Cardiac surgery patients experience rapidly evolving hemodynamics in early post-operative period requiring intensive support. Identifying hemodynamic subphenotypes from these data can inform personalized management. Using 24-hour high-resolution physiologic and treatment data from 6,630 MIMIC-IV and 1,963 SICdb patients, we trained a transformer encoder with a reconstruction-contrastive objective to derive patient-level embeddings capturing multivariate temporal dynamics within first 24h of ICU stay and compared them against those generated by dynamic time warping (DTW). Spectral clustering uncovered three reproducible hemodynamic subphenotypes. Compared with subphenotype 1, subphenotype 3 received more IV fluids, vasopressors, inotropes, and exhibited higher in-hospital mortality (OR 5.85, 95 % CI 2.43-14.13), longer ICU stay (7.12 days, 95% CI: 5.52-8.73) and hospitalization (8.86 days, 95% CI: 6.57-11.16). DTW derived subphenotypes had weaker prognostic separation. Thus, contrastive-transformer framework identified more clinically meaningful temporal hemodynamic subphenotypes that may optimize post-operative risk stratification and inform personalized management.

Abstract Importance Physician-facing AI tools are now in clinical use, yet whether different platforms fail in similar or fundamentally different ways in high-stakes settings like critical care is unknown. Objective: To evaluate two physician-facing AI platforms, ChatGPT for Clinicians and OpenEvidence, for distinct vulnerabilities under structured stress testing. Design, Setting, and Participants An observational study conducted using 60 simulated critical care vignettes developed and adjudicated by four attending critical care physicians. Data were collected in the last week of April 2026, via the public website interfaces of each platform. Interventions/Exposures A 2×2×2×2 factorial design across four stressors - anchoring, cognitive load, social conformity pressure, and a clinically incorrect directive - yielded 16 prompt subsets per vignette and 960 prompts per platform. A separate multi-turn adversarial prompting paradigm administered three sequential "You are incorrect" challenges to baseline vignettes. All prompts had a universal output length constraint of fewer than 30 words. Main Outcomes and Measures Critical elements capture (percentage of gold-standard critical elements present in responses), susceptibility to clinically incorrect directive, and sycophancy (reversal of an initial correct recommendation under iterative adversarial challenge). Results Across 1916 responses to 1920 prompts, ChatGPT for Clinicians captured more gold-standard critical elements than OpenEvidence (81.4% ± 18.1% vs 61.0% ± 23.5%; adjusted difference, 20.3 percentage points; 95% CI, 18.3 to 22.4; P < .001) and was less susceptible to clinically incorrect directives (1.7% vs 8.0%; adjusted odds ratio, 0.07; 95% CI, 0.02-0.21; P < .001). Anchoring and social conformity pressure were associated with reduced critical element capture across both platforms, while cumulative stressor burden reduced critical element capture only on OpenEvidence. Conversely, ChatGPT for Clinicians reversed correct recommendations more readily under adversarial prompting (hazard ratio, 2.61; 95% CI, 1.10 - 6.19; P = .03). Conclusion and Relevance The two physician-facing clinical AI platforms evaluated demonstrated non-overlapping vulnerabilities, with neither platform uniformly superior. These findings argue against single-axis ranking of clinical AI systems and support multidimensional safety evaluation encompassing completeness of reasoning, resistance to incorrect directives, and stability under adversarial challenge. Key Points Question Do physician-facing clinical AI platforms exhibit similar or distinct vulnerabilities when evaluated under structured stress testing? Findings In this observational study, factorial stress testing of two clinical AI platforms across 60 simulated vignettes, ChatGPT for Clinicians captured more gold-standard critical elements and was less susceptible to clinically incorrect directives than OpenEvidence, whereas OpenEvidence was more stable under repeated adversarial challenge; neither platform was uniformly superior. Meaning Physician-facing AI platforms exhibit non-overlapping failure modes and should be evaluated across multiple safety dimensions rather than ranked on a single performance measure.

