Dvir Aran

Moshe Zisser, Dvir Aran

Deep-learning techniques, particularly the transformer model, have shown great potential in enhancing the prediction performance of longitudinal health records. While previous methods have mainly focused on fixed-time risk prediction, time-to-event prediction (also known as survival analysis) is often more appropriate for clinical scenarios. Here, we present a novel deep-learning architecture we named STRAFE, a generalizable survival analysis transformer-based architecture for electronic health records. The performance of STRAFE was evaluated using a real-world claim dataset of over 130,000 individuals with stage 3 chronic kidney disease (CKD) and was found to outperform other time-to-event prediction algorithms in predicting the exact time of deterioration to stage 5. Additionally, STRAFE was found to outperform binary outcome algorithms in predicting fixed-time risk, possibly due to its ability to train on censored data. We show that STRAFE predictions can improve the positive predictive value of high-risk patients by 3-fold, demonstrating possible usage to improve targeting for intervention programs. Finally, we suggest a novel visualization approach to predictions on a per-patient basis. In conclusion, STRAFE is a cutting-edge time-to-event prediction algorithm that has the potential to enhance risk predictions in large claims datasets.

Joseph Bingham, Netanel Arussy, Dvir Aran

Unsupervised representations are widely assumed to be neutral with respect to sensitive attributes when those attributes are withheld from training. We show that this assumption is false. Using SOMtime, a topology-preserving representation method based on high-capacity Self-Organizing Maps, we demonstrate that sensitive attributes such as age and income emerge as dominant latent axes in purely unsupervised embeddings, even when explicitly excluded from the input. On two large-scale real-world datasets (the World Values Survey across five countries and the Census-Income dataset), SOMtime recovers monotonic orderings aligned with withheld sensitive attributes, achieving Spearman correlations of up to 0.85, whereas PCA and UMAP typically remain below 0.23 (with a single exception reaching 0.31), and against t-SNE and autoencoders which achieve at most 0.34. Furthermore, unsupervised segmentation of SOMtime embeddings produces demographically skewed clusters, demonstrating downstream fairness risks without any supervised task. These findings establish that \textit{fairness through unawareness} fails at the representation level for ordinal sensitive attributes and that fairness auditing must extend to unsupervised components of machine learning pipelines. We have made the code available at~ https://github.com/JosephBingham/SOMtime

Ortal Hirszowicz, Dvir Aran

We introduce RatchetEHR, a novel transformer-based framework designed for the predictive analysis of electronic health records (EHR) data in intensive care unit (ICU) settings, with a specific focus on bloodstream infection (BSI) prediction. Leveraging the MIMIC-IV dataset, RatchetEHR demonstrates superior predictive performance compared to other methods, including RNN, LSTM, and XGBoost, particularly due to its advanced handling of sequential and temporal EHR data. A key innovation in RatchetEHR is the integration of the Graph Convolutional Transformer (GCT) component, which significantly enhances the ability to identify hidden structural relationships within EHR data, resulting in more accurate clinical predictions. Through SHAP value analysis, we provide insights into influential features for BSI prediction. RatchetEHR integrates multiple advancements in deep learning which together provide accurate predictions even with a relatively small sample size and highly imbalanced dataset. This study contributes to medical informatics by showcasing the application of advanced AI techniques in healthcare and sets a foundation for further research to optimize these capabilities in EHR data analysis.

Moran Sorka, Alon Gorenshtein, Dvir Aran et al.

Large language models (LLMs) have shown promise in medical domains, but their ability to handle specialized neurological reasoning requires systematic evaluation. We developed a comprehensive benchmark using 305 questions from Israeli Board Certification Exams in Neurology, classified along three complexity dimensions: factual knowledge depth, clinical concept integration, and reasoning complexity. We evaluated ten LLMs using base models, retrieval-augmented generation (RAG), and a novel multi-agent system. Results showed significant performance variation. OpenAI-o1 achieved the highest base performance (90.9% accuracy), while specialized medical models performed poorly (52.9% for Meditron-70B). RAG provided modest benefits but limited effectiveness on complex reasoning questions. In contrast, our multi-agent framework, decomposing neurological reasoning into specialized cognitive functions including question analysis, knowledge retrieval, answer synthesis, and validation, achieved dramatic improvements, especially for mid-range models. The LLaMA 3.3-70B-based agentic system reached 89.2% accuracy versus 69.5% for its base model, with substantial gains on level 3 complexity questions. The multi-agent approach transformed inconsistent subspecialty performance into uniform excellence, addressing neurological reasoning challenges that persisted with RAG enhancement. We validated our approach using an independent dataset of 155 neurological cases from MedQA. Results confirm that structured multi-agent approaches designed to emulate specialized cognitive processes significantly enhance complex medical reasoning, offering promising directions for AI assistance in challenging clinical contexts.