Hard-to-heal wounds are a growing human and financial concern, constituting approximately 1–3% of the healthcare budget. Wound care is not a medical specialty and is often not prioritized within healthcare. A large portion of the cost and suffering caused by wounds has the potential to be mediated through improved knowledge and effectivised workflows. One potential way to achieve this is through the implementation of AI-tools to support clinicians in planning and executing wound care. Information-driven care is a framework for implementing AI-technology in healthcare. Wound Monitor is a Dutch database containing data collected from home-care visits conducted by wound specialists during 2005 to 2022, mostly in Limburg. It contains data of more than 17000 patients. Region Halland, Sweden, created a platform of integrated clinical, financial and operational data called “The Regional Healthcare Information Platform” (RHIP). The platform contains data on over 500 000 patients during 2008–2021. Within this data, a subset of almost 39000 patients have been diagnosed with wounds or wound related conditions. This subset of patients are defined as the RHIP Wound Cohort. This article characterizes the two wound cohorts in terms of demographics and wound types. Further, it examines the quality, quantity and granularity of the respective databases. The discussion section evaluates the strengths and weaknesses of the datasets in terms of the perspective they provide on the patient and wound journey. Lastly, the discussion section also explores how the cohorts may be utilized for predictive modeling and other machine learning-based applications in order to enable information-driven wound care. © 2026 Gustafsson et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

The Cox Proportional Hazards (CPH) model has long been the preferred survival model for its explainability. However, to increase its predictive power beyond its linear log-risk, it was extended to utilize deep neural networks, sacrificing its explainability. In this work, we explore the potential of self-explaining neural networks (SENN) for survival analysis. We propose a new locally explainable Cox proportional hazards model, named CoxSE, by estimating a locally-linear log-hazard function using the SENN. We also propose a modification to the Neural additive (NAM) model, hybrid with SENN, named CoxSENAM, which enables the control of the stability and consistency of the generated explanations. Several experiments using synthetic and real datasets are presented, benchmarking CoxSE and CoxSENAM against a NAM-based model, a DeepSurv model explained with SHAP, and a linear CPH model. The results show that, unlike the NAM-based model, the SENN-based model can provide more stable and consistent explanations while maintaining the predictive power of the black-box model. The results also show that, due to their structural design, NAM-based models demonstrate better robustness to non-informative features. Among the models, the hybrid model exhibits the best robustness. Full implementation is available on GitHub. © 2025 The Authors

Objectives: The objective was to develop an artificial intelligence (AI) model for predicting acute coronary occlusion myocardial infarction (OMI) in patients with chest pain at the emergency department (ED), using information that is widely available early in the ED assessment. Methods: In a cohort of 24,511 consecutive adult ED patients with chest pain from 5 Swedish hospitals, OMI cases were identified through register data and manual review of health records and angiographies. Ambulance patients bypassing the ED due to ST-elevation myocardial infarction (STEMI) were not included in the cohort. A deep-learning AI model was created to predict OMI using the electrocardiogram, optionally combined with other early ED data, including medical history and initial lab values. The model was internally validated on held-out data and compared with the STEMI criteria. Results: A total of 467 patients (1.9%) were identified as OMI, corresponding to 29% of all acute myocardial infarction cases. The 30-day mortality rate was 6.6% for OMI, compared with 3.3% for non-OMI. Only 5.4% of the OMI cases received angiography within the guideline-recommended maximum of 90 minutes after ED arrival. The AI model achieved an area under the receiver operating characteristic (AUC) of 95.3% (95% CI, 93.8%-97.3%), with a sensitivity of 62% compared with 27% for the STEMI criteria (difference 34.5%; 95% CI, 22.9%-45.2%) at the same specificity (97.4%). Conclusion: Our AI model identified OMI in ED patients with chest pain with an AUC of 95%, doubling sensitivity compared with the STEMI criteria at the same specificity. Using the model could reduce time to intervention, as only about 1 in 20 OMI cases currently receive timely angiography. © 2025 The Author(s)

We propose a new sequential decision-making setting, combining key aspects of two established online learning problems with bandit feedback. The optimal action to play at any given moment is contingent on an underlying changing state that is not directly observable by the agent. Each state is associated with a context distribution, possibly corrupted, allowing the agent to identify the state. Furthermore, states evolve in a Markovian fashion, providing useful information to estimate the current state via state history. In the proposed problem setting, we tackle the challenge of deciding on which of the two sources of information the agent should base its action selection. We present an algorithm that uses a referee to dynamically combine the policies of a contextual bandit and a multi-armed bandit. We capture the time-correlation of states through iteratively learning the action-reward transition model, allowing for efficient exploration of actions. Our setting is motivated by adaptive mobile health (mHealth) interventions. Users transition through different, time-correlated, but only partially observable internal states, determining their current needs. The side information associated with each internal state might not always be reliable, and standard approaches solely rely on the context risk of incurring high regret. Similarly, some users might exhibit weaker correlations between subsequent states, leading to approaches that solely rely on state transitions risking the same. We analyze our setting and algorithm in terms of regret lower bound and upper bounds and evaluate our method on simulated medication adherence intervention data and several real-world data sets, showing improved empirical performance compared to several popular algorithms. © The Author(s) 2024.

