Environ Geochem Health. 2025 Mar 5;47(4):101. doi: 10.1007/s10653-025-02413-z.

ABSTRACT

With increasing scholarly focus on the ramifications of land use changes on water quality, although substantial research has been undertaken, the findings demonstrate pronounced spatial variability and the heterogeneity of research methodologies. To address this critical gap, this review offers a rigorous evaluation of the strengths and limitations of current research methodologies, providing targeted recommendations for refinement. It systematically assesses the existing body of literature concerning the influence of land use changes on water quality, with particular emphasis on the spatial heterogeneity of research results and the uniformity of employed methodologies. Despite variations in geographical contexts and research subjects, the methodological paradigms remain largely consistent, typically encompassing the acquisition and analysis of water quality and land use data, the delineation of buffer zones, and the application of correlation and regression analyses. However, these approaches encounter limitations in addressing regional disparities, nonlinear interactions, and real-time monitoring complexities. The review advocates for methodological advancements, such as the integration of automated monitoring systems and IoT technologies, alongside the fusion of deep learning algorithms with remote sensing techniques, to enhance both the precision and efficiency of data collection. Furthermore, it recommends the standardization of buffer zone delineation, the reinforcement of foundational water quality assessments, and the utilization of catchment-scale analyses to more accurately capture the influence of land use changes on water quality. Future inquiries should prioritize the development of interdisciplinary ecological models to elucidate the interaction and feedback mechanisms between land use, water quality, and climate change.

PMID:40042544 | DOI:10.1007/s10653-025-02413-z

Alzheimers Dement. 2025 Mar;21(3):e14601. doi: 10.1002/alz.14601.

ABSTRACT

INTRODUCTION: The utility of retinal photography-derived aging biomarkers for predicting cognitive decline remains under-explored.

METHODS: A memory-clinic cohort in Singapore was followed-up for 5 years. RetiPhenoAge, a retinal aging biomarker, was derived from retinal photographs using deep-learning. Using competing risk analysis, we determined the associations of RetiPhenoAge with cognitive decline and dementia, with the UK Biobank utilized as the replication cohort. The associations of RetiPhenoAge with MRI markers(cerebral small vessel disease [CSVD] and neurodegeneration) and its underlying plasma proteomic profile were evaluated.

RESULTS: Of 510 memory-clinic subjects(N = 155 cognitive decline), RetiPhenoAge associated with incident cognitive decline (subdistribution hazard ratio [SHR] 1.34, 95% confidence interval [CI] 1.10-1.64, p = 0.004), and incident dementia (SHR 1.43, 95% CI 1.02-2.01, p = 0.036). In the UK Biobank (N = 33 495), RetiPhenoAge similarly predicted incident dementia (SHR 1.25, 95% CI 1.09-1.41, p = 0.008). RetiPhenoAge significantly associated with CSVD, brain atrophy, and plasma proteomic signatures related to aging.

DISCUSSION: RetiPhenoAge may provide a non-invasive prognostic screening tool for cognitive decline and dementia.

HIGHLIGHTS: RetiPhenoAge, a retinal aging marker, was studied in an Asian memory clinic cohort. Older RetiPhenoAge predicted future cognitive decline and incident dementia. It also linked to neuropathological markers, and plasma proteomic profiles of aging. UK Biobank replication found that RetiPhenoAge predicted 12-year incident dementia. Future studies should validate RetiPhenoAge as a prognostic biomarker for dementia.

PMID:40042460 | DOI:10.1002/alz.14601

Microb Biotechnol. 2025 Mar;18(3):e70121. doi: 10.1111/1751-7915.70121.

