Large language models (LLMs) with retrieval-augmented generation (RAG) have improved information extraction over previous methods, yet their reliance on embeddings often leads to inefficient retrieval. We introduce CLinical Entity Augmented Retrieval (CLEAR), a RAG pipeline that retrieves information using entities. We compared CLEAR to embedding RAG and full-note approaches for extracting 18 variables using six LLMs across 20,000 clinical notes. Average F1 scores were 0.90, 0.86, and 0.79; inference times were 4.95, 17.41, and 20.08 s per note; average model queries were 1.68, 4.94, and 4.18 per note; and average input tokens were 1.1k, 3.8k, and 6.1k per note for CLEAR, embedding RAG, and full-note approaches, respectively. In conclusion, CLEAR utilizes clinical entities for information retrieval and achieves >70% reduction in token usage and inference time with improved performance compared to modern methods.

Red teaming, the practice of adversarially exposing unexpected or undesired model behaviors, is critical towards improving equity and accuracy of large language models, but non-model creator-affiliated red teaming is scant in healthcare. We convened teams of clinicians, medical and engineering students, and technical professionals (80 participants total) to stress-test models with real-world clinical cases and categorize inappropriate responses along axes of safety, privacy, hallucinations/accuracy, and bias. Six medically-trained reviewers re-analyzed prompt-response pairs and added qualitative annotations. Of 376 unique prompts (1504 responses), 20.1% were inappropriate (GPT-3.5: 25.8%; GPT-4.0: 16%; GPT-4.0 with Internet: 17.8%). Subsequently, we show the utility of our benchmark by testing GPT-4o, a model released after our event (20.4% inappropriate). 21.5% of responses appropriate with GPT-3.5 were inappropriate in updated models. We share insights for constructing red teaming prompts, and present our benchmark for iterative model assessments.

Because meat is more resource intensive than vegetal protein sources, replacing it with efficient plant alternatives is potentially desirable, provided these alternatives prove nutritionally sound. We show that protein conserving plant alternatives to meat that rigorously satisfy key nutritional constraints while minimizing cropland, nitrogen fertilizer (Nr) and water use and greenhouse gas (GHG) emissions exist, and could improve public health. We develop a new methodology for identifying nutritional constraints whose satisfaction by plant eaters is challenging, disproportionately shaping the optimal diets, singling out energy, mass, monounsaturated fatty acids, vitamins B_3,6,12 and D, choline, zinc, and selenium. By replacing meat with the devised plant alternatives—dominated by tofu, soybeans, peanuts, and lentils—Americans can collectively eliminate pastureland use while saving 35–50% of their diet related needs for cropland, Nr, and GHG emission, but increase their diet related irrigation needs by 15%. While widely replacing meat with plants is logistically and culturally challenging, few competing options offer comparable multidimensional resource use reduction.

Patients experiencing mental health crises often seek help through messaging-based platforms, but may face long wait times due to limited message triage capacity. Here we build and deploy a machine-learning-enabled system to improve response times to crisis messages in a large, national telehealth provider network. We train a two-stage natural language processing (NLP) system with key word filtering followed by logistic regression on 721 electronic medical record chat messages, of which 32% are potential crises (suicidal/homicidal ideation, domestic violence, or non-suicidal self-injury). Model performance is evaluated on a retrospective test set (4/1/21–4/1/22, N = 481) and a prospective test set (10/1/22–10/31/22, N = 102,471). In the retrospective test set, the model has an AUC of 0.82 (95% CI: 0.78–0.86), sensitivity of 0.99 (95% CI: 0.96–1.00), and PPV of 0.35 (95% CI: 0.309–0.4). In the prospective test set, the model has an AUC of 0.98 (95% CI: 0.966–0.984), sensitivity of 0.98 (95% CI: 0.96–0.99), and PPV of 0.66 (95% CI: 0.626–0.692). The daily median time from message receipt to crisis specialist triage ranges from 8 to 13 min, compared to 9 h before the deployment of the system. We demonstrate that a NLP-based machine learning model can reliably identify potential crisis chat messages in a telehealth setting. Our system integrates into existing clinical workflows, suggesting that with appropriate training, humans can successfully leverage ML systems to facilitate triage of crisis messages.

