Independent Test Set Research Articles

A multicenter study of neurofibromatosis type 1 utilizing deep learning for whole body tumor identification

Deep-learning models have shown promise in differentiating between benign and malignant lesions. Previous studies have primarily focused on specific anatomical regions, overlooking tumors occurring throughout the body with highly heterogeneous whole-body backgrounds. Using neurofibromatosis type 1 (NF1) as an example, this study developed highly accurate MRI-based deep-learning models for the early automated screening of malignant peripheral nerve sheath tumors (MPNSTs) against complex whole-body background. In a Chinese seven-center cohort, data from 347 subjects were analyzed. Our one-step model incorporated normal tissue/organ labels to provide contextual information, offering a solution for tumors with complex backgrounds. To address privacy concerns, we utilized a lightweight deep neural network suitable for hospital deployment. The final model achieved an accuracy of 85.71% for MPNST diagnosis in the validation cohort and 84.75% accuracy in the independent test set, outperforming another classic two-step model. This success suggests potential for AI models in screening other whole-body primary/metastatic tumors.

A prospective and comparative study on improving the diagnostic accuracy of early gastric cancer based on deep convolutional neural network real-time diagnosis system (with video).

Endoscopic diagnosis of early gastric cancer (EGC) is a challenge. It is not clear whether deep convolutional neural network (DCNN) model could improve the endoscopists' diagnostic performance. We established a DCNN-assisted system and found that accuracy of diagnosis is higher than endoscopists. We prospectively collected an independent image test set of 1289 images and a video test set of 130 patients from three endoscopic centers to compare the diagnostic efficacy of 12 endoscopists before and after DCNN model assistance. Accuracy, sensitivity, specificity, time, and AUC were the main indicators for comparison. The DCNN model discriminated EGC from the control group (including ulcers and chronic gastritis) with an AUC of 0.917, a sensitivity of 93.38% (95% CI 91.09-95.12%), and a specificity of 90.07% (95% CI 87.59-92.10%) in the image dataset. The video test dataset have an AUC of 0.930, a sensitivity of 96.92% (95% CI 88.83-99.78%), and a specificity of 89.23% (95% CI 79.11-94.98%). The diagnostic performance of novice endoscopists was comparable to those of expert endoscopists with the DCNN model's assistance (accuracy: 95.22 vs. 96.16%) in image test dataset. In the video test, the novice endoscopists, accuracy after DCNN assistance was also improved from 79.36 to 86.41%, and from 86.28 to 91.03% for expert endoscopists. The mean pairwise kappa value of endoscopists was increased significantly with the DCNN model's assistance (0.705-0.753 vs.0.767-0.890) in image testing, and (0.657-0.793 vs. 0.738-0.905) in video testing. The diagnostic duration reduced considerably with the assistance of the DCNN model from 7.09 ± 0.6s to 5.05 ± 0.55s in image test, and from 2392.17 ± 7.77s to2378.34 ± 23.51s in video test. We developed a DCNN-assisted diagnostic system. And the system can improve the diagnostic performance of endoscopists and help novice endoscopists achieve diagnostic accuracy comparable to that of expert endoscopists.

Constructing a fall risk prediction model for hospitalized patients using machine learning

Study objectivesThis study aimed to identify the risk factors associated with falls in hospitalized patients, develop a predictive risk model using machine learning algorithms, and evaluate the validity of the model’s predictions.Study designA cross-sectional design was employed using data from the DRYAD public database.Research methodsThe study utilized data from the Fukushima Medical University Hospital Cohort Study, obtained from the DRYAD public database. 20% of the dataset was allocated as an independent test set, while the remaining 80% was utilized for training and validation. To address data imbalance in binary variables, the Synthetic Minority Oversampling Technique combined with Edited Nearest Neighbors (SMOTE-ENN) was applied. Univariate analysis and least absolute shrinkage and selection operator (LASSO) regression were used to analyze and screen variables. Predictive models were constructed by integrating key clinical features, and eight machine learning algorithms were evaluated to identify the most effective model. Additionally, SHAP (Shapley Additive Explanations) was used to interpret the predictive models and rank the importance of risk factors.ResultsThe final model included the following variables: Adl_standing, Adl_evacuation, Age_group, Planned_surgery, Wheelchair, History_of_falls, Hypnotic_drugs, Psychotropic_drugs, and Remote_caring_system. Among the evaluated models, the Random Forest algorithm demonstrated superior performance, achieving an AUC of 0.814 (95% CI: 0.802–0.827) in the training set, 0.781 (95% CI: 0.740–0.821) in the validation set, and 0.795 (95% CI: 0.770–0.820) in the test set.ConclusionMachine learning algorithms, particularly Random Forest, are effective in predicting fall risk among hospitalized patients. These findings can significantly enhance fall prevention strategies within healthcare settings.

