427. Online ahead of print.

ABSTRACT

The combined application of near-infrared spectroscopy (NIRS) and X-ray fluorescence spectroscopy (XRF) has achieved remarkable results in coal quality analysis by leveraging NIRS's sensitivity to organic compounds and XRF's reliability for inorganic composition. However, variations in particle size distribution negatively affect the diffuse reflectance of NIRS and the fluorescence signal intensities of XRF, leading to decreased accuracy and repeatability in predictions. To address this issue, this study innovatively proposes a particle size correction method that integrates image processing and deep learning. The method first captures micro-images of the coal sample surface using a microscope camera and employs the Segment Anything Model (SAM) for binarization to represent particle size distribution. Subsequently, a Spatial Transformer Network (STN) is applied for geometric correction, followed by feature extraction using a Convolutional Neural Network (CNN) to establish a correlation model between particle size distribution and ash measurement errors. In experiments involving 56 coal samples, including 48 at 0.2 mm for the standard ash prediction model and 8 within a 0∼1 mm range for correction, the results showed significant improvements: standard deviation (SD), mean absolute error (MAE), and root mean square error of prediction (RMSEP) decreased from 0.321%, 0.317%, and 0.335% to 0.229%, 0.225%, and 0.257%, respectively. Using the accuracy of the 0.2 mm particle size validation set as a reference, compared to before correction, the errors in these metrics were reduced by 64.06%, 50%, and 60.80%, respectively. This study demonstrates that integrating deep learning and image analysis significantly enhances the repeatability and accuracy of NIRS-XRF measurements, effectively mitigating sub-millimeter particle size effects on spectral detection results and improving model adaptability. This method, through automated particle size distribution analysis and real-time result correction, holds promise for providing essential technical support for the development of online quality detection technologies for conveyor belt materials.

PMID:39709828 | DOI:10.1016/j.talanta.2024.127427

NPJ Precis Oncol. 2024 Dec 21;8(1):287. doi: 10.1038/s41698-024-00766-9.

ABSTRACT

Endometrial cancer (EC) diagnosis traditionally relies on tumor morphology and nuclear grade, but personalized therapy demands a deeper understanding of tumor mutational burden (TMB), i.e., a key biomarker for immune checkpoint inhibition and immunotherapy response. Traditional TMB prediction methods, such as sequencing exomes or whole genomes, are costly and often unavailable in clinical settings. We present the first TR-MAMIL deep learning framework to predict TMB status and classify the EC cancer subtype directly from H&E-stained WSIs, enabling effective personalized immunotherapy planning and prognostic refinement of EC patients. Our models were evaluated on a large dataset from The Cancer Genome Atlas. TR-MAMIL performed exceptionally well in classifying aggressive and non-aggressive EC, as well as predicting TMB, outperforming seven state-of-the-art approaches. It also performed well in classifying normal and abnormal p53 mutations in EC using H&E WSIs. Kaplan-Meier analysis further demonstrated TR-MAMIL's ability to differentiate patients with longer survival in the aggressive EC.

PMID:39709501 | DOI:10.1038/s41698-024-00766-9

Microbiome. 2024 Dec 21;12(1):272. doi: 10.1186/s40168-024-01985-9.

ABSTRACT

BACKGROUND: The overuse of antibiotics has led to lethal multi-antibiotic-resistant microorganisms around the globe, with restricted availability of novel antibiotics. Compared to conventional antibiotics, evolutionarily originated antimicrobial peptides (AMPs) are promising alternatives to address these issues. The gut microbiome of Blattella germanica represents a previously untapped resource of naturally evolving AMPs for developing antimicrobial agents.

RESULTS: Using the in-house designed tool "AMPidentifier," AMP candidates were mined from the gut microbiome of B. germanica, and their activities were validated both in vitro and in vivo. Among filtered candidates, AMP1, derived from the symbiotic microorganism Blattabacterium cuenoti, demonstrated broad-spectrum antibacterial activity, low cytotoxicity towards mammalian cells, and a lack of hemolytic effects. Mechanistic studies revealed that AMP1 rapidly permeates the bacterial cell and accumulates intracellularly, resulting in a gradual and mild depolarization of the cell membrane during the initial incubation period, suggesting minimal direct impact on membrane integrity. Furthermore, observations from fluorescence microscopy and scanning electron microscopy indicated abnormalities in bacterial binary fission and compromised cell structure. These findings led to the hypothesis that AMP1 may inhibit bacterial cell wall synthesis. Furthermore, AMP1 showed potent antibacterial and wound healing effects in mice, with comparable performances of vancomycin.

