Christopher B. C. Koo, M.D.
Founder – Koo Medical Consulting PLLC
Anesthesiologist & Intensivist · Critical Care & Perioperative Operations Consulting
Helping hospitals and surgical centers deliver safer, more efficient perioperative and critical care — guided by a physician who is in the operating room and ICU every week.
Through Koo Medical Consulting PLLC, I advise hospitals and health systems, ambulatory surgery centers, clinical groups, and health-technology companies on perioperative and critical-care operations, quality and patient safety, and the adoption of AI and machine learning in clinical care. The full list of services is on the Consulting page.
A curated research digest from three topics, automated by an agent I built:
- Critical care medicine
- Healthcare machine learning/AI
- Hospital operations & logistics
- Weekly Research Digest — September 21, 2026: Critical Care, Healthcare AI & Hospital Operations
14 articles · 3 topics
Critical care medicine6 articles
Score 0.86 · ⓘ Reference
Multicomponent Pulmonary Rehabilitation for Prolonged Mechanical Ventilation in Patients Aged 80 Years or Older: A Randomized Controlled TrialTingting Liu et al. — CHEST · September 1, 2026
Why it’s here: This randomized controlled trial investigates pulmonary rehabilitation for patients on prolonged mechanical ventilation, directly relevant to ICU management.
Abstract unavailable — listed as a pointer; not summarized.
● Preprint · Score 0.76 · ✓ Verified
The Causal Artificial Intelligence Clinician for early haemodynamic management of septic shock in ICUAngelotti, G. et al. — medRxiv Preprint · health informatics · September 14, 2026
Why it’s here: This article uses causal AI for hemodynamic management of septic shock in the ICU, directly matching the researcher's interests.
Study Design
This study applied graphical causal inference models, grounded in expert clinical knowledge, to estimate heterogeneous treatment effects for septic shock management. The model was trained on 1,706 MIMIC admissions and externally validated on 1,450 eICU admissions, focusing on the first six hours of ICU admission.
Key Results
Deviation from vasopressor recommendations was associated with failed clinical improvement (median OR 1.26, 95% CI 1.21-1.32) and, less consistently, with in-hospital mortality (median OR 1.16, 95% CI 1.10-1.24), while fluid deviations showed weaker associations (OR 1.12, 95% CI 1.08-1.18 for improvement; OR 1.02, 95% CI .96-1.07 for mortality). External validation AUROCs were 0.72 for survival and 0.70 for improvement, comparable to predictive baselines, with treatment response varying by baseline physiology.
Why It Matters
This expert-grounded causal framework produced recommendations using approximately one-third of the variables required by conventional predictive models, allowing for inspection and revision of assumptions by clinicians. While these recommendations require prospective evaluation before clinical use, they offer an informative policy for early fluid and vasopressor management in septic shock.
● Preprint · Score 0.68 · ✓ Verified
Microbiome Profiling Reveals Prognostic Heterogeneity in Staphylococcus aureus PneumoniaKitsios, G. et al. — medRxiv Preprint · infectious diseases · September 20, 2026
Why it’s here: This article directly addresses Staphylococcus aureus pneumonia in mechanically ventilated patients in the ICU, examining microbiome profiling and its association with mortality, highly relevant to critical care.
Study Design
This prospective study investigated microbial ecology in mechanically ventilated patients with culture-confirmed Staphylococcus aureus pneumonia using 16S rRNA gene sequencing and shotgun nanopore metagenomics on endotracheal aspirate samples. The study enrolled 109 patients and assessed associations between microbial composition, host inflammatory biomarkers, and 60-day mortality.
Key Results
16S sequencing revealed significant heterogeneity in Staphylococcus relative abundance, with dominance (>50%) observed in only 33% of patients and associated with worse 60-day survival (50% vs. 80%, p=0.013). Nanopore metagenomics validated these findings, showing high absolute S. aureus read counts independently predicted mortality (adjusted HR 11.23 [95%CI 2.25-55.9], p=0.003).