Benchmarking protein language models for protein crystallization

The problem of protein structure determination is usually solved by X-ray crystallography. Several in silico deep learning methods have been developed to overcome the high attrition rate, cost of experiments and extensive trial-and-error settings, for predicting the crystallization propensities of proteins based on their sequences. In this work, we benchmark the power of open protein language models (PLMs) through the TRILL platform, a be-spoke framework democratizing the usage of PLMs for the task of predicting crystallization propensities of proteins. By comparing LightGBM / XGBoost classifiers built on the average embedding representations of proteins learned by different PLMs, such as ESM2, Ankh, ProtT5-XL, ProstT5, xTrimoPGLM, SaProt with the performance of state-of-the-art sequence-based methods like DeepCrystal, ATTCrys and CLPred, we identify the most effective methods for predicting crystallization outcomes. The LightGBM classifiers utilizing embeddings from ESM2 model with 30 and 36 transformer layers and 150 and 3000 million parameters respectively have performance gains by 3- than all compared models for various evaluation metrics, including AUPR (Area Under Precision-Recall Curve), AUC (Area Under the Receiver Operating Characteristic Curve), and F1 on independent test sets. Furthermore, we fine-tune the ProtGPT2 model available via TRILL to generate crystallizable proteins. Starting with 3000 generated proteins and through a step of filtration processes including consensus of all open PLM-based classifiers, sequence identity through CD-HIT, secondary structure compatibility, aggregation screening, homology search and foldability evaluation, we identified a set of 5 novel proteins as potentially crystallizable.

Deep Learning-Derived Quantitative Scores for Chronic Rhinosinusitis Assessment: Correlation With Quality of Life Outcomes.

Computed tomography (CT) plays a crucial role in assessing chronic rhinosinusitis, but lacks objective quantifiable indicators. This study aimed to use deep learning for automated sinus segmentation to generate distinct quantitative scores and explore their correlations with disease-specific quality of life. From July 2021 to August 2022, 445 CT data were collected from 2 medical centers. A deep learning model based on nnU-Net was trained for automatic sinus segmentation and internally validated using 300 cases. The remaining 145 cases were split into an external testing set (74 cases) and an independent testing set (71 cases). Two quantitative scores, the quantitative Lund-MacKay score and the quantitative opacification score (QOS), were derived from the segmentation results. The quantitative scores' efficacy was assessed by comparing them with the Lund-MacKay score (LMS), the 22-item Sinonasal Outcome Test score (SNOT-22), and other clinical variables through correlation analyses. Furthermore, the relationship between quantitative scores and postoperative quality of life improvement was explored using single-factor logistic regression. The segmentation model achieved average Dice similarity coefficients of 0.993, 0.978, 0.958, and 0.871 for the training, validation, external testing, and independent testing sets, respectively. Both quantitative scores significantly correlated with the LMS (rho = 0.87 and rho = 0.70, P < .001). Neither score correlated with the total SNOT-22 score, although the modified QOS showed significant correlations with the nasal and sleep subdomains (rho = 0.26 and rho = 0.27, P < .05). No significant association was found between quantitative score and postoperative improvement in quality of life. Deep learning enables the automated segmentation of sinuses on CT scans, producing quantitative scores of sinus opacification. These automatic quantitative scores may serve as tools for chronic rhinosinusitis assessment.

