PLOS digital health最新文献

{"title":"Generative artificial intelligence-assisted medical self-care and associated factors among undergraduate students at Arsi University, South-Eastern Ethiopia.","authors":"Jibril Bashir Adem, Anas Ali Alhur","doi":"10.1371/journal.pdig.0001748","DOIUrl":"10.1371/journal.pdig.0001748","url":null,"abstract":"<p><p>The rapid expansion of social media and the availability of Generative Artificial Intelligence (GenAI) technologies like ChatGPT and Gemini have significantly increased access to personalized health information. This has led to a rise in self-directed health behaviors such as self-diagnosis and self-medication. While GenAI can support informed self-care decisions, over-reliance on these technologies may result in incorrect self-medication and delayed consultations with medical professionals if not managed appropriately. However, there is a notable lack of systematic empirical data regarding GenAI-assisted medical self-care practices, particularly in Ethiopia. This study aimed to assess the prevalence of GenAI-assisted medical self-care and its influencing factors among undergraduate students at Arsi University in South-Eastern Ethiopia. A cross-sectional study was conducted with 414 randomly selected students using a self-administered questionnaire. Data were collected through Kobo Toolbox, and descriptive analyses alongside binary logistic regression were performed to identify factors associated with GenAI-assisted self-care. The study achieved a response rate of 97.87%, with participants having a mean age of 22.2 years. Results showed that 70.05% of students engaged in GenAI-assisted medical self-care. Factors positively associated with this behavior included being a second-year student (AOR=2.98), possessing good digital health literacy (AOR=1.71), awareness of GenAI technologies (AOR=2.18), a positive attitude towards these technologies (AOR=1.63), and recent health facility visits (AOR=2.46). Conversely, good health-seeking behavior (AOR=0.61) and infrequent health facility visits (AOR=0.49) were linked to lower odds of engaging in GenAI-assisted self-care. The findings suggest a high prevalence of GenAI-assisted self-care among university students and highlight the need for guidelines and training on digital health literacy for effective and safe technology use in healthcare.</p>","PeriodicalId":74465,"journal":{"name":"PLOS digital health","volume":"5 9","pages":"e0001748"},"PeriodicalIF":7.7,"publicationDate":"2026-09-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13606965/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149042546","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Large-scale patterns of self-reported mood and sleep across the day and week on an online cognitive-training platform.","authors":"Reina A Mendoza, Andrew S Tubbs, Michael L Perlis, Fabian-Xosé Fernandez, Michael A Grandner","doi":"10.1371/journal.pdig.0001754","DOIUrl":"10.1371/journal.pdig.0001754","url":null,"abstract":"<p><p>Mood follows a well-documented circadian rhythm. Sleep timing and duration are similarly shaped by biological and social factors. However, large-scale, real-world data on how these patterns vary across demographic groups remain limited. Prior studies often rely on laboratory data or social media linguistic analysis, which lack direct self-reports. This study examines hourly, weekly, and demographic trends in mood and sleep duration reports using one of the largest online datasets of directly self-reported behavioral health: 468,092 Lumosity users (aged 20-79) who reported daily mood (-2 to +2 scale) and sleep duration (≤5-9 + hours). We assessed variation by time of day, day of week, and interactions of age, gender, and race/ethnicity using ANOVA and linear regression. Mood was higher among users reporting in the afternoon and evening, and lower from late night through morning. Mood showed a U-shaped pattern with age, consistent throughout the week, and longer sleep duration was linked to better mood reports throughout the day and week. Sleep duration varied by time of report and by day of week, with longer durations in the afternoon/evening and on weekends. Women reported longer sleep overall, Black users reported shorter sleep, and sleep duration declined from young adulthood through midlife before rising slightly among older users. Higher mood was associated with longer sleep, and the weekday-weekend mood difference was larger among users reporting longer sleep. These findings demonstrate the value of large-scale digital data for public health insight, highlighting demographic disparities in sleep and the importance of midlife as a key intervention period. Because the sample is likely skewed toward higher educational attainment and health engagement, disparity estimates may understate true population-level differences. The late-night peak in negative affect aligns with literature on nocturnal suicide risk, though this nonclinical sample cannot directly speak to clinical risk or crisis-service need.