CONCLUSIONS: This study exemplifies an interdisciplinary approach to screening safe and effective AMPs from natural biological tissues, and our identified AMP 1 holds promising potential for clinical application.

PMID:39709489 | DOI:10.1186/s40168-024-01985-9

Ultrasound Med Biol. 2024 Dec 20:S0301-5629(24)00438-1. doi: 10.1016/j.ultrasmedbio.2024.11.014. Online ahead of print.

ABSTRACT

OBJECTIVE: Breast ultrasound (BUS) is used to classify benign and malignant breast tumors, and its automatic classification can reduce subjectivity. However, current convolutional neural networks (CNNs) face challenges in capturing global features, while vision transformer (ViT) networks have limitations in effectively extracting local features. Therefore, this study aimed to develop a deep learning method that enables the interaction and updating of intermediate features between CNN and ViT to achieve high-accuracy BUS image classification.

METHODS: This study introduced the CNN and transformer multi-stage fusion network (CTMF-Net) consisting of two branches: a CNN branch and a transformer branch. The CNN branch employs visual geometry group as its backbone, while the transformer branch utilizes ViT as its base network. Both branches were divided into four stages. At the end of each stage, a proposed feature interaction module facilitated feature interaction and fusion between the two branches. Additionally, the convolutional block attention module was employed to enhance relevant features after each stage of the CNN branch. Extensive experiments were conducted using various state-of-the-art deep-learning classification methods on three public breast ultrasound datasets (SYSU, UDIAT and BUSI).

RESULTS: For the internal validation on SYSU and UDIAT, our proposed method CTMF-Net achieved the highest accuracy of 90.14 ± 0.58% on SYSU and 92.04 ± 4.90% on UDIAT, which showed superior classification performance over other state-of-art networks (p < 0.05). Additionally, for external validation on BUSI, CTMF-Net showed outstanding performance, achieving the highest area under the curve score of 0.8704 when trained on SYSU, marking a 0.0126 improvement over the second-best visual geometry group attention ViT method. Similarly, when applied to UDIAT, CTMF-Net achieved an area under the curve score of 0.8505, surpassing the second-best global context ViT method by 0.0130.

CONCLUSION: Our proposed method, CTMF-Net, outperforms all existing methods and can effectively assist doctors in achieving more accurate classification performance of breast tumors.

PMID:39709289 | DOI:10.1016/j.ultrasmedbio.2024.11.014

Methods. 2024 Dec 19:S1046-2023(24)00278-0. doi: 10.1016/j.ymeth.2024.12.005. Online ahead of print.

ABSTRACT

Identifying angiotensin-I-converting enzyme (ACE) inhibitory peptides accurately is crucial for understanding the primary factor that regulates the renin-angiotensin system and for providing guidance in developing new potential drugs. Given the inherent experimental complexities, using computational methods for in silico peptide identification could be indispensable for facilitating the high-throughput characterization of ACE inhibitory peptides. In this paper, we propose a novel deep stacking-based ensemble learning framework, termed Deepstack-ACE, to precisely identify ACE inhibitory peptides. In Deepstack-ACE, the input peptide sequences are fed into the word2vec embedding technique to generate sequence representations. Then, these representations were employed to train five powerful deep learning methods, including long short-term memory, convolutional neural network, multi-layer perceptron, gated recurrent unit network, and recurrent neural network, for the construction of base-classifiers. Finally, the optimized stacked model was constructed based on the best combination of selected base-classifiers. Benchmarking experiments showed that Deepstack-ACE attained a more accurate and robust identification of ACE inhibitory peptides compared to its base-classifiers and several conventional machine learning classifiers. Remarkably, in the independent test, our proposed model significantly outperformed the current state-of-the-art methods, with a balanced accuracy of 0.916, sensitivity of 0.911, and Matthews correlation coefficient scores of 0.826. Moreover, we developed a user-friendly web server for Deepstack-ACE, which is freely available at https://pmlabqsar.pythonanywhere.com/Deepstack-ACE. We anticipate that our proposed Deepstack-ACE model can provide a faster and reasonably accurate identification of ACE inhibitory peptides.

PMID:39709069 | DOI:10.1016/j.ymeth.2024.12.005

Magn Reson Imaging. 2024 Dec 19:110313. doi: 10.1016/j.mri.2024.110313. Online ahead of print.