Whole lung radiomic features are associated with overall survival in patients with locally advanced non-small cell lung cancer treated with definitive radiotherapy

BackgroundSeveral studies have suggested that lung tissue heterogeneity is associated with overall survival (OS) in lung cancer. However, the quantitative relationship between the two remains unknown. The purpose of this study is to investigate the prognostic value of whole lung-based and tumor-based radiomics for OS in LA-NSCLC treated with definitive radiotherapy.MethodsA total of 661 patients with LA-NSCLC treated with definitive radiotherapy in combination with chemotherapy were enrolled in this study, with 292 patients in the training set, 57 patients from the same hospital from January to December 2017 as an independent test set (test-set-1), 83 patients from a multi-institutional prospective clinical trial data set (RTOG0617) as test-set-2, and 229 patients from a Dutch radiotherapy center as test-set-3. Tumor-based radiomic features and whole lung-based radiomic features were extracted from primary tumor and whole lungs (excluding the primary tumor) delineations in planning CT images. Feature selection of radiomic features was done by the least absolute shrinkage (LASSO) method embedded with a Cox proportional hazards (CPH) model with 5-fold cross-internal validation, with 1000 bootstrap samples. Radiomics prognostic scores (RS) were calculated by CPH regression based on selected features. Three models based on a tumor RS, and a lung RS separately and their combinations were constructed. The Harrell concordance index (C-index) and calibration curves were used to evaluate the discrimination and calibration performance. Patients were stratified into high and low risk groups based on median RS, and a log-rank test was performed.ResultsThe discrimination ability of lung- and tumor-based radiomics model was similar in terms of C-index, 0.69 vs. 0.68 in training set, 0.68 vs. 0.66 in test-set-1, 0.61 vs. 0.62 in test-set-2, 0.65 vs. 0.64 in test-set-3. The combination of tumor- and lung-based radiomics model performed best, with C-index of 0.71 in training set, 0.70 in test-set-1, 0.69 in test-set-2, and 0.68 in test-set-3. The calibration curve showed good agreement between predicted values and actual values. Patients were well stratified in training set, test-set-1 and test-set-3. In test-set-2, it was only whole lung-based RS that could stratify patients well and tumor-based RS performed bad.ConclusionLung- and tumor-based radiomic features have the power to predict OS in LA-NSCLC. The combination of tumor- and lung-based radiomic features can achieve optimal performance.

Discovery of Methylated DNA Biomarkers for Potential Non-Endoscopic Detection of Barrett's Esophagus and Esophageal Adenocarcinoma.

We sought to develop a minimally-invasive, robust, accessible nonendoscopic strategy to diagnose Barrett's esophagus (BE), esophageal adenocarcinoma (EAC), and its immediate precursor lesion, high-grade dysplasia (HGD) based on methylated DNA biomarkers applied to a retrievable sponge-capsule device in a cohort representative of the BE population (i.e., mostly short-segment, non-dysplastic BE, NDBE). We identified 12 candidate methylation markers to distinguish normal vs. abnormal esophagus. These 12 markers were first assayed in 21-paired matched NDBE-normal esophageal tissues, then assessed in a case-control study of 234 esophageal samples collected using a sponge-capsule device. A classification algorithm was developed using Least Absolute Shrinkage and Selection Operator (LASSO) in a 199-patient training set and tested in an independent 35-patient test set. Twelve markers (A1BG, C9orf50, cg00720137, FLI1, GRAMD1B, HOXB13, IRF4, KCNQ3, NTNG1, SPX, TBC1D30, and USP44) were significantly hypermethylated (i.e., all P<0.05) in BE vs. matched normal esophageal biopsies. A discriminatory three-gene LASSO panel (USP44, TBCD1D30, NELL1), adjusted for age and sex, accurately distinguished HGD or EAC from normal control patients in both training (AUC=0.911, 95%CI=0.863-0.959) and test (AUC=0.969, 95%CI=0.911-1.00) sets. In normal vs. NDBE/LGD/HGD/EAC patients, this algorithm exhibited AUCs of 0.862 (95%CI=0.812-0.912) and 0.864 (95%CI=0.745-0.982) in training and test sets, respectively. In normal vs. NDBE patients, the algorithm yielded AUCs of 0.819 (95%CI=0.748-0.889) and 0.776 (95%CI=0.583-0.968) in training and test sets, respectively. This discriminatory biomarker panel algorithm exemplifies a practical nonendoscopic strategy to diagnose BE, HGD, and EAC using a minimally-invasive sponge-capsule device coupled with DNA methylation markers.