</p>","PeriodicalId":74465,"journal":{"name":"PLOS digital health","volume":"5 9","pages":"e0001754"},"PeriodicalIF":7.7,"publicationDate":"2026-09-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13606931/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149042631","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Early malaria risk screening in Nigerian minors using AutoML and cluster-based analysis of non-clinical survey data.","authors":"Ashiqur Rahman Khan, Nafisa Mahbub, Ridwan Al Aziz, Mahamudul Hassan Siddique, Zahidul Islam Sayeem","doi":"10.1371/journal.pdig.0001736","DOIUrl":"10.1371/journal.pdig.0001736","url":null,"abstract":"<p><p>Malaria remains a major public health challenge in sub-Saharan Africa, with Nigeria accounting for a substantial proportion of the global malaria burden, particularly among children under five years of age. Although rapid diagnostic tests (RDTs) enable timely screening, false-negative results and delays in confirmatory microscopy can hinder early treatment in resource-limited settings. This study developed and evaluated an automated machine learning (AutoML)-based framework for early malaria risk prediction using demographic, household, and socioeconomic data from the nationally representative Nigeria Malaria Indicator Survey (MIS) 2021. Initially, 43 candidate variables were selected and subsequently reduced to 13 statistically significant predictors through preprocessing and statistical dependency testing. Five machine learning models, namely Logistic Regression (LR), Decision Tree (DT), Support Vector Classifier (SVC), Extreme Gradient Boosting (XGBoost), and H2O AutoML, were developed and evaluated using stratified train-test splitting and 10-fold cross-validation. To reduce the influence of regional information, Agglomerative Hierarchical Clustering (AHC) was performed using the eight most important predictors after excluding regional identifiers, yielding two optimal population subgroups (Silhouette Score = 0.3939). For the overall dataset, H2O AutoML Generalized Linear Model (GLM) achieved the highest F2-score (79.38%) with a recall of 83.08%, whereas XGBoost achieved the highest test accuracy (75.54%). Cluster-specific analyses demonstrated that H2O GLM provided the best recall-oriented performance in Cluster 1 (F2-score = 83.31%), while XGBoost performed best in Cluster 2 (F2-score = 83.25%). These findings demonstrate that H2O AutoML provides a robust and automated framework for malaria risk prediction, while cluster-specific modeling further improves predictive performance by accounting for population heterogeneity. The proposed framework has the potential to support early malaria risk assessment and complement conventional diagnostic strategies in resource-limited public health settings.</p>","PeriodicalId":74465,"journal":{"name":"PLOS digital health","volume":"5 9","pages":"e0001736"},"PeriodicalIF":7.7,"publicationDate":"2026-09-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13600517/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149032306","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Large language models for pneumonia detection in radiology reports via text analysis.","authors":"Weisheng Chen, Xinya Li, Zhongkai Qu, Ting Liu, Bo Chen, Liying Huang, Ningxia He, Zhigang Wang, Jun Lyu","doi":"10.1371/journal.pdig.0001104","DOIUrl":"10.1371/journal.pdig.0001104","url":null,"abstract":"<p><p>Pneumonia is a common infection in critically ill patients with poor outcomes, and its identification relies heavily on radiological evidence. Most MIMIC-based pneumonia studies rely on structured codes rather than radiological text, which may limit the precision of pneumonia-related case identification. This study evaluated large language models (LLMs) and retrieval-augmented generation (RAG) frameworks for automated detection of radiographic evidence of pneumonia from radiology reports. Radiological reports from the MIMIC-IV database were analyzed using pre-trained models in an inference-only setting without additional training, fine-tuning, or dataset splitting. Model performance was evaluated using accuracy, precision, sensitivity, specificity, and F1-score, with point estimates calculated from the full dataset and corresponding 95% confidence