AIIM Research

Plan A blocks in regional anaesthesia: a narrative review

Sat, 20 Jun 2026 12:00:00 GMT

Plan A blocks have become a core in regional anesthesia in locations such as interscalene brachial plexus, axillary brachial plexus, femoral nerve, adductor canal, sciatic nerve, erector spinae plane, and rectus sheath. The goal of this article is to look at the importance of these blocks, along with looking at the current educational methods in aims to identify the best practices for hospitals to adopt for novice anesthesiologists. Plan A blocks act as a stepping stone into more complicated nerve blocks, offering novice anesthesiologists a chance to learn with “high-value and versatile nerve blocks with the goal of creating a core foundation of regional anesthesia skills.” A Delphi study was initiated to determine the core regional anesthesia curriculum, and of this, interscalene and axillary brachial plexus blocks, femoral nerve blocks, and sciatic nerve blocks in the popliteal fossa were the four noted to be of major importance, reaching the cutoff of 70% in a panel of 469. Additionally, a pediatric Delphi project was performed and a similar result emerged, supporting the original Delphi project of incorporating regional Plan A nerve blocks into the curriculum for anesthesiologists. Each of these nerve blocks provides specific value when considering patient outcome, potential harm, and effectiveness of the block. With this information from the study, this article calls for a change to the current UK guidelines in trainees anesthesiologists training, pushing for more Plan A nerve block centered approaches in the later years of residency, as 38% of UK anesthesiologist trainees said they felt comfortable performing a regional nerve block. This incorporation includes breaking the current barriers that exist, including a lack of confidence in training. To counteract this, learning opportunities such as workshops and teaching materials, practice ultrasound devices, have begun to become available to provide more opportunities to learn. Finally, the article addresses the current controversies exist with the incorporation of Plan A blocks into the learning curriculum. These include lack of agreement between professionals, “choice overload”, and the inconsistencies in the type of block to perform in specific situations.

VitalDB Arrhythmia Database: An Anesthesiologist-Validated Large- scale Intraoperative Arrhythmia Dataset with Beat and Rhythm Labels

Fri, 19 Jun 2026 12:00:00 GMT

The VitalDB Arrhythmia Database serves as a publicly accessible database containing labeled, intraoperative ECG recordings that can serve as source to develop further arrhythmia detection tools. Navigating the extensive source data for the creation of the VitalDB database involved using the UniMS-ECGNet deep learning algorithm to classify ECG beats and identify arrhythmia segments. Particular features of importance included beat annotation, rhythm annotation, signal quality assessment, transient conduction block events, patterned arrhythmias, and sinus node dysfunction. All segments and annotations presented by the deep learning algorithm were further reviewed and validated by five anesthesiologists prior to incorporation into the database.

Circulating DNA reveals nucleosome occupancy patterns that are associated with nucleosome-DNA affinity and are affected in cancer

Thu, 18 Jun 2026 12:00:00 GMT

Cell-free circulating DNA fragments form liquid biopsies has been increasingly studies for its potential use as a biomarker as they can originate from cells in many tissues, most often hematopoietic cells. The researchers obtained genomic positions of the circulating DNA fragments from patients with and without cancer to analyze these fragments across the two populations. They found that nucleosome occupancy was associated with histone-DNA affinity, differing across healthy and cancer samples. In addition, they found that the fragment biomarkers for cancer were generally similar, with distinct cancers having their own specific features. The discovery of well-positioned nucleosomes at transcription factor binding sites showed pan-cancer regulation of these programs affecting the main circulating DNA sources. This research allows for future use of circulating DNA fragments as a physical readout of cancer from liquid biopsies.

Engineering biomarker representations of vital signs data enhances deep learning mortality prediction

Fri, 12 Jun 2026 12:00:00 GMT

The ability to accurately predict inpatient mortality in patients admitted to the intensive care unit (ICU) is crucial to approaching patient care. This study differentiated between three methods of predicting patient mortality after 24 hours of ICU admission. These methods included raw data collected every 5 minutes, preprocessed data averaged over an hour, and pre-processed biomarker representations of vital sign data using an expanded version of PhysioZoo Pulse Oximetry Benchmarking (POBM). The expanded PhysioZoo POBM is a platform that analyzes data collected over time including temperature, heart rate, blood pressure, respiratory rate, and SpO2. This vital signs data can be used to train a bidirectional long short-term memory (BiLSTM) classifier that has the ability to predict mortality. The superior AUROC, AUPRC, and Brier scores suggested that biomarker representations are superior in training deep learning models to accurately predict ICU mortality following the initial 24 hours after admission.