ABSTRACT

Antimicrobial peptides (AMPs) face stability and toxicity challenges in clinical use. Stapled modification enhances their stability and effectiveness, but its application in peptide design is rarely reported. This study built ten prediction models for stapled AMPs using deep and machine learning, tested their accuracy with an independent data set and wet lab experiments, and characterised stapled loop structures using structural, sequence and amino acid descriptors. AlphaFold improved stapled peptide structure prediction. The support vector machine model performed best, while two deep learning models achieved the highest accuracy of 1.0 on an external test set. Designed cysteine- and lysine-stapled peptides inhibited various bacteria with low concentrations and showed good serum stability and low haemolytic activity. This study highlights the potential of the deep learning method in peptide modification and design.

PMID:40042163 | DOI:10.1111/1751-7915.70121

Adv Sci (Weinh). 2025 Mar 5:e2415808. doi: 10.1002/advs.202415808. Online ahead of print.

ABSTRACT

Current imaging technologies are limited in their capability to simultaneously capture intracellular and extracellular dynamics in a spatially and temporally resolved manner. This study presents a multimodal imaging system that integrates nanoplasmonic sensing with multichannel fluorescence imaging to concomitantly analyze intracellular and extracellular processes in space and time at the single-cell level. Utilizing a highly sensitive gold nanohole array biosensor, the system provides label-free and real-time monitoring of extracellular secretion, while implementing nanoplasmonic-compatible multichannel fluorescence microscopy enables to visualize the interconnected intracellular activities. Combined with deep-learning-assisted image processing, this integrated approach allows multiparametric and simultaneous study of various cellular constituents in hundreds of individual cells with subcellular spatial and minute-level temporal resolution over extended periods of up to 20 h. The system's utility is demonstrated by characterizing a range of secreted biomolecules and fluorescence toolkits across three distinct applications: visualization of secretory behaviors along with subcellular organelles and metabolic processes, concurrent monitoring of protein expression and secretion, and assessment of cell cycle phases alongside their corresponding secretory profiles. By offering comprehensive insights, the multifunctional approach is expected to enhance holistic readouts of biological systems, facilitating new discoveries in both fundamental and translational sciences.

PMID:40042114 | DOI:10.1002/advs.202415808

Clin Chem Lab Med. 2025 Mar 6. doi: 10.1515/cclm-2024-1219. Online ahead of print.

NO ABSTRACT

PMID:40042089 | DOI:10.1515/cclm-2024-1219

ACS Sens. 2025 Mar 5. doi: 10.1021/acssensors.4c03472. Online ahead of print.

ABSTRACT

Anterior talofibular ligament (ATFL) sprain is one of the most prevalent sports-related injuries, so proper evaluation of ligament sprains is critical for treatment options. However, existing tests suffer from a lack of standardized quantitative evaluation criteria, interindividual variability, incompatible materials, or risks of infection. Although advanced medical diagnostic methods already have been using noninvasive, portable, and wearable diagnostic electronics, these devices have insufficient adhesion to accurately respond to internal body injuries. Therefore, we propose a high-adhesive hydrogel-based strain sensor made from gelatin, cellulose nanofiber (CNF), and cross-linked poly(acrylic acid) grafted with N-hydrosuccinimide ester. The adhesive strain sensor, with excellent conformability and stretchability, firmly adheres to the skin, making it suitable for accurately evaluating the severity of anterior talofibular ligament sprain. Its strong adhesive (up to 192 kPa) can adapt to the surface characterization of ankles. The high-adhesive hydrogel-based strain sensor has a high tensile strength (680%) and achieves a high gauge factor (GF) of 8.29. Simultaneously, it also presents a 40 μm ultralow detection limit. Additionally, after a deep learning model was integrated to improve sensing accuracy, the system achieved a diagnostic accuracy of 95%, significantly surpassing the magnetic resonance imaging (MRI) gold standard of 81.1%.

PMID:40042081 | DOI:10.1021/acssensors.4c03472

Patterns (N Y). 2025 Feb 6;6(2):101177. doi: 10.1016/j.patter.2025.101177. eCollection 2025 Feb 14.

