Eric Karl Oermann

Krithik Vishwanath, Mrigayu Ghosh, Anton Alyakin et al.

Specialized clinical AI assistants are rapidly entering medical practice, often framed as safer or more reliable than general-purpose large language models (LLMs). Yet, unlike frontier models, these clinical tools are rarely subjected to independent, quantitative evaluation, creating a critical evidence gap despite their growing influence on diagnosis, triage, and guideline interpretation. We assessed two widely deployed clinical AI systems (OpenEvidence and UpToDate Expert AI) against three state-of-the-art generalist LLMs (GPT-5, Gemini 3 Pro, and Claude Sonnet 4.5) using a 1,000-item mini-benchmark combining MedQA (medical knowledge) and HealthBench (clinician-alignment) tasks. Generalist models consistently outperformed clinical tools, with GPT-5 achieving the highest scores, while OpenEvidence and UpToDate demonstrated deficits in completeness, communication quality, context awareness, and systems-based safety reasoning. These findings reveal that tools marketed for clinical decision support may often lag behind frontier LLMs, underscoring the urgent need for transparent, independent evaluation before deployment in patient-facing workflows.

Hongyi Zheng, Yixin Zhu, Lavender Yao Jiang et al.

Recent advances in large language models have led to renewed interest in natural language processing in healthcare using the free text of clinical notes. One distinguishing characteristic of clinical notes is their long time span over multiple long documents. The unique structure of clinical notes creates a new design choice: when the context length for a language model predictor is limited, which part of clinical notes should we choose as the input? Existing studies either choose the inputs with domain knowledge or simply truncate them. We propose a framework to analyze the sections with high predictive power. Using MIMIC-III, we show that: 1) predictive power distribution is different between nursing notes and discharge notes and 2) combining different types of notes could improve performance when the context length is large. Our findings suggest that a carefully selected sampling function could enable more efficient information extraction from clinical notes.

Krithik Vishwanath, Anton Alyakin, Mrigayu Ghosh et al.

The Congress of Neurological Surgeons Self-Assessment for Neurological Surgeons (CNS-SANS) questions are widely used by neurosurgical residents to prepare for written board examinations. Recently, these questions have also served as benchmarks for evaluating large language models' (LLMs) neurosurgical knowledge. This study aims to assess the performance of state-of-the-art LLMs on neurosurgery board-like questions and to evaluate their robustness to the inclusion of distractor statements. A comprehensive evaluation was conducted using 28 large language models. These models were tested on 2,904 neurosurgery board examination questions derived from the CNS-SANS. Additionally, the study introduced a distraction framework to assess the fragility of these models. The framework incorporated simple, irrelevant distractor statements containing polysemous words with clinical meanings used in non-clinical contexts to determine the extent to which such distractions degrade model performance on standard medical benchmarks. 6 of the 28 tested LLMs achieved board-passing outcomes, with the top-performing models scoring over 15.7% above the passing threshold. When exposed to distractions, accuracy across various model architectures was significantly reduced-by as much as 20.4%-with one model failing that had previously passed. Both general-purpose and medical open-source models experienced greater performance declines compared to proprietary variants when subjected to the added distractors. While current LLMs demonstrate an impressive ability to answer neurosurgery board-like exam questions, their performance is markedly vulnerable to extraneous, distracting information. These findings underscore the critical need for developing novel mitigation strategies aimed at bolstering LLM resilience against in-text distractions, particularly for safe and effective clinical deployment.

Yanbing Chen, Ruilin Wang, Zihao Yang et al.

Packing and shuffling tokens is a common practice in training auto-regressive language models (LMs) to prevent overfitting and improve efficiency. Typically documents are concatenated to chunks of maximum sequence length (MSL) and then shuffled. However setting the atom size, the length for each data chunk accompanied by random shuffling, to MSL may lead to contextual incoherence due to tokens from different documents being packed into the same chunk. An alternative approach is to utilize padding, another common data packing strategy, to avoid contextual incoherence by only including one document in each shuffled chunk. To optimize both packing strategies (concatenation vs padding), we investigated the optimal atom size for shuffling and compared their performance and efficiency. We found that matching atom size to MSL optimizes performance for both packing methods (concatenation and padding), and padding yields lower final perplexity (higher performance) than concatenation at the cost of more training steps and lower compute efficiency. This trade-off informs the choice of packing methods in training language models.

Salman Rahman, Lavender Yao Jiang, Saadia Gabriel et al.