Deep learning approach based on a patch residual for pediatric supracondylar subtle fracture detection.

Supracondylar humerus fractures in children are among the most common elbow fractures in pediatrics. However, their diagnosis can be particularly challenging due to the anatomical characteristics and imaging features of the pediatric skeleton. In recent years, convolutional neural networks (CNNs) have achieved notable success in medical image analysis, though their performance typically relies on large-scale, high-quality labeled datasets. Unfortunately, labeled samples for pediatric supracondylar fractures are scarce and difficult to obtain. To address this issue, this paper introduces a deep learning-based multi-scale patch residual network (MPR) for the automatic detection and localization of subtle pediatric supracondylar fractures. The MPR framework combines a CNN for automatic feature extraction with a multi-scale generative adversarial network to model skeletal integrity using healthy samples. By leveraging healthy images to learn the normal skeletal distribution, the approach reduces the dependency on labeled fracture data and effectively addresses the challenges posed by limited pediatric datasets. Datasets from two different hospitals were used, with data augmentation techniques applied during both training and validation. On an independent test set, the proposed model achieves an accuracy of 90.5%, with 89% sensitivity, 92% specificity, and an F1 score of 0.906-outperforming the diagnostic accuracy of emergency medicine physicians and approaching that of pediatric radiologists. Furthermore, the model demonstrates a fast inference speed of 1.1 s per sheet, underscoring its substantial potential for clinical application.

A novel hybrid deep learning framework based on biplanar X-ray radiography images for bone density prediction and classification.

To explore the feasibility of using a hybrid deep learning framework (HDLF) to establish a model for BMD prediction and classification based on BPX images. This study aimed to establish an automated tool for screening patients at a high risk of osteoporosis. A total of 906 BPX scans from 453 subjects were included in this study, with QCT results serving as the reference standard. The training-validation set:independent test set ratio was 4:1. The L1-L3 vertebral bodies were manually annotated by experienced radiologists, and the HDLF was established to predict BMD and diagnose abnormality based on BPX images and clinical information. The performance metrics of the models were calculated and evaluated. The values of the BMD prediction regression model in the independent test set based on BPX images and multimodal data (BPX images and clinical information) were 0.77 and 0.79, respectively. The Pearson correlation coefficients were 0.88 and 0.89, respectively, with P-values < 0.001. Bland-Altman analysis revealed no significant difference between the predictions of the models and QCT results. The classification model achieved the highest AUC of 0.97 based on multimodal data in the independent test set, with an accuracy of 0.93, sensitivity of 0.84, specificity of 0.96, and F1 score of 0.93. This study demonstrates that deep learning neural networks applied to BPX images can accurately predict BMD and perform classification diagnoses, which can reduce the radiation risk, economic consumption, and time consumption associated with specialized BMD measurement.

MDFGNN-SMMA: prediction of potential small molecule-miRNA associations based on multi-source data fusion and graph neural networks

