Nigam Shah

Alejandro Lozano, Scott L Fleming, Chia-Chun Chiang et al.

The quickly-expanding nature of published medical literature makes it challenging for clinicians and researchers to keep up with and summarize recent, relevant findings in a timely manner. While several closed-source summarization tools based on large language models (LLMs) now exist, rigorous and systematic evaluations of their outputs are lacking. Furthermore, there is a paucity of high-quality datasets and appropriate benchmark tasks with which to evaluate these tools. We address these issues with four contributions: we release Clinfo.ai, an open-source WebApp that answers clinical questions based on dynamically retrieved scientific literature; we specify an information retrieval and abstractive summarization task to evaluate the performance of such retrieval-augmented LLM systems; we release a dataset of 200 questions and corresponding answers derived from published systematic reviews, which we name PubMed Retrieval and Synthesis (PubMedRS-200); and report benchmark results for Clinfo.ai and other publicly available OpenQA systems on PubMedRS-200.

Ethan Steinberg, Jason Fries, Yizhe Xu et al.

We present a self-supervised, time-to-event (TTE) foundation model called MOTOR (Many Outcome Time Oriented Representations) which is pretrained on timestamped sequences of events in electronic health records (EHR) and health insurance claims. TTE models are used for estimating the probability distribution of the time until a specific event occurs, which is an important task in medical settings. TTE models provide many advantages over classification using fixed time horizons, including naturally handling censored observations, but are challenging to train with limited labeled data. MOTOR addresses this challenge by pretraining on up to 55M patient records (9B clinical events). We evaluate MOTOR's transfer learning performance on 19 tasks, across 3 patient databases (a private EHR system, MIMIC-IV, and Merative claims data). Task-specific models adapted from MOTOR improve time-dependent C statistics by 4.6% over state-of-the-art, improve label efficiency by up to 95% ,and are more robust to temporal distributional shifts. We further evaluate cross-site portability by adapting our MOTOR foundation model for six prediction tasks on the MIMIC-IV dataset, where it outperforms all baselines. MOTOR is the first foundation model for medical TTE predictions and we release a 143M parameter pretrained model for research use at [redacted URL].

Lin Lawrence Guo, Jason Fries, Ethan Steinberg et al.

Foundation models hold promise for transforming AI in healthcare by providing modular components that are easily adaptable to downstream healthcare tasks, making AI development more scalable and cost-effective. Structured EHR foundation models, trained on coded medical records from millions of patients, demonstrated benefits including increased performance with fewer training labels, and improved robustness to distribution shifts. However, questions remain on the feasibility of sharing these models across different hospitals and their performance for local task adaptation. This multi-center study examined the adaptability of a recently released structured EHR foundation model ($FM_{SM}$), trained on longitudinal medical record data from 2.57M Stanford Medicine patients. Experiments were conducted using EHR data at The Hospital for Sick Children and MIMIC-IV. We assessed both adaptability via continued pretraining on local data, and task adaptability compared to baselines of training models from scratch at each site, including a local foundation model. We evaluated the performance of these models on 8 clinical prediction tasks. In both datasets, adapting the off-the-shelf $FM_{SM}$ matched the performance of GBM models locally trained on all data while providing a 13% improvement in settings with few task-specific training labels. With continued pretraining on local data, label efficiency substantially improved, such that $FM_{SM}$ required fewer than 1% of training examples to match the fully trained GBM's performance. Continued pretraining was also 60 to 90% more sample-efficient than training local foundation models from scratch. Our findings show that adapting shared EHR foundation models across hospitals provides improved prediction performance at less cost, underscoring the utility of base foundation models as modular components to streamline the development of healthcare AI.

Ting Fang Tan, Kabilan Elangovan, Jasmine Ong et al.

