Frontiers in artificial intelligence · 2024
Arkaitz Zubiaga
暂无摘要
J. Am. Medical Informatics Assoc. · 2025
Katherine E. Brown、Chao Yan、Zhuo-Hang Li 等
Abstract Objectives To determine the extent to which current large language models (LLMs) can serve as substitutes for traditional machine learning (ML) as clinical predictors using data from electronic health records (EHRs), we investigated various factors that can impact their adoption, including overall performance, calibration, fairness, and resilience to privacy protections that reduce data fidelity. Materials and Methods We evaluated GPT-3.5, GPT-4, and traditional ML (as gradient-boosting trees) on clinical prediction tasks in EHR data from Vanderbilt University Medical Center (VUMC) and MIMIC IV. We measured predictive performance with area under the receiver operating characteristic (AUROC) and model calibration using Brier Score. To evaluate the impact of data privacy protections, we assessed AUROC when demographic variables are generalized. We evaluated algorithmic fairness using equalized odds and statistical parity across race, sex, and age of patients. We also considered the impact of using in-context learning by incorporating labeled examples within the prompt. Results Traditional ML [AUROC: 0.847, 0.894 (VUMC, MIMIC)] substantially outperformed GPT-3.5 (AUROC: 0.537, 0.517) and GPT-4 (AUROC: 0.629, 0.602) (with and without in-context learning) in predictive performance and output probability calibration [Brier Score (ML vs GPT-3.5 vs GPT-4): 0.134 vs 0.384 vs 0.251, 0.042 vs 0.06 vs 0.219)]. Discussion Traditional ML is more robust than GPT-3.5 and GPT-4 in generalizing demographic information to protect privacy. GPT-4 is the fairest model according to our selected metrics but at the cost of poor model performance. Conclusion These findings suggest that non-fine-tuned LLMs are less effective and robust than locally trained ML for clinical prediction tasks, but they are improving across releases.
NEJM evidence · 2023
Michael Fralick、Chana A Sacks、Daniel Muller 等
Large Language ModelsIn the latest edition of Stats, STAT!, Fralick and colleagues explain the statistics behind large language models - used in chat bots like ChatGPT and Bard. While these new tools may seem remarkably intelligent, at their core they just assemble sentences based on statistics from large amounts of text.
International journal of rheumatic diseases · 2025
Fabian Lechner、Sebastian Kuhn、Johannes Knitza
暂无摘要
IEEE Access · 2025
Rahul Mundlamuri、G. Gunnam、Nikhil Kumar Mysari 等
This paper provides a comprehensive review of the evolution of artificial intelligence from early symbolic, rule-based systems to modern large language models (LLMs) and retrieval-augmented generation (RAG) architectures. Early AI was dominated by symbolic reasoning and expert systems using hand-crafted rules, before a paradigm shift toward data-driven learning occurred with the advent of machine learning, notably the backpropagation algorithm that enabled training of multi-layer neural networks to learn complex tasks. This breakthrough ushered in the deep learning era: convolutional neural networks (CNNs) achieved landmark results in computer vision, recurrent neural networks (RNNs such as LSTM) excelled at sequential data processing, and deep reinforcement learning (RL) systems mastered challenging decision-making tasks. In natural language processing (NLP), distributed word embedding models like Word2Vec and GloVe replaced sparse representations, capturing semantic relationships in vector space. To handle rare and out-of-vocabulary words, subword tokenization techniques such as Byte Pair Encoding (BPE) and WordPiece were introduced. These advances culminated in the Transformer architecture, based on self-attention mechanisms, which now underpins most state-of-the-art LLMs. Transformer-based LLMs, pre-trained on large text corpora, demonstrate advanced capabilities in language generation and understanding. Beyond model architectures, the review highlights how LLMs are augmented with external knowledge via RAG. Retrieval-augmented generation combines an LLM with a vector database of text embeddings, allowing the model to retrieve relevant external information to ground its outputs. This approach significantly improves factual accuracy and relevance, mitigating issues like hallucination, though ensuring complete truthfulness remains challenging. Additional obstacles include the limited context windows of LLMs (on the order of a few thousand tokens) and the computational complexity and latency of inference. The review also examines societal implications of advanced AI, such as bias amplification from training data and other ethical considerations in deployment. Finally, it outlines future directions, including extended context lengths, improved interpretability and alignment, greater efficiency, and neuro-symbolic hybrid approaches, aiming to develop more robust and trustworthy AI systems.