Artificial intelligence-assisted risk prediction of postoperative pulmonary complications in non-small cell lung cancer surgery

Wed, 10 Jun 2026 12:00:00 GMT

Surgery for non-small cell lung cancer (NSCLC) is the most common treatment option, and, with this, the primary post-operative complication involves the lungs, ranging from 20-60%. The pulmonary problems involve air leaks, pneumonia, atelectasis, secretion retention, bronchopleural fistula, and empyema, all of which primarily stem from previous lung health, smoking history, muscle weakness, and pain. With the use of a combination of artificial intelligence (AI) and deep learning, the aim of this study is to accurately predict patient who will undergo postoperative pulmonary complications using a Fully Connected Neural Network (FCNN) deep learning model. The data gathered from the patients’ included demographics, laboratory data, respiratory function, type of surgery, PET-CT scan, and type of pulmonary disease outcome which totaled to 24 variables being incorporated into the FCNN model. The data of 953 patients was used to train the model to an sensitivity rate of 66.4%, a positive predictive value of 89.8%, and an accuracy rate of 88.6%. Meanwhile, the test dataset was used to determine the same metrics which recorded a score of 65.4% sensitivity, 100% PPV, and 90.4% accuracy with an area under the curve of .84, which indicated a strong ability to predict negative pulmonary outcomes. An F1 score was calculated which combines both sensitivity and specificity, which resulted in a rate of 84.4% for the training data and 86.4% for the test data, indicating a very good performance. Overall, this model strongly predicted outcomes and can be used to identify high risk patients’ post-operation which can lead to better pulmonary health and survival outcomes.

Beyond Human Error: Building Intelligent Resilience for Medication Safety in the ICU

Sat, 06 Jun 2026 12:00:00 GMT

The intensive care unit (ICU) is a care setting vulnerable to medical errors in relation to medication safety. There exists a surveillance gap where many errors go unnoticed in the demanding workflow of the ICU. This review proposes the implementation of a five-layer Intelligent Safety Stack to address the surveillance gap by integrating artificial intelligence (AI) tools with human oversight. The first layer of the safety stack oversees the addition of standardized error classifications into existing electronic health record systems to allow for machine learning (ML) training. The second layer involves using ML to proactively identify patients at high risk for medical errors, which transforms the AI tools from being a reactive system to a system that has the capability of identifying errors before they produce harm. Layer 3 addresses alarm fatigue and proposes using ML to analyze and filter out inappropriate warnings so that clinicians can focus on high priority warnings. The fourth layer of this intelligent safety stack involves overseeing transition of care processes by using generative AI to compile patient records into organized reconciliation lists that can be reviewed by the human user in a more efficient manner. Finally, the last layer of the stack involves using a bi-directional smart pump for infusion safety monitoring to prompt digital alerts in the case of error during medication administration.

Tech-based Evaluation of Healthcare Quality During the COVID-19 Pandemic

Fri, 29 May 2026 12:00:00 GMT

This study evaluated healthcare service quality during the COVID-19 pandemic using machine learning techniques. Researchers analyzed 52,490 responses from the 2021 Global Burden of Disease COVID-19 Health Service Disruption Survey and developed a four-dimensional framework based on World Health Organization principles for people-centered care. Six machine learning models were tested, with Support Vector Machine and Random Forest demonstrating the strongest performance. After optimization, the SVM model achieved 96% accuracy in predicting healthcare service quality. The analysis revealed significant disparities in care quality, with individuals of lower socioeconomic status, lower educational attainment, and those living in rural areas more likely to experience lower-quality healthcare services during the pandemic.

Beyond Human Error: Building Intelligent Resilience for Medication Safety in the ICU

Sat, 06 Jun 2026 12:00:00 GMT

Deep-learning time-series anomaly detection of acute kidney injury from creatinine–eGFR trajectories in the ICU

Thu, 04 Jun 2026 12:00:00 GMT

The study develops and evaluates a deep learning model to detect acute kidney injury (AKI) in ICU patients by analyzing how kidney function changes over time. Researchers used two large ICU databases encompassing over 80,000 and 140,000 admissions from MIMIC-III/IV and eICU Collaborative Research Database, respectively. Daily measurements of serum creatinine and estimated glomerular filtration rate (eGFR) were assembled into rolling seven-day windows, and a deep learning algorithm called the Anomaly Transformer was trained to learn what typical and abnormal kidney function trajectories look like. An anomaly score was generated based on the creatinine-eGFR trajectories. Anomaly scores largely reflected the real clinical deterioration seen in patients, even in cases that did not yet meet conventional AKI diagnostic criteria, which was a phenomenon termed by authors as “hidden AKI”. Anomaly scores increased in a stepwise fashion across KDIGO kidney injury severity categories and were higher in patients who required dialysis or died within 24-96 hours. The model achieved an area under the receiver operating characteristic (AUROCs) of 0.83, 0.82, 0.81, and 0.80 for predicting kidney replacement therapy at 1, 2, 3, and 4 days, respectively. Mortality prediction had AUROCs of 0.62-0.66, which was a modest value and expected given that creatinine is a substandard predictor of death. Overall, the study suggests that AI can be trained to detect red flags in lab trajectory patterns and bring to attention clinically meaningful kidney injury before it is clinically apparent.