A comprehensive qualitative evaluation framework for large language models (LLM) in healthcare that expands beyond traditional accuracy and quantitative metrics needed. We propose 5 key aspects for evaluation of LLMs: Safety, Consensus, Objectivity, Reproducibility and Explainability (S.C.O.R.E.). We suggest that S.C.O.R.E. may form the basis for an evaluation framework for future LLM-based models that are safe, reliable, trustworthy, and ethical for healthcare and clinical applications.

Rebecca Handler, Suhana Bedi, Nigam Shah

Premature closure, or committing to a conclusion before sufficient information is available, is a recognized contributor to diagnostic error but remains underexamined in large language models (LLMs). We define LLM premature closure as inappropriate commitment under uncertainty: providing an answer, recommendation, or clinical guidance when the safer response would be clarification, abstention, escalation, or refusal. We evaluated five frontier LLMs across structured and open-ended medical tasks. In MedQA (n = 500) and AfriMed-QA (n = 490) questions where the correct choice had been removed, models still selected an answer at high rates, with baseline false-action rates of 55-81% and 53-82%, respectively. In open-ended evaluation, models gave inappropriate answers on an average of 30% of 861 HealthBench questions and 78% of 191 physician-authored adversarial queries. Safety-oriented prompting reduced premature closure across models, but residual failure persisted, highlighting the need to evaluate whether medical LLMs know when not to answer.

Philip Chung, Akshay Swaminathan, Alex J. Goodell et al.

Methods to ensure factual accuracy of text generated by large language models (LLM) in clinical medicine are lacking. VeriFact is an artificial intelligence system that combines retrieval-augmented generation and LLM-as-a-Judge to verify whether LLM-generated text is factually supported by a patient's medical history based on their electronic health record (EHR). To evaluate this system, we introduce VeriFact-BHC, a new dataset that decomposes Brief Hospital Course narratives from discharge summaries into a set of simple statements with clinician annotations for whether each statement is supported by the patient's EHR clinical notes. Whereas highest agreement between clinicians was 88.5%, VeriFact achieves up to 92.7% agreement when compared to a denoised and adjudicated average human clinican ground truth, suggesting that VeriFact exceeds the average clinician's ability to fact-check text against a patient's medical record. VeriFact may accelerate the development of LLM-based EHR applications by removing current evaluation bottlenecks.

Hejie Cui, Alyssa Unell, Bowen Chen et al. · stanford

Large language models (LLMs) have emerged as promising tools for assisting in medical tasks, yet processing Electronic Health Records (EHRs) presents unique challenges due to their longitudinal nature. While LLMs' capabilities to perform medical tasks continue to improve, their ability to reason over temporal dependencies across multiple patient visits and time frames remains unexplored. We introduce TIMER (Temporal Instruction Modeling and Evaluation for Longitudinal Clinical Records), a framework that incorporate instruction-response pairs grounding to different parts of a patient's record as a critical dimension in both instruction evaluation and tuning for longitudinal clinical records. We develop TIMER-Bench, the first time-aware benchmark that evaluates temporal reasoning capabilities over longitudinal EHRs, as well as TIMER-Instruct, an instruction-tuning methodology for LLMs to learn reasoning over time. We demonstrate that models fine-tuned with TIMER-Instruct improve performance by 7.3% on human-generated benchmarks and 9.2% on TIMER-Bench, indicating that temporal instruction-tuning improves model performance for reasoning over EHR.

Monica Munnangi, Akshay Swaminathan, Jason Alan Fries et al.

Verifying and attributing factual claims is essential for the safe and effective use of large language models (LLMs) in healthcare. A core component of factuality evaluation is fact decomposition, the process of breaking down complex clinical statements into fine-grained atomic facts for verification. Recent work has proposed fact decomposition, which uses LLMs to rewrite source text into concise sentences conveying a single piece of information, to facilitate fine-grained fact verification. However, clinical documentation poses unique challenges for fact decomposition due to dense terminology and diverse note types and remains understudied. To address this gap and explore these challenges, we present FactEHR, an NLI dataset consisting of document fact decompositions for 2,168 clinical notes spanning four types from three hospital systems, resulting in 987,266 entailment pairs. We assess the generated facts on different axes, from entailment evaluation of LLMs to a qualitative analysis. Our evaluation, including review by the clinicians, reveals substantial variability in LLM performance for fact decomposition. For example, Gemini-1.5-Flash consistently generates relevant and accurate facts, while Llama-3 8B produces fewer and less consistent outputs. The results underscore the need for better LLM capabilities to support factual verification in clinical text.