intervals estimated using 1,000 patient-level clustered bootstrap replicates. Independent evaluation was subsequently performed using a GZ dataset from a tertiary hospital in Guangzhou. RAG generally improved model performance, with gemma3-27b achieving the highest accuracy of 96.4% on the MIMIC-IV dataset and maintaining an accuracy of 87.6% on the GZ dataset. Model performance did not increase linearly with parameter size, suggesting that task-specific model suitability may be more important than parameter scale for this application. Although larger models required longer inference times, their increased computational costs did not consistently translate into proportional performance gains. Overall, LLMs demonstrated high accuracy in interpreting radiology reports for detecting radiographic evidence of pneumonia. The RAG architecture further enhanced robustness and generalizability through knowledge augmentation, providing a methodological basis for automated radiology report analysis and large-scale database curation.</p>","PeriodicalId":74465,"journal":{"name":"PLOS digital health","volume":"5 9","pages":"e0001104"},"PeriodicalIF":7.7,"publicationDate":"2026-09-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13600486/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149032322","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"PREFER-IT: A transdisciplinary co-created framework to realise inclusive health AI.","authors":"Patrícia Pita Ferreira, Sara Soriano Longarón, Wiam Bouisaghouane, Jetse Goris, Anne H Hoekman, Balázs Markos, Benjamin Maus, Giorgia Pozzi, Hadi Hasan, Indre Kalinauskaité, Jonáh Stunt, Joosje D Kist, Judith van der Elst, Katell Maguet, Liv Ziegfeld, Maarten Cuypers, Megan Milota, Michelle Habets, Sara Colombo, Špela Petrič, Steff Groefsema, Steven Warmelink, Elja Daae, Giovanni Briganti, Ildikó Vajda, Matias Valdenegro-Toro, Matthias Braun, Pieter Jeekel, Simone Goosen, Alex Schepel, Laxmie Ester, Riane Kuzee, Sophie de Klerk, Claudine Lamoth, Lisa Ballard, Mirjam Plantinga","doi":"10.1371/journal.pdig.0001096","DOIUrl":"10.1371/journal.pdig.0001096","url":null,"abstract":"<p><p>Artificial intelligence (AI) in healthcare holds transformative potential but risks exacerbating existing health disparities if inclusivity is not explicitly accounted for. This study addresses the disconnected discussions on inclusive health AI by developing a comprehensive framework, PREFER-IT. This framework is based on the outcomes of a five-day transdisciplinary co-creation workshop that involved 37 experts from diverse backgrounds, including healthcare, ethics, law, social sciences, AI, and patient advocacy, held in the Netherlands. For this workshop, we used design thinking and participatory methodologies to develop a framework for realising inclusive health AI. We identified three key challenges for realising inclusive health AI: integrating the lived experiences and stakeholder voices across the AI lifecycle, designing data collection practices that promote fairness and prevent inequalities, and fostering regulatory frameworks to uphold human rights and promote inclusivity. The analysis of participants' perspectives informed the development of eight key thematic clusters of PREFER-IT: Participatory and co-design approaches (P), Representative and diverse data (R), Education and digital literacy (E), Fairness (F), Ethical and legal accountability (E), Real-world validation and feedback (R), Inclusive communication (I), and Technical interoperability (T). These elements were mapped across structural layers of AI (humans, data, process, system and governance) and the AI lifecycle to guide inclusive design, development, validation, implementation, monitoring, and governance. This framework fosters stakeholder engagement and systemic change, positioning inclusion as a guiding principle in practice. PREFER-IT offers a practical and conceptual contribution for how to include ethical, legal, and societal aspects when aiming to foster responsible and inclusive AI in healthcare.