BackgroundMicroRNAs (miRNAs) are pivotal in the initiation and progression of complex human diseases and have been identified as targets for small molecule (SM) drugs. However, the expensive and time-intensive characteristics of conventional experimental techniques for identifying SM-miRNA associations highlight the necessity for efficient computational methodologies in this field.ResultsIn this study, we proposed a deep learning method called Multi-source Data Fusion and Graph Neural Networks for Small Molecule-MiRNA Association (MDFGNN-SMMA) to predict potential SM-miRNA associations. Firstly, MDFGNN-SMMA extracted features of Atom Pairs fingerprints and Molecular ACCess System fingerprints to derive fusion feature vectors for small molecules (SMs). The K-mer features were employed to generate the initial feature vectors for miRNAs. Secondly, cosine similarity measures were computed to construct the adjacency matrices for SMs and miRNAs, respectively. Thirdly, these feature vectors and adjacency matrices were input into a model comprising GAT and GraphSAGE, which were utilized to generate the final feature vectors for SMs and miRNAs. Finally, the averaged final feature vectors were utilized as input for a multilayer perceptron to predict the associations between SMs and miRNAs.ConclusionsThe performance of MDFGNN-SMMA was assessed using 10-fold cross-validation, demonstrating superior compared to the four state-of-the-art models in terms of both AUC and AUPR. Moreover, the experimental results of an independent test set confirmed the model’s generalization capability. Additionally, the efficacy of MDFGNN-SMMA was substantiated through three case studies. The findings indicated that among the top 50 predicted miRNAs associated with Cisplatin, 5-Fluorouracil, and Doxorubicin, 42, 36, and 36 miRNAs, respectively, were corroborated by existing literature and the RNAInter database.

Semi-Supervised Learning Allows for Improved Segmentation With Reduced Annotations of Brain Metastases Using Multicenter MRI Data.

Deep learning-based segmentation of brain metastases relies on large amounts of fully annotated data by domain experts. Semi-supervised learning offers potential efficient methods to improve model performance without excessive annotation burden. This work tests the viability of semi-supervision for brain metastases segmentation. Retrospective. There were 156, 65, 324, and 200 labeled scans from four institutions and 519 unlabeled scans from a single institution. All subjects included in the study had diagnosed with brain metastases. 1.5 T and 3 T, 2D and 3D T1-weighted pre- and post-contrast, and fluid-attenuated inversion recovery (FLAIR). Three semi-supervision methods (mean teacher, cross-pseudo supervision, and interpolation consistency training) were adapted with the U-Net architecture. The three semi-supervised methods were compared to their respective supervised baseline on the full and half-sized training. Evaluation was performed on a multinational test set from four different institutions using 5-fold cross-validation. Method performance was evaluated by the following: the number of false-positive predictions, the number of true positive predictions, the 95th Hausdorff distance, and the Dice similarity coefficient (DSC). Significance was tested using a paired samples t test for a single fold, and across all folds within a given cohort. Semi-supervision outperformed the supervised baseline for all sites with the best-performing semi-supervised method achieved an on average DSC improvement of 6.3% ± 1.6%, 8.2% ± 3.8%, 8.6% ± 2.6%, and 15.4% ± 1.4%, when trained on half the dataset and 3.6% ± 0.7%, 2.0% ± 1.5%, 1.8% ± 5.7%, and 4.7% ± 1.7%, compared to the supervised baseline on four test cohorts. In addition, in three of four datasets, the semi-supervised training produced equal or better results than the supervised models trained on twice the labeled data. Semi-supervised learning allows for improved segmentation performance over the supervised baseline, and the improvement was particularly notable for independent external test sets when trained on small amounts of labeled data. Artificial intelligence requires extensive datasets with large amounts of annotated data from medical experts which can be difficult to acquire due to the large workload. To compensate for this, it is possible to utilize large amounts of un-annotated clinical data in addition to annotated data. However, this method has not been widely tested for the most common intracranial brain tumor, brain metastases. This study shows that this approach allows for data efficient deep learning models across multiple institutions with different clinical protocols and scanners. 3 TECHNICAL EFFICACY: Stage 2.

Utilizing machine-learning techniques on MRI radiomics to identify primary tumors in brain metastases.

