Juan M. Banda

While large language models (LLMs) achieve near-perfect scores on medical licensing exams, these evaluations inadequately reflect the complexity and diversity of real-world clinical practice. Here we introduce MedHELM, an extensible evaluation framework with three contributions. First, a clinician-validated taxonomy organizing medical AI applications into five categories that mirror real clinical tasks—clinical decision support (diagnostic decisions, treatment planning), clinical note generation (visit documentation, procedure reports), patient communication (education materials, care instructions), medical research (literature analysis, clinical data analysis) and administration (scheduling, workflow coordination). These encompass 22 subcategories and 121 specific tasks reflecting daily medical practice. Second, a comprehensive benchmark suite of 37 evaluations covering all subcategories. Third, systematic comparison of nine frontier LLMs—Claude 3.5 Sonnet, Claude 3.7 Sonnet, DeepSeek R1, Gemini 1.5 Pro, Gemini 2.0 Flash, GPT-4o, GPT-4o mini, Llama 3.3 and o3-mini—using an automated LLM-jury evaluation method. Our LLM-jury uses multiple AI evaluators to assess model outputs against expert-defined criteria. Advanced reasoning models (DeepSeek R1, o3-mini) demonstrated superior performance with win rates of 66%, although Claude 3.5 Sonnet achieved comparable results at 15% lower computational cost. These results not only highlight current model capabilities but also demonstrate how MedHELM could enable evidence-based selection of medical AI systems for healthcare applications.

Electronic phenotyping involves a detailed analysis of both structured and unstructured data, employing rule-based methods, machine learning, natural language processing, and hybrid approaches. Currently, the development of accurate phenotype definitions demands extensive literature reviews and clinical experts, rendering the process time-consuming and inherently unscalable. Large language models offer a promising avenue for automating phenotype definition extraction but come with significant drawbacks, including reliability issues, the tendency to generate non-factual data (“hallucinations”), misleading results, and potential harm. To address these challenges, our study embarked on two key objectives: (1) defining a standard evaluation set to ensure large language models outputs are both useful and reliable and (2) evaluating various prompting approaches to extract phenotype definitions from large language models, assessing them with our established evaluation task. Our findings reveal promising results that still require human evaluation and validation for this task. However, enhanced phenotype extraction is possible, reducing the amount of time spent in literature review and evaluation.

The COVID-19 pandemic has presented significant challenges to the healthcare industry and society as a whole. With the rapid development of COVID-19 vaccines, social media platforms have become a popular medium for discussions on vaccine-related topics. Identifying vaccine-related tweets and analyzing them can provide valuable insights for public health research-ers and policymakers. However, manual annotation of a large number of tweets is time-consuming and expensive. In this study, we evaluate the usage of Large Language Models, in this case GPT-4 (March 23 version), and weak supervision, to identify COVID-19 vaccine-related tweets, with the purpose of comparing performance against human annotators. We leveraged a manu-ally curated gold-standard dataset and used GPT-4 to provide labels without any additional fine-tuning or instructing, in a single-shot mode (no additional prompting).

Objective: Biases within probabilistic electronic phenotyping algorithms are largely unexplored. In this work, we characterize differences in subgroup performance of phenotyping algorithms for Alzheimer’s disease and related dementias (ADRD) in older adults.Materials and methods: We created an experimental framework to characterize the performance of probabilistic phenotyping algorithms under different racial distributions allowing us to identify which algorithms may have differential performance, by how much, and under what conditions. We relied on rule-based phenotype definitions as reference to evaluate probabilistic phenotype algorithms created using the Automated PHenotype Routine for Observational Definition, Identification, Training and Evaluation framework.Results: We demonstrate that some algorithms have performance variations anywhere from 3% to 30% for different populations, even when not using race as an input variable. We show that while performance differences in subgroups are not present for all phenotypes, they do affect some phenotypes and groups more disproportionately than others.Discussion: Our analysis establishes the need for a robust evaluation framework for subgroup differences. The underlying patient populations for the algorithms showing subgroup performance differences have great variance between model features when compared with the phenotypes with little to no differences.Conclusion: We have created a framework to identify systematic differences in the performance of probabilistic phenotyping algorithms specifically in the context of ADRD as a use case. Differences in subgroup performance of probabilistic phenotyping algorithms are not widespread nor do they occur consistently. This highlights the great need for careful ongoing monitoring to evaluate, measure, and try to mitigate such differences.

This study presents the outcomes of the shared task competition BioCreative VII (Task 3) focusing on the extraction of medication names from a Twitter user’s publicly available tweets (the user’s ‘timeline’). In general, detecting health-related tweets is notoriously challenging for natural language processing tools. The main challenge, aside from the informality of the language used, is that people tweet about any and all topics, and most of their tweets are not related to health. Thus, finding those tweets in a user’s timeline that mention specific health-related concepts such as medications requires addressing extreme imbalance. Task 3 called for detecting tweets in a user’s timeline that mentions a medication name and, for each detected mention, extracting its span. The organizers made available a corpus consisting of 182 049 tweets publicly posted by 212 Twitter users with all medication mentions manually annotated. The corpus exhibits the natural distribution of positive tweets, with only 442 tweets (0.2%) mentioning a medication. This task was an opportunity for participants to evaluate methods that are robust to class imbalance beyond the simple lexical match. A total of 65 teams registered, and 16 teams submitted a system run. This study summarizes the corpus created by the organizers and the approaches taken by the participating teams for this challenge. The corpus is freely available at https://biocreative.bioinformatics.udel.edu/tasks/biocreative-vii/track-3/. The methods and the results of the competing systems are analyzed with a focus on the approaches taken for learning from class-imbalanced data.

