Background: The efficacy of large language models (LLMs) in domain-specific medicine, particularly for managing complex diseases such as osteoarthritis (OA), remains largely unexplored. Objective: This study focused on evaluating and enhancing the clinical capabilities and explainability of LLMs in specific domains, using OA management as a case study. Methods: A domain-specific benchmark framework was developed to evaluate LLMs across a spectrum from domain-specific knowledge to clinical applications in real-world clinical scenarios. DocOA, a specialized LLM designed for OA management integrating retrieval-augmented generation and instructional prompts, was developed. It can identify the clinical evidence upon which its answers are based through retrieval-augmented generation, thereby demonstrating the explainability of those answers. The study compared the performance of GPT-3.5, GPT-4, and a specialized assistant, DocOA, using objective and human evaluations. Results: Results showed that general LLMs such as GPT-3.5 and GPT-4 were less effective in the specialized domain of OA management, particularly in providing personalized treatment recommendations. However, DocOA showed significant improvements. Conclusions: This study introduces a novel benchmark framework that assesses the domain-specific abilities of LLMs in multiple aspects, highlights the limitations of generalized LLMs in clinical contexts, and demonstrates the potential of tailored approaches for developing domain-specific medical LLMs.

Children’s health is a worldwide concern, and eye health in children has received a lot of attention. Approximately 50% of schoolchildren under the age of 18 years in China are affected by myopia.^1,2 Strabismus occurs in 1.19% to 5.65% of children^3-7 and ptosis affects around 0.18% to 1.41% of children.^8-10

Background: Surgical site infection (SSI) is one of the most common types of health care–associated infections. It increases mortality, prolongs hospital length of stay, and raises health care costs. Many institutions developed risk assessment models for SSI to help surgeons preoperatively identify high-risk patients and guide clinical intervention. However, most of these models had low accuracies. Objective: We aimed to provide a solution in the form of an Artificial intelligence–based Multimodal Risk Assessment Model for Surgical site infection (AMRAMS) for inpatients undergoing operations, using routinely collected clinical data. We internally and externally validated the discriminations of the models, which combined various machine learning and natural language processing techniques, and compared them with the National Nosocomial Infections Surveillance (NNIS) risk index. Methods: We retrieved inpatient records between January 1, 2014, and June 30, 2019, from the electronic medical record (EMR) system of Rui Jin Hospital, Luwan Branch, Shanghai, China. We used data from before July 1, 2018, as the development set for internal validation and the remaining data as the test set for external validation. We included patient demographics, preoperative lab results, and free-text preoperative notes as our features. We used word-embedding techniques to encode text information, and we trained the LASSO (least absolute shrinkage and selection operator) model, random forest model, gradient boosting decision tree (GBDT) model, convolutional neural network (CNN) model, and self-attention network model using the combined data. Surgeons manually scored the NNIS risk index values. Results: For internal bootstrapping validation, CNN yielded the highest mean area under the receiver operating characteristic curve (AUROC) of 0.889 (95% CI 0.886-0.892), and the paired-sample t test revealed statistically significant advantages as compared with other models (P<.001). The self-attention network yielded the second-highest mean AUROC of 0.882 (95% CI 0.878-0.886), but the AUROC was only numerically higher than the AUROC of the third-best model, GBDT with text embeddings (mean AUROC 0.881, 95% CI 0.878-0.884, P=.47). The AUROCs of LASSO, random forest, and GBDT models using text embeddings were statistically higher than the AUROCs of models not using text embeddings (P<.001). For external validation, the self-attention network yielded the highest AUROC of 0.879. CNN was the second-best model (AUROC 0.878), and GBDT with text embeddings was the third-best model (AUROC 0.872). The NNIS risk index scored by surgeons had an AUROC of 0.651. Conclusions: Our AMRAMS based on EMR data and deep learning methods—CNN and self-attention network—had significant advantages in terms of accuracy compared with other conventional machine learning methods and the NNIS risk index. Moreover, the semantic embeddings of preoperative notes improved the model performance further. Our models could replace the NNIS risk index to provide personalized guidance for the preoperative intervention of SSIs. Through this case, we offered an easy-to-implement solution for building multimodal RAMs for other similar scenarios.