Yuyin Zhou, Shih-Cheng Huang, Jason Alan Fries et al.

Despite the routine use of electronic health record (EHR) data by radiologists to contextualize clinical history and inform image interpretation, the majority of deep learning architectures for medical imaging are unimodal, i.e., they only learn features from pixel-level information. Recent research revealing how race can be recovered from pixel data alone highlights the potential for serious biases in models which fail to account for demographics and other key patient attributes. Yet the lack of imaging datasets which capture clinical context, inclusive of demographics and longitudinal medical history, has left multimodal medical imaging underexplored. To better assess these challenges, we present RadFusion, a multimodal, benchmark dataset of 1794 patients with corresponding EHR data and high-resolution computed tomography (CT) scans labeled for pulmonary embolism. We evaluate several representative multimodal fusion models and benchmark their fairness properties across protected subgroups, e.g., gender, race/ethnicity, age. Our results suggest that integrating imaging and EHR data can improve classification performance and robustness without introducing large disparities in the true positive rate between population groups.

Steve Yadlowsky, Scott Fleming, Nigam Shah et al.

There are a number of available methods for selecting whom to prioritize for treatment, including ones based on treatment effect estimation, risk scoring, and hand-crafted rules. We propose rank-weighted average treatment effect (RATE) metrics as a simple and general family of metrics for comparing and testing the quality of treatment prioritization rules. RATE metrics are agnostic as to how the prioritization rules were derived, and only assess how well they identify individuals that benefit the most from treatment. We define a family of RATE estimators and prove a central limit theorem that enables asymptotically exact inference in a wide variety of randomized and observational study settings. RATE metrics subsume a number of existing metrics, including the Qini coefficient, and our analysis directly yields inference methods for these metrics. We showcase RATE in the context of a number of applications, including optimal targeting of aspirin to stroke patients.

Somalee Datta, Jose Posada, Garrick Olson et al.

Stanford Medicine is building a new data platform for our academic research community to do better clinical data science. Hospitals have a large amount of patient data and researchers have demonstrated the ability to reuse that data and AI approaches to derive novel insights, support patient care, and improve care quality. However, the traditional data warehouse and Honest Broker approaches that are in current use, are not scalable. We are establishing a new secure Big Data platform that aims to reduce time to access and analyze data. In this platform, data is anonymized to preserve patient data privacy and made available preparatory to Institutional Review Board (IRB) submission. Furthermore, the data is standardized such that analysis done at Stanford can be replicated elsewhere using the same analytical code and clinical concepts. Finally, the analytics data warehouse integrates with a secure data science computational facility to support large scale data analytics. The ecosystem is designed to bring the modern data science community to highly sensitive clinical data in a secure and collaborative big data analytics environment with a goal to enable bigger, better and faster science.

Anitha Kannan, Jason Alan Fries, Eric Kramer et al.

A third of adults in America use the Internet to diagnose medical concerns, and online symptom checkers are increasingly part of this process. These tools are powered by diagnosis models similar to clinical decision support systems, with the primary difference being the coverage of symptoms and diagnoses. To be useful to patients and physicians, these models must have high accuracy while covering a meaningful space of symptoms and diagnoses. To the best of our knowledge, this paper is the first in studying the trade-off between the coverage of the model and its performance for diagnosis. To this end, we learn diagnosis models with different coverage from EHR data. We find a 1\% drop in top-3 accuracy for every 10 diseases added to the coverage. We also observe that complexity for these models does not affect performance, with linear models performing as well as neural networks.

