PodGPT is a language model trained on thousands of hours of medical podcasts, capturing expert insights often missing in text-only models
It outperforms standard open-source models by 2-4% on medical question-answering benchmarks and science-related tests, demonstrating a better understanding of complex medical knowledge and interdisciplinary topics
Using Retrieval Augmented Generation (RAG), PodGPT stays updated with recent research without costly retraining
PodGPT can be customized for medical specialties to aid research and clinical tasks but still faces challenges like biases and occasional errors
Varus malalignment is common among persons with knee OA and it increases load across the medial knee compartment, the most affected compartment.
Various nonsurgical attempts to realign the knee such as knee braces and modified shoes have not been successful in reducing pain substantially. Their treatment effects are small. Major load reducing surgeries have not been examined in RCTs. The trial discussed tested whether cartilage loss was reduced in knees of patients undergoing high tibial osteotomy. The trial showed that this operation both slowed cartilage loss and substantially reduced pain.
Trials testing the efficacy of surgical treatments need to achieve a large effect size compared with nonsurgical treatments and therefore the needed sample size is often more modest than trials of nonsurgical treatments.
Chronic pain prevalence is highest in rural areas but has increased across all regions since 1998, with suburban areas increasing sharply since 2010
Increases affect all demographics, driven by factors like obesity, mental health, and social changes; a temporary dip during COVID-19 may reflect policy protections and remote work
The study used longitudinal HRS data converted to repeated cross sections to analyze geographic pain trends over time
Generalized estimating equations accounted for repeated measures, controlling for confounders; limitations include self-re[ported pain and sociocultural reporting biases
The findings highlight the need for expanded research into structural and environmental drivers of pain and targeted policy interventions to address geographic and demographic disparities
Mobile health (mHealth) technologies can be used to collect subjective and objective data on symptoms, behaviors, and other information in real-time.
This data can be used to answer complex research questions, such as exploring how to predict the beginning of a pain flare, using data on symptoms, experiences, sleep, and physical activity.
There are multiple ways to characterize dynamic patterns of change across individual participants, which can range in complexity from a within-person mean or standard deviation to an algorithm-derived beginning of a symptom flare.
To select the best measures and methods for scoring data, it is important to carefully consider your research question and relevant past research.
Collecting high-quality mHealth data requires high engagement from participants, especially for self-reported outcomes.
For long-term follow-up analyses of RCTs, need to consider adjustment to account for loss to follow-up and competing events. Otherwise the intention to treat analysis may give biased results.
A framework for variable selection includes identification of “core variables” (RCT treatment assignment, outcome), a literature review to identify potential confounders, an expert panel to review identified variables, suggest additional variables, and determine whether they are causes and/or consequences of the treatment assignment and outcome(s) of interest.
Models to predict the risk of Total Knee Replacement (TKR) surgery have traditionally used standard statistical approaches (e.g., logistic regression with stepwise predictor selection), which typically assume linear relationships and have limited ability to incorporate interactions.
Newer machine learning approaches offer improved predictor selection, allowance for non-linear associations, and greater potential to model interacting predictors, but they are also more prone to overfitting.
All risk prediction models—both traditional statistical and machine learning—require rigorous evaluation of prediction accuracy (discrimination, calibration, positive predictive value, etc.) on independent data not used for model development to avoid overly optimistic performance estimates.
Blanco et al. (RMD Open, 2026) applied six machine learning methods to develop TKR risk models using demographic, clinical, and genetic predictors, reporting AUCs between 0.71 and 0.75 with good calibration and reasonable positive predictive values.
Michelle Yau, PhD, MPH
Assistant Professor of Medicine, Hebrew SeniorLife and Harvard Medical School
GWAS are powerful for identifying genetic variants in complex traits but often require large sample sizes; network-based GWAS can enhance detection by leveraging functional relationships between genes.
Integrating cell- and tissue-specific molecular data (e.g., gene expression) helps validate the biological relevance of gene modules identified through network-based approaches.
Because the underlying networks used in the network-based GWAS are not disease-specific, they may miss RA-specific interactions while incorporating broader biological connections, reducing the specificity of the findings. Consequently, whether CCP+/RF+ and CCP−/RF− cases lie on the same disease spectrum or represent distinct biological entities remains unresolved, although the approach remains valuable for studying traits with limited sample sizes.
