Reliable multimodal prognosis after acute ischemic stroke requires methods that integrate heterogeneous information under rigorous leakage control and are evaluated against strong simpler alternatives. We developed QIMCA-Net, a multimodal residual self-attention framework integrating six-channel three-dimensional computed tomography, 17 baseline clinical variables with 17 missingness indicators, and 77 quantitative whole-brain imaging biomarkers. Poor 90-day functional outcome was defined as a modified Rankin Scale score of 3–6. The locked ISLES’24 cohort comprised 119 patients, including 41 with poor outcomes. Patient-level five-fold outer cross-validation with an inner training–validation hold-out split confined preprocessing and model fitting to inner-training data, while checkpoint and threshold selection and, where applicable, temperature scaling used only inner-validation data. The frozen primary QIMCA-Net achieved an out-of-fold AUROC of 0.7301 (95% CI, 0.6304–0.8280), an AUPRC of 0.5869, and a Brier score of 0.2038. Strong structured-data benchmarks remained competitive, and no model consistently dominated discrimination, positive-class retrieval, and probabilistic accuracy. Across five alternative patient-level five-fold assignments, QIMCA-Net showed a mean AUROC difference of +0.0421 relative to parameter-matched direct concatenation; however, the Nadeau–Bengio corrected interval included zero (95% CI, −0.0373 to 0.1215; p = .215). CT permutation produced negligible performance change, center-related signal was substantial, and uncertainty measures only weakly discriminated prediction errors. Lesion-auxiliary learning, Bayesian optimization, and uncertainty-aware extensions did not replace the frozen primary model. These findings support the feasibility of leakage-controlled multimodal self-attention while showing uncertain incremental architectural benefit and the need for independent external validation before clinical use.
Financial fraud poses an evolving threat to modern economies, requiring advanced detection solutions beyond traditional rule-based systems. This systematic literature review follows established guidelines, analyzing 103 selected studies published between 2021 and 2025 to capture recent advancements in artificial intelligence for banking transaction fraud detection. The review synthesizes 20 representative methods across four primary computational paradigms: enhanced machine learning, deep learning, ensemble techniques, and hybrid frameworks. While AI models report standard accuracy rates of 95–99%, such metrics can be misleading due to severe class imbalance; evaluated models demonstrate more meaningful performance with F1-scores and AUC-ROC values ranging between 85% and 99% depending on experimental setups. Financial institutions should transition from traditional systems to AI-powered solutions, address class imbalance with specialized resampling or generative strategies, prioritize real-time detection infrastructure, and adopt hybrid or ensemble architectures to optimize performance and lower misclassification costs. This technological evolution marks a fundamental shift in modern financial security and fraud‑prevention efficacy.
Large language models (LLMs) have shown promise for supporting clinical decision-making from unstructured health records. This study provides a systematic comparison of feature-extraction pipelines and end-to-end fine-tuned LLMs for the early prediction of major adverse cardiovascular events (MACE) using the earliest available clinical notes from patients with acute coronary syndrome (ACS). Clinical notes from 2,650 patients with ST-elevation myocardial infarction (STEMI) and non-ST-elevation myocardial infarction (NSTEMI) were extracted from the MIMIC-III database. Feature-extraction approaches based on TF-IDF, word embeddings, and transformer representations combined with machine learning classifiers were compared against end-to-end fine-tuned LLMs in a multilabel prediction setting. Fine-tuned LLMs consistently outperformed feature-extraction pipelines. The best-performing model achieved AUC improvements of 18.2%, 17.3%, 9.1%, 12.4%, and 15.3% in STEMI, and 17.2%, 7.5%, 9.2%, 14.5%, and 10.0% in NSTEMI for predicting in-hospital mortality, reinfarction, heart failure, cardiogenic shock, and MACE, respectively. These findings suggest that fine-tuned LLMs can effectively leverage admission clinical narratives for early cardiovascular risk stratification and identification of high-risk patients.
Cardiovascular disease (CVD) risk assessment underpins preventive care. Although machine learning models often outperform traditional risk prediction methods, adoption of artificial intelligence (AI)-based clinical decision support (CDS) tools remains limited, especially in resource-constrained settings. This study explored the sociotechnical context of CVD risk assessment in resource-constrained settings and identified potential applications for AI-based CDS. Semi-structured interviews were conducted with 27 Indonesian doctors, including general practitioners and cardiologists. Interviews examined clinical decision-making processes, challenges associated with resource constraints, and potential use cases for AI-based CDS. Data were analyzed using a deductive thematic approach informed by the SEIPS (Systems Engineering Initiative for Patient Safety) framework. Doctors reported managing high patient volumes, relying on paper-based medical records, and making largely subjective decisions due to limited access to confirmatory investigations and specialist support. Acute coronary syndrome resulting from atherosclerosis was identified as the most common and important clinical problem. Participants highlighted atherosclerotic cardiovascular disease (ASCVD) risk assessment as a promising application for AI-based CDS because it uses widely available clinical information and can support preventive care decisions. Implementation of AI-based CDS should be guided by local clinical needs and sociotechnical context. Further research is needed to evaluate effects on clinical decision-making and patient outcomes.
Grasping the latest knowledge in any field has become increasingly challenging given the rapid pace of modern publishing. Recent advances in artificial intelligence (AI), particularly embedding models and large language models (LLMs), provide promising tools to aid researchers in the filtration and identification of the most relevant literature. However, there is little practical guidance on how to incorporate LLMs into literature reviews, and there are not yet direct comparisons of embedding models versus generative approaches. In this work, we compare these methods using small local LLMs for literature screening on both a simple dataset containing distinct categories and a more complex example with abstracts obtained from the Web of Science. We find that embedding models perform strongly for simple, well-separated queries due to their robustness and computational efficiency. For more complex filtering tasks, generative approaches based on token probabilities can achieve competitive or superior performance, but only when properly calibrated and when token-level outputs are interpreted as continuous scores rather than binary decisions. However, we show these methods are highly sensitive to prompt formulation, token conventions, and threshold selection.