ABSTRACT

Balancing prediction accuracy, model interpretability, and domain generalization (also known as [a.k.a.] out-of-distribution testing/evaluation) is a central challenge in machine learning. To assess this challenge, we took 120 interpretable and 166 opaque models from 77,640 tuned configurations, complemented with ChatGPT, 3 probabilistic language models, and Vec2Read. The models first performed text classification to derive principles of textual complexity (task 1) and then generalized these to predict readers' appraisals of processing difficulty (task 2). The results confirmed the known accuracy-interpretability trade-off on task 1. However, task 2's domain generalization showed that interpretable models outperform complex, opaque models. Multiplicative interactions further improved interpretable models' domain generalization incrementally. We advocate for the value of big data for training, complemented by (1) external theories to enhance interpretability and guide machine learning and (2) small, well-crafted out-of-distribution data to validate models-together ensuring domain generalization and robustness against data shifts.

PMID:40041855 | PMC:PMC11873011 | DOI:10.1016/j.patter.2025.101177

Health Sci Rep. 2025 Mar 2;8(3):e70498. doi: 10.1002/hsr2.70498. eCollection 2025 Mar.

ABSTRACT

BACKGROUND AND AIMS: The increasing elderly population presents significant challenges for healthcare systems, necessitating innovative solutions for continuous health monitoring. This study develops and validates an IoT-based elderly monitoring system designed to enhance the quality of life for elderly people. The system features a robust Android-based user interface integrated with the Firebase cloud platform, ensuring real-time data collection and analysis. In addition, a supervised machine learning technology is implemented to conduct prediction task of the observed user whether in "stable" or "not stable" condition based on real-time parameter.

METHODS: The system architecture adopts the IoT layer including physical layer, network layer, and application layer. Device validation is conducted by involving six participants to measure the real-time data of heart-rate, oxygen saturation, and body temperature, then analysed by mean average percentage error (MAPE) to define error rate. A comparative experiment is conducted to define the optimal supervised machine learning model to be deployed into the system by analysing evaluation metrics. Meanwhile, the user satisfaction aspect evaluated by the terms of usability, comfort, security, and effectiveness.

RESULTS: IoT-based elderly health monitoring system has been constructed with a MAPE of 0.90% across the parameters: heart-rate (1.68%), oxygen saturation (0.57%), and body temperature (0.44%). In machine learning experiment indicates XGBoost model has the optimal performance based on the evaluation metrics of accuracy and F1 score which generates 0.973 and 0.970, respectively. In user satisfaction aspect based on usability, comfort, security, and effectiveness achieving a high rating of 86.55%.

CONCLUSION: This system offers practical applications for both elderly users and caregivers, enabling real-time monitoring of health conditions. Future enhancements may include integration with artificial intelligence technologies such as machine learning and deep learning to predict health conditions from data patterns, further improving the system's capabilities and effectiveness in elderly care.

PMID:40041774 | PMC:PMC11873372 | DOI:10.1002/hsr2.70498

PNAS Nexus. 2025 Feb 20;4(2):pgaf054. doi: 10.1093/pnasnexus/pgaf054. eCollection 2025 Feb.

ABSTRACT

Discovering manufacturing processes has been largely experienced-based. We propose a shift to a systematic approach driven by dependencies between energy inputs and performance outputs. Uncovering these dependencies across diverse process classes requires a universal language that characterizes process inputs and performances. Traditional manufacturing languages, with their individualized syntax and terminology, hinder the characterization across varying length scales and energy inputs. To enable the evaluation of process dependencies, we propose a broad manufacturing language that facilitates the characterization of diverse process classes, which include energy inputs, tool-material interactions, material compatibility, and performance outputs. We analyze the relationships between these characteristics by constructing a dataset of over 50 process classes, which we use to train a variational autoencoder (VAE) model. This generative model encodes our dataset into a 2D latent space, where we can explore, select, and generate processes based on desired performances and retrieve the corresponding process characteristics. After verifying the dependencies derived from the VAE model match with existing knowledge on manufacturing processes, we demonstrate the usefulness of using the model to discover new potential manufacturing processes through three illustrative cases.