Background: Pathway-based analysis of transcriptomic data has shown greater stability and better performance than traditional gene-based analysis. Until now, some pathway-based deep learning models have been developed for bioinformatic analysis, but these models have not fully considered the topological features of pathways, which limits the performance of the final prediction result. Results: To address this issue, we propose a novel model, called PathGNN, which constructs a Graph Neural Networks (GNNs) model that can capture topological features of pathways. As a case, PathGNN was applied to predict long-term survival of four types of cancer and achieved promising predictive performance when compared to other common methods. Furthermore, the adoption of an interpretation algorithm enabled the identification of plausible pathways associated with survival. Conclusion: PathGNN demonstrates that GNN can be effectively applied to build a pathway-based model, resulting in promising predictive power.

Background: Large language models (LLMs) have achieved great progress in natural language processing tasks and demonstrated the potential for use in clinical applications. Despite their capabilities, LLMs in the medical domain are prone to generating hallucinations (not fully reliable responses). Hallucinations in LLMs’ responses create substantial risks, potentially threatening patients’ physical safety. Thus, to perceive and prevent this safety risk, it is essential to evaluate LLMs in the medical domain and build a systematic evaluation. Objective: We developed a comprehensive evaluation system, MedGPTEval, composed of criteria, medical data sets in Chinese, and publicly available benchmarks. Methods: First, a set of evaluation criteria was designed based on a comprehensive literature review. Second, existing candidate criteria were optimized by using a Delphi method with 5 experts in medicine and engineering. Third, 3 clinical experts designed medical data sets to interact with LLMs. Finally, benchmarking experiments were conducted on the data sets. The responses generated by chatbots based on LLMs were recorded for blind evaluations by 5 licensed medical experts. The evaluation criteria that were obtained covered medical professional capabilities, social comprehensive capabilities, contextual capabilities, and computational robustness, with 16 detailed indicators. The medical data sets include 27 medical dialogues and 7 case reports in Chinese. Three chatbots were evaluated: ChatGPT by OpenAI; ERNIE Bot by Baidu, Inc; and Doctor PuJiang (Dr PJ) by Shanghai Artificial Intelligence Laboratory. Results: Dr PJ outperformed ChatGPT and ERNIE Bot in the multiple-turn medical dialogues and case report scenarios. Dr PJ also outperformed ChatGPT in the semantic consistency rate and complete error rate category, indicating better robustness. However, Dr PJ had slightly lower scores in medical professional capabilities compared with ChatGPT in the multiple-turn dialogue scenario. Conclusions: MedGPTEval provides comprehensive criteria to evaluate chatbots by LLMs in the medical domain, open-source data sets, and benchmarks assessing 3 LLMs. Experimental results demonstrate that Dr PJ outperforms ChatGPT and ERNIE Bot in social and professional contexts. Therefore, such an assessment system can be easily adopted by researchers in this community to augment an open-source data set.

The availability of high-throughput sequencing data creates opportunities to comprehensively understand human diseases as well as challenges to train machine learning models using such high dimensions of data. Here, we propose a denoised multi-omics integration framework, which contains a distribution-based feature denoising algorithm, Feature Selection with Distribution (FSD), for dimension reduction and a multi-omics integration framework, Attention Multi-Omics Integration (AttentionMOI) to predict cancer prognosis and identify cancer subtypes. We demonstrated that FSD improved model performance either using single omic data or multi-omics data in 15 The Cancer Genome Atlas Program (TCGA) cancers for survival prediction and kidney cancer subtype identification. And our integration framework AttentionMOI outperformed machine learning models and current multi-omics integration algorithms with high dimensions of features. Furthermore, FSD identified features that were associated to cancer prognosis and could be considered as biomarkers.