</p>","PeriodicalId":74465,"journal":{"name":"PLOS digital health","volume":"5 9","pages":"e0001096"},"PeriodicalIF":7.7,"publicationDate":"2026-09-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13600515/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149032341","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Large language models and artificial intelligence for generating, simplifying, and enhancing outpatient clinic letters: A systematic review.","authors":"Rocco Sheldon, Morya Wadodkar, Andrew Gan, Rachael Yip, Yimeng Zhang, Jyoti Baharani","doi":"10.1371/journal.pdig.0001745","DOIUrl":"10.1371/journal.pdig.0001745","url":null,"abstract":"<p><p>Outpatient clinic letters are a cornerstone of clinical communication but frequently exceed recommended reading levels, limiting patient comprehension and engagement. Large Language Models (LLMs) and Artificial Intelligence (AI) have been proposed as scalable tools for generating and simplifying these documents, but the evidence regarding their clinical accuracy, safety, and effectiveness remains fragmented. We conducted a systematic review of studies evaluating AI or LLMs for generating, simplifying, or enhancing outpatient clinic letters. Five databases (PubMed, EMBASE, Web of Science, CENTRAL, and CINAHL) were searched from inception to 1 November 2025. Two independent reviewers screened studies, extracted data, assessed quality using the Mixed Methods Appraisal Tool (MMAT), and certainty using the GRADE appraisal tool. Seven studies were included, comprising two studies using real-world clinic data and five using synthetic or hypothetical scenarios. Findings regarding readability were mixed; while some AI models improved readability scores compared to human-authored letters, most AI-generated content still failed to meet the recommended US Grade 6 reading level. The two studies that measured patient-reported outcomes reported high satisfaction and comprehension with AI-simplified letters, but these studies relied on subjective measures of understanding or lacked a human-authored control group. Clinical accuracy varied substantially, with information fidelity ranging from 10% to 100%. Risks of 'hallucinations', inappropriate medical advice, and tactless patient descriptions were documented. One study reported a ten-fold reduction in drafting time compared to human dictation. Overall certainty of evidence was rated very low across all primary outcomes due to methodological limitations, small sample sizes, and reliance on hypothetical scenarios. The current evidence supporting AI and LLMs use for outpatient clinic letters is preliminary and limited, providing insufficient evidence to support routine clinical implementation beyond supervised experimental and quality-improvement settings. Concerns regarding accuracy, safety, and generalisability remain. Further real-world patient-centred research is required before widespread adoption. PROSPERO registration: CRD420251181303.</p>","PeriodicalId":74465,"journal":{"name":"PLOS digital health","volume":"5 9","pages":"e0001745"},"PeriodicalIF":7.7,"publicationDate":"2026-09-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13596789/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149026876","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Development and external evaluation of an interpretable machine-learning model for early prediction of organ failure in higher-risk acute pancreatitis patients: A multicentre cohort study.","authors":"Di Wu, Wenhao Cai, Chunmei Chen, Minting Chen, Yang Lv, Yilin Huang, Anthony Evans, Juan Lin, Diane Latawiec, Arjun Kattakayam, Rajarshi Mukherjee, Wei Huang, Qing Xia, Jie Xiao, Chunqiu Su, Jie Peng, Kuirong Jiang, Robert Sutton","doi":"10.1371/journal.pdig.0001735","DOIUrl":"10.1371/journal.pdig.0001735","url":null,"abstract":"<p><p>Organ failure (OF) is the most important determinant of prognosis in acute pancreatitis (AP) and its duration defines disease severity. Early identification of individuals at high risk of developing OF is crucial. We aimed to develop machine-learning models, benchmarked against a feedforward multilayer-perceptron (MLP) neural network, to predict new-onset OF at admission using static admission-day variables. In this study, data were extracted from MIMIC-IV (development cohort) and a multicentre AP cohort from two Chinese tertiary teaching hospitals (evaluation cohort), which excluded mild AP and therefore comprised patients requiring ICU-level or high-dependency care. Features were selected with Boruta and LASSO. Five machine-learning models and one multilayer-perceptron benchmark were developed in MIMIC-IV and externally evaluated in the Xiangya cohort without updating. SHapley Additive exPlanations (SHAP) were used to visualize decision-making patterns and individual prediction interpretations. The best-performing model was deployed as an interactive, web-based tool. We found that in the discovery cohort, 341 (23.9%) of 1429 AP patients developed new-onset OF within 28 days of admission while 45 of 216 patients (20.8%) in the validation cohort developed OF. Boruta and LASSO algorithms identified six key predictors including blood urea nitrogen, platelets, triglyceride-glucose index, albumin, white blood cells, and partial thromboplastin time, which were used to construct the ML and MLP models. XGBoost gave the best discrimination on external evaluation (AUC 0.837, 95% CI 0.771-0.902). Logistic recalibration improved calibration, and the recalibrated XGBoost model was implemented as a web-based research prototype (Decent app). In conclusion, XGBoost predicted new-onset OF with good discrimination in a severity-enriched AP cohort, and SHAP made individual predictions interpretable. Model selection, threshold selection and recalibration all used the evaluation cohort, so the deployed model still requires an independent cohort. Prospective usability and clinical-impact studies are needed before clinical use.