Objective Patients increasingly use large language models (LLMs) for menopause-related health information, yet the quality and readability of their responses have not been systematically compared across widely used platforms. This study compared the information quality and readability of responses from popular LLMs to frequently asked menopause questions. Methods In this comparative study, 50 menopause questions were compiled from four sources (ChatGPT, Google Gemini, Microsoft Copilot, and Google’s “People also ask”) after removing duplicates and less relevant items. Each was entered into the three tools, and responses were categorized with a modified Rothwell system and independently rated by two obstetrics and gynecology specialists using DISCERN, a Likert scale, PEMAT, Flesch Reading Ease, and Flesch-Kincaid Grade Level. Results Of the 50 questions, 70% were fact, 20% policy, and 10% value. Total DISCERN scores and overall information quality did not differ among platforms. Gemini scored higher on Likert and PEMAT actionability than ChatGPT and Copilot; ChatGPT and Gemini had higher Flesch Reading Ease than Copilot, which had a higher grade level. Overall readability differed (p = 0.022), with Gemini most often sufficient. Final overall sufficiency, meeting both information quality and readability, was low across all platforms (24.7%). Conclusion Popular LLMs produced menopause responses generally consistent with the literature but often limited in reliability structure and reading level, with low overall sufficiency across all three tools. Performance varied by platform and question type. Current consumer LLMs may complement rather than substitute for clinician-guided counseling; whether they can safely substitute requires patient-centered evaluation.
Background Smartphones have become a primary channel through which older adults obtain health information, yet technology-induced resistance to change may be associated with lower engagement in health information acquisition. Objective Drawing on social influence theory and social cognitive theory, this study examines the association between smartphone-induced change resistance and older adults’ smartphone-based health information acquisition and whether this association varies according to family support and peer influence. Method Using a sequential explanatory mixed-methods approach, this study combines a survey of 218 older adults and follow-up in-depth interviews with 19 older adults, all residing in China. Results Smartphone self-efficacy was positively associated with smartphone-based health information acquisition, as were smartphone health outcome expectations. Smartphone-induced change resistance was negatively associated with smartphone-based health information acquisition. Peer influence was positively associated with both smartphone self-efficacy and smartphone health outcome expectations, whereas family support was not significantly associated with smartphone self-efficacy and showed only a marginal positive association with smartphone health outcome expectations. In the moderation analysis, the negative association between smartphone-induced change resistance and smartphone-based health information acquisition was weaker at higher levels of peer influence, whereas this association did not vary significantly across levels of family support. Conclusion This study advances research on older adults’ health information acquisition by examining smartphone-induced change resistance in relation to everyday smartphone-based health information acquisition. It shows that peer influence was more strongly related to smartphone self-efficacy than family support. For smartphone health outcome expectations, the peer-influence coefficient was numerically larger, but its difference from the family-support coefficient was not significant. The negative relationship between change resistance and health information acquisition was also weaker when peer influence was higher, whereas family support did not show the same pattern. These findings highlight the distinctive role of peer networks and may offer insights into digital health inequities.
Objective To compare four patient-facing AI assistants for chronic constipation education across expert-rated quality, objective readability, and patient-reported acceptability. Methods In this two-phase study, 15 common questions and two clinical scenarios were submitted once to ChatGPT-4o, DeepSeek V3, Qwen3, and MedGo V1 using the same standardized Chinese prompt. Six blinded experts rated 68 responses across six domains. Objective readability was assessed using Chinese character count, reading difficulty score, and recommended reading age. In Phase 2, 30 patients each posed three questions; the investigator entered each question into all four assistants in a fixed sequence, and patients rated eight acceptability domains. Friedman tests were used; significant expert and patient results underwent Holm-adjusted Wilcoxon signed-rank tests. Results Accuracy ( p = .175) and safety ( p = .068) did not differ significantly across models. Domain specificity, expert-rated readability, humanity, and actionability differed significantly. Qwen3 scored higher than the other models for domain specificity and actionability in adjusted comparisons. MedGo V1 produced the shortest responses and had the lowest mean reading difficulty and recommended reading age. Patient ratings showed a different pattern: DeepSeek V3 had the highest mean overall acceptability (4.22 ± 0.19), exceeding Qwen3 (adjusted p < .001) and MedGo V1 (adjusted p = .044), but not ChatGPT-4o (adjusted p = .053). Conclusion Expert-rated quality, text complexity, and short-term patient acceptability were not fully aligned. No single assistant consistently outperformed the others. These findings describe model-specific performance profiles but do not establish clinical effectiveness, implementation readiness, or superiority of general-purpose versus medical-specialized systems.
