JMIR AI最新文献

Digital health interventions often use machine learning (ML) models to make predictions of repeated adverse health events. For example, models may be used to analyze patient data to identify patterns that can anticipate the likelihood of disease exacerbations, enabling timely interventions and personalized treatment plans. However, many digital health applications require the prediction of highly heterogeneous and nuanced health events. The cross-subject variability of these events makes traditional ML approaches, where a single generalized model is trained to classify a particular condition, unlikely to generalize to patients outside of the training set. A natural solution is to train a separate model for each individual or subgroup, essentially overfitting the model to the unique characteristics of the individual without negatively overfitting in terms of the desired prediction task. Such an approach has traditionally required extensive data labels from each individual, a reality that has rendered personalized ML infeasible for precision health care. The recent popularization of self-supervised learning, however, provides a solution to this issue: by pretraining deep learning models on the vast array of unlabeled data streams arising from patient-generated health data, personalized models can be fine-tuned to predict the health outcome of interest with fewer labels than purely supervised approaches, making personalization of deep learning models much more achievable from a practical perspective. This perspective describes the current state-of-the-art in both self-supervised learning and ML personalization for health care as well as growing efforts to combine these two ideas by conducting self-supervised pretraining on an individual's data. However, there are practical challenges that must be addressed in order to fully realize this potential, such as human-computer interaction innovations to ensure consistent labeling practices within a single participant.

Background: Efficient allocation of health care resources is essential for long-term hospital operation. Effective intensive care unit (ICU) management is essential for alleviating the financial strain on health care systems. Accurate prediction of length-of-stay in ICUs is vital for optimizing capacity planning and resource allocation, with the challenge of achieving early, real-time predictions.

Objective: This study aimed to develop a predictive model, namely wavelet long short-term memory model (WT-LSTM), for ICU length-of-stay using only real-time vital sign data. The model is designed for urgent care settings where demographic and historical patient data or laboratory results may be unavailable; the model leverages real-time inputs to deliver early and accurate ICU length-of-stay predictions.

Methods: The proposed model integrates discrete wavelet transformation and long short-term memory (LSTM) neural networks to filter noise from patients' vital sign series and improve length-of-stay prediction accuracy. Model performance was evaluated using the electronic ICU database, focusing on 10 common ICU admission diagnoses in the database.

Results: The results demonstrate that WT-LSTM consistently outperforms baseline models, including linear regression, LSTM, and bidirectional long short-term memory, in predicting ICU length-of-stay using vital sign data, achieving significant improvements in mean square error. Specifically, the wavelet transformation component of the model enhances the overall performance of WT-LSTM. Removing this component results in an average decrease of 3.3% in mean square error; such a phenomenon is particularly pronounced in specific patient cohorts. The model's adaptability is highlighted through real-time predictions using only 3-hour, 6-hour, 12-hour, and 24-hour input data. Using only 3 hours of input data, the WT-LSTM model delivers competitive results across the 10 most common ICU admission diagnoses, often outperforming Acute Physiology and Chronic Health Evaluation IV, the leading ICU outcome prediction system currently implemented in clinical practice. WT-LSTM effectively captures patterns from vital signs recorded during the initial hours of a patient's ICU stay, making it a promising tool for early prediction and resource optimization in the ICU.

Conclusions: Our proposed WT-LSTM model, based on real-time vital sign data, offers a promising solution for ICU length-of-stay prediction. Its high accuracy and early prediction capabilities hold significant potential for enhancing clinical practice, optimizing resource allocation, and supporting critical clinical and administrative decisions in ICU management.

Background: The visual similarity of melanoma and seborrheic keratosis has made it difficult for older patients with disabilities to know when to seek medical attention, contributing to the metastasis of melanoma.

Objective: This study aimed to present a novel multimodal deep learning-based technique to distinguish between melanoma and seborrheic keratosis.

Methods: Our strategy is three-fold: (1) use patient image data to train and test three deep learning models using transfer learning (ResNet50, InceptionV3, and VGG16) and one author-designed model, (2) use patient metadata to train and test a deep learning model, and (3) combine the predictions of the image model with the best accuracy and the metadata model, using nonlinear least squares regression to specify ideal weights to each model for a combined prediction.

Results: The accuracy of the combined model was 88% (195/221 classified correctly) on test data from the HAM10000 dataset. Model reliability was assessed by visualizing the output activation map of each model and comparing the diagnosis patterns to that of dermatologists. The addition of metadata to the image dataset was key to reducing the false-negative and false-positive rates simultaneously, thereby producing better metrics and improving overall model accuracy.

Conclusions: Results from this experiment could be used to eliminate late diagnosis of melanoma via easy access to an app. Future experiments can use text data (subjective data pertaining to how the patient felt over a certain period of time) to allow this model to reflect the real hospital setting to a greater extent.