PMID:40041620 | PMC:PMC11878556 | DOI:10.1093/pnasnexus/pgaf054

Hortic Res. 2024 Nov 20;12(2):uhae319. doi: 10.1093/hr/uhae319. eCollection 2025 Feb.

ABSTRACT

Genomic prediction for multiple environments can aid the selection of genotypes suited to specific soil and climate conditions. Methodological advances allow effective integration of phenotypic, genomic (additive, nonadditive), and large-scale environmental (enviromic) data into multi-environmental genomic prediction models. These models can also account for genotype-by-environment interaction, utilize alternative relationship matrices (kernels), or substitute statistical approaches with deep learning. However, the application of multi-environmental genomic prediction in apple remained limited, likely due to the challenge of building multi-environmental datasets and structurally complex models. Here, we applied efficient statistical and deep learning models for multi-environmental genomic prediction of eleven apple traits with contrasting genetic architectures by integrating genomic- and enviromic-based model components. Incorporating genotype-by-environment interaction effects into statistical models improved predictive ability by up to 0.08 for nine traits compared to the benchmark model. This outcome, based on Gaussian and Deep kernels, shows these alternatives can effectively substitute the standard genomic best linear unbiased predictor (G-BLUP). Including nonadditive and enviromic-based effects resulted in a predictive ability very similar to the benchmark model. The deep learning approach achieved the highest predictive ability for three traits with oligogenic genetic architectures, outperforming the benchmark by up to 0.10. Our results demonstrate that the tested statistical models capture genotype-by-environment interactions particularly well, and the deep learning models efficiently integrate data from diverse sources. This study will foster the adoption of multi-environmental genomic prediction to select apple cultivars adapted to diverse environmental conditions, providing an opportunity to address climate change impacts.

PMID:40041603 | PMC:PMC11879405 | DOI:10.1093/hr/uhae319

Comput Struct Biotechnol J. 2025 Feb 14;27:631-638. doi: 10.1016/j.csbj.2025.02.007. eCollection 2025.

ABSTRACT

DNA methylations, such as 5-methylcytosine (5mC), are crucial in biological processes, and aberrant methylations are strongly linked to various human diseases. Genomic 5mC is not randomly distributed but exhibits a strong association with genomic sequences. Thus, various computational methods were developed to predict 5mC status based on DNA sequences. These methods generated promising achievements and overcome the limitations of experimental approaches. However, few studies have comprehensively investigated the dependency of 5mC on genomic sequences, and most existing methods focus on specific genomic regions. In this work, we introduce Deep5mC, a deep learning transformer-based method designed to predict 5mC methylations. Deep5mC leverages long-range dependencies within genomic sequences to estimate the probability of cytosine methylations. Through cross-chromosome evaluation, Deep5mC achieves Matthew's correlation coefficient over 0.86 and F1-score over 0.93, substantially outperforming state-of-the-art methods. Deep5mC not only confirms the influence of long-range sequence context on 5mC prediction but also paves the way for further studying 5mC-sequence dependency across species and in human diseases.

PMID:40041569 | PMC:PMC11879672 | DOI:10.1016/j.csbj.2025.02.007

Front Med (Lausanne). 2025 Feb 18;12:1561485. doi: 10.3389/fmed.2025.1561485. eCollection 2025.

NO ABSTRACT

PMID:40041465 | PMC:PMC11876553 | DOI:10.3389/fmed.2025.1561485

Proc Mach Learn Res. 2024 Jun;248:182-197.

ABSTRACT

Electronic Health Records (EHRs) contain rich patient information and are crucial for clinical research and practice. In recent years, deep learning models have been applied to EHRs, but they often rely on massive features, which may not be readily available for all patients. We propose HTP-Star, which leverages hypergraph structures with a pretrain-then-finetune framework for modeling EHR data, enabling seamless integration of additional features. Additionally, we design two techniques, namely (1) Smoothness-inducing Regularization and (2) Group-balanced Reweighting, to enhance the model's robustness during finetuning. Through experiments conducted on two real EHR datasets, we demonstrate that HTP-Star consistently outperforms various baselines while striking a balance between patients with basic and extra features.