Frontiers in Cellular and Infection Microbiology · 2025
Yupeng Han、Xi-Yuan Xie、Jia-Peng Qiu 等
Background Sepsis associated encephalopathy (SAE) is prevalent among elderly patients in the ICU and significantly affects patient prognosis. Due to the symptom similarity with other neurological disorders and the absence of specific biomarkers, early clinical diagnosis remains challenging. This study aimed to develop a predictive model for SAE in elderly ICU patients. Methods The data of elderly sepsis patients were extracted from the MIMIC IV database (version 3.1) and divided into training and test sets in a 7:3 ratio. Feature variables were selected using the LASSO-Boruta combined algorithm, and five machine learning (ML) models, including Extreme Gradient Boosting (XGBoost), Categorical Boosting (CatBoost),Light Gradient Boosting Machine(LGBM), Multilayer Perceptron (MLP), and Support Vector Machines (SVM), were subsequently developed using these variables. A comprehensive set of performance metrics was used to assess the predictive accuracy, calibration, and clinical applicability of these models. For the machine learning model with the best performance, we employed the SHapley Additive Explanations(SHAP) method to visualize the model. Results Based on strict inclusion and exclusion criteria, a total of 3,156 elderly sepsis patients were enrolled in the study, with an SAE incidence rate of 48.7%. The mortality rate of elderly sepsis patients who developed SAE was significantly higher than that of patients in the non-SAE group (28.78% vs. 12.59%, P < 0.001). A total of 18 feature variables were selected for the construction of the ML model using the LASSO-Boruta combined algorithm. Compared to the other four models and traditional scoring systems, the XGBoost model demonstrated the best overall predictive performance, with Area Under the Curve(AUC)=0.898, accuracy=0.830, recall=0.819, F1-Score=0.820, specificity=0.840, and Precision=0.821. Furthermore, the results from the Decision Curve Analysis (DCA) and calibration curves demonstrated that the XGBoost model has significant clinical value and stable predictive performance. The ten-fold cross-validation method further confirmed the robustness and generalizability of the model. In addition, we simplified the model based on the SHAP feature importance ranking, and the results indicated that the simplified XGBoost model retains excellent predictive ability (AUC=0.858). Conclusions The XGBoost model effectively predicts SAE in elderly ICU patients and may serve as a reliable tool for clinicians to identify high-risk patients.
Journal of plastic, reconstructive & aesthetic surgery : JPRAS · 2025
Robert Browne、Kevin C Cahill
This correspondence serves to highlight the potential ethical issues associated with the use of large language models in plastic surgery.
Archives of dermatological research · 2024
Alyssa Breneman、Megan H Trager、Emily R Gordon 等
暂无摘要
BMC Infectious Diseases · 2025
Muhammad Zubair、Irfanud Din、N. Sarwar 等
Sepsis is a life-threatening condition resulting from a dysregulated immune response to infection, often leading to organ failure and death. Early detection is vital, as delays significantly worsen outcomes. In recent years, the integration of artificial intelligence (AI), particularly Machine Learning (ML) and Deep Learning (DL), has shown great promise in enhancing early sepsis detection by identifying digital biomarkers from large-scale clinical datasets. This systematic review analyzes and synthesizes existing ML/DL approaches applied to sepsis prediction, with an emphasis on intensive care unit (ICU) settings. A total of 80 studies were included, covering diverse data sources (e.g., MIMIC-III, eICU), feature selection methods, algorithm types, preprocessing techniques, and evaluation metrics. The models ranged from traditional techniques like logistic regression and decision trees to advanced architectures such as LSTM, transformers, and ensemble methods. A key contribution of this review is the inclusion of a forest plot summarizing reported AUC and sensitivity values from selected studies, offering a comparative visual of diagnostic performance. This helps highlight the relative effectiveness of different models and provides insights into their generalizability across clinical datasets. The review also discusses challenges related to model interpretability, ethical considerations, and the lack of external and temporal validation in many studies. It further identifies trends such as the use of real-time EHR data, patient-specific model development, and explainability tools like SHAP for clinician trust. By mapping out methodological strengths and limitations in current research, this work provides actionable recommendations for future studies and clinical deployment. The review contributes to the development of more robust, interpretable, and clinically relevant ML/DL models for early sepsis detection and improved patient care outcomes.
IEEE Access · 2025
D. Panagoulias、Evangelia-Aikaterini Tsichrintzi、Dionisios N. Sotiropoulos 等
This paper presents LYRICEL, a framework integrating Knowledge Graph (KG) representation learning, Large Language Models (LLMs), and machine learning for reliable, explainable, and validatable cross-cultural lyric analysis. The core component, Sequential Language Model Integration (SLMI), enhances the interpretability and reliability of transformer-based LLMs by addressing explainability and validation challenges through Retrieval-Augmented Generation (RAG), hybrid search, and rule-based evaluation. An important feature of LYRICEL is its use of KG visualizations, which serve as dynamic links to improve interpretability and validatability by structuring data relationships and sources. These visualizations are central to advancements in four areas: KG representation learning, knowledge acquisition, temporal KGs, and knowledge-aware applications. Tested on Greek folk music with models like GPT-4o and BERT, LYRICEL’s trustworthiness is assessed using the VIRTSI model, which quantifies cognitive trust in human-computer interactions. The framework shows strong potential for cross-cultural applications, particularly in languages such as Modern Greek which encompasses a rich cultural heritage spanning centuries of history and traditions resulting in a complex study. The outcomes of GPT-enabled LYRICEL are compared to ChatGPT alone and show a significant improvement in the reliability and efficiency of interactions that can reach a global audience, enhancing the accessibility and understanding of diverse cultural heritages.
为什么值得关注
AI 辅助短评，基于标题和摘要整理该研究构建交互式基准，评估语言模型在序贯诊断中的表现，AI系统在NEJM CPC病例和MIMIC-IV-ED数据上诊断准确率显著高于全科医生。但结果限于受控基准环境，未涉及真实临床验证，且未报告检查成本的具体比较数据。其模拟医生小组和成本效益选择的思路值得关注，但向临床转化仍需前瞻性研究支持。