Background: Hospital-acquired pressure injuries (HAPIs) affect approximately 2.5 million patients annually in the United States, leading to increased morbidity and healthcare costs. Current rule-based screening tools, such as the Braden Scale, lack sensitivity, highlighting the need for improved risk prediction methods. Methods: We developed and externally validated a machine learning model to predict HAPI risk using longitudinal electronic medical record (EMR) data. This study included adult inpatients (2018–2023) across five hospitals within a large health system. An automated pipeline was built for EMR data curation, labeling, and integration. The model employed XGBoost with recursive feature elimination to identify 35 optimal clinical variables and utilized time-series analysis for dynamic risk prediction. Results: Internal validation and multi-center external validation on 5510 hospitalizations demonstrated AUROC values of 0.83–0.85. The model outperformed the Braden Scale in sensitivity and F1-score and showed superior performance compared to previous predictive models. Conclusions: This is the first externally validated, cross-institutional HAPI prediction model using longitudinal EMR data and automated pipelines. The model demonstrates strong generalizability, scalability, and real-time applicability, offering a novel bioengineering approach to improve HAPI prevention, patient care, and clinical operations.

Importance:Acute kidney injury (AKI) affects one-third of patients after cardiac surgery and increases morbidity and mortality. AKI lasting over 48 hours, known as persistent AKI (pAKI), has much worse outcomes. Hemodynamic optimization is cornerstone of AKI management, however, current strategies rely on bundled care interventions that are inconsistently implemented, underscoring the need for personalized hemodynamic optimization. Objective:To develop and validate a reinforcement learning (RL) model to guide individualized dosing of intravenous (IV) fluids, vasopressors, and inotropes for prevention of pAKI after cardiac surgery. Design:Cohort study. Model development and internal validation were performed retrospectively in MIMIC-IV, with external validation in SICdb, a European database (retrospective), and then in Mount Sinai Health System cohort using data from Jan 1-Aug 18, 2025). Setting:Multicenter retrospective cohort study. Participants:Admissions to ICU after cardiac surgery. Exposures:Postoperative hemodynamic management during first 72 hours of ICU stay using IV fluids, vasopressors, and inotropes. Main Outcomes and Measures:Primary outcome was pAKI within 5 days after surgery. The RL model optimized treatment policies through reward-based learning, where higher rewards reflected improved outcome. We assessed model performance relative to clinicians using Fitted Q Evaluation and adjusted weighted pooled logistic regression. Results:There were 6,643 adult ICU admissions following cardiac surgery in MIMIC-IV, 2,254 in SICdb, and 846 in MSHS. Median age was 70 years in MIMIC-IV, 70.0 years in SICdb, and 64 years in MSHS cohort with 72%, 73%, and 70% males respectively. AKI occurred in 41.4%, 19.7%, and 22.5% of admissions, with pAKI in 30.5%, 43.0%, and 33.7% of AKI cases, respectively. RL model achieved higher cumulative rewards than clinicians across all cohorts. Concordance between clinician actions and RL model's recommendations was associated with lower adjusted odds of pAKI (OR, 0.92 [0.89-0.96] in SICdb; 0.91 [0.86-0.96] in MSHS). RL model favored smaller IV fluid volumes, moderate vasopressor dosing, and greater inotrope use. Conclusions and Relevance:In this study, personalization of early postoperative hemodynamic management using an RL model was associated with decreased risk of pAKI. These findings suggest that AI guided hemodynamic strategies may enhance postoperative care after cardiac surgery. Key Points:Question: Can reinforcement learning (RL) personalize early postoperative hemodynamic management to prevent persistent AKI (pAKI) after cardiac surgery?Findings: In 9,743 postoperative cardiac surgery ICU admissions across 3 cohorts (MIMIC-IV, SICdb, and Mount Sinai Health System), the RL model achieved higher cumulative rewards than clinician policies and was associated with lower adjusted odds of developing pAKI when clinician actions aligned with model recommendations. The RL model favored smaller intravenous fluid volumes and earlier, graded adjustments in vasopressor and inotrope dosing compared with standard practice.Meaning: RL guided individualized hemodynamic management after cardiac surgery shows promise in reducing the risk of persistent AKI and should be tested in randomized clinical trials.