Unlabelled: Shared decision-making is central to patient-centered care but is often hampered by artificial intelligence (AI) systems that focus on technical transparency rather than delivering context-rich, clinically meaningful reasoning. Although AI explainability methods elucidate how decisions are made, they fall short of addressing the "why" that supports effective patient-clinician dialogue. To bridge this gap, we introduce artificial intelligence-supported shared decision-making (AI-SDM), a conceptual framework designed to integrate AI-based reasoning into shared decision-making to enhance care quality while preserving patient autonomy. AI-SDM is a structured, multimodel framework that synthesizes predictive modeling, evidence-based recommendations, and generative AI techniques to produce adaptive, context-sensitive explanations. The framework distinguishes conventional AI explainability from AI reasoning-prioritizing the generation of tailored, narrative justifications that inform shared decisions. A hypothetical clinical scenario in stroke management is used to illustrate how AI-SDM facilitates an iterative, triadic deliberation process between health care providers, patients, and AI outputs. This integration is intended to transform raw algorithmic data into actionable insights that directly support the decision-making process without supplanting human judgment.

Background: Overcrowded emergency rooms might degrade the quality of care and overload the clinic staff. Assessing unscheduled return visits (URVs) to the emergency department (ED) is a quality assurance procedure to identify ED-discharged patients with a high likelihood of bounce-back, to ensure patient safety, and ultimately to reduce medical costs by decreasing the frequency of URVs. The field of machine learning (ML) has evolved considerably in the past decades, and many ML applications have been deployed in various contexts.

Objective: This study aims to develop an ML-assisted framework that identifies high-risk patients who may revisit the ED within 72 hours after the initial visit. Furthermore, this study evaluates different ML models, feature sets, and feature encoding methods in order to build an effective prediction model.

Methods: This study proposes an ML-assisted system that extracts the features from both structured and unstructured medical data to predict patients who are likely to revisit the ED, where the structured data includes patients' electronic health records, and the unstructured data is their medical notes (subjective, objective, assessment, and plan). A 5-year dataset consisting of 184,687 ED visits, along with 324,111 historical electronic health records and the associated medical notes, was obtained from Kaohsiung Veterans General Hospital, a tertiary medical center in Taiwan, to evaluate the proposed system.

Results: The evaluation results indicate that incorporating convolutional neural network-based feature extraction from unstructured ED physician narrative notes, combined with structured vital signs and demographic data, significantly enhances predictive performance. The proposed approach achieves an area under the receiver operating characteristic curve of 0.705 and a recall of 0.718, demonstrating its effectiveness in predicting URVs. These findings highlight the potential of integrating structured and unstructured clinical data to improve predictive accuracy in this context.

Conclusions: The study demonstrates that an ML-assisted framework may be applied as a decision support tool to assist ED clinicians in identifying revisiting patients, although the model's performance may not be sufficient for clinic implementation. Given the improvement in the area under the receiver operating characteristic curve, the proposed framework should be further explored as a workable decision support tool to pinpoint ED patients with a high risk of revisit and provide them with appropriate and timely care.

Background: Ultrasound examinations, while valuable, are time-consuming and often limited in availability. Consequently, many hospitals implement reservation systems; however, these systems typically lack prioritization for examination purposes. Hence, our hospital uses a waitlist system that prioritizes examination requests based on their clinical value when slots become available due to cancellations. This system, however, requires a manual review of examination purposes, which are recorded in free-form text. We hypothesized that artificial intelligence language models could preliminarily estimate the priority of requests before manual reviews.

Objective: This study aimed to investigate potential challenges associated with using language models for estimating the priority of medical examination requests and to evaluate the performance of language models in processing Japanese medical texts.

Methods: We retrospectively collected ultrasound examination requests from the waitlist system at Keio University Hospital, spanning January 2020 to March 2023. Each request comprised an examination purpose documented by the requesting physician and a 6-tier priority level assigned by a radiologist during the clinical workflow. We fine-tuned JMedRoBERTa, Luke, OpenCalm, and LLaMA2 under two conditions: (1) tuning only the final layer and (2) tuning all layers using either standard backpropagation or low-rank adaptation.

Results: We had 2335 and 204 requests in the training and test datasets post cleaning. When only the final layers were tuned, JMedRoBERTa outperformed the other models (Kendall coefficient=0.225). With full fine-tuning, JMedRoBERTa continued to perform best (Kendall coefficient=0.254), though with reduced margins compared with the other models. The radiologist's retrospective re-evaluation yielded a Kendall coefficient of 0.221.

Conclusions: Language models can estimate the priority of examination requests with accuracy comparable with that of human radiologists. The fine-tuning results indicate that general-purpose language models can be adapted to domain-specific texts (ie, Japanese medical texts) with sufficient fine-tuning. Further research is required to address priority rank ambiguity, expand the dataset across multiple institutions, and explore more recent language models with potentially higher performance or better suitability for this task.

Background: People who use drugs (PWUD) are at heightened risk of severe injection-related infections. Current research relies on billing codes to identify PWUD-a methodology with suboptimal accuracy that may underestimate the economic, racial, and ethnic diversity of hospitalized PWUD.

Objective: The goal of this study is to examine the impact of natural language processing (NLP) on enhancing identification of PWUD in electronic medical records, with a specific focus on determining improved systems of identifying populations who may previously been missed, including people who have low income or those from racially and ethnically minoritized populations.