Scott L. Fleming, Kuhan Jeyapragasan, Tony Duan et al.

There is an emerging trend in the reinforcement learning for healthcare literature. In order to prepare longitudinal, irregularly sampled, clinical datasets for reinforcement learning algorithms, many researchers will resample the time series data to short, regular intervals and use last-observation-carried-forward (LOCF) imputation to fill in these gaps. Typically, they will not maintain any explicit information about which values were imputed. In this work, we (1) call attention to this practice and discuss its potential implications; (2) propose an alternative representation of the patient state that addresses some of these issues; and (3) demonstrate in a novel but representative clinical dataset that our alternative representation yields consistently better results for achieving optimal control, as measured by off-policy policy evaluation, compared to representations that do not incorporate missingness information.

Alejandro Schuler, Michael Baiocchi, Robert Tibshirani et al.

Practitioners in medicine, business, political science, and other fields are increasingly aware that decisions should be personalized to each patient, customer, or voter. A given treatment (e.g. a drug or advertisement) should be administered only to those who will respond most positively, and certainly not to those who will be harmed by it. Individual-level treatment effects can be estimated with tools adapted from machine learning, but different models can yield contradictory estimates. Unlike risk prediction models, however, treatment effect models cannot be easily evaluated against each other using a held-out test set because the true treatment effect itself is never directly observed. Besides outcome prediction accuracy, several metrics that can leverage held-out data to evaluate treatment effects models have been proposed, but they are not widely used. We provide a didactic framework that elucidates the relationships between the different approaches and compare them all using a variety of simulations of both randomized and observational data. Our results show that researchers estimating heterogenous treatment effects need not limit themselves to a single model-fitting algorithm. Instead of relying on a single method, multiple models fit by a diverse set of algorithms should be evaluated against each other using an objective function learned from the validation set. The model minimizing that objective should be used for estimating the individual treatment effect for future individuals.

Alejandro Schuler, Ken Jung, Robert Tibshirani et al.

Many decisions in healthcare, business, and other policy domains are made without the support of rigorous evidence due to the cost and complexity of performing randomized experiments. Using observational data to answer causal questions is risky: subjects who receive different treatments also differ in other ways that affect outcomes. Many causal inference methods have been developed to mitigate these biases. However, there is no way to know which method might produce the best estimate of a treatment effect in a given study. In analogy to cross-validation, which estimates the prediction error of predictive models applied to a given dataset, we propose synth-validation, a procedure that estimates the estimation error of causal inference methods applied to a given dataset. In synth-validation, we use the observed data to estimate generative distributions with known treatment effects. We apply each causal inference method to datasets sampled from these distributions and compare the effect estimates with the known effects to estimate error. Using simulations, we show that using synth-validation to select a causal inference method for each study lowers the expected estimation error relative to consistently using any single method.

Sebastien Dubois, Nathanael Romano, David C. Kale et al.

Clinical notes are a rich source of information about patient state. However, using them to predict clinical events with machine learning models is challenging. They are very high dimensional, sparse and have complex structure. Furthermore, training data is often scarce because it is expensive to obtain reliable labels for many clinical events. These difficulties have traditionally been addressed by manual feature engineering encoding task specific domain knowledge. We explored the use of neural networks and transfer learning to learn representations of clinical notes that are useful for predicting future clinical events of interest, such as all causes mortality, inpatient admissions, and emergency room visits. Our data comprised 2.7 million notes and 115 thousand patients at Stanford Hospital. We used the learned representations, along with commonly used bag of words and topic model representations, as features for predictive models of clinical events. We evaluated the effectiveness of these representations with respect to the performance of the models trained on small datasets. Models using the neural network derived representations performed significantly better than models using the baseline representations with small ($N < 1000$) training datasets. The learned representations offer significant performance gains over commonly used baseline representations for a range of predictive modeling tasks and cohort sizes, offering an effective alternative to task specific feature engineering when plentiful labeled training data is not available.