Objective: Biases within probabilistic electronic phenotyping algorithms are largely unexplored. In this work, we characterize differences in sub-group performance of phenotyping algorithms for Alzheimer’s Disease and Related Dementias (ADRD) in older adults. Materials and methods: We created an experimental framework to characterize the performance of probabilistic phenotyping algorithms under different racial distributions allowing us to identify which algorithms may have differential performance, by how much, and under what conditions. We relied on rule-based phenotype definitions as reference to evaluate probabilistic phenotype algorithms created using the Automated PHenotype Routine for Observational Definition, Identification, Training and Evaluation (APHRODITE) framework. Results: We demonstrate that some algorithms have performance variations anywhere from 3 to 30% for different populations, even when not using race as an input variable. We show that while performance differences in subgroups are not present for all phenotypes, they do affect some phenotypes and groups more disproportionately than others. Discussion: Our analysis establishes the need for a robust evaluation framework for subgroup differences. The underlying patient populations for the algorithms showing subgroup performance differences have great variance between model features when compared to the phenotypes with little to no differences. Conclusion: We have created a framework to identify systematic differences in the performance of probabilistic phenotyping algorithms specifically in the context of ADRD as a use case. Differences in subgroup performance of probabilistic phenotyping algorithms are not widespread nor do they occur consistently. This highlights the great need for careful ongoing monitoring to evaluate, measure, and try to mitigate such differences.

Supervised learning algorithms are heavily reliant on annotated datasets to train machine learning models. However, the curation of the annotated datasets is laborious and time consuming due to the manual effort involved and has become a huge bottleneck in supervised learning. In this work, we apply the theory of noisy learning to generate weak supervision signals instead of manual annotation. We curate a noisy labeled dataset using a labeling heuristic to identify epidemic related tweets. We evaluated the performance using a large epidemic corpus and our results demonstrate that models trained with noisy data in a class imbalanced and multi-classification weak supervision setting achieved performance greater than 90%.

Colombia announced the first case of severe acute respiratory syndrome coronavirus 2 on March 6, 2020. Since then, the country has reported a total of 5,002,387 cases and 127,258 deaths as of October 31, 2021. The aggressive transmission dynamics of SARS-CoV-2 motivate an investigation of COVID-19 at the national and regional levels in Colombia. We utilize the case incidence and mortality data to estimate the transmission potential and generate short-term forecasts of the COVID-19 pandemic to inform the public health policies using previously validated mathematical models. The analysis is augmented by the examination of geographic heterogeneity of COVID-19 at the departmental level along with the investigation of mobility and social media trends. Overall, the national and regional reproduction numbers show sustained disease transmission during the early phase of the pandemic, exhibiting sub-exponential growth dynamics. Whereas the most recent estimates of reproduction number indicate disease containment, with R_t<1.0 as of October 31, 2021. On the forecasting front, the sub-epidemic model performs best at capturing the 30-day ahead COVID-19 trajectory compared to the Richards and generalized logistic growth model. Nevertheless, the spatial variability in the incidence rate patterns across different departments can be grouped into four distinct clusters. As the case incidence surged in July 2020, an increase in mobility patterns was also observed. On the contrary, a spike in the number of tweets indicating the stay-at-home orders was observed in November 2020 when the case incidence had already plateaued, indicating the pandemic fatigue in the country.

Purpose: Routinely collected real world data (RWD) have great utility in aiding the novel coronavirus disease (COVID-19) pandemic response. Here we present the international Observational Health Data Sciences and Informatics (OHDSI) Characterizing Health Associated Risks and Your Baseline Disease In SARS-COV-2 (CHARYBDIS) framework for standardisation and analysis of COVID-19 RWD. Patients and Methods: We conducted a descriptive retrospective database study using a federated network of data partners in the United States, Europe (the Netherlands, Spain, the UK, Germany, France and Italy) and Asia (South Korea and China). The study protocol and analytical package were released on 11th June 2020 and are iteratively updated via GitHub. We identified three non-mutually exclusive cohorts of 4,537,153 individuals with a clinical COVID-19 diagnosis or positive test, 886,193 hospitalized with COVID-19, and 113,627 hospitalized with COVID-19 requiring intensive services. Results: We aggregated over 22,000 unique characteristics describing patients with COVID-19. All comorbidities, symptoms, medications, and outcomes are described by cohort in aggregate counts and are readily available online. Globally, we observed similarities in the USA and Europe: more women diagnosed than men but more men hospitalized than women, most diagnosed cases between 25 and 60 years of age versus most hospitalized cases between 60 and 80 years of age. South Korea differed with more women than men hospitalized. Common comorbidities included type 2 diabetes, hypertension, chronic kidney disease and heart disease. Common presenting symptoms were dyspnea, cough and fever. Symptom data availability was more common in hospitalized cohorts than diagnosed. Conclusion: We constructed a global, multi-centre view to describe trends in COVID-19 progression, management and evolution over time. By characterising baseline variability in patients and geography, our work provides critical context that may otherwise be misconstrued as data quality issues. This is important as we perform studies on adverse events of special interest in COVID-19 vaccine surveillance.