PMID:40041452 | PMC:PMC11876795

Proc IEEE Int Conf Big Data. 2024 Dec;2024:4941-4945. doi: 10.1109/bigdata62323.2024.10825319.

ABSTRACT

Predicting seizures ahead of time will have a significant positive clinical impact for people with epilepsy. Advances in machine learning/artificial intelligence (ML/AI) has provided us the tools needed to perform such predictive tasks. To date, advanced deep learning (DL) architectures such as the convolutional neural network (CNN) and long short-term memory (LSTM) have been used with mixed results. However, highly connected activity exhibited by epileptic seizures necessitates the design of more complex ML techniques which can better capture the complex interconnected neurological processes. Other challenges include the variability of EEG sensor data quality, different epilepsy and seizure profiles, lack of annotated datasets and absence of ML-ready benchmarks. In addition, successful models will need to perform inference in almost real-time using limited hardware compute-capacity. To address these challenges, we propose a lightweight architecture, called ESPFormer, whose novelty lies in the simple and smaller model-size and a lower computational load footprint needed to infer in real-time compared to other works in the literature. To quantify the performance of this lightweight model, we compared its performance with a custom-designed residual neural network (ResNet), a pre-trained vision transformer (ViT) and a pre-trained large-language model (LLM). We tested ESPFormer on MLSPred-Bench which is the largest patient-independent seizure prediction dataset comprising 12 benchmarks. Our results demonstrate that ESPFormer provides the best performance in terms of prediction accuracy for 4/12 benchmarks with an average improvement of 2.65% compared to the LLM, 3.35% compared to the ViT and 17.65% compared to the ResNet - and comparable results for other benchmarks. Our results indicate that lightweight transformer architecture may outperform resource-intensive LLM based models for real-time EEG-based seizure predictions.

PMID:40041397 | PMC:PMC11877310 | DOI:10.1109/bigdata62323.2024.10825319

Front Plant Sci. 2025 Feb 18;16:1527508. doi: 10.3389/fpls.2025.1527508. eCollection 2025.

ABSTRACT

Currently, potato defect sorting primarily relies on manual labor, which is not only inefficient but also prone to bias. Although automated sorting systems offer a potential solution by integrating potato detection models, real-time performance remains challenging due to the need to balance high accuracy and speed under limited resources. This study presents an enhanced version of the YOLO v5s model, named YOLO v5s-ours, specifically designed for real-time detection of potato defects. By integrating Coordinate Attention (CA), Adaptive Spatial Feature Fusion (ASFF), and Atrous Spatial Pyramid Pooling (ASPP) modules, the model significantly improves detection accuracy while maintaining computational efficiency. The model achieved 82.0% precision, 86.6% recall, 84.3% F1-Score and 85.1% mean average precision across six categories - healthy, greening, sprouting, scab, mechanical damage, and rot - marking improvements of 24.6%, 10.5%, 19.4%, and 13.7%, respectively, over the baseline model. Although memory usage increased from 13.7 MB to 23.3 MB and frame rate slightly decreased to 30.7 fps, the accuracy gains ensure the model's suitability for practical applications. The research provides significant support for the development of automated potato sorting systems, advancing agricultural efficiency, particularly in real-time applications, by overcoming the limitations of traditional methods.

PMID:40041023 | PMC:PMC11876418 | DOI:10.3389/fpls.2025.1527508

Front Plant Sci. 2025 Feb 18;16:1539068. doi: 10.3389/fpls.2025.1539068. eCollection 2025.