Beyond One-Size-Fits-All: Precision Mechanical Ventilation in ARDS

Wed, 03 Jun 2026 12:00:00 GMT

Acute Respiratory Distress Syndrome (ARDS) is a disease that has a lot of variability to presentation. Traditionally, ARDS is managed through mechanical ventilation and low tidal volume, however, these strategies do not address the underlying causes of inflammation, lung injury, and overall recovery. This is due to the complex nature of the disease involving a broad range of complex biological, anatomical, and mechanical mechanisms. As a result of this, many variables exist in deciphering the best treatment plan available. In this narrative review study, the authors aim to investigate numerous factors such as lung size, driving pressure, and mechanical power among others to look at current treatment plans and determine if heterogeneity can be predicted to make more tailored treatment plans for individual patients. The first step in this is viewing ARDS as a spectrum rather than a disease. This involves many different subphenotypes for ARDS, including mechanical subphenotypes, where the mechanical stress of the lung is measured and categorized based on existing data. Next, there are different biological subphenotypes characterized by inflammatory markers which lead to various outcomes such as shock and mortality. Additionally, there are radiological subphenotypes which identify specific lung patterns, and finally, there are etiological subphenotypes, which cateogrize the root cause of the problem, either the lungs or extrapulmonary. The idea is that the specific disease of the individual lies in the complex web of these subphenotypes, and a correlated treatment plans lies with it. One key treatment of ARDS is a low tidal volume, however, there exists errors in determining the recommended tidal volume through the use of calculators as they rely on factors such as predicted body weight compared to lung function. These limitations call for a shift towards more “energy-based” frameworks, which look at the existing pulmonary function and base it off that. The final goal of this study was to explore the future of the treatment of ARDS, especially by using a potential machine learning model to look for pattern recognition. Through the use of data from pulmonary volumes, flow, and patient effort, the goal would be to have a machine model predict the best treatment. This can prove important in determining when to use positive end-expiratory pressure and recruitment, which when used incorrectly can lead to further lung damage. This machine learning use alongside a clinician can lead to better overall outcomes for ARDS, including increased survival rates and recovery rates. This termed “precision ventilation” can lead to an overall increase in the treatment of ARDS through the use of complex pattern recognition of ARDS that treats it is as a spectrum, not a disease.

Development and Validation of an Interpretable Machine Learning Model for Predicting Distant Metastasis in Tongue Squamous Cell Carcinoma: A Multicentre Study

Sun, 31 May 2026 12:00:00 GMT

This retrospective cohort study aimed to develop and validate a machine learning model (MLM) to predict distant metastasis in postoperative tongue squamous cell carcinoma (TSCC). Distant metastasis in TSCC has a poor prognosis, with a median overall survival typically of about 10 months. Early identification of patients with high risk of distant metastasis is vital to optimizing postoperative surveillance and guiding treatment; however, current assessment of this risk fails to capture the multifactorial nature of tumor metastasis. Due to this gap, the authors aimed to develop MLMs that could consider the complexity of this pathology. This study analyzed 1,091 patients, with inclusion criteria being over 18 years old, confirmed TSCC, and no distant metastasis prior to surgery. Patients were excluded if they had prior head or neck treatment, received nonsurgical therapy, or if they died from non-TSCC causes within 2 years of treatment. Demographic variables and postoperative pathological features were collected as well. The primary endpoint was defined as the occurrence of distant metastasis during the two-year follow-up period after curative surgery. Six ML algorithms were tested: Light Gradient Boosting Machine (LightGBM), logistic regression (LR), random forest (RF), Extreme Gradient Boosting (XGBoost), k-Nearest Neighbours (KNN) and Elastic Net (ENet). Model performance was evaluated using the area under the receiver operating characteristic curve (ROC-AUC) and the SHAP method was used to rank importance of variables collected. Overall, the Enet model demonstrated the highest ROC-AUC (0.935) and PR-AUC (0.754) in both the internal and external validation cohorts. SHAP identified number of regional lymph node metastases, maximum tumor diameter, monocyte-to-lymphocyte ratio, neutrophil-to-lymphocyte ratio, depth of invasion, and lymphovascular invasion as key predictors of increased risk of distant metastasis. Additionally, SHAP identified lymphovascular invasion and poorly differentiated histological grade as variables that had a positive contribution to predicting distant metastasis risk, while well-differentiated tumors had a negative contribution.