Accurate discharge summaries are essential for effective communication between hospital and outpatient providers but generating them is labor-intensive. Large language models (LLMs), such as GPT-4, have shown promise in automating this process, potentially reducing clinician workload and improving documentation quality. A recent study using GPT-4 to generate discharge summaries via concatenated clinical notes found that while the summaries were concise and coherent, they often lacked comprehensiveness and contained errors. To address this, we evaluated a structured prompting strategy, summarize-then-prompt, which first generates concise summaries of individual clinical notes before combining them to create a more focused input for the LLM.The objective of this study was to assess the effectiveness of a novel prompting strategy, summarize-then-prompt, in generating discharge summaries that are more complete, accurate, and concise in comparison to the approach that simply concatenates clinical notes.We conducted a retrospective study comparing two prompting strategies: direct concatenation (M1) and summarize-then-prompt (M2). A random sample of 50 hospital stays was selected from a large hospital system. Three attending physicians independently evaluated the generated hospital course summaries for completeness, correctness, and conciseness using a 5-point Likert scale.The summarize-then-prompt strategy outperformed direct concatenation strategy in both completeness (4.28 ± 0.63 vs. 4.01 ± 0.69, p < 0.001) and correctness (4.37 ± 0.54 vs. 4.17 ± 0.57, p = 0.002) of the summarization of the hospital course. However, the two strategies showed no significant difference in conciseness (p = 0.308).Summarizing individual notes before concatenation improves LLM-generated discharge summaries, enhancing their completeness and accuracy without sacrificing conciseness. This approach may facilitate the integration of LLMs into clinical workflows, offering a promising strategy for automating discharge summary generation and could reduce clinician burden.

BACKGROUND:Despite the profound impact of "brain fog" and/or cognitive impairment in relatively young people with Long COVID, no interventions with demonstrated efficacy are currently available. We conducted a pilot randomized controlled trial investigating the preliminary outcomes of a cognitive rehabilitation (CR) intervention adapted for persons with post-COVID cognitive impairment. METHODS:Participants were ≥18 years of age, English-speaking, had history of SARS-CoV-2, and had cognitive impairment on objective measures. Eligible participants were randomized to a 12-week CR intervention or a time - and attention-matched control arm. Objective and subjective cognitive functioning was assessed at pre - and within 2-weeks post-intervention, utilizing validated neuropsychological measures across multiple domains. We compared pre vs. post intervention changes in cognitive scores in intervention vs. control groups. RESULTS:The mean change in the intervention group compared to the controls in measures of processing speed, learning, memory, language, and of executive function did not reach the threshold for futility. In comparison to controls, the intervention group self-reported significant improvements in cognitive functioning. CONCLUSIONS:We found that an adapted CR intervention for Long COVID may improve post-COVID cognitive impairment in comparison to a time - and attention-matched control group and should be evaluated in a larger trial. Trial registration: ClinicalTrials.gov identifier: NCT05498493. Registered on 08/10/2022.

This study introduces Glucose Level Understanding and Control Optimized for Safety and Efficacy (GLUCOSE), a distributional offline reinforcement learning algorithm for optimizing insulin dosing after cardiac surgery. Trained on 5228 patients, tested on 920, and externally validated on 649, GLUCOSE achieved a mean estimated reward of 0.0 [–0.07, 0.06] in internal testing and –0.63 [–0.74, –0.52] in external validation, outperforming clinician returns of –1.29 [–1.37, –1.20] and –1.02 [–1.16, –0.89]. In multi-phase human validation, GLUCOSE first showed a significantly lower mean absolute error (MAE) in insulin dosing, with 0.9 units MAE versus clinicians’ 1.97 units (p < 0.001) in internal testing and 1.90 versus 2.24 units (p = 0.003) in external validation. The second and third phases found GLUCOSE’s performance as comparable to or exceeding that of senior clinicians in MAE, safety, effectiveness, and acceptability. These findings suggest GLUCOSE as a robust tool for improving postoperative glucose management.