Why It Matters
Metagenomic profiling demonstrates clinically significant heterogeneity within S. aureus pneumonia that is masked by conventional diagnostics, identifying a high-risk phenotype characterized by Staphylococcus dominance. This challenges the assumption that culture positivity represents a uniform clinical entity and highlights the prognostic value of microbial abundance and composition.
Score 0.64 · ⓘ Reference
Impact of pre-ICU balanced crystalloids versus saline on mortality in sepsis: a propensity score-matched multicenter cohort studyTae Wan Kim et al. — CHEST · September 1, 2026
Why it’s here: This study investigates fluid management in sepsis, a core topic in critical care, and uses a propensity score-matched cohort, relevant to ICU outcomes.
Abstract unavailable — listed as a pointer; not summarized.
Score 0.60 · ✓ Verified
The NLRP3 inflammasome in physiological and dysfunctional host response in human sepsis and critical illness: a narrative reviewCatherine M. Neumann et al. — Critical Care · September 19, 2026
Why it’s here: This review directly addresses the NLRP3 inflammasome in sepsis and critical illness, aligning well with the researcher's interests.
Study Design
This narrative review examines the role of the NLRP3 inflammasome in human sepsis and critical illness. The review synthesizes accumulating evidence regarding its function in innate immunity and host response.
Key Results
The NLRP3 inflammasome, essential for host defense, can become dysregulated in critical illness, leading to hyperinflammation, endothelial dysfunction, and multi-organ failure. Conversely, insufficient inflammasome responses can impair microbial clearance and increase susceptibility to secondary infections.
Why It Matters
Targeted modulation of inflammasome pathways, such as IL-1 blockade, shows promise for restoring immune homeostasis in critical illness. However, clinical translation is challenging due to the need for improved patient stratification, biomarker-guided approaches, and a deeper understanding of disease heterogeneity.
✦ Something Different · Score 0.56 · ⓘ Reference
Post intensive care burden on family and healthcare unit: a prospective observational studyArnavjyoti Das et al. — BMC Health Services Research · September 21, 2026
✦ A change of pace: This article shifts focus from direct patient physiology to the significant, often overlooked, impact of critical care on families and healthcare units, offering a crucial psychosocial and systemic perspective.
Abstract unavailable — listed as a pointer; not summarized.
Healthcare machine learning6 articles
Score 0.78 · ✓ Verified
Outcome-grounded effect of clinically stigmatizing information on large language model emergency triage prioritizationPhilip Jarrett et al. — npj Digital Medicine · September 19, 2026
Why it’s here: This article directly investigates the impact of large language models on clinical triage prioritization, aligning with the interest in LLMs in medicine and clinical decision support.
Study Design
This study employed a controlled, outcome-grounded experiment across two academic emergency departments. Emergency Severity Index-matched pairs of patients, one deteriorating within 6 hours and one not, were evaluated by three open-weight LLMs (Gemma, Qwen, DeepSeek). Models were tested before and after inserting demographic, social, or stigma-related attributes, including neutral and stigmatizing formulations of the same concepts.
Key Results
Across 221,556 comparisons, the stigmatizing frequent-emergency-department-use formulation caused significant harmful reprioritization in all six model-dataset cells, with increases of up to 9.4% compared to neutral formulations. Psychiatric history also led to significant shifts (up to 9.5%), but race, language, and insurance did not show consistent harmful shifts.
Why It Matters
These findings demonstrate that stigmatizing formulations of clinical information can bias LLM triage prioritization against deteriorating patients, highlighting input selection and formulation as safety-critical design choices. The study did not find consistent harmful shifts for race, language, or insurance, suggesting these attributes may have less impact on LLM triage prioritization in this context.
Score 0.78 · ✓ Verified
External validation of AI assisted colposcopy using WHO dataset for cervical precancer and cancer detectionTong Wu et al. — npj Digital Medicine · September 19, 2026
Why it’s here: This study externally validates an AI system for colposcopy, demonstrating its utility in clinical decision support and bridging diagnostic gaps, fitting the criteria for prospective validation and deployment.