ABSTRACT

Intelligent and accurate evaluation of KASP primer typing effect is crucial for large-scale screening of excellent markers in molecular marker-assisted breeding. However, the efficiency of both manual discrimination methods and existing algorithms is limited and cannot match the development speed of molecular markers. To address the above problems, we proposed a typing evaluation method for KASP primers by integrating deep learning and traditional machine learning algorithms, called TAL-SRX. First, three algorithms are used to optimize the performance of each model in the Stacking framework respectively, and five-fold cross-validation is used to enhance stability. Then, a hybrid neural network is constructed by combining ANN and LSTM to capture nonlinear relationships and extract complex features, while the Transformer algorithm is introduced to capture global dependencies in high-dimensional feature space. Finally, the two machine learning algorithms are fused through a soft voting integration strategy to output the KASP marker typing effect scores. In this paper, the performance of the model was tested using the KASP test results of 3399 groups of cotton variety resource materials, with an accuracy of 92.83% and an AUC value of 0.9905, indicating that the method has high accuracy, consistency and stability, and the overall performance is better than that of a single model. The performance of the TAL-SRX method is the best when compared with the different integrated combinations of methods. In summary, the TAL-SRX model has good evaluation performance and is very suitable for providing technical support for molecular marker-assisted breeding and other work.

PMID:40041015 | PMC:PMC11876144 | DOI:10.3389/fpls.2025.1539068

J Imaging Inform Med. 2025 Mar 4. doi: 10.1007/s10278-025-01458-x. Online ahead of print.

ABSTRACT

BACKGROUND: The increasing rates of lung cancer emphasize the need for early detection through computed tomography (CT) scans, enhanced by deep learning (DL) to improve diagnosis, treatment, and patient survival. This review examines current and prospective applications of 2D- DL networks in lung cancer CT segmentation, summarizing research, highlighting essential concepts and gaps; Methods: Following the Preferred Reporting Items for Systematic Reviews and Meta-Analysis guidelines, a systematic search of peer-reviewed studies from 01/2020 to 12/2024 on data-driven population segmentation using structured data was conducted across databases like Google Scholar, PubMed, Science Direct, IEEE (Institute of Electrical and Electronics Engineers) and ACM (Association for Computing Machinery) library. 124 studies met the inclusion criteria and were analyzed.

RESULTS: The LIDC-LIDR dataset was the most frequently used; The finding particularly relies on supervised learning with labeled data. The UNet model and its variants were the most frequently used models in medical image segmentation, achieving Dice Similarity Coefficients (DSC) of up to 0.9999. The reviewed studies primarily exhibit significant gaps in addressing class imbalances (67%), underuse of cross-validation (21%), and poor model stability evaluations (3%). Additionally, 88% failed to address the missing data, and generalizability concerns were only discussed in 34% of cases.

CONCLUSIONS: The review emphasizes the importance of Convolutional Neural Networks, particularly UNet, in lung CT analysis and advocates for a combined 2D/3D modeling approach. It also highlights the need for larger, diverse datasets and the exploration of semi-supervised and unsupervised learning to enhance automated lung cancer diagnosis and early detection.

PMID:40038137 | DOI:10.1007/s10278-025-01458-x

J Imaging Inform Med. 2025 Mar 4. doi: 10.1007/s10278-025-01449-y. Online ahead of print.

ABSTRACT

Accurate segmentation of adrenal glands from CT images is essential for enhancing computer-aided diagnosis and surgical planning. However, the small size, irregular shape, and proximity to surrounding tissues make this task highly challenging. This study introduces a novel pipeline that significantly improves the segmentation of left and right adrenal glands by integrating advanced pre-processing techniques and a robust post-processing framework. Utilising a 2D UNet architecture with various backbones (VGG16, ResNet34, InceptionV3), the pipeline leverages test-time augmentation (TTA) and targeted removal of unconnected regions to enhance accuracy and robustness. Our results demonstrate a substantial improvement, with a 38% increase in the Dice similarity coefficient for the left adrenal gland and an 11% increase for the right adrenal gland on the AMOS dataset, achieved by the InceptionV3 model. Additionally, the pipeline significantly reduces false positives, underscoring its potential for clinical applications and its superiority over existing methods. These advancements make our approach a crucial contribution to the field of medical image segmentation.

