JMIR Medical Informatics最新文献

Background: Electronic health records (EHRs) can aid in provider efficiency, but may also lead to unintended consequences, such as documentation burden and increased length of notes. To combat issues related to documentation, copying and pasting (CP) and copying or carrying forward (CF) are tools that have been used to aid in documentation burden. Multiple studies have identified the benefits and challenges of using these tools; however, few studies have identified the unintended consequences of CP and CF, and how the adoption of these tools may affect users.

Objective: The objective was to describe providers' perceptions and use of copying tools available in the EHR and describe their suggestions for improvement on these copying tools.

Methods: Research team members conducted semistructured interviews with faculty members, advanced practice providers, residents or fellow trainees, and medical students at a single academic health sciences center. The Diffusion of Innovations Theory of Unintended Consequences guided the analysis and interpretation of interview results.

Results: A total of 22 semistructured interviews were conducted in 2023 and analyzed during 2024. The findings showed that respondents use and value these tools for efficiency and communication purposes. The negative unintended consequences include inaccuracies and errors in documentation and increased patient safety risks. Some respondents experience inner angst or moral injury related to using CP/CF, but they feel that they must use them to satisfy organizational requirements surrounding documentation. The respondents suggested that artificial intelligence will likely help improve documentation tools, as would further training around these types of documentation tools.

Conclusions: Some respondents noted feeling both internal and external pressures that influenced when and how they use CP/CF. Respondents noted that they value EHR copying tools for efficiency purposes, but they also understand the risks involved. This tension may lead to moral angst or moral injury. They offered numerous suggestions for lowering the risk, especially by improving the documentation capabilities of the EHR through artificial intelligence. Future research should investigate both technical and educational solutions to relieve the documentation burden and moral angst they are experiencing.

Background: Accurately predicting the survival outcomes of patients with lung cancer receiving chemotherapy remains challenging.

Objective: To improve clinical management of this population, this study developed a multivariate machine learning (ML) model to assess all-cause mortality risk in chemotherapy-treated patients with lung cancer.

Methods: This study retrospectively recruited 1278 postchemotherapy patients with lung cancer from Guangzhou Chest Hospital between 2017 and 2019. Candidate features such as demographic characteristics, environmental exposures, clinical information, and patient-reported symptoms were collected via questionnaires and the electronic medical record system. The survival status and the deceased date were investigated twice a year. A total of 84 predictive models were constructed on the training set using 5 ML algorithms either individually or in pairwise combinations. The concordance index was used to identify the optimal model on the testing set, with performance validated via receiver operating characteristic curves, calibration curves, and decision curve analysis. Additionally, Shapley Additive Explanations and restricted cubic splines were applied for feature attribution analysis.

Results: The optimal model ultimately retained 21 prognosis-association features, including age, sex, BMI, smoking status, environmental smoke, the MD Anderson Symptom Inventory for Lung Cancer total score trajectories, cluster of differentiation 56, TNM stage, histology, and prechemotherapy blood biomarkers. On the testing set, the model acquired a concordance index of 0.702 (95% CI 0.652-0.753). The decision curves demonstrated positive clinical benefit when the risk thresholds were 0.40-0.69, 0.62-0.99, and 0.72-0.99 for 1-, 3-, and 5-year mortality predictions, respectively. The calibration curves showed that the predicted mortality probabilities fluctuated around the observed probabilities, and the Brier scores for 1-, 3-, and 5-year predictions were 0.20, 0.18, and 0.11, respectively. The area under the curve of the model was 0.740, 0.777, and 0.915 for 1-, 3-, and 5-year mortality predictions, respectively. Interpretability feature attribution analysis revealed that the significant features could predict all-cause mortality risk in chemotherapy-treated patients with lung cancer.

Conclusions: Our ML models exhibited acceptable discrimination, calibration, and clinical benefit in predicting the mortality risk of chemotherapy-treated patients with lung cancer, which could help clinicians in personalized prognostic management.

Background: Knee cartilage injury (KCI) poses significant challenges in the early clinical diagnosis process, primarily due to its high incidence, the complexity of healing, and the limited sensitivity of initial imaging modalities.

Objective: This study aims to employ magnetic resonance imaging and machine learning methods to enhance the classification accuracy of the classifier for KCI, improve the existing network structure, and demonstrate important clinical application value.

Methods: The proposed methodology is a multidimensional feature cross-level fusion classification network driven by the large separable kernel attention, which enables high-precision hierarchical diagnosis of KCI through deep learning. The network first fuses shallow high-resolution features with deep semantic features via the cross-level fusion module. Then, the large separable kernel attention module is embedded in the YOLOv8 network. This network utilizes the combined optimization of depth-separable and point-by-point convolutions to enhance features at multiple scales, thereby dramatically improving the hierarchical characterization of cartilage damage. Finally, five classifications of knee cartilage injuries are performed by classifiers.

