Transformer models pre-trained on self-supervised tasks and fine-tuned on downstream objectives have achieved remarkable results across a variety of domains. However, fine-tuning these models for clinical predictions from longitudinal medical data, such as electronic health records (EHR), remains challenging due to limited labeled data and the complex, event-driven nature of medical sequences. While self-attention mechanisms are powerful for capturing relationships within sequences, they may underperform when modeling subtle dependencies between sparse clinical events under limited supervision. We introduce a simple yet effective fine-tuning technique, Adaptive Noise-Augmented Attention (ANAA), which injects adaptive noise directly into the self-attention weights and applies a 2D Gaussian kernel to smooth the resulting attention maps. This mechanism broadens the attention distribution across tokens while refining it to emphasize more informative events. Unlike prior approaches that require expensive modifications to the architecture and pre-training phase, ANAA operates entirely during fine-tuning. Empirical results across multiple clinical prediction tasks demonstrate consistent performance improvements. Furthermore, we analyze how ANAA shapes the learned attention behavior, offering interpretable insights into the model's handling of temporal dependencies in EHR data. © 2025 Amirahmadi, Etminani and Ohlsson.

Integrated Gradients (IG) is a popular method for explaining clinical deep models—including widely used multimodal, pretrained Transformers—but its utility on EHR code sequences is hampered by (i) the lack of principled baselines for sequence of discrete tokens and (ii) dense, hard-to-interpret generated attributions. To address both, first, we introduce a manifold-aware baseline: the expected value under the empirical dist—implemented as the position-wise empirical mean of pre-Transformer token embeddings on held-out validation data, which keeps IG interpolants near the data manifold. Second, we introduce {GS-IG}, which preserves the straight path geometry but re-parameterizes the schedule (\alpha(t)=t^{\theta}) and selects (\theta) per input by minimizing a token-level (\ell_{2,1}) (group-sparsity) objective, producing concise, practitioner-friendly explanations. On MIMIC-IV (incident heart failure) and MDC (early mortality), the manifold-aware baseline improves faithfulness (higher Comprehensiveness, lower Sufficiency), and GS-IG reduces token-level (\ell_{2,1}) by 9–18% with negligible change in those metrics on the manifold-aware baseline. The method is lightweight and yields faithful, sparse, and actionable. © 2025 A. Amirahmadi, F. Etminani & M. Ohlsson.

Electronic Health Records (EHRs) provide valuable source of information detailing patient encounters over time and representing patient trajectories. By leveraging machine learning (ML) techniques, researchers can develop models to support healthcare professionals in decision-making for better patient care. However, EHRs exhibit a complex temporal structure, with diverse entities such as diagnostic codes and medications posing challenges when representing patient trajectory over time. Several methods are used in literature to represent EHRs as time-series data. Some represent patient trajectory as an irregular time series with the sequence of patient encounters. Others use fixed time intervals and aggregate patient encounters within a particular time window. This paper examines the optimization of temporal granularity by exploring two aggregation methods: the established time-based aggregation and a newly introduced similarity-based approach, which aggregates consecutive encounters with a high overlap of clinical codes. Additionally, this study utilizes graph and recurrent neural network models to represent patient trajectories and assess how these aggregation techniques affect model performance. Furthermore, we propose a hybrid ML model combining graph neural networks with a recurrent model. We evaluated model performance using two clinical prediction tasks, the first is to predict the top-k diagnoses codes for the last patient encounter, while the second is to predict top k codes at every future encounter. Results show that time-based aggregation enhances performance of recurrent models, while similarity-based aggregation allows hybrid and graph neural models to reach higher performance than recurrent model. © 2025 The Author(s). Published by Elsevier Inc. on behalf of American College of Emergency Physicians.

Background: Understanding non-clinical factors that influence how women choose obstetricians and gynecologists (OB/GYNs) is essential for delivering patient-centered care. This study aimed to identify and rank the non-clinical preferences when selecting OB/GYN specialists in the private healthcare sector in Mashhad, Iran. Methods: This cross-sectional study, conducted from January to February 2018, 462 patients completed a validated 45-item questionnaire (CVI = 0.80, Cronbach’s alpha = 0.88) assessing their non-clinical preferences. Preferences were rated on a 5-point scale and ranked using Friedman’s test. Associations between demographic factors and preferences were analyzed using the Kruskal-Wallis test and ordinal logistic regression. Results: The highest-rated criteria included physicians’ attentiveness and respect for patients, respectful staff behavior, short waiting times, and ensuring privacy during examinations. The latest important criteria were physician age, university affiliation, and office proximity to patient’s home. Education level, pregnancy experience, and number of prior OB/GYN visits were significantly associated with certain preferences. Multivariate regression revealed that higher education and more prior OB/GYN visits independently predicted greater importance placed on short waiting time and respectful staff behavior. Conclusion: Beyond clinical competence, non-clinical factors-particularly those related to interpersonal behavior, communication, and privacy-are central to patient-centered care in OB/GYN settings. Recognizing and integrating these preferences into service delivery can strengthen trust, enhance satisfaction, and support ethical, patient-centered care in the private healthcare sector. © The Author(s) 2025.