Advances in large language models (LLMs) provide new opportunities in healthcare for improved patient care, clinical decision-making, and enhancement of physician and administrator workflows. However, the potential of these models importantly depends on their ability to generalize effectively across clinical environments and populations, a challenge often underestimated in early development. To better understand reasons for these challenges and inform mitigation approaches, we evaluated ClinicLLM, an LLM trained on [HOSPITAL]'s clinical notes, analyzing its performance on 30-day all-cause readmission prediction focusing on variability across hospitals and patient characteristics. We found poorer generalization particularly in hospitals with fewer samples, among patients with government and unspecified insurance, the elderly, and those with high comorbidities. To understand reasons for lack of generalization, we investigated sample sizes for fine-tuning, note content (number of words per note), patient characteristics (comorbidity level, age, insurance type, borough), and health system aspects (hospital, all-cause 30-day readmission, and mortality rates). We used descriptive statistics and supervised classification to identify features. We found that, along with sample size, patient age, number of comorbidities, and the number of words in notes are all important factors related to generalization. Finally, we compared local fine-tuning (hospital specific), instance-based augmented fine-tuning and cluster-based fine-tuning for improving generalization. Among these, local fine-tuning proved most effective, increasing AUC by 0.25% to 11.74% (most helpful in settings with limited data). Overall, this study provides new insights for enhancing the deployment of large language models in the societally important domain of healthcare, and improving their performance for broader populations.

Krithik Vishwanath, Anton Alyakin, Daniel Alexander Alber et al.

Large language models (LLMs) have the potential to transform medicine, but real-world clinical scenarios contain extraneous information that can hinder performance. The rise of assistive technologies like ambient dictation, which automatically generates draft notes from live patient encounters, has the potential to introduce additional noise making it crucial to assess the ability of LLM's to filter relevant data. To investigate this, we developed MedDistractQA, a benchmark using USMLE-style questions embedded with simulated real-world distractions. Our findings show that distracting statements (polysemous words with clinical meanings used in a non-clinical context or references to unrelated health conditions) can reduce LLM accuracy by up to 17.9%. Commonly proposed solutions to improve model performance such as retrieval-augmented generation (RAG) and medical fine-tuning did not change this effect and in some cases introduced their own confounders and further degraded performance. Our findings suggest that LLMs natively lack the logical mechanisms necessary to distinguish relevant from irrelevant clinical information, posing challenges for real-world applications. MedDistractQA and our results highlights the need for robust mitigation strategies to enhance LLM resilience to extraneous information.

Shrutika Singh, Anton Alyakin, Daniel Alexander Alber et al.

The performance of Large Language Models (LLMs) on multiple-choice question (MCQ) benchmarks is frequently cited as proof of their medical capabilities. We hypothesized that LLM performance on medical MCQs may in part be illusory and driven by factors beyond medical content knowledge and reasoning capabilities. To assess this, we created a novel benchmark of free-response questions with paired MCQs (FreeMedQA). Using this benchmark, we evaluated three state-of-the-art LLMs (GPT-4o, GPT-3.5, and LLama-3-70B-instruct) and found an average absolute deterioration of 39.43% in performance on free-response questions relative to multiple-choice (p = 1.3 * 10-5) which was greater than the human performance decline of 22.29%. To isolate the role of the MCQ format on performance, we performed a masking study, iteratively masking out parts of the question stem. At 100% masking, the average LLM multiple-choice performance was 6.70% greater than random chance (p = 0.002) with one LLM (GPT-4o) obtaining an accuracy of 37.34%. Notably, for all LLMs the free-response performance was near zero. Our results highlight the shortcomings in medical MCQ benchmarks for overestimating the capabilities of LLMs in medicine, and, broadly, the potential for improving both human and machine assessments using LLM-evaluated free-response questions.

Gabriel R. Rosenbaum, Lavender Yao Jiang, Ivaxi Sheth et al.

Large language models (LLMs) have recently emerged as powerful tools, finding many medical applications. LLMs' ability to coalesce vast amounts of information from many sources to generate a response-a process similar to that of a human expert-has led many to see potential in deploying LLMs for clinical use. However, medicine is a setting where accurate reasoning is paramount. Many researchers are questioning the effectiveness of multiple choice question answering (MCQA) benchmarks, frequently used to test LLMs. Researchers and clinicians alike must have complete confidence in LLMs' abilities for them to be deployed in a medical setting. To address this need for understanding, we introduce a knowledge graph (KG)-based method to evaluate the biomedical reasoning abilities of LLMs. Essentially, we map how LLMs link medical concepts in order to better understand how they reason. We test GPT-4, Llama3-70b, and PalmyraMed-70b, a specialized medical model. We enlist a panel of medical students to review a total of 60 LLM-generated graphs and compare these graphs to BIOS, a large biomedical KG. We observe GPT-4 to perform best in our human review but worst in our ground truth comparison; vice-versa with PalmyraMed, the medical model. Our work provides a means of visualizing the medical reasoning pathways of LLMs so they can be implemented in clinical settings safely and effectively.

