Background: When patients are admitted to the hospital, clinical care teams must enter an expected discharge date (CT-EDD) into the electronic health record (EHR). This is often challenging for neonates admitted to the neonatal intensive care unit (NICU) due to extended lengths of stay. Pieces, a generative artificial intelligence (AI) program, uses EHR data and machine learning (ML) algorithms to predicted discharge dates (AI-PDD) for patients anticipated to discharge home in the next 5 days. Objective: To compare CT-EDD and AI-PDD for patients at a level 3 (L3) delivery NICU and level 4 (L4) referral NICU with actual discharge dates. Methods: Retrospective chart review from Sept 1, 2024 to December 31, 2024 including all NICU patients discharged home. Patients who were transferred or who died before discharge were excluded. The CT-EDD and AI-PDD were recorded starting with the date of discharge and up to six days prior, for up to 7 total days. Days with missing CT-EDD or AI-PDD were excluded. Differences between CT-EDD/AI-PDD and actual discharge date were evaluated. Results: At the L3 delivery NICU, the CT-EDD demonstrated better accuracy compared to the AI-PDD for each of the 6 days preceding discharge and the day of discharge (p<0.001). The CT-EDD performed better than the AI-PDD for 61.1% of total patient days and 69.4% of patients. Conversely, at the L4 referral NICU, the AI-PDD demonstrated better accuracy compared to the CT-EDD for each of the 6 days preceding discharge (p<0.001). The AI-PDD performed better than the CT-EDD for 61.8% of total patient days and 62.6% of patients. Conclusion: AI-PDD was more accurate than CT-EDD in a L4 referral NICU setting but underperformed in a L3 delivery NICU setting. Additional studies are warranted to evaluate factors that may contribute to these differences.
Abstract Background Patient portal messaging connects patients with their health care teams but is underused by Spanish-speaking patients. Objectives This study aims to evaluate content and use patterns of portal messaging among Spanish-speaking patients to inform future implementation strategies to increase portal messaging. Methods We evaluated messaging use patterns among a random sample of adults with an electronic health record-designated Spanish language preference across three diverse health care systems in Texas: an academic medical center, a public safety-net institution, and a regional, integrated community health system. Eligible participants initiated one or more medical advice request conversations via the portal between April 5, 2021, and April 4, 2022. We thematically coded messages to capture characteristics including message sender and receiver role, language, topic, conversation length, and concordance. Results We analyzed 883 patient-initiated conversations from 297 patients, including n = 1,557 messages sent by patients and n = 1,123 sent to patients. Most (77.3%, 1,204/1,557) messages from patient accounts were initiated in English, and 26.7% (414/1,557) were written by a self-identified care partner. Messages were most frequently about prescriptions (40.0%, 353/883), appointments (26.6%, 235/883), and symptoms (21.5%, 190/883). Most conversations were brief, with 56.4% consisting of one (21.4%) or two (35.0%) messages. Nearly all messages sent by patient accounts received language-concordant responses (94.9%, 976/1,028), although some patients and care partners used automated translation tools. Conclusion Spanish-speaking patients who use the patient portal usually message in English, and more than a quarter of messages are written by care partners. Vendors and health systems could increase portal messaging use among Spanish-speaking patients via Spanish-language portal configuration and educational resources, especially for those without access to an English-speaking care partner. The most common messaging topics (appointment scheduling and medication refill requests) have existing, but underused, discrete portal functions. Generative artificial intelligence (AI) solutions in Spanish could facilitate portal messaging triage and language-concordant responses to Spanish speakers.