Artificial intelligence for keratosis characterization and identification of lichenoid lesions in histological samples of oral leukoplakia

Fri, 15 May 2026 12:00:00 GMT

In this paper, the application of artificial intelligence (AI) and computational pathology techniques to automatically assess the histology of oral leukoplakia (OL) and oral lichenoid lesions (OLL) was investigated. OL and OLL are examples of potentially malignant disorders linked to the development of oral squamous cell carcinoma (OSCC). In this study, 240 hematoxylin and eosin–stained slides from 192 patients were analyzed retrospectively using deep learning-based algorithms for tissue segmentation, morphometry, and classification of lesions. The nnU-Net framework provided accurate tissue segmentation of epithelium, subepithelium, keratin, and nuclei, reaching Dice scores above 0.92 for the majority of tissues. Keratin thickness, epithelial histology, and nuclei density were quantified, enabling identification of keratinization types (orthokeratosis and parakeratosis), with an accuracy close to 92%. Furthermore, a Tiny Vision Transformer (TinyViT) was able to classify OL and OLL with lesion-level accuracy reaching nearly 93%.

Longitudinal Plasma Proteomics Reveals an Immuno-thrombotic Signature that Predicts Radiation Pneumonitis in Lung Cancer

Fri, 29 May 2026 12:00:00 GMT

Radiation pneumonitis is a toxicity in lung cancer therapy, which is often poorly predicted by current clinical models. The researchers identified a blood-based signature rooted in biological response to radiation that can be used for possible risk stratification. Using 267 longitudinal samples from 57 lung cancer patients, they conducted a prospective study focused on identifying protein trajectories associated with symptomatic radiation pneumonitis. The researchers identified dysregulated immuno-thrombotic axis as a central driver of radiation pneumonitis, characterized by significant enrichment in "Platelet Activation" and "Serpin" pathways. This signature was found to be a significant independent predictor and was distinct from markers of overall survival. This can be potentially utilized as a clinical tool for early risk stratification and personalized toxicity mitigation.

From Static Diagnosis to Dynamic Guidance : Evolution of Artificial Intelligence in Pediatric Neuroimaging

Fri, 29 May 2026 12:00:00 GMT

The field of pediatric neuroimaging has evolved in its reliance on artificial intelligence (AI) from it being a static to now a more dynamic tool. Image preprocessing steps such as bias-field correction, intensity normalization, and skull stripping enabled an early analytical approach to neuroimaging. Age-specific atlases emerged which were representative of pediatric and infant developmental stages. Progress in diffusion tensor imaging and resting-state functional MRI allowed for visualizing tracts and networks rather than discrete slices and signals. Automated ventricle segmentation allowed for prediction with greater accuracy. Current advancements in deep learning include Convolutional Neural Networks and Vision Transformers. Variational auto encoders and memory-augmented networks improve the ability to detect anomalies on neuroimaging. Generative AI frameworks like generative adversarial networks and diffusion-based frameworks aid in addressing fundamental constraints such as data scarcity and limitations in modality by various methods such as harmonizing datasets and supplementing for missing imaging sequences. AI tools such as Neuro-GPT use an autoencoder approach to review EEG data which is especially helpful for seizure classification. AI allows for immersive visualization techniques such as virtual reality-based rehearsals and mixed reality systems along with computer vision and robotic platforms to enhance and optimize surgical performance. As AI continues to advance so too will it expand its reach in the field of pediatric neuroimaging.

Fully automated, deep learning, cardiac CT-based multimodal network for cardiovascular risk stratification in high-risk perioperative patients

Tue, 26 May 2026 12:00:00 GMT

This article focuses on creating a multimodal deep learning system based on patient demographics and risk factors to identify at risk patients for major adverse cardiac events (MACE). While other clinical risk assessment tools exist to screen for high risk patients for a major cardiac event, these scores often underestimate specific populations, leading to underrepresentation. The aim of this deep learning model is to more accurately represent each patient through the use of data obtained from a coronary computed tomography angiography (CCTA) as well as other data points such as CAD-RADS score which is an AI tool developed to analyze the results of a CCTA, demographics, past medical history, medication, and pre-operative eGFR. This study included 639 patients who were already previously determined to be at risk of a major cardiac event and had atherosclerotic disease. The data that was gathered from the CCTA, such as lumen occlusion, was categorized based on CAD-RADS scores. Overall, 80% of the data was used to train the model, while the other 20% was used to test the accuracy of the model. From the population, over the next 30 days, 45 patients experiences a MACE, and from this, 47% had 3 or more CV risk factors. When only patient data and RCRI was used, the deep learning model had an area under the curve of .62, while expert analysis and traditional CAD-RADS scores had an area under the curve of .61 and .63 respectively. When combining other data points such as patient data, CAD-RADS, and structural data from the CCTA scan, the deep learning model had an AUC of .82, higher than the combination of patient data and CAD-RADS score alone. The most important data points were found to be BMI, age, eGFR, and CAD-RADS score. The overall findings indicate that while the CAD-RADS tool is comparable to expert analysis, the aid of a deep learning model helps it to become more accurate for very little extra cost. The use of the data obtained from the CCTA such as left ventricle function was found to also be a major piece of information, which is not included in the traditional CAD-RADS tool, which is where this deep learning model excels and aids this previously existing tool. Through the use of multiple tools, a more personalized risk assessment is generated, leading to better overall outcomes from surgery, while avoiding more negative outcomes that can result in permanent health issues and/or death. However, this deep learning model was also seen to overestimate the risk in some individuals, leading to delays or cancellations of surgeries, which can ultimately cause more harm than good.