</p>","PeriodicalId":74465,"journal":{"name":"PLOS digital health","volume":"5 9","pages":"e0001735"},"PeriodicalIF":7.7,"publicationDate":"2026-09-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13596832/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149026795","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"User needs and design opportunities for a conversational agent for tuberculosis treatment: A mixed-methods study.","authors":"Joon Sang Baek, Sehwa Choi, Seojin Sung, Vasuki Rajaguru, Youngmok Park","doi":"10.1371/journal.pdig.0001720","DOIUrl":"10.1371/journal.pdig.0001720","url":null,"abstract":"<p><p>Tuberculosis (TB) remains a major global health challenge requiring prolonged treatment that is often complicated by adverse drug reactions (ADRs), stigma, and poor treatment adherence. Although digital health interventions show promise in supporting TB care, patients' needs, expectations, and design requirements for conversational agents remain poorly understood. This study identified challenges experienced by healthcare professionals and patients with TB and explored opportunities for designing a user-centered conversational agent for TB care. A mixed-methods approach was adopted, comprising surveys of 107 healthcare professionals and 31 patients with TB, followed by in-depth interviews with 10 patients. The survey assessed treatment challenges, adherence barriers, and digital health needs. The interview data explored participants' treatment experiences and expectations for conversational agent-based support. Quantitative data were analyzed using descriptive statistics, and qualitative data were analyzed using inductive thematic analysis. Healthcare professionals reported ADR (25.1%), multidrug-resistant TB (17.1%), drug interaction management (15.6%), and low treatment adherence (15.6%) as key challenges. Patients' needs focused on ADR consultation (35%), TB information (21.7%), and diagnostics (20%). The majority wanted to use mHealth (62.1%), with specific interests in side effect guidance (20.4%), self-diagnosis (13%), and treatment information (12.5%). The interviews revealed patients' multifaceted needs related to TB treatment, including consistent communication with healthcare providers, effective ADR management, accurate and accessible TB information, social and psychological support during treatment, usable and accessible digital health solutions, and affordable services. These findings identified key design opportunities, including TB awareness, timely and personalized information, treatment adherence support, transparent treatment progress, psychosocial support, accessible digital services, and affordability. Healthcare professionals and patients reported unmet needs that conversational agents could potentially address. These findings provide user-centered guidance for developing conversational agents and contribute to evidence on digital health interventions for TB treatment support.</p>","PeriodicalId":74465,"journal":{"name":"PLOS digital health","volume":"5 9","pages":"e0001720"},"PeriodicalIF":7.7,"publicationDate":"2026-09-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13596825/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149026848","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Mental health monitoring in the workplace using artificial intelligence: A scoping review.","authors":"Kalee Jack, Julie Tian, Viktoriia Kurkova, Setayesh Modanloo, Liz Dennett, Jasmine Noble, Andrew Greenshaw, Jake Hayward","doi":"10.1371/journal.pdig.0001743","DOIUrl":"10.1371/journal.pdig.0001743","url":null,"abstract":"<p><p>Workplace mental health issues are a prevalent concern. A potential solution is remote health monitoring. In this scoping review, we map evidence on artificial intelligence (AI) for workplace mental health monitoring. MEDLINE, PsycINFO, Scopus, Embase, and Web of Science were searched until March 2024. Studies that evaluated AI for monitoring of a common