ABSTRACT

OBJECTIVE: The total examination time can be reduced if high-quality two-dimensional (2D) cine images can be collected post-contrast to minimize non-scanning time prior to late gadolinium-enhanced imaging. This study aimed to assess the equivalency of the pre-and post-contrast performance of 2D deep learning-based highly accelerated cardiac cine (DL cine) imaging by evaluating the image quality and the quantification of biventricular volumes and function in the clinical setting.

MATERIAL AND METHODS: Thirty patients (20 men, mean age 53.7 ± 17.8 years) underwent cardiac magnetic resonance on a 1.5 T scanner for clinical indications, and pre- and post-contrast DL cine images were acquired with a short-axis view. Image-quality was scored according to three main criteria: the blood-to-myocardial contrast, endocardial edge delineation, and presence of motion artifacts throughout the cardiac cycle. Biventricular end-diastolic volume (EDV), end-systolic volume (ESV), stroke volume (SV), ejection fraction (EF), and left ventricular mass (LVM) were analyzed and compared between the pre- and post-contrast DL cine images.

RESULTS: The actual median time of 2D DL cine acquisition was 38.4 ± 9.1 s. There were no significant differences in the image quality scores between pre- and post-contrast DL cine images (p > 0.05). In the volume and functional analysis, there was no significant difference in terms of biventricular EDV, ESV, SV, EF, and LVM (p > 0.05).

CONCLUSIONS: The performance of 2D DL cine is equivalent before and after contrast injection for the assessment of image quality and ventricular function in the clinical setting.

PMID:39708928 | DOI:10.1016/j.mri.2024.110313

Magn Reson Imaging. 2024 Dec 19:110314. doi: 10.1016/j.mri.2024.110314. Online ahead of print.

ABSTRACT

PURPOSE: To explore the feasibility of Deep learning radiomics nomograms (DLRN) in predicting IDH genotype.

METHODS: A total of 402 glioma patients from two independent centers were retrospectively included, and the data from center I was randomly divided into a training cohort (n = 239) and an internal validation cohort (n = 103) on a 7:3 basis. Center II served as an independent external validation cohort (n = 60). We developed a DLRN for IDH classification of gliomas based on T2 images. This hybrid model integrates deep learning features, radiomics features, and clinical features most relevant to IDH genotypes and finally classifies them using multivariate logistic regression analysis. We used the area under the curve (AUC) of the receiver operating characteristic (ROC) to evaluate the performance of the model and applied the DLRN score to the survival analysis of some of the follow-up glioma patients.

RESULTS: The proposed model had an area under the curve (AUC) of 0.98 in an externally validated cohort, and DLRN scores were significantly associated with the overall survival of glioma patients.

CONCLUSIONS: Deep learning radiomics nomograms performed well in non-invasively predicting IDH mutation status in gliomas, assisting stratified management and targeted therapy for glioma patients.

PMID:39708927 | DOI:10.1016/j.mri.2024.110314

Brief Bioinform. 2024 Nov 22;26(1):bbae662. doi: 10.1093/bib/bbae662.

ABSTRACT

Cell type annotation is a critical step in analyzing single-cell RNA sequencing (scRNA-seq) data. A large number of deep learning (DL)-based methods have been proposed to annotate cell types of scRNA-seq data and have achieved impressive results. However, there are several limitations to these methods. First, they do not fully exploit cell-to-cell differential features. Second, they are developed based on shallow features and lack of flexibility in integrating high-order features in the data. Finally, the low-dimensional gene features may lead to overfitting in neural networks. To overcome those limitations, we propose a novel DL-based model, cell type annotation of single-cell RNA-seq data using residual graph convolutional neural network with contrastive learning (scRGCL), based on residual graph convolutional neural network and contrastive learning for cell type annotation of single-cell RNA-seq data. scRGCL mainly consists of a residual graph convolutional neural network, contrastive learning, and weight freezing. A residual graph convolutional neural network is utilized to extract complex high-order features from data. Contrastive learning can help the model learn meaningful cell-to-cell differential features. Weight freezing can avoid overfitting and help the model discover the impact of specific gene expression on cell type annotation. To verify the effectiveness of scRGCL, we compared its performance with six methods (three shallow learning algorithms and three state-of-the-art DL-based methods) on eight single-cell benchmark datasets from two species (seven in human and one in mouse). Experimental results not only show that scRGCL outperforms competing methods but also demonstrate the generalizability of scRGCL for cell type annotation. scRGCL is available at https://github.com/nathanyl/scRGCL.

PMID:39708840 | DOI:10.1093/bib/bbae662

Brief Bioinform. 2024 Nov 22;26(1):bbae651. doi: 10.1093/bib/bbae651.