Climate and socioeconomic factors drive heterogeneous dengue risk escalation in the Chinese population

Fri, 22 May 2026 12:00:00 GMT

This study, presented by Guang and colleagues, examined how climate and socioeconomic changes may influence current and future dengue fever risk across China and whether these effects differ geographically. The researchers developed a geographically explainable artificial intelligence (GeoXAI) model using a hazard–exposure–vulnerability framework. They integrated environmental variables, population distribution, and socioeconomic indicators to estimate dengue risk under future climate and development scenarios through 2050 and 2100. The study found that dengue risk is projected to increase across China but unevenly between regions. Rising minimum winter temperatures emerged as the strongest predictor of dengue expansion, while population density and urbanization also contributed to increasing risk. High-risk areas are expected to expand northward beyond historically affected regions, with southwestern and southeastern China showing particularly notable increases and additional relative growth occurring in lower-risk northwestern areas. Under the highest emissions scenario (SSP585), overall dengue risk increased substantially by both mid- and late-century.

Research priorities for improved pandemic and epidemic intelligence

Wed, 20 May 2026 12:00:00 GMT

This study conducted a global research priority-setting exercise for pandemic and epidemic intelligence using a modified Child Health and Nutrition Research Initiative approach. Led by the WHO Hub for Pandemic and Epidemic Intelligence in collaboration with global partners, the process involved 213 international experts and resulted in 23 prioritized research statements across 8 thematic domains: AI and technological advances, data preparedness, quality standards, analytical frameworks, multisectoral approaches, community-centred approaches, governance, and evidence-to-policy translation. The top priorities identified were developing multipathogen point-of-care diagnostics, improving multidisciplinary data integration for epidemic detection, and establishing rapid evaluation methods for public health interventions. Despite broad expert input, participation skewed toward high-income countries and European institutions.

Integrating the interpretable machine learning Score For Emergency Risk Prediction (SERP) with emergency department triage to better predict 30-Day mortality

Mon, 11 May 2026 12:00:00 GMT

Wong Qi Feng et al. investigated whether integrating a machine learning-based Score for Emergency Risk Prediction (SERP) into the Patient Acuity Category Scale (PACS) could improve ED triage accuracy. Using data from over 190,000 ED visits, two enhanced models were developed: one that adjusted triage in both directions: upgrading and downgrading patients based on 30-day mortality risk, and one that exclusively upgraded high-risk patients. Both models outperformed standard PACS in predicting 30-day mortality, with higher AUC values reflecting superior discriminative accuracy. The bidirectional model demonstrated the greatest overall clinical benefit. These findings suggest that augmenting traditional triage systems with ML has the potential to improve risk stratification and support more accurate, data-driven clinical decision-making in emergency care.

Quantifying immune dysregulation in pneumonia and sepsis with a parsimonious machine-learning model: a multicohort analysis across care settings and reanalysis of a hydrocortisone randomised controlled trial

Mon, 04 May 2026 12:00:00 GMT

Michels et al. developed and validated a machine-learning (ML) tool to quantify immune dysregulation in community-acquired pneumonia (CAP) and sepsis patients. This multicohort analysis and reanalysis of a randomised controlled trial placed 398 CAP patients on a continuum of three Dysregulated Immune Profile stages (DIP1–3) and a continuous score (cDIP) using 35 plasma biomarkers. Researchers then built a parsimonious three-biomarker machine-learning tool (procalcitonin, soluble TREM-1, and IL-6) that was validated across five external cohorts (total n=1191). Higher dysregulation (DIP/cDIP) was associated with increased mortality (p<0.0001) and secondary infections (p=0.0005), regardless of clinical severity. Additionally, in a post-hoc reanalysis of the CAPE COD randomized trial in severe CAP, hydrocortisone’s survival benefit appeared limited to patients classified as severely dysregulated (DIP3 and above a cDIP threshold, p=0.011) alongside faster resolution of dysregulation. Stratification by clinical severity failed to demonstrate a similar effect.

Machine learning prediction of hospitalization outcomes in critically Ill emergency department patients transported by ambulance: A Retrospective Single Center Cohort Study

Mon, 04 May 2026 12:00:00 GMT

Durmuş and Eke Kurt investigated whether machine-learning models can predict the need for hospitalization among critically ill emergency department (ED) patients transported by ambulance. In this retrospective single-center cohort study, demographic, clinical and early laboratory parameters in 2,338 adult patients were used to develop machine learning models, These included logistic regression, random forest, and gradient boosting models, and were evaluated using five-fold cross validation. The primary outcome was defined as admission vs. no admission. Model performance was assessed using receiver operating characteristic area under the curve (ROC AUC), accuracy, sensitivity, and specificity. The study found that 37.0% of patients required hospitalization, and the random forest model showed the strongest performance (cross-validation ROC AUC 0.850; test ROC AUC 0.776). Important predictors included troponin, altered mental status, lactate, age (including ≥65 years), creatinine, leukocyte count, and pH. Meanwhile, the random forest vs gradient boosting ROC AUC difference was not statistically significant (mean difference 0.008; p=0.052).