Results: To overcome the limitations of network models trained with single-plane images, this study presents the first hospital-based multidimensional magnetic resonance imaging real dataset for KCI, on which the classification accuracy is 99.7%, the Kappa statistic is 99.6%, the F-measure is 99.7%, the sensitivity is 99.7%, and the specificity is 99.9%. The experimental results validate the feasibility of the proposed method.

Conclusions: The experimental outcomes confirm that the proposed methodology not only achieves exceptional performance in classifying knee cartilage injuries but also offers substantial improvements over existing techniques. This underscores its potential for clinical deployment in enhancing diagnostic precision and efficiency.

Background: We designed learning assignments for students to develop knowledge, skills, and professional attitudes about generative artificial intelligence (AI) in 2 different Master's level courses in health informatics. Our innovative approach assumed that the students had no technical background or experience in using generative AI tools.

Objective: This study aims to offer generalizable methods and experiences on integration and assessment of generative AI content into the higher education's health informatics curricula. The study's central driver is the preparation of graduate students with generative AI tools, skills, ethical discernment, and critical thinking capacities aligned with the rapidly shifting job-market requirements, independent of graduate students' backgrounds and technical expertise.

Methods: During the semester, students completed a pretest and posttest to assess knowledge about generative AI. Reflections explored their expectations and experiences using generative AI to complete their assignments and projects during the semester. Strong emphasis was placed on building skills and professional attitudes by using generative AI. Student engagement in behavioral, emotional, and cognitive domains was explored via detailed analysis of student reflections by faculty.

Results: Students at the University of Illinois Chicago increased their knowledge about generative AI from 81% to 93% through research of the basic generative AI concepts, as evidenced from outcomes of the open-book pre-and posttests given at the beginning and end of the capstone course. University of San Francisco students also improved from 77% to 80% by the end of the semester. Faculty analysis of student reflections upon completion of the course revealed primary interests in the essentials of generative AI, AI transformations to information and knowledge, and organizational changes influenced by AI adoption in the health care organizations, with ethics being a primary driver of students' interests and engagement.

Conclusions: Data from student reflections provided insight into generative AI skills that students developed and that health informatics programs can consider incorporating into their curricula. Building competencies in generative AI will prepare students for the 21st century workforce and enable them to build skills employers are seeking in the new digital health environment.

Background: Metabolic dysfunction-associated fatty liver disease (MAFLD) is a leading cause of chronic disease and can progress to liver fibrosis or hepatocellular carcinoma. Its subtypes-obese, diabetic, and lean-are associated with varying degrees of fibrotic burden and different complications, yet the existing analytics methods often overlook its multisystem nature, intraphenotype variability, and disease dynamics. These limitations hinder accurate risk stratification and restrict personalized intervention planning.

Objective: This study developed a novel, 2-stage, contrastive learning-based method to predict the phenotype of MAFLD among adults. This method leverages multiview contrastive learning; it models individual heterogeneities and important relationships in clinical and survey-based data to predict phenotypes among adults, thus supporting clinical decision-making and personalized care.

Methods: Demographic, clinical, lifestyle, and genetic family history data of 4408 adults revealed how capturing essential relationships in patient data from different sources can transform individual-level representations into multiple, complementary views. Evaluation of the predictive efficacy of the proposed method in comparison with 8 prevalent methods relied on recall, precision, F1-score, and area under the curve values. Moreover, a Shapley additive explanation analysis was performed for interpretability.

Results: The proposed method consistently and significantly outperformed all benchmark methods. It attained the highest F1-score, showing a 32.8% improvement for nondiabetic MAFLD (0.531 vs 0.400) and 30.4% improvement for diabetic MAFLD (0.519 vs 0.398) over the respective best-performing benchmark. The results underscore the clinical value and utility of integrating clinical and survey-based data in the prediction of MAFLD phenotypes among adults.

Conclusions: The proposed method is a viable approach for MAFLD phenotype prediction. It is more effective in identifying at-risk adults than many prevalent data-driven analytics methods and thereby can enhance clinical decision-making and support patient-centric care and management.

Background: Acute care use (ACU) represents a major economic burden in oncology, which can ideally be prevented. Existing models effectively predict such events.

Objective: We aimed to quantify the cost savings achieved by implementing a model to predict ACU in oncology patients undergoing systemic therapy.

Methods: This retrospective cohort study analyzed patients with cancer at an academic medical center from 2010 to 2022. We included patients who received systemic therapy and identified ACU events occurring after treatment initiation, excluding those with known death dates within the study period. Data on ACU-related expenses were gathered from Medicare claims and mapped to service codes in electronic health records, yielding average daily costs for each patient over 180 days following the start of therapy. The exposure was an ACU event.