Study Design
This study externally validated an AI-guided colposcopy system using an independent WHO open-source dataset of 187 patients with 855 colposcopic images. Forty-five colposcopists with varying experience levels from 12 regions in China participated in the validation.
Key Results
The AI system achieved a standalone sensitivity of 84.2% for CIN2+ detection, and with AI assistance, colposcopist sensitivity increased from 84.8% to 90.6%. This improvement was most pronounced in less experienced colposcopists, with a 6.5% sensitivity increase and an AUC improvement from 0.72 to 0.76 (p = 0.043). Additionally, AI-guided biopsy support reduced the mean number of biopsies per case from 2.48 to 2.02.
Why It Matters
These findings demonstrate the AI system's robust performance on an external dataset and its potential to mitigate diagnostic experience gaps among colposcopists. The AI system supports its clinical implementation as a valuable adjunct for colposcopic assessment.
● Preprint · Score 0.66 · ✓ Verified
Leveraging Large Language Models for Colorectal Cancer Symptom Extraction from MIMIC-IV Clinical NotesLee, Y. et al. — medRxiv Preprint · health informatics · September 20, 2026
Why it’s here: This article directly benchmarks Large Language Models for extracting clinical symptoms from patient notes, demonstrating their application in clinical care and information extraction.
Study Design
This study benchmarked rule-based, named entity recognition (NER), and zero-shot large language model (LLM) methods for extracting 46 cancer-related symptoms from 2,704 colorectal cancer (CRC) patient discharge notes in MIMIC-IV. A 46-symptom target list was derived from established symptom assessment scales. Performance was evaluated against a 200-note gold standard adjudicated by two raters.
Key Results
Gemini 3.5 Flash demonstrated the best performance with a Macro F1 score of 0.70 and Micro F1 of 0.86, significantly outperforming rule-based (Macro F1=0.44) and NER (Macro F1=0.38) methods. Claude Haiku followed with a Macro F1 of 0.63, also substantially exceeding traditional NLP approaches. Notably, post-hoc negation filtering paradoxically degraded LLM performance.
Why It Matters
Zero-shot LLMs offer a scalable and accurate alternative to manual chart review and traditional NLP for oncology symptom surveillance, eliminating the need for institution-specific rule development or model training. However, the study cautions against applying post-hoc negation correction to LLM outputs without syntactic scope validation, as it can reduce accuracy.
● Preprint · Score 0.62 · ✓ Verified
Distinguishing Social Isolation, Social Support, and Contextual False Positives in Clinical Notes Using Fine-Tuned Language Models: Multisite Validation StudyChinthala, L. K. et al. — medRxiv Preprint · health informatics · September 16, 2026
Why it’s here: This study develops and validates fine-tuned language models for classifying social isolation from clinical notes, demonstrating the application of LLMs for clinical information extraction and validation in healthcare settings.
Study Design
This multisite retrospective study developed and evaluated fine-tuned language models for classifying social isolation, social support, and irrelevant clinical text from electronic health records. The study included 326,847 adults aged 50 years or older across three Tennessee health systems from 2020 to 2023, using uncertainty-based active learning to annotate 9,748 spans from clinical notes.
Key Results
Fully fine-tuned FLAN-T5-Large achieved the highest performance with a mean macro-F1 of 0.92 (SD 0.04), outperforming BERT (0.77) and RoBERTa (0.80), with class-specific F1 scores of 0.91 for social isolation and 0.90 for social support. Gemma-2-2B achieved a macro-F1 of 0.89 (SD 0.10), and full fine-tuning was superior to parameter-efficient methods.
Why It Matters
Fine-tuned language models can effectively distinguish nuanced social connectedness concepts from heterogeneous clinical settings, converting unstructured documentation into structured data. The study's 3-class framework and multisite validation provide a practical approach, though remaining errors highlight complexities like ambiguous support and non-social uses of isolation terminology.