Evaluating large language models for specialist referral triage in primary care: a quantitative study using otolaryngology scenarios

Mon, 04 May 2026 12:00:00 GMT

Hack et al. assessed whether two artificial intelligence (AI) models could effectively and reliably assist primary care providers in deciding when and where to refer patients to ear, nose, and throat (ENT) specialists. Researchers presented 16 simulated clinical scenarios (such as sudden hearing loss or neck masses) to both ChatGPT and Gemini in two formats: structured clinical prompts written by doctors, and informal queries written in patient language. Outputs were rated by five ENT specialists and ten lay reviewers based on appropriateness, clarity, safety, and usefulness. The performance of each model was compared as well as the agreement among reviewers and impact of prompt structure on output quality was assessed. The study found that when given structured clinical prompts, both models produced safe and clinically appropriate recommendations, with no statistically significant differences between models across any assessment metric. Notably, performance of both models decreased when given informal prompts. Reviewer agreement was strong (Intraclass Correlation Coefficient (ICC) 0.66–0.79) suggesting high reliability.

Machine learning improves prediction of pulmonary thromboembolism and reduces unnecessary computed tomography scans in the emergency department

Mon, 04 May 2026 12:00:00 GMT

Yoon et al. investigated whether machine learning (ML) models could more accurately predict pulmonary thromboembolism (PTE) in emergency department (ED) patients than the revised Geneva score. In this single-center study, data from 2,525 patients with suspected PTE who underwent computed tomography pulmonary angiography (CTPA) were analyzed and divided into training and test sets. Six ML models were then compared against the revised Geneva score using the area under the receiver operating characteristic curve (AUC). Permutation was applied to assess the importance of individual variables. The study found that the XGBoost achieved the highest AUC of 0.814 (95% confidence interval [CI]: 0.759–0.862), while all models significantly outperformed the revised Geneva score (AUC: 0.622, 95% CI: 0.563–0.675). The XGBoost model could reduce unnecessary CTPA imaging by up to 33.2 % at 90% sensitivity (p<0.001). In addition, D-dimer and activated partial thromboplastin time (aPTT) were identified as the most critical predictors across all models.

Economic value of AI-based MRI triage for Parkinson’s disease: a cost-benefit study in South Korea and the United States

Mon, 04 May 2026 12:00:00 GMT

Kim et al. evaluated the economic value of an AI-based MRI triage strategy for the early detection of Parkinson’s Disease (PD) using a cost-benefit analysis conducted in South Korea and the United States. A simulation model was constructed to represent diagnostic and cost pathways of individuals within a theoretical national cohort. Based on national demographics and reported incidence rates, the cohort comprised 48,888 adults aged 65 years and older with clinically suspected PD in South Korea and 90,000 in the United States. The AI model utilized was Heuron IPD (Idiopathic Parkinson’s Disease), a commercially available software designed for automated nigrosome-1 evaluation, which generated bilateral nigral hyperintensity maps, standardized Z-scores, and volumetric measures, providing quantitative support for IPD assessment. Two diagnostic strategies were compared: conventional Positron Emission Tomography (PET) imaging for diagnostic confirmation versus an AI-based MRI model used to determine the need for PET imaging. This approach facilitated early detection in high-confidence AI-MRI cases while reducing unnecessary PET procedures. Each cohort member was assigned to one of 24 unique patient types defined by binary variables: 1) presence or absence of PD, 2) presence or absence of economic barriers to PET access, 3) MRI detection outcome, 4) PET detection outcome, and 5) AI utilization. Cost-benefit analyses for each patient type were simulated along their respective diagnostic pathways, accounting for early diagnosis, misdiagnosis, delayed diagnosis, and PET avoidance. In South Korea, the net benefit was estimated at $9.29 million at 30% AI adoption, rising to $30.97 million at full adoption, with a cost-benefit ratio of 1.48. In the United States, the net benefit reached $75.96 million at 30% adoption and $253.19 million at full adoption, with a cost-benefit ratio of 1.36. The break-even AI unit cost was approximately $226 in South Korea and $1,506 in the United States, suggesting economic viability despite high implementation costs. Long-term projections based on PD prevalence trends and population aging in South Korea estimated an annual net benefit of $2.1 million in 2025, projected to grow to $264.7 million by 2050.