To develop a machine learning-based clinical and/or radiomics model for predicting the primary site of brain metastases using multiparametric magnetic resonance imaging (MRI). A total of 202 patients (87 males, 115 females) with 439 brain metastases were retrospectively included, divided into training sets (brain metastases of lung cancer [BMLC] n = 194, brain metastases of breast cancer [BMBC] n = 108, brain metastases of gastrointestinal tumor [BMGiT] n = 48) and test sets (BMLC n = 50, BMBC n = 27, BMGiT n = 12). A total of 3,404 quantitative image features were obtained through semi-automatic segmentation from MRI images (T1WI, T2WI, FLAIR, and T1-CE). Intra-class correlation coefficient (ICC) was used to examine segmentation stability between two radiologists. Radiomics features were selected using analysis of variance (ANOVA), recursive feature elimination (RFE), and Kruskal-Wallis test. Three machine learning classifiers were used to build the radiomics model, which was validated using five-fold cross-validation on the training set. A comprehensive model combining radiomics and clinical features was established, and the diagnostic performance was compared by area under the curve (AUC) and evaluated in an independent test set. The radiomics model differentiated BMGiT from BMLC (13 features, AUC = 0.915 ± 0.071) or BMBC (20 features, AUC = 0.954 ± 0.064) with high accuracy, while the classification between BMLC and BMBC was unsatisfactory (11 features, AUC = 0.729 ± 0.114). However, the combined model incorporating radiomics and clinical features improved the predictive performance, with AUC values of 0.965 for BMLC vs. BMBC, 0.991 for BMLC vs. BMGiT, and 0.935 for BMBC vs. BMGiT. The machine learning-based radiomics model demonstrates significant potential in distinguishing the primary sites of brain metastases, and may assist screening of primary tumor when brain metastasis is suspected whereas history of primary tumor is absent.

S-DCNN: prediction of ATP binding residues by deep convolutional neural network based on SMOTE.

The realization of many protein functions requires binding with ligands. As a significant protein-binding ligand, ATP plays a crucial role in various biological processes. Currently, the precise prediction of ATP binding residues remains challenging. Based on the sequence information, this paper introduces a method called S-DCNN for predicting ATP binding residues, utilizing a deep convolutional neural network (DCNN) enhanced with the synthetic minority over-sampling technique (SMOTE). The incorporation of additional feature parameters such as dihedral angles, energy, and propensity factors into the standard parameter set resulted in a significant enhancement in prediction accuracy on the ATP-289 dataset. The S-DCNN achieved the highest Matthews correlation coefficient value of 0.5031 and an accuracy rate of 97.06% on an independent test set. Furthermore, when applied to the ATP-221 and ATP-388 datasets for validation, the S-DCNN outperformed existing methods on ATP-221 and performed comparably to other methods on ATP-388 during independent testing. Our experimental results underscore the efficacy of the S-DCNN in accurately predicting ATP binding residues, establishing it as a potent tool in the prediction of ATP binding residues.

MVSLLnc: LncRNA subcellular localization prediction based on multi-source features and two-stage voting strategy.

The subcellular localization of long non-coding RNAs (lncRNAs) is crucial for understanding the function of lncRNAs. Since the traditional biological experimental methods are time-consuming and some existing computational methods rely on high computing power, we are committed to finding a simple and easy-to-implement method to achieve more efficient prediction of the subcellular localization of lncRNAs. In this work, we proposed a model based on multi-source features and two-stage voting strategy for predicting the subcellular localization of lncRNAs (MVSLLnc). The multi-source features include k-mer frequency, features based on the coordinate values of Chaos Game Representation (CGR) and features based on physicochemical property (PhyChe). We feed the multi-source features into the traditional machine learning classifiers RF, SVM and XGBoost, respectively, and perform the final prediction task with two-stage voting strategy. Experimental results on three benchmark datasets show that the accuracy can reach 0.829, 0.793 and 0.968, respectively. The accuracy on three independent test sets is 0.642, 0.737 and 0.518, respectively, which are competitive with the existing methods. Our ablation analyses show that the two-stage voting strategy can make full use of the advantages of multi-source features and multiple classifiers, and obtain more robust results.