1 Importance: Large Language Models (LLMs) can assist in a wide range of healthcare-related activities. Current approaches to evaluating LLMs make it difficult to identify the most impactful LLM application areas. Objective: To summarize the current evaluation of LLMs in healthcare in terms of 5 components: evaluation data type, healthcare task, Natural Language Processing (NLP)/Natural Language Understanding (NLU) task, dimension of evaluation, and medical specialty. Data Sources: A systematic search of PubMed and Web of Science was performed for studies published between 01-01-2022 and 02-19-2024. Study Selection: Studies evaluating one or more LLMs in healthcare. Data Extraction and Synthesis: Three independent reviewers categorized 519 studies in terms of data used in the evaluation, the healthcare tasks (the what) and the NLP/NLU tasks (the how) examined, the dimension(s) of evaluation, and the medical specialty studied. Results: Only 5% of reviewed studies utilized real patient care data for LLM evaluation. The most popular healthcare tasks were assessing medical knowledge (e.g. answering medical licensing exam questions, 44.5%), followed by making diagnoses (19.5%), and educating patients (17.7%). Administrative tasks such as assigning provider billing codes (0.2%), writing prescriptions (0.2%), generating clinical referrals (0.6%) and clinical notetaking (0.8%) were less studied. For NLP/NLU tasks, the vast majority of studies examined question answering (84.2%). Other tasks such as summarization (8.9%), conversational dialogue (3.3%), and translation (3.1%) were infrequent. Almost all studies (95.4%) used accuracy as the primary dimension of evaluation; fairness, bias and toxicity (15.8%), robustness (14.8%), deployment considerations (4.6%), and calibration and uncertainty (1.2%) were infrequently measured. Finally, in terms of medical specialty area, most studies were in internal medicine (42%), surgery (11.4%) and ophthalmology (6.9%), with nuclear medicine (0.6%), physical medicine (0.4%) and medical genetics (0.2%) being the least represented. Conclusions and Relevance: Existing evaluations of LLMs mostly focused on accuracy of question answering for medical exams, without consideration of real patient care data. Dimensions like fairness, bias and toxicity, robustness, and deployment considerations received limited attention. To draw meaningful conclusions and improve LLM adoption, future studies need to establish a standardized set of LLM applications and evaluation dimensions, perform evaluations using data from routine care, and broaden testing to include administrative tasks as well as multiple medical specialties. Key Points: Question: How are healthcare applications of large language models (LLMs) currently evaluated? Findings: Studies rarely used real patient care data for LLM evaluation. Administrative tasks such as generating provider billing codes and writing prescriptions were understudied. Natural Language Processing (NLP)/Natural Language Understanding (NLU) tasks like summarization, conversational dialogue, and translation were infrequently explored. Accuracy was the predominant dimension of evaluation, while fairness, bias and toxicity assessments were neglected. Evaluations in specialized fields, such as nuclear medicine and medical genetics were rare. Meaning: Current LLM assessments in healthcare remain shallow and fragmented. To draw concrete insights on their performance, evaluations need to use real patient care data across a broad range of healthcare and NLP/NLU tasks and medical specialties with standardized dimensions of evaluation.

Over 250 million children in developing countries are at risk of not achieving their developmental potential, and unlikely to receive timely interventions because existing developmental assessments that help identify children who are faltering are prohibitive for use in low resource contexts. To bridge this “detection gap,” we developed a tablet-based, gamified cognitive assessment tool named DEvelopmental assessment on an E-Platform (DEEP), which is feasible for delivery by non-specialists in rural Indian households and acceptable to all end-users. Here we provide proof-of-concept of using a supervised machine learning (ML) approach benchmarked to the Bayley’s Scale of Infant and Toddler Development, 3rd Edition (BSID-III) cognitive scale, to predict a child’s cognitive development using metrics derived from gameplay on DEEP. Two-hundred children aged 34–40 months recruited from rural Haryana, India were concurrently assessed using DEEP and BSID-III. Seventy percent of the sample was used for training the ML algorithms using a 10-fold cross validation approach and ensemble modeling, while 30% was assigned to the “test” dataset to evaluate the algorithm’s accuracy on novel data. Of the 522 features that computationally described children’s performance on DEEP, 31 features which together represented all nine games of DEEP were selected in the final model. The predicted DEEP scores were in good agreement (ICC [2,1] > 0.6) and positively correlated (Pearson’s r = 0.67) with BSID-cognitive scores, and model performance metrics were highly comparable between the training and test datasets. Importantly, the mean absolute prediction error was less than three points (<10% error) on a possible range of 31 points on the BSID-cognitive scale in both the training and test datasets. Leveraging the power of ML which allows iterative improvements as more diverse data become available for training, DEEP, pending further validation, holds promise to serve as an acceptable and feasible cognitive assessment tool to bridge the detection gap and support optimum child development.