PMID:40038136 | DOI:10.1007/s10278-025-01449-y

J Am Med Inform Assoc. 2025 Mar 4:ocaf021. doi: 10.1093/jamia/ocaf021. Online ahead of print.

ABSTRACT

OBJECTIVE: This study aimed to develop a novel multi-stage self-supervised learning model tailored for the accurate classification of optical coherence tomography (OCT) images in ophthalmology reducing reliance on costly labeled datasets while maintaining high diagnostic accuracy.

MATERIALS AND METHODS: A private dataset of 2719 OCT images from 493 patients was employed, along with 3 public datasets comprising 84 484 images from 4686 patients, 3231 images from 45 patients, and 572 images. Extensive internal, external, and clinical validation were performed to assess model performance. Grad-CAM was employed for qualitative analysis to interpret the model's decisions by highlighting relevant areas. Subsampling analyses evaluated the model's robustness with varying labeled data availability.

RESULTS: The proposed model outperformed conventional supervised or self-supervised learning-based models, achieving state-of-the-art results across 3 public datasets. In a clinical validation, the model exhibited up to 17.50% higher accuracy and 17.53% higher macro F-1 score than a supervised learning-based model under limited training data.

DISCUSSION: The model's robustness in OCT image classification underscores the potential of the multi-stage self-supervised learning to address challenges associated with limited labeled data. The availability of source codes and pre-trained models promotes the use of this model in a variety of clinical settings, facilitating broader adoption.

CONCLUSION: This model offers a promising solution for advancing OCT image classification, achieving high accuracy while reducing the cost of extensive expert annotation and potentially streamlining clinical workflows, thereby supporting more efficient patient management.

PMID:40037789 | DOI:10.1093/jamia/ocaf021

J Orthop Surg Res. 2025 Mar 4;20(1):236. doi: 10.1186/s13018-025-05620-7.

ABSTRACT

BACKGROUND: Accurate measurement of the spinal alignment parameters is crucial for diagnosing and evaluating adolescent idiopathic scoliosis (AIS). Manual measurement is subjective and time-consuming. The recently developed artificial intelligence models mainly focused on measuring the coronal Cobb angle (CA) and ignored the evaluation of the sagittal plane. We developed a deep-learning model that could automatically measure spinal alignment parameters in biplanar radiographs.

METHODS: In this study, our model adopted ResNet34 as the backbone network, mainly consisting of keypoint detection and CA measurement. A total of 600 biplane radiographs were collected from our hospital and randomly divided into train and test sets in a 3:1 ratio. Two senior spinal surgeons independently manually measured and analyzed spinal alignment and recorded the time taken. The reliabilities of automatic measurement were evaluated by comparing them with the gold standard, using mean absolute difference (MAD), intraclass correlation coefficient (ICC), simple linear regression, and Bland-Altman plots. The diagnosis performance of the model was evaluated through the receiver operating characteristic (ROC) curve and area under the curve (AUC). Severity classification and sagittal abnormalities classification were visualized using a confusion matrix.

RESULTS: Our AI model achieved the MAD of coronal and sagittal angle errors was 2.15° and 2.72°, and ICC was 0.985, 0.927. The simple linear regression showed a strong correction between all parameters and the gold standard (p < 0.001, r2 ≥ 0.686), the Bland-Altman plots showed that the mean difference of the model was within 2° and the automatic measurement time was 9.1 s. Our model demonstrated excellent diagnostic performance, with an accuracy of 97.2%, a sensitivity of 96.8%, a specificity of 97.6%, and an AUC of 0.972 (0.940-1.000).For severity classification, the overall accuracy was 94.5%. All accuracy of sagittal abnormalities classification was greater than 91.8%.

CONCLUSIONS: This deep learning model can accurately and automatically measure spinal alignment parameters with reliable results, significantly reducing diagnostic time, and might provide the potential to assist clinicians.

PMID:40038733 | DOI:10.1186/s13018-025-05620-7