Aims Atrial fibrillation (AF), often asymptomatic and underdiagnosed, is an independent risk factor for ischaemic stroke. A knowledge gap remains regarding the optimal target population and method to use for AF screening. We aimed to test whether screening for AF using a machine learning–based risk prediction model (RPM) and 14-day continuous patch electrocardiogram (ECG) (Philips ePatch) in high-risk individuals ≥ 65 years is more effective than standard care. Methods and results Individuals ≥ 65 years were assigned to general or RPM cohort. The general cohort was randomized to control or invitation. In the RPM cohort, high-risk individuals, identified by RPM, were randomized to control or invitation. The primary outcome was 6-month AF incidence, analysed as intention-to-invite, comparing RPM + invitation with general + control. Of the 2960 randomized individuals, participation was 43% (632/1480) in invitation arms. Atrial fibrillation incidence was higher in RPM + invitation than in general + control arm (3.8%, 28/740 vs. 0.7%, 5/740; P < 0.001), yielding a risk ratio of 5.6, [95% confidence interval (2.2, 14.4)], and a number needed to invite of 32. Atrial fibrillation was more often detected in RPM + invitation than in general + invitation arm (1.1%, 8/740; P < 0.001), but not more often than in RPM + control arm (2.2%, 16/740; P = 0.07). No difference was found between general + invitation and general + control arms (1.1%, 8/740 vs. 0.7%, 5/740; P = 0.40). Conclusion Among high-risk individuals ≥ 65 years, the combination of a machine learning–based RPM and long-term ECG recording was superior to standard care in identifying new AF cases. © The Author(s) 2025.

Background: The growing availability of electronic health records (EHRs) presents an opportunity to enhance patient care by uncovering hidden health risks and improving informed decisions through advanced deep learning methods. However, modeling EHR sequential data, that is, patient trajectories, is challenging due to the evolving relationships between diagnoses and treatments over time. Significant progress has been achieved using transformers and self-supervised learning. While BERT-inspired models using masked language modeling (MLM) capture EHR context, they often struggle with the complex temporal dynamics of disease progression and interventions.

Objective: This study aims to improve the modeling of EHR sequences by addressing the limitations of traditional transformer-based approaches in capturing complex temporal dependencies.

Methods: We introduce Trajectory Order Objective BERT (Bidirectional Encoder Representations from Transformers; TOO-BERT), a transformer-based model that advances the MLM pretraining approach by integrating a novel TOO to better learn the complex sequential dependencies between medical events. TOO-Bert enhanced the learned context by MLM by pretraining the model to distinguish ordered sequences of medical codes from permuted ones in a patient trajectory. The TOO is enhanced by a conditional selection process that focus on medical codes or visits that frequently occur together, to further improve contextual understanding and strengthen temporal awareness. We evaluate TOO-BERT on 2 extensive EHR datasets, MIMIC-IV hospitalization records and the Malmo Diet and Cancer Cohort (MDC)-comprising approximately 10 and 8 million medical codes, respectively. TOO-BERT is compared against conventional machine learning methods, a transformer trained from scratch, and a transformer pretrained on MLM in predicting heart failure (HF), Alzheimer disease (AD), and prolonged length of stay (PLS).

Results: TOO-BERT outperformed conventional machine learning methods and transformer-based approaches in HF, AD, and PLS prediction across both datasets. In the MDC dataset, TOO-BERT improved HF and AD prediction, increasing area under the receiver operating characteristic curve (AUC) scores from 67.7 and 69.5 with the MLM-pretrained Transformer to 73.9 and 71.9, respectively. In the MIMIC-IV dataset, TOO-BERT enhanced HF and PLS prediction, raising AUC scores from 86.2 and 60.2 with the MLM-pretrained Transformer to 89.8 and 60.4, respectively. Notably, TOO-BERT demonstrated strong performance in HF prediction even with limited fine-tuning data, achieving AUC scores of 0.877 and 0.823, compared to 0.839 and 0.799 for the MLM-pretrained Transformer, when fine-tuned on only 50% (442/884) and 20% (176/884) of the training data, respectively.

Conclusions: These findings demonstrate the effectiveness of integrating temporal ordering objectives into MLM-pretrained models, enabling deeper insights into the complex temporal relationships inherent in EHR data. Attention analysis further highlights TOO-BERT's capability to capture and represent sophisticated structural patterns within patient trajectories, offering a more nuanced understanding of disease progression.