Objectives The objective of this study is to evaluate the effect of a continuous remote patient monitoring (CRPM) program on reducing 30-day readmissions in a case-versus-retrospective-propensity-matched-control study of heart failure (HF) patients. Methods The study was conducted at Endeavor Health (formerly known as NorthShore University HealthSystem), Evanston, IL, in three post-soft launch phases: Phase IIa, IIb, and Phase III. HF patients were monitored for 30 days postdischarge with wearable biosensors to collect continuous ambulatory physiological data and a study phone for capturing patient-reported outcomes via daily surveys. Sensor data were analyzed by rules-based and machine learning algorithms to alert on physiologic perturbation. Platform alerts and survey data were monitored by home health nurses (HHNs) who assessed patients and, following a structured escalation pathway, referred them to HF advanced practice providers (APPs) or physicians for further management. Results A total of 39 patients completed the study. Baseline characteristics were appropriately matched in all categories with the exception of the New York Heart Association (NYHA) classification (intervention vs. control, class II: 10.3 vs. 41.0%, class III: 76.9 vs. 53.8%, class IV: 12.8 vs. 5.1%; p = 0.006). The intervention group had more patients who received HF APP calls (66.7 vs. 7.7%; p <= 0.001), HF APP visits (43.6 vs. 5.1%; p < 0.001), HF physician visits (64.1 vs. 41.0%; p = 0.041), diuretic escalation (43.6 vs. 12.8%; p = 0.006), and laboratories within 30 days of discharge (76.9 vs. 43.6%; p = 0.005). There was no significant difference between the adjusted 30-day readmission rates for the intervention and control groups (0.31, 95% confidence interval [CI]: 0.06-1.38; p = 0.138). All-cause readmissions occurred, but none were due to HF in phases IIa, IIb, and III. Conclusion Integrating CRPM directly into clinical workflows is feasible but challenging. Further research with larger patient cohorts and an adequately powered randomized trial is required to robustly evaluate the role of CRPM, machine learning algorithms, and cascading workflows in reducing 30-day readmissions.
Background Open foundation models designed for on-device and edge deployment, exemplified by the Qwen 3.5 small series (March 2, 2026) and Gemma 4 (April 2, 2026), challenge the default of routing all clinical natural language processing through commercial cloud application programming interfaces (APIs). Objectives This study aimed to examine the implications of these models for clinical informatics, for clinical informaticians, healthcare chief information officers, governance committees and NLP researchers. Methods Narrative (non-systematic) perspective. We summarized the technical advances of both releases and situated them within the 2024 to 2026 peer-reviewed evidence on locally deployed clinical large language models. We searched PubMed, Scopus and Google Scholar (January 2024 to June 2026) and added citation tracking. Results Both releases narrow the gap between local and frontier cloud capability for structured extraction and template-based documentation. The existing evidence comes from 3 to 70 billion parameter models on institutional servers. Clinical benchmarks at 0.8 to 4 billion parameters are absent. Hallucination, quantization, endpoint security, liability and unvalidated multimodal and multilingual claims remain open concerns. Frontier cloud models retain clear advantages for complex clinical reasoning. Conclusion For a defined subset of tasks, cloud-by-default may no longer be the most defensible choice. The field needs reporting standards, governance and liability frameworks, task-appropriate routing, and edge-aware benchmarks.
Objective Assess inter-rater agreement on clinician- and clinical decision support (CDS)-driven diuretic titrations and risk classifications to benchmark a semi-autonomous CDS (OTTO-FM). Methods Secondary analysis of prospectively collected porcine data modeling postoperative fluid overload. Three pediatric cardiac intensivists rated items (reasonable/unsure/unreasonable) for clinician-driven (human phase; 29 items) and CDS-driven titrations with risk classification (CDS phase; 44 items). Inter-rater agreement was estimated with ordinal-weighted Gwet's AC2. Results Dosing inter-rater agreement was similar between phases (human AC2: 0.79 [0.61-0.92]; CDS: 0.83 [0.68-0.92]), though unanimity was reached on only 19 of 29 (65.5%) human-phase and 31 of 44 (70.5%) CDS-phase items. Risk-label inter-rater agreement stratified sharply by label: high-risk, 0.92 (0.74-0.95) versus low-risk, 0.54 (0.28-0.75). Disagreement also differed in structure between tasks: 19 of 20 nonunanimous risk items were minority dissent (one rater against a two-rater majority), including 16 of 17 among low-risk items, whereas 7 of 13 nonunanimous CDS dosing items were three-way splits. Conclusion Inter-rater agreement on dosing was substantial and similar across phases, but not unanimous, suggesting that perfect reasonableness consensus with CDS may be unrealistic. Agreement on risk classification was lower, and the structure of disagreement differed between CDS dose and risk. Empirical inter-rater agreement baselines may contextualize CDS evaluation, and these differences may have implications for the design of semi-autonomous CDS if confirmed in larger studies.