Unveiling the Impact of Occupational Therapy on Acute Care Outcomes: A Machine Learning Approach

Mon, 04 May 2026 12:00:00 GMT

Lee et al. conducted a retrospective cohort study to characterize optimal occupational therapy (OT) service patterns associated with improved health outcomes in orthopedic patients, using decision tree models. Eligible patients met all of the following criteria: 1) admission to and discharge from one of eleven Cleveland Clinic hospitals between January 1, 2017 and December 1, 2021; 2) receipt of at least one OT session; 3) length of stay (LOS) of less than 30 days; and 4) an orthopedic surgeon listed as their primary provider. OT service patterns were defined by 1) total number of OT sessions; 2) total OT session time; and 3) the proportion of hospital days that included an OT session. Three outcome variables were assessed: 1) 30-day readmission to a Cleveland Clinic hospital; 2) discharge to a home setting; and 3) minimal detectable change in daily activities at discharge, as measured by the Activity Measure for Post-Acute Care (AM-PAC) 6-Clicks daily activity short form, scored by occupational therapists at each visit. The study comprised a total of 36,300 orthopedic patients. Prediction models demonstrated strong performance (AUC: 0.71-0.94; specificity: 0.66-0.91; sensitivity: 0.67-0.92; accuracy: 0.69-0.91). The pruned decision tree model for 30-day readmission identified the following risk patterns: 1) patients with LOS less than 3 days, a Charlson Comorbidity Index (CCI) of 4 or greater, and OT coverage on less than 67% of hospital days were at increased readmission risk; 2) patients with LOS less than 3 days, CCI below 4, and OT coverage below 39% were similarly at elevated risk; and 3) patients with LOS of 3 days or more and CCI below 2 exhibited complex multifactorial risk patterns. Thresholds associated with minimal detectable change in daily activities included: 1) three or more OT sessions; 2) OT sessions on more than 82% of hospital days; and 3) a cumulative OT session time of 39 minutes or more.

Human vs. artificial intelligence in medical charting: a comparative study in the simulated emergency medicine context

Sat, 02 May 2026 12:00:00 GMT

Lipinski et al. evaluated whether AI scribes improve documentation efficiency in emergency medicine compared to traditional physician voice dictation. In this study, 16 physicians and residents completed standardized patient encounters recorded by the AI scribe Mutuo, then documented the same encounters using Dragon dictation. Participants subsequently edited the AI-generated notes, enabling comparison across key factors including documentation time, editing burden, and user preference. Editing AI-generated notes was significantly faster than dictation (1:15 vs. 2:05 minutes), representing a 39% reduction in documentation time. Notably, 69% of note sections required no edits, and most changes were minor, typically limited to a few words or a single sentence, with edits most frequently occurring in the history and plan sections. Furthermore, 63% of participants preferred AI-generated notes overall. These findings suggest that AI scribes can meaningfully reduce documentation burden with minimal editing required, supporting their potential to improve physician efficiency, though real-world clinical validation remains necessary.

Prediction of Respiratory Decompensation in Patients Receiving Home Mechanical Ventilation: Machine Learning Model Development and Validation Study

Mon, 30 Mar 2026 12:00:00 GMT

Casado et al. aimed to determine whether a machine learning-based model could predict respiratory decompensation events in patients using home mechanical ventilation based on device usage data. Data collected included device usage patterns, compliance with utilization, mask leakage, and ventilator settings. Respiratory decompensation was defined as emergency department visits or hospitalizations due to acute respiratory worsening. Multiple machine learning models were trained to recognize the parameters and were evaluated by 10-fold cross validation. Logistic regression possessed the highest recall (mean 0.94, SD 0.06, 95% CI 0.90-0.98) despite having limited accuracy (mean 0.60, SD 0.05, 95% CI 0.56-0.64). The random forest classifier accomplished the most superior balance across the measures (accuracy: mean 0.66, SD 0.10, 95% CI 0.59-0.73; recall: mean 0.78, SD 0.15, 95% CI 0.67-0.89; F1-score: mean 0.70, SD 0.10, 95% CI 0.63-0.77). SHAP analysis indicated that device usage rates, mask leakage levels, and adherence patterns in the week preceding a decompensation event were the most important predictive features.

Comparative feasibility of reasoning and non-reasoning large language models for gynecologic cancer emergency care

Fri, 06 Mar 2026 12:00:00 GMT

Kim et al. sought out to determine the efficacy of large language models, more specifically the accessibility of the generative pretrained transformer (GPT)-4o and o3-mini-high and its role in assisting physicians diagnostically with patients diagnosed with gynecologic cancer. 15 cases were selected to evaluate these large language models, and their performances were compared to two gynecologic oncology fellows, two obstetrics and gynecology residents. They assessed the cases based on diagnostic measures utilized, conclusions drawn, diagnoses accuracy, and treatment options. Responses were assessed based on precision and speed. GPT-4o and o3-mini-high both responded more quickly than the physicians and were determined to be more accurate, with mean accuracy scores of 3.55 (95% CI, 2.98-4.10; P < 0.001) and 3.05 (95% CI, 1.98-3.88; P < 0.001), respectively. These findings suggest that large language models may have clinical utility in diagnostic evaluation when used cautiously and with appropriate oversight.