Lavender Y. Jiang, Angelica Chen, Xu Han et al.

Hospitals and healthcare systems rely on operational decisions that determine patient flow, cost, and quality of care. Despite strong performance on medical knowledge and conversational benchmarks, foundation models trained on general text may lack the specialized knowledge required for these operational decisions. We introduce Lang1, a family of models (100M-7B parameters) pretrained on a specialized corpus blending 80B clinical tokens from NYU Langone Health's EHRs and 627B tokens from the internet. To rigorously evaluate Lang1 in real-world settings, we developed the REalistic Medical Evaluation (ReMedE), a benchmark derived from 668,331 EHR notes that evaluates five critical tasks: 30-day readmission prediction, 30-day mortality prediction, length of stay, comorbidity coding, and predicting insurance claims denial. In zero-shot settings, both general-purpose and specialized models underperform on four of five tasks (36.6%-71.7% AUROC), with mortality prediction being an exception. After finetuning, Lang1-1B outperforms finetuned generalist models up to 70x larger and zero-shot models up to 671x larger, improving AUROC by 3.64%-6.75% and 1.66%-23.66% respectively. We also observed cross-task scaling with joint finetuning on multiple tasks leading to improvement on other tasks. Lang1-1B effectively transfers to out-of-distribution settings, including other clinical tasks and an external health system. Our findings suggest that predictive capabilities for hospital operations require explicit supervised finetuning, and that this finetuning process is made more efficient by in-domain pretraining on EHR. Our findings support the emerging view that specialized LLMs can compete with generalist models in specialized tasks, and show that effective healthcare systems AI requires the combination of in-domain pretraining, supervised finetuning, and real-world evaluation beyond proxy benchmarks.

Ioana Marinescu, Kyunghyun Cho, Eric Karl Oermann

In-context learning (ICL) is the ability of a large language model (LLM) to learn a new task from a few demonstrations presented as part of the context. Past studies have attributed a large portion of the success of ICL to the way these in-context demonstrations are represented, particularly to how labels are represented in classification tasks. On the other hand, observations of the learning capacity of ICL (i.e., the extent to which more in-context demonstrations can lead to higher performance) have been mixed, and ICL is often thought to occur only under specific conditions. The interaction between these two aspects in ICL, representation and learning, has not been studied in depth until now. We hypothesize that they are largely independent of one another, such that the representation of demonstrations determines the baseline accuracy of ICL, while learning from additional demonstrations improves only on top of this baseline. We validate this hypothesis by developing an optimization algorithm that can enumerate a spectrum of possible label sets (representations) varying in semantic relevance. We then perform ICL with varying numbers of in-context demonstrations for each of these label sets. We observed that learning happens regardless of the quality of the label set itself, although its efficiency, measured by the slope of improvement over in-context demonstrations, is conditioned on both the label set quality and the parameter count of the underlying language model. Despite the emergence of learning, the relative quality (accuracy) of the choice of a label set (representation) is largely maintained throughout learning, confirming our hypothesis and implying their orthogonality. Our work reveals a previously underexplored aspect of ICL: the independent effects of learning from demonstrations and their representations on ICL performance.

Radhika Dua, Young Joon, Kwon et al.

Radiological imaging is central to diagnosis, treatment planning, and clinical decision-making. Vision-language foundation models have spurred interest in automated radiology report generation (RRG), but safe deployment requires reliable clinical evaluation of generated reports. Existing metrics often rely on surface-level similarity or behave as black boxes, lacking interpretability. We introduce ICARE (Interpretable and Clinically-grounded Agent-based Report Evaluation), an interpretable evaluation framework leveraging large language model agents and dynamic multiple-choice question answering (MCQA). Two agents, each with either the ground-truth or generated report, generate clinically meaningful questions and quiz each other. Agreement on answers captures preservation and consistency of findings, serving as interpretable proxies for clinical precision and recall. By linking scores to question-answer pairs, ICARE enables transparent, and interpretable assessment. Clinician studies show ICARE aligns significantly more with expert judgment than prior metrics. Perturbation analyses confirm sensitivity to clinical content and reproducibility, while model comparisons reveal interpretable error patterns.

Chi Hang, Ruiqi Deng, Lavender Yao Jiang et al.