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: 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.

At the emergency department, it is important to quickly and accurately identify patients at risk of acute myocardial infarction (AMI). One of the main tools for detecting AMI is the electrocardiogram (ECG), which can be difficult to interpret manually. There is a long history of applying machine learning algorithms to ECGs, but such algorithms are quite data hungry, and correctly labeled high-quality ECGs are difficult to obtain. Transfer learning has been a successful strategy for mitigating data requirements in other applications, but the benefits for predicting AMI are understudied. Here we show that a straightforward application of transfer learning leads to large improvements also in this domain. We pre-train models to classify sex and age using a collection of 840 k ECGs from non-chest-pain patients, and fine-tune the resulting models to predict AMI using 44 k ECGs from chest-pain patients. The results are compared with models trained without transfer learning. We find a considerable improvement from transfer learning, consistent across multiple state-of-the-art ResNet architectures and data sizes, with the best performing model improving from 0.79 AUC to 0.85 AUC. This suggests that even a simple form of transfer learning from a moderately sized dataset of non-chest-pain ECGs can lead to major improvements in predicting AMI. © 2025 Nyström et al.

Objective: Half of all adult emergency department (ED) visits with a complaint of dyspnea involve acute heart failure (AHF), exacerbation of chronic obstructive pulmonary disease (eCOPD), or pneumonia, which are often misdiagnosed. We aimed to create an artificial intelligence (AI) diagnostic decision support tool to detect patients with AHF, eCOPD, and pneumonia among dyspneic adults at the beginning of their ED visit. Methods: In this cross-sectional study, we included all ED visits of patients 18 years or older with dyspnea at two regional Swedish EDs 07/01/2017–12/31/2019. In-hospital or ED discharge notes were used as outcome labels, with a subset manually reviewed by experts. We analyzed data from a complete regional healthcare system, along with socioeconomic factors, using Hierarchical Attention Networks. Each patient displayed a unique set of variables important for diagnosing dyspnea. All patients’ unique variable sets were aggregated into a variable list. The top 100, 50, and 20 variables were tested in a simpler CatBoost model. Finally, performance was compared after adding medical expertise to the AI model. Results: We included 10,869 visits, with 15.1% having AHF, 13.6% eCOPD, and 13.1% pneumonia. The median number of variables per unique ED visit was 187 (IQR 111–307). Aggregating the unique sets of variables resulted in a cohort list of 2,064 variables. The median micro AUROC was 87.8% (2.5–97.5 percentile; 86.4–89.3%). Age, ECGs, previous diagnoses, and medication were considered important by the AI model, while sex, vital signs, and socioeconomic factors were deemed almost non-predictive. Using the top 20 AI-selected variables, the AUROC was 86.6% (85.1–88.1%). Adding human medical expertise did not significantly change the AUROC. Conclusion: Based on the analysis of a high-dimensional dataset, we designed a lightweight 20-variable machine learning model that can early and effectively diagnose AHF, eCOPD, and pneumonia among ED patients with dyspnea. © 2025 The Authors. Published by Elsevier B.V.

Background: Dyspnoea is one of the emergency department’s (ED) most common and deadly chief complaints, but frequently misdiagnosed and mistreated. We aimed to design a diagnostic decision support which classifies dyspnoeic ED visits into acute heart failure (AHF), exacerbation of chronic obstructive pulmonary disease (eCOPD), pneumonia and “other diagnoses” by using deep learning and complete, unselected data from an entire regional health care system.

Methods: In this cross-sectional study, we included all dyspnoeic ED visits of patients ≥ 18 years of age at the two EDs in the region of Halland, Sweden, 07/01/2017–12/31/2019. Data from the complete regional health care system within five years prior to the ED visit were analysed. Gold standard diagnoses were defined as the subsequent in-hospital or ED discharge notes, and a subsample was manually reviewed by emergency medicine experts. A novel deep learning model, the clinical attention-based recurrent encoder network (CareNet), was developed. Cohort performance was compared to a simpler CatBoost model. A list of all variables and their importance for diagnosis was created. For each unique patient visit, the model selected the most important variables, analysed them and presented them to the clinician interpretably by taking event time and clinical context into account. AUROC, sensitivity and specificity were compared.

Findings: The most prevalent diagnoses among the 10,315 dyspnoeic ED visits were AHF (15.5%), eCOPD (14.0%) and pneumonia (13.3%). Median number of unique events, i.e., registered clinical data with time stamps, per ED visit was 1,095 (IQR 459–2,310). CareNet median AUROC was 87.0%, substantially higher than the CatBoost model´s (81.4%). CareNet median sensitivity for AHF, eCOPD, and pneumonia was 74.5%, 92.6%, and 54.1%, respectively, with a specificity set above 75.0, slightly inferior to that of the CatBoost baseline model. The model assembled a list of 1,596 variables by importance for diagnosis, on top were prior diagnoses of heart failure or COPD, daily smoking, atrial fibrillation/flutter, life management difficulties and maternity care. Each patient visit received their own unique attention plot, graphically displaying important clinical events for the diagnosis.

Interpretation: We designed a novel interpretable deep learning model for diagnosis in emergency department dyspnoea patients by analysing unselected data from a complete regional health care system. © 2024 Heyman et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.