Background Early detection of lung nodules using low-dose computed tomography (LDCT) is essential for reducing lung cancer mortality. Although deep-learning-based segmentation models have demonstrated promising performances, centralized training across multiple medical institutions is often infeasible because of patient privacy regulations and data-sharing constraints. Heterogeneous data distribution among hospitals may also degrade the robustness of collaboratively trained models. A privacy-preserving and robust multicenter segmentation framework is required for real-world clinical deployment. Methods We developed a privacy-preserving federated framework based on a 3D Transformer-UNet. Multi-head self-attention was applied at the network bottleneck to capture long-range volumetric dependencies, whereas multi-head cross-attention was incorporated into the skip connections to regulate the fusion of encoder and decoder features. Three physically independent client workstations participated in every communication round. Each local model was evaluated using a server-side LUNA16 reference set, and the resulting nodule-level F1 scores were normalized to determine the aggregation weights. The initially aggregated model was subsequently refined using LUNA16-assisted morphological correction. Results The complete framework achieved nodule-level precision, recall, and F1 scores of 84.5%, 85.1%, and 84.8%, respectively. At the voxel level, it achieved a Dice similarity coefficient of 78.9%, an Intersection-over-Union of 65.2%, and an average surface distance of 1.36 mm. The proposed method reached the predefined convergence criterion after 45 communication rounds, compared with 85 rounds for FedAvg, corresponding to a 47.1% reduction in the estimated cumulative model-transfer payload required to reach convergence. Conclusions The proposed framework combined attention-based volumetric feature learning, F1-score-weighted aggregation, and server-side morphological correction. Under the evaluated three-client simulation, this combination improved nodule-level detection, voxel-level segmentation, and convergence efficiency while avoiding the exchange of private LDCT images. Further validation using prospective and naturally distributed multicenter clinical data is required before clinical deployment.
Objective Accurate prediction of chronic disease progression is essential for improving clinical decision-making, personalized treatment, and healthcare resource allocation. However, longitudinal healthcare data are characterized by temporal dependencies, patient-specific heterogeneity, and complex nonlinear relationships that often limit the effectiveness of conventional predictive models. To address these challenges, this study proposes a hybrid Linear Mixed Model (LMM)–XGBoost framework for predicting chronic disease progression using simulated longitudinal healthcare data. Methods A large-scale synthetic dataset comprising 10,000 patients observed over 10 follow-up visits under slow, moderate, and rapid disease progression scenarios was generated through Monte Carlo simulation, yielding 300,000 longitudinal observations. Missing Completely at Random (MCAR) and Missing at Random (MAR) mechanisms were incorporated and handled using Multiple Imputation by Chained Equations (MICE). The proposed framework integrates LMM’s ability to capture longitudinal and patient-specific effects with Extreme Gradient Boosting (XGBoost)’s nonlinear predictive capability. Its performance was compared with Generalized Estimating Equations (GEE), Random Forest (RF), Light Gradient Boosting Machine (LightGBM) and XGBoost models. Results demonstrated that the proposed Hybrid LMM–XGBoost framework achieved the best predictive performance among all competing models, yielding an Root Mean Squared Error (RMSE) of 2.84, Mean Absolute Error (MAE) of 2.00, and coefficient of determination (R²) of 0.980. For disease-stage classification, the model attained an accuracy of 0.93, precision of 0.92, recall of 0.92, F1-score of 0.92, and Area Under the Curve (AUC) of 0.981. SHapley Additive exPlanations (SHAP) analysis identified previous disease severity (DS_Lag), Body Mass Index (BMI), glucose level (GLU), and blood pressure (BP) as the most influential predictors. Monte Carlo validation (B = 1000) confirmed the model’s robustness, with a mean RMSE of 2.85 (95% CI: 2.67–3.03). Conclusion The findings demonstrate the effectiveness of integrating longitudinal statistical modeling, machine learning, and explainable artificial intelligence for chronic disease progression prediction.