Accurately predicting optimal conditions for microorganism proteins through geometric graph learning and language model

Proteins derived from microorganisms that survive in the harshest environments on Earth have stable activity under extreme conditions, providing rich resources for industrial applications and enzyme engineering. Due to the time-consuming nature of experimental determinations, it is imperative to develop computational models for fast and accurate prediction of protein optimal conditions. Previous studies were limited by the scarcity of data and the neglect of protein structures. To solve these problems, we constructed an up-to-date dataset with 175,905 non-redundant proteins and proposed a new model GeoPoc based on geometric graph learning for the protein optimal temperature, pH, and salt concentration prediction. GeoPoc leverages protein structures and sequence embeddings extracted from pre-trained language model, and further employs a geometric graph transformer network to capture the sequence and spatial information. We first focused on in-house validation for optimal temperature prediction for robustness assessment, and achieved a PCC of 0.78. The algorithm is further confirmed in an independent test set, where GeoPoc surpasses the state-of-the-art method by 2.3% in AUC. Additionally, GeoPoc was extended to pH and salt concentration prediction, and obtained AUC scores of 0.78 and 0.77, respectively. Through further interpretable analysis, GeoPoc elucidates the critical physicochemical properties that contribute to enhancing protein thermostability.

Development of an accurate breast cancer detection classifier based on platelet RNA

Platelets possess cancer-induced reprogramming properties, thereby contributing to RNA profile alterations and further cancer progression, while the former is considered a promising biosource for cancer detection. Hence, tumor-educated platelets (TEP) are considered a prospective novel method for early breast cancer (BC) screening. Our study integrated the data from 276 patients with untreated BC, 95 with benign disease controls, 214 healthy controls, and 2 who underwent mastectomy in Chinese and European cohorts to develop a 10-biomarker diagnostic model. The model demonstrated high diagnostic performance for BC in an independent test set (n = 177) with an area under the curve of 0.957. The sensitivity for BC diagnosis was 89.2%, with 100% specificity in asymptomatic controls, while that for the symptomatic group, including benign tumors and inflammatory diseases, was 62.1%. The model demonstrated substantial accuracy for stages 0–III BC (80% for stage 0 [n = 5], 83.3% for stage I [n = 12], 94.6% for stage II [n = 37], and 88.9% for stage III [n = 9]) and precisely helped determine residual cancer in two patients who underwent mastectomy. Moreover, our developed classifiers distinguish different BC subtypes properly. In summary, we created and tested a new TEP-RNA-based BC diagnostic model that was confirmed valid and demonstrated high efficiency in detecting early-stage BC and heterogeneous subtypes, including recurrent tumors. However, these results warrant more validation in larger population-based prospective studies before clinical implementation.

Porter 6: Protein Secondary Structure Prediction by Leveraging Pre-Trained Language Models (PLMs)

Accurately predicting protein secondary structure (PSSP) is crucial for understanding protein function, which is foundational to advancements in drug development, disease treatment, and biotechnology. Researchers gain critical insights into protein folding and function within cells by predicting protein secondary structures. The advent of deep learning models, capable of processing complex sequence data and identifying meaningful patterns, offer substantial potential to enhance the accuracy and efficiency of protein structure predictions. In particular, recent breakthroughs in deep learning—driven by the integration of natural language processing (NLP) algorithms—have significantly advanced the field of protein research. Inspired by the remarkable success of NLP techniques, this study harnesses the power of pre-trained language models (PLMs) to advance PSSP prediction. We conduct a comprehensive evaluation of various deep learning models trained on distinct sequence embeddings, including one-hot encoding and PLM-based approaches such as ProtTrans and ESM-2, to develop a cutting-edge prediction system optimized for accuracy and computational efficiency. Our proposed model, Porter 6, is an ensemble of CBRNN-based predictors, leveraging the protein language model ESM-2 as input features. Porter 6 achieves outstanding performance on large-scale, independent test sets. On a 2022 test set, the model attains an impressive 86.60% accuracy in three-state (Q3) and 76.43% in eight-state (Q8) classifications. When tested on a more recent 2024 test set, Porter 6 maintains robust performance, achieving 84.56% in Q3 and 74.18% in Q8 classifications. This represents a significant 3% improvement over its predecessor, outperforming or matching state-of-the-art approaches in the field.