 ©Ali Amirahmadi, Farzaneh Etminani, Jonas Björk, Olle Melander, Mattias Ohlsson.

Synthetic data generation has been proposed as a potential solution to accessing Electronic Health Records (EHRs) while minimizing the privacy risks associated with real EHRs. Nevertheless, the practical use of synthetic EHRs rests on their ability to resemble the quality of real EHRs. Existing evaluations of synthetic EHRs often focus on assessing them as static snapshots frozen in time, neglecting temporal dependencies and varying temporal patterns. Moreover, some of these methods rely on subjective judgments, are limited to segmentable time-series, and employ methods that adopt a one-to-one approach. This study employs a comprehensive approach to evaluating fidelity in synthetic time-series EHRs to address these challenges. We extend the functionality of time-series analysis methods such as temporal clustering, time-series similarity measures, Sample Entropy, and trend analysis, to evaluate varying temporal patterns in synthetic time-series EHRs. Our findings provide valuable insights into how synthetic EHRs align with real EHRs in the temporal context, considering aspects such as patient groupings, temporal dynamics, predictability, and directional change. We empirically demonstrate the feasibility of assessing temporal fidelity with these methods, offering an understanding of the quality of synthetic EHRs in capturing the varying temporal patterns inherent in EHRs. © 2024 IEEE.

Recent studies have shown how synthetic data generation methods can be applied to electronic health records (EHRs) to obtain synthetic versions that do not violate privacy rules. This growing body of research has resulted in the emergence of numerous methods for evaluating the quality of generated data, with new publications often introducing novel evaluation methods. This work presents a detailed review of synthetic EHRs, focusing on the various evaluation methods used to assess the quality of the generated EHRs. We discuss the existing evaluation methods, offering insights into their use as well as providing an interpretation of the evaluation metrics from the perspectives of achieving fidelity, utility and privacy. Furthermore, we highlight the key factors influencing the selection of evaluation methods, such as the type of data (e.g., categorical, continuous, or discrete) and the mode of application (e.g., patient level, cohort level, and feature level). To assess the effectiveness of current evaluation measures, we conduct a series of experiments to shed light on the potential limitations of these measures. The findings from these experiments reveal notable shortcomings, including the need for meticulous application of methods to the data to reduce inconsistent evaluations, the qualitative nature of some assessments subject to individual judgment, the need for clinical validations, and the absence of techniques to evaluate temporal dependencies within the data. This highlights the need to place greater emphasis on evaluation measures, their application, and the development of comprehensive evaluation frameworks as it is crucial for advancing progress in this field. © 2024 The Author(s)

Objective: This study aims to comprehensively review the use of graph neural networks (GNNs) for clinical risk prediction based on electronic health records (EHRs). The primary goal is to provide an overview of the state-of-the-art of this subject, highlighting ongoing research efforts and identifying existing challenges in developing effective GNNs for improved prediction of clinical risks. Methods: A search was conducted in the Scopus, PubMed, ACM Digital Library, and Embase databases to identify relevant English-language papers that used GNNs for clinical risk prediction based on EHR data. The study includes original research papers published between January 2009 and May 2023. Results: Following the initial screening process, 50 articles were included in the data collection. A significant increase in publications from 2020 was observed, with most selected papers focusing on diagnosis prediction (n = 36). The study revealed that the graph attention network (GAT) (n = 19) was the most prevalent architecture, and MIMIC-III (n = 23) was the most common data resource. Conclusion: GNNs are relevant tools for predicting clinical risk by accounting for the relational aspects among medical events and entities and managing large volumes of EHR data. Future studies in this area may address challenges such as EHR data heterogeneity, multimodality, and model interpretability, aiming to develop more holistic GNN models that can produce more accurate predictions, be effectively implemented in clinical settings, and ultimately improve patient care. © 2024 The Authors

In recent years, knowledge graphs (KGs) have been widely applied in various domains for different purposes. The semantic model of KGs can represent knowledge through a hierarchical structure based on classes of entities, their properties, and their relationships. The construction of large KGs can enable the integration of heterogeneous information sources and help Artificial Intelligence (AI) systems be more explainable and interpretable. This systematic review examines a selection of recent publications to understand how KGs are currently being used in eXplainable AI systems. To achieve this goal, we design a framework and divide the use of KGs into four categories: extracting features, extracting relationships, constructing KGs, and KG reasoning. We also identify where KGs are mostly used in eXplainable AI systems (pre-model, in-model, and post-model) according to the aforementioned categories. Based on our analysis, KGs have been mainly used in pre-model XAI for feature and relation extraction. They were also utilised for inference and reasoning in post-model XAI. We found several studies that leveraged KGs to explain the XAI models in the healthcare domain. © The Author(s) 2022.