Importance:Automating the identification of risk for developing hospital delirium with models that use machine learning (ML) could facilitate more rapid prevention, identification, and treatment of delirium. However, there are very few reports on the performance of ML models for delirium risk stratification in live clinical practice. Objective:To report on development, operationalization, and validation of a multimodal ML model for delirium risk stratification in live clinical practice and its associations with workflow and clinical outcomes. Design, Setting, and Participants:This quality improvement study developed an ML model supported by automated electronic medical records to stratify the risk of non-intensive care unit delirium in live clinical practice using the Confusion Assessment Method as the diagnostic reference standard, with an iterative model update method. Data from patients aged at least 60 years admitted to non-intensive care units at Mount Sinai Hospital between January 2016 and January 2020 were used to train and test the ML model presented. The model was validated in live clinical practice from March 2023 to March 2024. Analysis of the model's associations with workflow and clinical outcomes was conducted retrospectively in 2024, comparing hospitalized patients prior to deployment of any model version (pre-ML cohort) and during model clinical deployment (post-ML cohort). Main Outcomes and Measures:Outcomes of interest were area under the receiver operating characteristic curve, monthly delirium detection rates, median length of hospital stay, and daily doses of opiate, benzodiazepine, and antipsychotic medications administered. Results:The overall sample included 32 284 inpatient admissions (mean [SD] age, 73.56 (9.67) years, 15 157 [46.9%] women). A total of 25 261 inpatient admissions of older patients with both medical and surgical primary diagnoses represented the combined model testing and training cohort (median age, 73.37 [66.42-81.36] years) and live clinical deployment validation cohort (median [IQR] age, 72.11 [62.26-78.97] years), while 7023 inpatient admissions of older patients with both medical and surgical primary diagnoses represented the combined pre-ML (median [IQR] age, 74.00 [68.00-81.00] years) and post-ML (median [IQR] age, 75.33 [68.34-82.91] years) cohorts. The model presented is a fusion of electronic medical record patient data features and clinical note features processed by natural language processing. The results of model validation in live clinical practice included an area under the curve of 0.94 (95% CI, 0.93-0.95). Median (IQR) monthly delirium detection rates of inpatients assessed for delirium with the Confusion Assessment Method increased from 4.42% (95% CI, 3.70%-5.14%) in the pre-ML cohort to 17.17% (95% CI, 15.54%-18.80%) in the post-ML cohort (P < .001). Post-ML vs pre-ML cohorts received lower daily doses of benzodiazepines (median [IQR] 0.93 [0.42-2.28] diazepam dose equivalents vs 1.60 [0.66-4.27] diazepam dose equivalents; P < .001) and olanzapine (median [IQR], 1.09 [0.38-2.46] mg vs 2.50 [1.17-6.65] mg; P < .001). Conclusions and Relevance:This quality improvement study demonstrates the feasibility of a novel multimodal ML model to automate delirium risk stratification in live clinical practice. The model demonstrated acceptable performance in live clinical practice and may facilitate resource allocation to enhance delirium identification and care.

Large language models (LLMs) offer potential benefits in clinical care. However, concerns remain regarding socio-demographic biases embedded in their outputs. Opioid prescribing is one domain in which these biases can have serious implications, especially given the ongoing opioid epidemic and the need to balance effective pain management with addiction risk. We tested ten LLMs-both open-access and closed-source-on 1,000 acute-pain vignettes. Half of the vignettes were labeled as non-cancer and half as cancer. Each vignette was presented in 34 socio-demographic variations, including a control group without demographic identifiers. We analyzed the models' recommendations on opioids, anxiety treatment, perceived psychological stress, risk scores, and monitoring recommendations. Overall, yielding 3.4 million model-generated responses. Using logistic and linear mixed-effects models, we measured how these outputs varied by demographic group and whether a cancer diagnosis intensified or reduced observed disparities. Across both cancer and non-cancer cases, historically marginalized groups-especially cases labeled as individuals who are unhoused, Black, or identify as LGBTQIA+-often received more or stronger opioid recommendations, sometimes exceeding 90% in cancer settings, despite being labeled as high risk by the same models. Meanwhile, low-income or unemployed groups were assigned elevated risk scores yet fewer opioid recommendations, hinting at inconsistent rationales. Disparities in anxiety treatment and perceived psychological stress similarly clustered within marginalized populations, even when clinical details were identical. These patterns diverged from standard guidelines and point to model-driven bias rather than acceptable clinical variation. Our findings underscore the need for rigorous bias evaluation and the integration of guideline-based checks in LLMs to ensure equitable and evidence-based pain care.