Clinical measurements such as blood pressures and respiration rates are critical in diagnosing and monitoring patient outcomes. It is an important component of biomedical data, which can be used to train transformer-based language models (LMs) for improving healthcare delivery. It is, however, unclear whether LMs can effectively interpret and use clinical measurements. We investigate two questions: First, can LMs effectively leverage clinical measurements to answer related medical questions? Second, how to enhance an LM's performance on medical question-answering (QA) tasks that involve measurements? We performed a case study on blood pressure readings (BPs), a vital sign routinely monitored by medical professionals. We evaluated the performance of four LMs: BERT, BioBERT, MedAlpaca, and GPT-3.5, on our newly developed dataset, BPQA (Blood Pressure Question Answering). BPQA contains $100$ medical QA pairs that were verified by medical students and designed to rely on BPs . We found that GPT-3.5 and MedAlpaca (larger and medium sized LMs) benefit more from the inclusion of BPs than BERT and BioBERT (small sized LMs). Further, augmenting measurements with labels improves the performance of BioBERT and Medalpaca (domain specific LMs), suggesting that retrieval may be useful for improving domain-specific LMs.

Anton Alyakin, Jaden Stryker, Daniel Alexander Alber et al.

General-purpose vision-language models (VLMs) demonstrate impressive capabilities, but their opaque training on uncurated internet data posse critical limitations for high-stakes decision-making, such as in neurosurgery. We present CNS-Obsidian, a neurosurgical VLM trained on peer-reviewed neurosurgical literature, and demonstrate its clinical utility compared with GPT-4o in a real-world setting. We compiled 23,984 articles from Neurosurgery Publications journals, yielding 78,853 figures and captions. Using GPT-4o and Claude Sonnet-3.5, we converted these image-text pairs into 263,064 training samples across three formats: instruction fine-tuning, multiple-choice questions, and differential diagnosis. We trained CNS-Obsidian, a fine-tune of the 34-billion parameter LLaVA-Next model. In a blinded, randomized deployment trial at NYU Langone Health (Aug 30-Nov 30, 2024), neurosurgeons were assigned to use either CNS-Obsidian or GPT-4o as a diagnostic co-pilot after patient consultations. Primary outcomes were diagnostic helpfulness and accuracy. CNS-Obsidian matched GPT-4o on synthetic questions (76.13% vs 77.54%, p=0.235), but only achieved 46.81% accuracy on human-generated questions versus GPT-4o's 65.70% (p<10-15). In the randomized trial, 70 consultations were evaluated (32 CNS-Obsidian, 38 GPT-4o) from 959 total consults. CNS-Obsidian received positive ratings in 40.62% of cases versus 57.89% for GPT-4o (p=0.230). Both models included correct diagnosis in approximately 60% of cases (59.38% vs 65.79%, p=0.626). Domain-specific VLMs trained on curated scientific literature can approach frontier model performance in specialized medical domains despite being orders of magnitude smaller and less expensive to train. However, low clinical utilization suggests chatbot interfaces may not align with specialist workflows, indicating need for alternative AI integration strategies.

Krithik Vishwanath, Jaden Stryker, Anton Alyakin et al.

Language models (LMs) have demonstrated expert-level reasoning and recall abilities in medicine. However, computational costs and privacy concerns are mounting barriers to wide-scale implementation. To address these significant limitations, we introduce a parsimonious adaptation of phi-3-mini, MedMobile, a 3.8 billion parameter LM capable of running on a mobile device, for medical applications. We perform a careful set of pipeline additions and demonstrate that chain of thought, ensembling, and fine-tuning lead to the greatest performance gains, while unexpectedly retrieval augmented generation fails to demonstrate significant improvements. We evaluate the efficiency of our pipeline on the MultiMedQA and MedBullets. We demonstrate that MedMobile scores 75.7% on the MedQA (USMLE), surpassing the passing mark for licensed physicians (~60%) and rivaling scores of models 100 times its size. Across the entirety of the MultiMedQA, MedMobile achieves SOTA performance for models with less than 5B parameters and represents the smallest model to pass the MedQA (USMLE). MedMobile holds promise to democratize access to language models in medicine, bolstering lower compute needs and fast inference speeds. With the ability to combat the biggest barriers to entry for language models in medicine, we hope that MedMobile is a critical step forward in developing clinically relevant language models.

Yifan Li, Garrett Yoon, Mustafa Nasir-Moin et al.

Numerous COVID-19 clinical decision support systems have been developed. However many of these systems do not have the merit for validity due to methodological shortcomings including algorithmic bias. Methods Logistic regression models were created to predict COVID-19 mortality, ventilator status and inpatient status using a real-world dataset consisting of four hospitals in New York City and analyzed for biases against race, gender and age. Simple thresholding adjustments were applied in the training process to establish more equitable models. Results Compared to the naively trained models, the calibrated models showed a 57% decrease in the number of biased trials, while predictive performance, measured by area under the receiver/operating curve (AUC), remained unchanged. After calibration, the average sensitivity of the predictive models increased from 0.527 to 0.955. Conclusion We demonstrate that naively training and deploying machine learning models on real world data for predictive analytics of COVID-19 has a high risk of bias. Simple implemented adjustments or calibrations during model training can lead to substantial and sustained gains in fairness on subsequent deployment.