Establishing age-group specific reference intervals of human salivary proteome and its preliminary application for epilepsy diagnosis.

Salivary proteins serve multifaceted roles in maintaining oral health and hold significant potential for diagnosing and monitoring diseases due to the non-invasive nature of saliva sampling. However, the clinical utility of current saliva biomarker studies is limited by the lack of reference intervals (RIs) to correctly interpret the testing result. Here, we developed a rapid and robust saliva proteome profiling workflow, obtaining coverage of >1,200 proteins from a 50-µL unstimulated salivary flow with 30 min gradients. With the workflow, we investigated protein variation in a cohort of 1,743 healthy individuals. The significant differences in non-linear saliva proteomes among age groups resulted in the establishment of age-related RIs covering 1,123 salivary protein variations. We then utilized a cohort of 30 epilepsy patients as a case study to illustrate the practical application of RIs in identifying disease-enriched outlier proteins, elucidating their cellular origins, determining disease diagnosis, and visualizing outlier proteins in each epilepsy patient. Our study showed the classification model based on the RI achieved PR-AUC of 0.815 (95%CI: 0.813-0.826). Additionally, we validated these results in an independent test set. Furthermore, the epilepsy cohort could be further stratified into 2 major subtypes, with one subtype characterized by disrupted metabolic proteins and containing mostly Focal Cortical Dysplasia (FCD) type III epilepsy patients. Overall, our study provided a proof-of-principle workflow for the use of salivary proteome for health monitoring, epilepsy diagnosis and subtyping.

Automated Detection and Differentiation of Stanford Type A and Type B Aortic Dissections in CTA Scans Using Deep Learning.

To develop and validate a model system using deep learning algorithms for the automatic detection of type A aortic dissection (AD), and differentiate it from normal and type B AD patients. In this retrospective study, a deep learning model is developed, based on aortic computed tomography angiography (CTA) scans of 498 patients using training, validation and test sets of 398, 50 and 50 patients, respectively. An independent test set of 316 patients is used to validate and evaluate its performance. Our model comprises two components. The first one is an objection detection model, which can identify the aorta from CTA. The second one is a dissection classification model, which can automatically detect the presence of aortic dissection and determine its type based on Stanford classification. Overall, the sensitivity and specificity for Type A AD were 0.969 and 0.982, for Type B AD were 0.946 and 0.996 and for normal cases were 0.988 and 1.000, respectively. The average processing time per CTA scan was 7.9 ± 2.8 s. (mean ± standard deviation). This deep learning automatic model can accurately and quickly detect type A AD patients, and could serve as an imaging triage in an emergency setting and facilitate early intervention and surgery to decrease the mortality rates of type A AD patients.

One-year mortality prediction of patients with hepatitis in Kazakhstan based on administrative health data: A machine learning approach

&lt;b&gt;Background and objective: &lt;/b&gt;Hepatitis B virus (HBV) and hepatitis C virus (HCV) are major contributors to chronic viral hepatitis (CVH), leading to significant global health mortality. This study aims to predict the one-year mortality in patients with CVH using their demographics and health records.&lt;br /&gt; &lt;b&gt;Methods: &lt;/b&gt;Clinical data from 82,700 CVH patients diagnosed with HBV or HCV between January 2014 and December 2019 was analyzed. We developed a machine learning (ML) platform based on six broad categories including linear, nearest neighbors, discriminant analysis, support vector machine, naïve Bayes, and ensemble (gradient boosting, AdaBoost, and random forest) models to predict the one-year mortality. Feature importance analysis was performed by computing SHapley Additive exPlanations (SHAP).&lt;br /&gt; &lt;b&gt;Results: &lt;/b&gt;The models achieved an area under the curve between 0.74 and 0.8 on independent test sets. Key predictors of mortality were age, sex, hepatitis type, and ethnicity.&lt;br /&gt; &lt;b&gt;Conclusion: &lt;/b&gt;ML with administrative health data can be utilized to accurately predict one-year mortality in CVH patients. Future integration with detailed laboratory and medical history data could further enhance model performance.