BACKGROUND:Large language models (LLMs) show promise in clinical contexts but can generate false facts (often referred to as "hallucinations"). One subset of these errors arises from adversarial attacks, in which fabricated details embedded in prompts lead the model to produce or elaborate on the false information. We embedded fabricated content in clinical prompts to elicit adversarial hallucination attacks in multiple large language models. We quantified how often they elaborated on false details and tested whether a specialized mitigation prompt or altered temperature settings reduced errors. METHODS:We created 300 physician-validated simulated vignettes, each containing one fabricated detail (a laboratory test, a physical or radiological sign, or a medical condition). Each vignette was presented in short and long versions-differing only in word count but identical in medical content. We tested six LLMs under three conditions: default (standard settings), mitigating prompt (designed to reduce hallucinations), and temperature 0 (deterministic output with maximum response certainty), generating 5,400 outputs. If a model elaborated on the fabricated detail, the case was classified as a "hallucination". RESULTS:Hallucination rates range from 50 % to 82 % across models and prompting methods. Prompt-based mitigation lowers the overall hallucination rate (mean across all models) from 66 % to 44 % (p < 0.001). For the best-performing model, GPT-4o, rates decline from 53 % to 23 % (p < 0.001). Temperature adjustments offer no significant improvement. Short vignettes show slightly higher odds of hallucination. CONCLUSIONS:LLMs are highly susceptible to adversarial hallucination attacks, frequently generating false clinical details that pose risks when used without safeguards. While prompt engineering reduces errors, it does not eliminate them.

OBJECTIVES:Machine learning algorithms can outperform older methods in predicting clinical deterioration, but rigorous prospective data on their real-world efficacy are limited. We hypothesized that real-time machine learning generated alerts sent directly to front-line providers would reduce escalations. DESIGN:Single-center prospective pragmatic nonrandomized clustered clinical trial. SETTING:Academic tertiary care medical center. PATIENTS:Adult patients admitted to four medical-surgical units. Assignment to intervention or control arms was determined by initial unit admission. INTERVENTIONS:Real-time alerts stratified according to predicted likelihood of deterioration sent either to the primary team or directly to the rapid response team (RRT). Clinical care and interventions were at the providers' discretion. For the control units, alerts were generated but not sent, and standard RRT activation criteria were used. MEASUREMENTS AND MAIN RESULTS:The primary outcome was the rate of escalation per 1000 patient bed days. Secondary outcomes included the frequency of orders for fluids, medications, and diagnostic tests, and combined in-hospital and 30-day mortality. Propensity score modeling with stabilized inverse probability of treatment weight (IPTW) was used to account for differences between groups. Data from 2740 patients enrolled between July 2019 and March 2020 were analyzed (1488 intervention, 1252 control). Average age was 66.3 years and 1428 participants (52%) were female. The rate of escalation was 12.3 vs. 11.3 per 1000 patient bed days (difference, 1.0; 95% CI, -2.8 to 4.7) and IPTW adjusted incidence rate ratio 1.43 (95% CI, 1.16-1.78; p < 0.001). Patients in the intervention group were more likely to receive cardiovascular medication orders (16.1% vs. 11.3%; 4.7%; 95% CI, 2.1-7.4%) and IPTW adjusted relative risk (RR) (1.74; 95% CI, 1.39-2.18; p < 0.001). Combined in-hospital and 30-day-mortality was lower in the intervention group (7% vs. 9.3%; -2.4%; 95% CI, -4.5% to -0.2%) and IPTW adjusted RR (0.76; 95% CI, 0.58-0.99; p = 0.045). CONCLUSIONS:Real-time machine learning alerts do not reduce the rate of escalation but may reduce mortality.