● Preprint · Score 0.56 · ✓ Verified
Machine Learning Approach to Identify Gut Microbiota Biomarkers in Patients with ST-Elevation Myocardial Infarction Presenting Primary Ventricular TachyarrhythmiasTseng, H.-P. et al. — medRxiv Preprint · cardiovascular medicine · September 20, 2026
Why it’s here: This study uses machine learning to identify gut microbiota biomarkers for risk stratification in ST-elevation myocardial infarction patients, which is relevant to clinical care.
Study Design
This study employed a machine learning approach to identify gut microbiome biomarkers associated with primary ventricular tachycardia/ventricular fibrillation (VT/VF) in ST-elevation myocardial infarction (STEMI) patients. Stool samples from 33 STEMI patients, including 7 with primary VT/VF, were analyzed using 16S rRNA sequencing during acute and recovery phases.
Key Results
Five acute-phase taxa, including Clostridium aldenense and Enterocloster bolteae, were reproducibly associated with primary VT/VF, with a support vector machine model achieving an AUC of 0.846. Functional analysis revealed enrichment of fucose degradation and purine catabolism pathways.
Why It Matters
These findings suggest a distinct gut microbiome-metabolic signature associated with primary VT/VF in STEMI, characterized by specific taxa and metabolic pathways. The authors note these are exploratory findings that warrant validation in larger cohorts and further mechanistic studies.
★ Flagship · ✦ Something Different · Score 0.51 · ⓘ Reference
Patient-facing generative artificial intelligence: interpretive influence and system-level evaluationApurva Parikh, Pushmeet Kohli — The Lancet Digital Health · September 1, 2026
✦ A change of pace: This article offers a unique perspective on patient-facing generative AI, shifting the focus from technical ML applications to the user experience and ethical considerations in healthcare, which is a departure from the more clinically focused or technical articles already selected.
Abstract unavailable — listed as a pointer; not summarized.
Hospital logistics and operations2 articles
Score 0.62 · ✓ Verified
Integrating urgent care centers into primary care networks to reduce emergency department overcrowding: a real-world implementation and evaluationDaniela Fortuna et al. — BMC Health Services Research · September 19, 2026
Why it’s here: Directly addresses ED overcrowding by integrating urgent care centers, impacting patient flow and resource utilization.
Study Design
This study employed a retrospective, population-based time-series design to evaluate the impact of implementing community-based Urgent Care Centers (UCCs) on Emergency Department (ED) overcrowding in the Emilia-Romagna Region, Italy. Using regional health databases from January 2022 to April 2025, the study analyzed changes in low-acuity ED visits across 32 health districts with staggered UCC openings. Interrupted time-series and segmented regression models were used to estimate these changes, stratified by urbanization and organizational models.
Key Results
The implementation of UCCs was associated with a significant reduction in daily low-acuity ED visits, which were 13.5% lower than expected on average since the first UCC opened, reaching -23.6% in later periods. The organizational model significantly influenced this effect, with shared pre-triage models reducing ED visits by 42.2%, while hospital-adjacent UCCs without shared pre-triage showed minimal reduction (-0.7%). The strongest predictor of UCC use was its presence in the patient's municipality of residence, increasing UCC likelihood sevenfold.
Why It Matters
These findings suggest that integrating UCCs into primary care networks can effectively reduce ED overcrowding, particularly when organized with shared access and pre-triage pathways. The study highlights that organizational design is critical for optimizing urgent-care delivery and achieving ED decongestion. However, the authors caution that observed differences should not be solely attributed to UCC implementation due to the absence of a contemporaneous untreated comparator.
Score 0.42 · ⓘ Reference
The application of lean methodology in the preoperative phase: a scoping reviewRui Cortes, Marília Silva Paulo, Paulo Sousa — BMC Health Services Research · September 21, 2026
Why it’s here: Applies lean methodology, which is relevant to operations improvement, but focuses on the preoperative phase specifically.
Abstract unavailable — listed as a pointer; not summarized.
Curated from OpenAlex, Crossref & medRxiv. Summaries are machine-generated and grounded in each article’s abstract or open-access text; ⚠ flags a summary that needs a second look, and ✦ marks a deliberate change-of-pace pick.