Methods: Health informatics specialists assisted in querying a cohort of likely PWUD hospital admissions at Tufts Medical Center between 2020-2022 using the following criteria: (1) ICD-10 codes indicative of drug use, (2) positive drug toxicology results, (3) prescriptions for medications for opioid use disorder, and (4) applying NLP-detected presence of "token" keywords in the electronic medical records likely indicative of the patient being a PWUD. Hospital admissions were split into two groups: highly documented (all four criteria present) and minimally documented (NLP-only). These groups were examined to assess the impact of race, ethnicity, and social vulnerability index. With chart review as the "gold standard," the positive predictive value was calculated.

Results: The cohort included 4548 hospitalization admissions, with broad heterogeneity in how people entered the cohort and subcohorts; a total of 288 hospital admissions entered the cohort through NLP token presence alone. NLP demonstrated a 54% positive predictive value, outperforming biomarkers, prescription for medications for opioid use disorder, and ICD codes in identifying hospitalizations of PWUD. Additionally, NLP significantly enhanced these methods when integrated into the identification algorithm. The study also found that people from racially and ethnically minoritized communities and those with lower social vulnerability index were significantly more likely to have lower rates of PWUD-related documentation.

Conclusions: NLP proved effective in identifying hospitalizations of PWUD, surpassing traditional methods. While further refinement is needed, NLP shows promising potential in minimizing health care disparities.

Unstructured: Shared decision-making is central to patient-centered care but is often hampered by AI systems that focus on technical transparency rather than delivering context-rich, clinically meaningful reasoning. Although XAI methods elucidate how decisions are made, they fall short in addressing the "why" that supports effective patient-clinician dialogue. To bridge this gap, we introduce AI-SDM, a conceptual framework designed to integrate AI-based reasoning into Shared decision-making to enhance care quality while preserving patient autonomy. AI-SDM is a structured, multi-model framework that synthesizes predictive modelling, evidence-based recommendations, and generative AI techniques to produce adaptive, context-sensitive explanations. The framework distinguishes conventional AI explainability from AI reasoning-prioritizing the generation of tailored, narrative justifications that inform shared decisions. A hypothetical clinical scenario in stroke management is used to illustrate how AI-SDM facilitates an iterative, triadic deliberation process between healthcare providers, patients, and AI outputs. This integration is intended to transform raw algorithmic data into actionable insights that directly support the decision-making process without supplanting human judgment.

Background: The visual similarity of melanoma and seborrheic keratosis has made it difficult for elderly patients with disabilities to know when to seek medical attention, contributing to the metastasis of melanoma.

Objective: In this paper, we present a novel multi-modal deep learning-based technique to distinguish between melanoma and seborrheic keratosis.

Methods: Our strategy is three-fold: (1) utilize patient image data to train and test three deep learning models using transfer learning (ResNet50, InceptionV3, and VGG16) and one author designed model, (2) use patient metadata to train and test a deep learning model, and (3) combine the predictions of the image model with the best accuracy and the metadata model, using nonlinear least squares regression to specify ideal weights to each model for a combined prediction.

Results: The accuracy of the combined model was 88% (195/221 classified correctly) on test data from the HAM10000 dataset. Model reliability was assessed by visualizing the output activation map of each model and comparing the diagnosis patterns to that of dermatologists. The addition of metadata to the image dataset was key to reducing the false negative and false positive rate simultaneously, thereby producing better metrics and improving overall model accuracy.

Conclusions: Results from this experiment could be used to eliminate late diagnosis of melanoma via easy access to an app. Future experiments can utilize text data (subjective data pertaining to how the patient felt over a certain period of time) to allow this model to reflect the real hospital setting to a greater extent.

Clinicaltrial:

Background: The digital transformation of health care has introduced both opportunities and challenges, particularly in managing and analyzing the vast amounts of unstructured medical data generated daily. There is a need to explore the feasibility of generative solutions in extracting data from medical reports, categorized by specific criteria.

Objective: This study aimed to investigate the application of large language models (LLMs) for the automated extraction of structured information from unstructured medical reports, using the LangChain framework in Python.

Methods: Through a systematic evaluation of leading LLMs-GPT-4o, Llama 3, Llama 3.1, Gemma 2, Qwen 2, and Qwen 2.5-using zero-shot prompting techniques and embedding results into a vector database, this study assessed the performance of LLMs in extracting patient demographics, diagnostic details, and pharmacological data.

Results: Evaluation metrics, including accuracy, precision, recall, and F₁-score, revealed high efficacy across most categories, with GPT-4o achieving the highest overall performance (91.4% accuracy).

Conclusions: The findings highlight notable differences in precision and recall between models, particularly in extracting names and age-related information. There were challenges in processing unstructured medical text, including variability in model performance across data types. Our findings demonstrate the feasibility of integrating LLMs into health care workflows; LLMs offer substantial improvements in data accessibility and support clinical decision-making processes. In addition, the paper describes the role of retrieval-augmented generation techniques in enhancing information retrieval accuracy, addressing issues such as hallucinations and outdated data in LLM outputs. Future work should explore the need for optimization through larger and more diverse training datasets, advanced prompting strategies, and the integration of domain-specific knowledge to improve model generalizability and precision.