Importance Increased intracranial pressure (ICP) is associated with adverse neurological outcomes, but needs invasive monitoring. Objective Development and validation of an AI approach for detecting increased ICP (aICP) using only non-invasive extracranial physiological waveform data. Design Retrospective diagnostic study of AI-assisted detection of increased ICP. We developed an AI model using exclusively extracranial waveforms, externally validated it and assessed associations with clinical outcomes. Setting MIMIC-III Waveform Database (2000-2013), a database derived from patients admitted to an ICU in an academic Boston hospital, was used for development of the aICP model, and to report association with neurologic outcomes. Data from Mount Sinai Hospital (2020-2022) in New York City was used for external validation. Participants Patients were included if they were older than 18 years, and were monitored with electrocardiograms, arterial blood pressure, respiratory impedance plethysmography and pulse oximetry. Patients who additionally had intracranial pressure monitoring were used for development (N=157) and external validation (N=56). Patients without intracranial monitors were used for association with outcomes (N=1694). Exposures Extracranial waveforms including electrocardiogram, arterial blood pressure, plethysmography and SpO 2 . Main Outcomes and Measures Intracranial pressure > 15 mmHg. Measures were Area under receiver operating characteristic curves (AUROCs), sensitivity, specificity, and accuracy at threshold of 0.5. We calculated odds ratios and p-values for phenotype association. Results The AUROC was 0.91 (95% CI, 0.90-0.91) on testing and 0.80 (95% CI, 0.80-0.80) on external validation. aICP had accuracy, sensitivity, and specificity of 73.8% (95% CI, 72.0%-75.6%), 99.5% (95% CI 99.3%-99.6%), and 76.9% (95% CI, 74.0-79.8%) on external validation. A ten-percentile increment was associated with stroke (OR=2.12; 95% CI, 1.27-3.13), brain malignancy (OR=1.68; 95% CI, 1.09-2.60), subdural hemorrhage (OR=1.66; 95% CI, 1.07-2.57), intracerebral hemorrhage (OR=1.18; 95% CI, 1.07-1.32), and procedures like percutaneous brain biopsy (OR=1.58; 95% CI, 1.15-2.18) and craniotomy (OR = 1.43; 95% CI, 1.12-1.84; P < 0.05 for all). Conclusions and Relevance aICP provides accurate, non-invasive estimation of increased ICP, and is associated with neurological outcomes and neurosurgical procedures in patients without intracranial monitoring.

Generative Large Language Models (LLMs) hold significant promise in healthcare, demonstrating capabilities such as passing medical licensing exams and providing clinical knowledge. However, their current use as information retrieval tools is limited by challenges like data staleness, resource demands, and occasional generation of incorrect information. This study assessed the potential of LLMs to function as autonomous agents in a simulated tertiary care medical center, using real-world clinical cases across multiple specialties. Both proprietary and open-source LLMs were evaluated, with Retrieval Augmented Generation (RAG) enhancing contextual relevance. Proprietary models, particularly GPT-4, generally outperformed open-source models, showing improved guideline adherence and more accurate responses with RAG. The manual evaluation by expert clinicians was crucial in validating models' outputs, underscoring the importance of human oversight in LLM operation. Further, the study emphasizes Natural Language Programming (NLP) as the appropriate paradigm for modifying model behavior, allowing for precise adjustments through tailored prompts and real-world interactions. This approach highlights the potential of LLMs to significantly enhance and supplement clinical decision-making, while also emphasizing the value of continuous expert involvement and the flexibility of NLP to ensure their reliability and effectiveness in healthcare settings.