Results: The main outcome was the average daily cost per patient at the end of the first 180 days of systemic therapy. We observed that expense accumulation flattened earlier and more rapidly among non-ACU patients. This study included 20,556 patients, of whom 3820 (18.58%) experienced at least 1 ACU. The average daily cost per patient for those with and without ACU was US $94.62 (SD US $72.54; 95% CI US $92.32-$96.92) and US $53.28 (SD US $59.92; 95% CI US $52.37-$54.19), respectively. The average total cost per ACU and non-ACU patient was US $17,031.92 (SD US $13,056.63; 95% CI US $16,616.74-$17,445.09) and US $9591.06 (SD US $10,785.83; 95% CI US $9427.64-$9754.48), respectively. To estimate the long-term financial impact of deploying the predictive model, we conducted a cost-benefit analysis based on an annual cohort size of 2177 patients. In the first year alone, the model yielded projected savings of US $910,000. By year 6, projected savings grew to US $9.46 million annually. The cumulative avoided costs over a 6-year deployment period totaled approximately US $31.11 million. These estimates compared the baseline cost model to the intervention model assuming a prevention rate of 35% for preventable ACU events and an average ACU cost of US $17,031.92 (SD US $13,037).

Conclusions: Predictive analytics can significantly reduce costs associated with ACU events, enhancing economic efficiency in cancer care. Further research is needed to explore potential health benefits.

Background: The use of large language models (LLMs) in radiology is expanding rapidly, offering new possibilities in report generation, decision support, and workflow optimization. However, a comprehensive evaluation of their applications, performance, and limitations across the radiology domain remains limited.

Objective: This review aimed to map current applications of LLMs in radiology, evaluate their performance across key tasks, and identify prevailing limitations and directions for future research.

Methods: A scoping review was conducted in accordance with the framework by Arksey and O'Malley framework and the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) guidelines. Three databases-PubMed, ScopusCOPUS, and IEEE Xplore-were searched for peer-reviewed studies published between January 2022 and December 2024. Eligible studies included empirical evaluations of LLMs applied to radiological data or workflows. Commentaries, reviews, and technical model proposals without evaluation were excluded. Two reviewers independently screened studies and extracted data on study characteristics, LLM type, radiological use case, data modality, and evaluation metrics. A thematic synthesis was used to identify key domains of application. No formal risk-of-bias assessment was performed, but a narrative appraisal of dataset representativeness and study quality was included.

Results: A total of 67 studies were included. (n/N, %)GPT-4 was the most frequently used model (n=28, 42%), with text-based corpora as the primary type of data used (n=43, 64%). Identified use cases fell into three thematic domains: (1) decision support (n=39, 58%), (2) report generation and summarization (n=16, 24%), and (3) workflow optimization (n=12, 18%). While LLMs demonstrated strong performance in structured-text tasks (eg, report simplification with >94% accuracy), diagnostic performance varied widely (16%-86%) and was limited by dataset bias, lack of fine tuning, and minimal clinical validation. Most studies (n=53, 79.1%) had single-center, proof-of-concept designs with limited generalizability.

Conclusions: LLMs show strong potential for augmenting radiological workflows, particularly for structured reporting, summarization, and educational tasks. However, their diagnostic performance remains inconsistent, and current implementations lack robust external validation. Future work should prioritize prospective, multicenter validation of domain-adapted and multimodal models to support safe clinical integration.

Background: Liver failure often results in significant coagulation dysfunction, which is a major complication. Artificial liver support systems (ALSS) have been used to ameliorate coagulation parameters, but the dynamic nature of these improvements and the development of predictive models remain insufficiently explored.

Objective: This study aimed to evaluate the effects of ALSS on coagulation function and to develop a dynamic prediction model using machine learning techniques to predict the improvement trends of coagulation parameters.

Methods: A systematic search was conducted in PubMed, Embase, and other databases to identify relevant studies, resulting in 18 studies comprising 1771 patients. A meta-analysis was performed to assess the impact of ALSS on coagulation parameters, including international normalized ratio (INR), prothrombin time (PT), activated partial thromboplastin time (APTT), and fibrinogen levels. In addition, clinical data from the Medical Information Mart for Intensive Care database were used to construct prediction models using logistic regression, extreme gradient boosting, random forest, and long short-term memory networks.

Results: Meta-analysis results showed that ALSS significantly improved INR, PT, APTT, and fibrinogen levels (all P<.05), with the treatment efficacy varying by modality. Among the machine learning models, the random forest model demonstrated the best performance, achieving an area under the curve of 92.12%. Dynamic INR was identified as the key predictor for coagulation abnormalities.

Conclusions: This study systematically evaluated the effects of ALSS on coagulation function in patients with liver failure, demonstrating significant improvements in key parameters such as INR, PT, and APTT, with efficacy varying across different treatment modalities. Simultaneously, a machine learning model built using intensive care unit clinical data exhibited strong predictive capability for identifying the risk of coagulation dysfunction, particularly useful in supporting early-stage clinical recognition of high-risk patients and guiding personalized coagulation management strategies. It is important to emphasize that this model is positioned as a dynamic risk alert and assessment tool, intended to assist clinical baseline evaluation and nursing interventions, rather than serving as direct validation of ALSS therapeutic efficacy.