Large Language Models (LLMs) are increasingly transforming medical applications. However, proprietary models such as GPT-4o face significant barriers to clinical adoption because they cannot be deployed on site within healthcare institutions, making them non-compliant with stringent privacy regulations. Recent advancements in open-source LLMs such as DeepSeek models offer a promising alternative since they allow efficient fine-tuning on local data in hospitals with advanced IT infrastructure. To demonstrate the clinical utility of DeepSeek-V3 and DeepSeek-R1, we benchmarked their performance on clinical decision support tasks against proprietary LLMs, including GPT-4o and Gemini-2.0 Flash Thinking Experimental. Using 125 patient cases with sufficient statistical power, covering a broad range of frequent and rare diseases, we found that DeepSeek models perform equally well and in some cases better than proprietary LLMs. Our study demonstrates that open-source LLMs can provide a scalable pathway for secure model training enabling real-world medical applications in accordance with data privacy and healthcare regulations.

Transposable elements (TEs) constitute a significant portion of eukaryotic genomes and play crucial roles in genome evolution, yet their diverse and complex sequences pose challenges for accurate classification. Existing tools often lack reliability in TE classification, limiting genomic analyses. Here, we present TEclass2, a software employing a deep learning approach based on a linear transformer architecture with k-mer tokenization and sequence-specific adaptations to classify TE consensus sequences into sixteen superfamilies. TEclass2 demonstrates improved classification performance and offers flexible model training on custom datasets. Accessible via a web interface with pre-trained models, TEclass2 facilitates rapid and reliable TE classification. These advancements provide a foundation for enhanced genomic annotation and support further bioinformatics research involving transposable elements.

Deep Learning architectures for time series require a large number of training samples, however traditional sample size estimation for sufficient model performance is not applicable for machine learning, especially in the field of electrocardiograms (ECGs). This paper outlines a sample size estimation strategy for binary classification problems on ECGs using different deep learning architectures and the large publicly available PTB-XL dataset, which includes 21801 ECG samples. This work evaluates binary classification tasks for Myocardial Infarction (MI), Conduction Disturbance (CD), ST/T Change (STTC), and Sex. All estimations are benchmarked across different architectures, including XResNet, Inception-, XceptionTime and a fully convolutional network (FCN). The results indicate trends for required sample sizes for given tasks and architectures, which can be used as orientation for future ECG studies or feasibility aspects.

Aims: Automated QTc measurements from commercial ECG systems often diverge from expert readings. We developed QTcNet, a deep learning model trained and validated on multiple large ECG datasets to improve automated QTc measurement accuracy. Methods and results: QTcNet employs a regression-based convolutional neural network architecture. It was trained on 120 300 algorithm-labelled ECGs (60 150 from an internal hospital cohort and 60 150 from the MIMIC-IV dataset) after correction for a vendor-specific +15 ms bias. Performance was evaluated against expert QTc measurements in three independent datasets: PTB Diagnostic ECG Database (n = 100 ECGs in validation set), QTcMS (n = 210), and ECGRDVQ (n = 5219). The effect of fine-tuning on cardiologist-annotated ECGs was tested in the PTB database (n = 449 in fine-tuning set). Model explainability analyses were performed with Integrated Gradient maps. QTcNet reduced cross-cohort mean absolute error (MAE) from 23.4 to 13.4 ms and root mean square error (RMSE) from 40.1 to 22.1 ms, almost halving large (>50 ms) outliers. Fine-tuning only reduced errors in the PTB dataset but did not improve cross-cohort performance. Integrated Gradient maps confirmed that the model concentrated on QRS onset and T wave offset, supporting physiological plausibility. Conclusion: QTcNet, trained on large-scale algorithmically labelled data, consistently outperformed conventional algorithms across three independent, external validation datasets. Fine-tuning of QTcNet may adapt the model to the characteristics of specific cohorts but reduces external validity in other cohorts. We openly release the full model and code, along with a ready-to-use online implementation at https://qtcnet.uni-muenster.de, facilitating further research and community-driven improvement.

Background: Electrocardiogram (ECG) data constitutes one of the most widely available biosignal data in clinical and research settings, providing critical insights into cardiovascular diseases as well as broader health conditions. Advancements in deep learning demonstrate high performance in diverse ECG classification tasks, ranging from arrhythmia detection to risk prediction for various diseases. However, the widespread adoption of deep learning for ECG analysis faces significant barriers, including the heterogeneity of file formats, restricted access to pretrained model weights, and complex technical workflows for out-of-domain users. Objective: This study aims to address major bottlenecks in ECG-based deep learning by introducing ExChanGeAI, an open-source, web-based platform designed to offer an integrated, user-friendly platform for ECG data analysis. Our objective is to streamline the entire workflow—from initial data ingestion (regardless of device or format) and intuitive visualization to privacy-preserving model training and task-specific fine-tuning—making advanced ECG deep learning accessible for both clinical researchers and practitioners without machine learning (ML) expertise. Methods: ExChanGeAI incorporates robust preprocessing modules for various ECG file types, a set of interactive visualization tools for exploratory data analysis, and multiple state-of-the-art deep learning architectures for ECGs. Users can choose to train models from scratch or fine-tune pretrained models using their own datasets, while all computations are performed locally to ensure data privacy. The platform is adaptable for deployment on personal computers as well as scalable to high-performance computing infrastructures. We demonstrate the platform’s performance on several clinically relevant classification tasks across 3 external and heterogeneous validation datasets, including a newly curated test set from routine care, evaluating both model generalizability and resource efficiency. Results: Our experiments show that de novo training with user-provided, task-specific data can outperform a leading foundation model, while requiring substantially fewer parameters and computational resources. The platform enables users to empirically determine the most suitable model for their specific tasks, based on systematic validations, while lowering technical barriers for out-of-domain experts and promoting open research. Conclusions: ExChanGeAI provides a comprehensive, privacy-aware platform that democratizes access to ECG analysis and model training. By simplifying complex workflows, ExChanGeAI empowers out-of-domain researchers to use state-of-the-art ML on diverse datasets, democratizing the access to ML in the field of ECG data. The platform is available as open-source code under the Massachusetts Institute of Technology (MIT) license.

Aims: Electrocardiogram (ECG) recordings are fundamental for diagnosing cardiac conditions. Recent advances in automatic ECG analysis have been dominated by deep learning, particularly convolutional neural networks (CNNs). CNNs excel in processing high-dimensional signal data, where their ability to automatically extract complex features has enabled significant progress. However, while CNNs are powerful for biomedical signal analysis, their application to ECG data also carries disadvantages, such as the requirement for large annotated datasets and limited explainability. To address these challenges, we aim to reintroduce shallow learning methods, such as linear classifiers, by leveraging the cyclic nature of ECG signals. Methods and results: We developed an adaptive transformation that restructures ECG signals into a fully structured format suitable for shallow learning algorithms. This method aligns R-peaks across all signals in a dataset and resamples the inter-QRS segments to match a predefined reference heart rate. The approach was systematically evaluated across tasks including classification, clustering, and explainability. Our transformation substantially improved the performance of shallow learning techniques. Compared with CNN approaches, shallow models trained on transformed ECGs achieved superior accuracy and interpretability in data-limited scenarios. Conclusion: We demonstrate that shallow machine learning methods, when combined with our alignment-based transformation, can reach CNN level performance in ECG analysis, especially under conditions of limited training data. This approach offers clear advantages in classification, clustering, and explainability and provides an accessible alternative to deep learning. To facilitate adoption and further research, we release a publicly available framework for ECG signal alignment at https://github.com/imi-ms/rlign.