Signal detection is essential for identifying potential safety issues with drugs and medical devices. Change point analysis (CPA) detects changes in longitudinal data for a medical device without incorporating comparisons with other devices. This study investigated the temporal relationship between endoleak reports for aortic stent grafts identified using CPA and the issuance of the "Pharmaceuticals and Medical Devices Agency (PMDA) Alert for Proper Use of Medical Devices" in Japan. Device malfunction and adverse event reports for aortic stent grafts submitted to the PMDA between April 2008 and September 2022 were analyzed. The reports were mapped to the Medical Device Problem Terminology from the Japan Federation of Medical Devices Associations. CPA was used to identify temporal changes in report counts, with change points having confidence levels ≥ 90% being considered significant. A total of 20,678 reports were analyzed, including 23,541 device malfunctions and 30,500 adverse events. Endoleaks accounted for 10,779 events. CPA identified a significant increase in endoleak events in April 2015 (2 years before the regulatory alert). Confidence levels exceeded 90% in 76 out of 174 evaluated time points, first in July 2008 (9 years before the alert) and next in April 2016 (1 year before the alert). Based on these findings, CPA may help prioritize investigations of medical-device-related malfunctions and adverse events. However, the results do not indicate earlier or superior signal detection relative to regulatory authorities. Further research is needed to assess CPA's applicability to other medical devices and malfunctions.

Methods This narrative review drew on targeted searches of PubMed/MEDLINE and arXiv (January 2020–October 2025), verification of peer-reviewed versions via ACL Anthology for selected preprints, and citation tracking of seminal literature. In this review, we trace the methodological evolution from rules to encoder-based models and LLMs, propose a multidimensional evaluation framework for real-world deployment—which includes accuracy, structural quality, human-in-the-loop effort, stability, and compliance—and develop an operational governance checklist to support auditable and reproducible implementations. Using representative tasks—diagnosis extraction, medication records, clinical trial data, and phenotype integration—we summarize the improvements and failure modes of LLM-based extraction and analyze key challenges, including domain shift, factual “hallucinations,” privacy and regulatory constraints, and cost/latency trade-offs. Finally, we outline future directions through which multimodal and cross-lingual extensions, human–machine collaborative annotation, and standardized reporting practices can advance precision medicine and sustainable, high-quality clinical research.

Blepharoptosis is a common eyelid disorder that impairs both vision and appearance, requiring accurate assessment for effective treatment. This study aimed to develop and evaluate a deep learning (DL)-based system for automatic measurement of eyelid and periorbital parameters and for classifying levator function (LF) in patients with blepharoptosis. We retrospectively included 1,177 patients who underwent ptosis surgery at a tertiary oculoplastic center from January 2016 to November 2021. LF was categorized into good (> 10mm), fair (4-10mm), and poor (≤ 4mm) based on clinical evaluation. The DL model segmented eyelid and eyebrow regions and measured key parameters; manual measurements were performed for comparison. A multinomial logistic regression model incorporating DL-derived features and demographic data was used to predict LF grades. The DL system achieved high segmentation performance (Dice coefficient = 0.910) and strong agreement with manual measurements (ICC = 0.988 for MRD1; 0.902 for CBH). The regression model classified LF grades with an overall accuracy of 0.760 and an AUC of 0.829, within the range of ophthalmologist assessments (highest clinician accuracy = 0.767). This DL-based system offers an efficient, objective tool for periorbital assessment and LF grading, supporting personalized diagnosis and surgical planning in blepharoptosis management.

Deep vein thrombosis (DVT) in fracture patients is often clinically silent, with a high incidence of thrombosis and associated mortality. Static machine learning methods struggle to address the challenge of early DVT diagnosis due to their inability to adapt to heterogeneous data across patients. In contrast, Dynamic Ensemble Selection (DES) improves clinical decision-making and therapeutic interventions by dynamically adapting to variations in data characteristics. Here, we developed and validated a risk prediction model for DVT using electronic medical record data from fracture patients upon admission. By employing the DES method to optimize the prediction process, the model generates patient-specific probabilities of DVT occurrence, enabling personalized clinical risk assessment. Validation results showed that the DES model achieved strong performance in predicting DVT, with an accuracy of 0.875 and an Area Under the Receiver Operating Characteristic Curve (AUC) of 0.906. Notably, it demonstrated a high recall of 0.918 for DVT. Furthermore, in the prospective test set, DES exhibited excellent generalization capability, maintaining robust performance with an accuracy of 0.813 and an AUC of 0.876. We further developed an interactive clinical tool based on the DES algorithm to facilitate model interpretation and implementation. By integrating this user-friendly solution into clinical workflows, DES not only improves early DVT detection but also optimizes the allocation of healthcare resources.

Software as a Medical Device (SaMD) has become indispensable in diagnostics, treatment planning, and patient monitoring. While high-income countries have introduced clear regulatory frameworks, Bangladesh and many low- and middle-income countries (LMICs) still lack tailored pathways for medical software approval (IMDRF. Software as a Medical Device (SaMD): Key Definitions (IMDRF/SaMD WG/N10FINAL:(2013)); IMDRF. Software as a Medical Device (SaMD): Clinical Evaluation (IMDRF/SaMD WG/N41FINAL:(2017)); U.S. Food and Drug Administration (FDA). Software as a Medical Device (SAMD): Clinical Evaluation Guidance for Industry and FDA Staff (2017)). The current reliance on manual processes designed for physical devices leads to inefficiencies, inconsistent decisions, and potential risks to patient safety. This Comment proposes a semi-automated, risk-based intake roadmap for Bangladesh's Directorate General of Drug Administration (DGDA). Drawing on IMDRF, EU MDCG, and U.S. FDA frameworks, it presents a tangible workflow showing which submissions can be automatically triaged, which require human review, and where human override is maintained (European Commission (MDCG). Guidance on Qualification and Classification of Software in Regulation (EU) 2017/745 (MDCG 2019-11) and World Health Organization (WHO) Global Model Regulatory Framework for medical devices including IVDs (draft; WHO) (n.d.)). Key intake fields, escalation rules, and measurable performance indicators are defined. Anchored to Bangladesh's current DGDA and national digital health context, the proposal identifies specific legal and infrastructural gaps and outlines steps for phased modernization that may guide other LMICs.

Electrocorticography (ECoG) signals provide a valuable window into neural activity, yet their complex structure makes reliable classification challenging. This study addresses the problem by proposing a feature-selective framework that integrates multiple feature extraction techniques with statistical feature selection to improve classification performance. Power spectral density, wavelet-based features, Shannon entropy, and Hjorth parameters were extracted from ECoG signals obtained during a visual task. The most informative features were then selected using analysis of variance (ANOVA), and classification was performed with several machine learning methods, including decision trees, support vector machines, neural networks, and long short-term memory (LSTM) networks. Experimental results show that the proposed framework achieves high accuracy across individual patients as well as the combined dataset, with clear separability between classes confirmed through t-SNE visualization. In addition, analysis of selected features highlights the prominent role of electrodes located near the visual cortex, providing insights into the spatial distribution of neural activity.