ABSTRACT

Advances in three-dimensional (3D) genomics have revealed the spatial characteristics of chromatin interactions in gene expression regulation, which is crucial for understanding molecular mechanisms in biological processes. High-throughput technologies like ChIA-PET, Hi-C, and their derivatives methods have greatly enhanced our knowledge of 3D chromatin architecture. However, the chromatin interaction mechanisms remain largely unexplored. Deep learning, with its powerful feature extraction and pattern recognition capabilities, offers a promising approach for integrating multi-omics data, to build accurate predictive models of chromatin interaction matrices. This review systematically summarizes recent advances in chromatin interaction matrix prediction models. By integrating DNA sequences and epigenetic signals, we investigate the latest developments in these methods. This article details various models, focusing on how one-dimensional (1D) information transforms into the 3D structure chromatin interactions, and how the integration of different deep learning modules specifically affects model accuracy. Additionally, we discuss the critical role of DNA sequence information and epigenetic markers in shaping 3D genome interaction patterns. Finally, this review addresses the challenges in predicting chromatin interaction matrices, in order to improve the precise mapping of chromatin interaction matrices and DNA sequence, and supporting the transformation and theoretical development of 3D genomics across biological systems.

PMID:39708837 | DOI:10.1093/bib/bbae651

Neural Netw. 2024 Dec 10;184:107016. doi: 10.1016/j.neunet.2024.107016. Online ahead of print.

ABSTRACT

Federated Learning (FL) is an efficient, distributed machine learning paradigm that enables multiple clients to jointly train high-performance deep learning models while maintaining training data locally. However, due to its distributed computing nature, malicious clients can manipulate the prediction of the trained model through backdoor attacks. Existing defense methods require significant computational and communication overhead during the training or testing phases, limiting their practicality in resource-constrained scenarios and being unsuitable for the Non-IID data distribution typical in general FL scenarios. To address these challenges, we propose the FedPD framework, in which servers and clients exchange prototypes rather than model parameters, preventing the implantation of backdoor channels by malicious clients during FL training and effectively eliminating the success of backdoor attacks at the source, significantly reducing communication overhead. Additionally, prototypes can serve as global knowledge to correct clients' local training. Experiments and performance analysis show that FedPD achieves superior and consistent defense performance compared to existing representative approaches against backdoor attacks. In specific scenarios, FedPD can reduce the success rate of attacks by 90.73% compared to FedAvg without defense while maintaining the main task accuracy above 90%.

PMID:39708704 | DOI:10.1016/j.neunet.2024.107016

Comput Med Imaging Graph. 2024 Dec 17;119:102479. doi: 10.1016/j.compmedimag.2024.102479. Online ahead of print.

ABSTRACT

Accurate preoperative grading of prostate cancer is crucial for assisted diagnosis. Multi-parametric magnetic resonance imaging (MRI) is a commonly used non-invasive approach, however, the interpretation of MRI images is still subject to significant subjectivity due to variations in physicians' expertise and experience. To achieve accurate, non-invasive, and efficient grading of prostate cancer, this paper proposes a deep learning method that adaptively fuses dual-view MRI images. Specifically, a dual-view adaptive fusion model is designed. The model employs encoders to extract embedded features from two MRI sequences: T2-weighted imaging (T2WI) and apparent diffusion coefficient (ADC). The model reconstructs the original input images using the embedded features and adopts a cross-embedding fusion module to adaptively fuse the embedded features from the two views. Adaptive fusion refers to dynamically adjusting the fusion weights of the features from the two views according to different input samples, thereby fully utilizing complementary information. Furthermore, the model adaptively weights the prediction results from the two views based on uncertainty estimation, further enhancing the grading performance. To verify the importance of effective multi-view fusion for prostate cancer grading, extensive experiments are designed. The experiments evaluate the performance of single-view models, dual-view models, and state-of-the-art multi-view fusion algorithms. The results demonstrate that the proposed dual-view adaptive fusion method achieves the best grading performance, confirming its effectiveness for assisted grading diagnosis of prostate cancer. This study provides a novel deep learning solution for preoperative grading of prostate cancer, which has the potential to assist clinical physicians in making more accurate diagnostic decisions and has significant clinical application value.

PMID:39708679 | DOI:10.1016/j.compmedimag.2024.102479

Cell. 2024 Dec 18:S0092-8674(24)01329-1. doi: 10.1016/j.cell.2024.11.012. Online ahead of print.