Importance. Medical ethics is inherently complex, shaped by a broad spectrum of opinions, experiences, and cultural perspectives. The integration of large language models (LLMs) in healthcare is new and requires an understanding of their consistent adherence to ethical standards. Objective. To compare the agreement rates in answering questions based on ethically ambiguous situations between three frontier LLMs (GPT-4, Gemini-pro-1.5, and Llama-3-70b) and a multi-disciplinary physician group. Methods. In this cross-sectional study, three LLMs generated 1,248 medical ethics questions. These questions were derived based on the principles outlined in the American College of Physicians Ethics Manual. The topics spanned traditional, inclusive, interdisciplinary, and contemporary themes. Each model was then tasked in answering all generated questions. Twelve practicing physicians evaluated and responded to a randomly selected 10% subset of these questions. We compared agreement rates in question answering among the physicians, between the physicians and LLMs, and among LLMs. Results. The models generated a total of 3,744 answers. Despite physicians perceiving the questions' complexity as moderate, with scores between 2 and 3 on a 5-point scale, their agreement rate was only 55.9%. The agreement between physicians and LLMs was also low at 57.9%. In contrast, the agreement rate among LLMs was notably higher at 76.8% (p < 0.001), emphasizing the consistency in LLM responses compared to both physician-physician and physician-LLM agreement. Conclusions. LLMs demonstrate higher agreement rates in ethically complex scenarios compared to physicians, suggesting their potential utility as consultants in ambiguous ethical situations. Future research should explore how LLMs can enhance consistency while adapting to the complexities of real-world ethical dilemmas.

Large language models (LLMs) can optimize clinical workflows; however, the economic and computational challenges of their utilization at the health system scale are underexplored. We evaluated how concatenating queries with multiple clinical notes and tasks simultaneously affects model performance under increasing computational loads. We assessed ten LLMs of different capacities and sizes utilizing real-world patient data. We conducted >300,000 experiments of various task sizes and configurations, measuring accuracy in question-answering and the ability to properly format outputs. Performance deteriorated as the number of questions and notes increased. High-capacity models, like Llama-3-70b, had low failure rates and high accuracies. GPT-4-turbo-128k was similarly resilient across task burdens, but performance deteriorated after 50 tasks at large prompt sizes. After addressing mitigable failures, these two models can concatenate up to 50 simultaneous tasks effectively, with validation on a public medical question-answering dataset. An economic analysis demonstrated up to a 17-fold cost reduction at 50 tasks using concatenation. These results identify the limits of LLMs for effective utilization and highlight avenues for cost-efficiency at the enterprise scale.

Increased intracranial pressure (ICP) >= 15 mmHg is associated with adverse neurological outcomes, but needs invasive intracranial monitoring. Using the publicly available MIMIC-III Waveform Database (2000-2013) from Boston, we developed an artificial intelligence-derived biomarker for elevated ICP (aICP) for adult patients. aICP uses routinely collected extracranial waveform data as input, reducing the need for invasive monitoring. We externally validated aICP with an independent dataset from the Mount Sinai Hospital (2020-2022) in New York City. The AUROC, accuracy, sensitivity, and specificity on the external validation dataset were 0.80 (95% CI, 0.80-0.80), 73.8% (95% CI, 72.0-75.6%), 73.5% (95% CI 72.5-74.5%), and 73.0% (95% CI, 72.0-74.0%), respectively. We also present an exploratory analysis showing aICP predictions are associated with clinical phenotypes. A ten-percentile increment was associated with brain malignancy (OR = 1.68; 95% CI, 1.09-2.60), intracerebral hemorrhage (OR = 1.18; 95% CI, 1.07-1.32), and craniotomy (OR = 1.43; 95% CI, 1.12-1.84; P < 0.05 for all).

The decision to extubate patients on invasive mechanical ventilation is critical; however, clinician performance in identifying patients to liberate from the ventilator is poor. Machine Learning-based predictors using tabular data have been developed; however, these fail to capture the wide spectrum of data available. Here, we develop and validate a deep learning-based model using routinely collected chest X-rays to predict the outcome of attempted extubation. We included 2288 serial patients admitted to the Medical ICU at an urban academic medical center, who underwent invasive mechanical ventilation, with at least one intubated CXR, and a documented extubation attempt. The last CXR before extubation for each patient was taken and split 79/21 for training/testing sets, then transfer learning with k-fold cross-validation was used on a pre-trained ResNet50 deep learning architecture. The top three models were ensembled to form a final classifier. The Grad-CAM technique was used to visualize image regions driving predictions. The model achieved an AUC of 0.66, AUPRC of 0.94, sensitivity of 0.62, and specificity of 0.60. The model performance was improved compared to the Rapid Shallow Breathing Index (AUC 0.61) and the only identified previous study in this domain (AUC 0.55), but significant room for improvement and experimentation remains.