workplace mental health concern in an actual workplace setting were included. Thirteen studies were included; most (8/13, 62%) were conducted in an office environment, and 4/13 (31%) were in healthcare. Most studies monitored stress (10/13, 77%). The most common method was wearable sensors (8/13, 62%). Studies monitored a wide variety of parameters; the most common were physical activity (7/13, 54%), heart rate or heart rate variability (7/13, 54%). A wide variety of participant-report surveys, including validated (7/13, 54%) and non-validated (6/13, 46%) surveys, were used to measure mental health outcomes. Studies employed diverse AI methods, including supervised, semi-supervised and other (Markov chain and Monte Carlo). The most common AI models used were support vector machines (5/13, 38%), decision trees (5/13, 38%), naive Bayes (4/13, 31%), and random forest (4/13, 38%). Model accuracy ranged from 62%-99%. Future work should identify standard metrics to enable cross-comparability and use larger, high-quality datasets for enriched training, advanced optimization techniques, and careful model tuning.</p>","PeriodicalId":74465,"journal":{"name":"PLOS digital health","volume":"5 9","pages":"e0001743"},"PeriodicalIF":7.7,"publicationDate":"2026-09-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13592608/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149019051","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"The economics of accuracy for medical reasoning with large language models.","authors":"Kiran Bhattacharyya, Sreeram Kamabattula","doi":"10.1371/journal.pdig.0001182","DOIUrl":"10.1371/journal.pdig.0001182","url":null,"abstract":"<p><p>Deploying large language models (LLMs) in clinical settings is limited by security, reliability, latency, and accessibility concerns that favor smaller, on-device or on-premise models. However, these smaller models may struggle to meet accuracy requirements. While fine-tuning and retrieval-augmented generation (RAG) can improve domain-specific accuracy, these methods require additional labeled data, technical skill, and infrastructure. In contrast, test-time scaling-allocating extra token-budget during inference-offers a training-free alternative to increasing accuracy. However, the trade-offs between these strategies and their interaction with model size remain poorly understood for medical reasoning. To address this gap, we compare three approaches-test-time scaling, fine-tuning, and context grounding-using the Gemma and MedGemma family of LLMs (Gemma-3 1B, Gemma-3 4B, Gemma-3 27B, MedGemma-4B, and MedGemma 27B) and evaluate these systems across common biomedical question-answering (QA) datasets and a set of recently released medical exam questions with the performance of practicing clinicians available for comparison. We test baseline prompts (direct answer, Chain-of-Thought, and self-consistency) while introducing a new prompting method we call \"prompt-chaining for continuous reflection\" (PCCR) that forces inference time minimum token-generation budgets. We assess accuracy and tokens-generated, allowing us to investigate the accuracy-efficiency trade-offs across prompting, context-grounding, fine-tuning, and model scales. We discover equivalency point configurations where a smaller model's accuracy falls within one 95% confidence interval of a larger model's (typically within 1-4 percentage points) reached through increased reasoning budgets, context-grounding, or fine-tuning. Specific effects are apparent and statistically supported: the benefit of medical fine-tuning grew from +4.6 to +15.7 percentage points (non-overlapping 95% CIs) when paired with self-consistency, and enforced extended reasoning raised MedGemma 27B accuracy from 58.1% to 80.1% (p < 10-5) in the absence of context. We also identify an \"overthinking\" inflection: when high-quality context is available, extended reasoning beyond roughly 128-256 tokens degrades accuracy by 7-13 percentage points. Using these empirical results, we formulate a general framework with equations to balance cost-benefit trade-offs when engineering LLM-based systems for medical reasoning and QA. We recommend generalizable configurations, designs, and patterns to achieve accuracy and efficiency objectives for example use-cases relevant to healthcare organizations.</p>","PeriodicalId":74465,"journal":{"name":"PLOS digital health","volume":"5 9","pages":"e0001182"},"PeriodicalIF":7.7,"publicationDate":"2026-09-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13588378/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149003043","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}