ABSTRACT

Psychiatric disorders are influenced by genetic and environmental factors. However, their study is hindered by limitations on precisely characterizing human behavior. New technologies such as wearable sensors show promise in surmounting these limitations in that they measure heterogeneous behavior in a quantitative and unbiased fashion. Here, we analyze wearable and genetic data from the Adolescent Brain Cognitive Development (ABCD) study. Leveraging >250 wearable-derived features as digital phenotypes, we show that an interpretable AI framework can objectively classify adolescents with psychiatric disorders more accurately than previously possible. To relate digital phenotypes to the underlying genetics, we show how they can be employed in univariate and multivariate genome-wide association studies (GWASs). Doing so, we identify 16 significant genetic loci and 37 psychiatric-associated genes, including ELFN1 and ADORA3, demonstrating that continuous, wearable-derived features give greater detection power than traditional case-control GWASs. Overall, we show how wearable technology can help uncover new linkages between behavior and genetics.

PMID:39706190 | DOI:10.1016/j.cell.2024.11.012

Physiol Meas. 2024 Dec 20. doi: 10.1088/1361-6579/ada246. Online ahead of print.

ABSTRACT

OBJECTIVE: We study the changes in morphology of the photoplethysmography (PPG) signals-acquired from a select group of South Asian origin-through a low-cost PPG sensor, and correlate it with healthy aging which allows us to reliably estimate the vascular age and chronological age of a healthy person as well as the age group he/she belongs to.

METHODS: Raw infrared PPG data is collected from the finger-tip of 173 appar- ently healthy subjects, aged 3-61 years, via a non-invasive low- cost MAX30102 PPG sensor. In addition, the following metadata is recorded for each subject: age, gender, height, weight, family history of cardiac disease, smoking history, vitals (heart rate and SpO2). The raw PPG data is conditioned and 62 features are then extracted based upon the first four PPG derivatives. Then, correlation-based feature-ranking is performed which retains 26 most important features. Finally, the feature set is fed to three machine learning (ML) classifiers, i.e., logistic regression, random forest, eXtreme Gradient Boosting (XGBoost), and two shallow neural networks: a feedforward neural network (FFNN) and a convolutional neural network (CNN).

MAIN RESULTS: For the age group classification problem, the ensemble method XGboost stands out with an accuracy of 99% for both binary classification (3-20 years vs. 20+ years) and three-class classification (3-18 years, 18-23 years, 23+ years). For the vascular/chronological age prediction problem, the ensemble random forest method stands out with a mean absolute error (MAE) of 6.97 years.

SIGNIFICANCE: The results demonstrate that PPG is indeed a promising (i.e., low-cost, non-invasive) biomarker to study the healthy aging phenomenon.

PMID:39706154 | DOI:10.1088/1361-6579/ada246

Int J Med Inform. 2024 Dec 17;195:105760. doi: 10.1016/j.ijmedinf.2024.105760. Online ahead of print.

ABSTRACT

OBJECTIVE: To evaluate the effectiveness and usability of an AI-assisted tool in providing real-time assistance to counselors during suicide prevention helpline conversations.

METHODS: In this RCT, the intervention group used an AI-assisted tool, which generated suggestions based on sentence embeddings (i.e. BERT) from previous successful counseling sessions. Cosine similarity was used to present the top 5 chat situation to the counsellors. The control group did not have access to the tool (care as usual). Both groups completed a questionnaire assessing their self-efficacy at the end of each shift. Counselors' usage of the tool was evaluated by measuring frequency, duration and content of interactions.

RESULTS: In total, 48 counselors participated in the experiment: 27 counselors in the experimental condition and 21 counselors in the control condition. Together they rated 188 shifts. No significant difference in self-efficacy was observed between the two groups (p=0.36). However, counselors that used the AI-assisted tool had marginally lower response time and used the tool more often during conversations that had a longer duration. A deeper analysis of usage showed that the tool was frequently used in inappropriate situations, e.g. after the counselor had already provided a response to the help-seeker, defeating the purpose of the information. When the tool was employed appropriately (64 conversations), it provided usable information in 53 conversations (83%). However, counselors used the tool less frequently at optimal moments, indicating their potential lack of proficiency with using AI-assisted tools during helpline conversations or initial trust issues with the system.

CONCLUSION: The study demonstrates benefits and pitfalls of integrating AI-assisted tools in suicide prevention for improving counselor support. Despite the lack of significant impact on self-efficacy, the support tool provided usable suggestions and the frequent use during long conversations suggests counsellors may wish to use the tool in complex or challenging interactions.

PMID:39705915 | DOI:10.1016/j.ijmedinf.2024.105760