Shifts in emergency physicians' attitudes toward large language model-based documentation: a pre- and post-implementation study

Thu, 05 Mar 2026 12:00:00 GMT

Lee et al. conducted a prospective longitudinal survey study examining the impact of a large language model (LLM) assistant on emergency department (ED) discharge notes documentation and provider attitudes toward artificial intelligence integration in clinical practice. In November 2024, the Y-KNOT-EDN (Your-Knowledgeable Navigator of Treatment-Emergency department Discharge Note assistant), an in-hospital LLM-based clinical documentation system, was deployed at Severance Hospital in Seoul, South Korea. The system automatically generates a draft discharge note from the patient’s electronic health record (EHR), which physicians then review, modify, and cosign to maintain full clinician oversight. Validated surveys were administered to eight emergency attending physicians at three time points: before implementation (T1), three days post-implementation (T2), and five weeks post-implementation (T3). On average, participants authored 10.9 LLM-assisted discharge notes during the study period. Four provider concerns: loss of control, worsening of patient care, generation of impersonal drafts, and legal and ethical issues, declined significantly over time (p = 0.002, 0.004, 0.010, and 0.028, respectively). Four additional concerns: generation of false information, privacy, impairment of physician reasoning, and data bias, also trended downward, though without reaching statistical significance (p = 0.089, 0.143, 0.156, and 0.268, respectively). Perceived documentation workload declined significantly by 37% from T1 to T3 (p = 0.040). Analysis of documentation time revealed a marked reduction in documentation time, from a mean of 127.5 seconds for manual notes at T1, to 42.8 seconds for LLM-assisted notes at T3 (p = 0.002). Qualitative responses to open-ended questions indicated broad agreement that the LLM assistant improved efficiency by offloading nonclinical administrative documentation tasks. While participants recognized generally high accuracy with fewer errors than anticipated, they identified notable gaps including insufficient patient-specific context for complex cases, absent explanations for abnormal test results, and omission of imaging and laboratory findings. Blinded physician review of the LLM-assisted discharge notes demonstrated high performance across five evaluated domains: fluency, coherence, relevance, safety, and consistency. Mean domain scores on a five-point scale (1 = lowest, 5 = highest) were 4.55, 4.60, 4.72, 4.73, and 4.78, respectively, indicating consistently strong performance.

An Algorithm to Avoid Missed Bowel Injuries in Blunt Abdominal Trauma Patients

Thu, 05 Mar 2026 12:00:00 GMT

Clement et al. developed and validated the point-based scoring Algorithm to Avoid Missed Bowel Injury (AMBI) to help clinicians identify bowel injuries more quickly and accurately in patients with blunt abdominal trauma (BAT). Over a 9-month period, the algorithm evaluated 141 BAT patients at a Level 1 trauma center. The AMBI awarded points to imaging, physical examination and laboratory findings that suggested bowel injury. The study found that for patients evaluated using the AMBI algorithm, the median hospital length of stay (HLOS) decreased significantly from 15 to 6 days (P = 0.0152) compared with pre-AMBI implementation. Patients with an AMBI score ≥1.5 were significantly more likely to have a confirmed bowel injury (odds ratio (OR): 262; 95% CI, 31.7-2785; P < 0.0001) compared with those who had an AMBI score <1.5. Furthermore, the algorithm’s sensitivity was greater than two previously developed scoring models to detect bowel injury.

Machine learning for risk stratification in the emergency department (MARS-ED) study protocol for a randomized controlled pilot trial on the implementation of a prediction model based on machine learning technology predicting 31-day mortality in the emergency department

Mon, 23 Feb 2026 12:00:00 GMT

Van Dam et al. conducted a clinical trial to explore the clinical implications of a machine learning-based prediction model. The RISK index predicts 31-day mortality risk based on laboratory tests and clinical data. Prior studies have shown that the RISK index outperforms internal medicine physicians in mortality prediction. Participants were randomly assigned to control or intervention groups. Both groups received standard care, with the intervention group's physicians additionally receiving RISK index predictions. Attending physicians in the intervention group could incorporate RISK index results into their clinical decision-making at their discretion. This is a study protocol; results have not yet been published.

Brief narrative interventions for adults with chronic illness or psychosocial distress: a scoping review protocol

Fri, 22 May 2026 12:00:00 GMT

Narrative interventions are an under-recognized therapeutic modality for individuals experiencing chronic medical conditions and/or psychosocial stressors. This review protocol provides a framework for understanding existing narrative interventions, which are defined as brief intervention sessions involving the patient expressing personal stories of distress. The goal of this therapeutic technique is to gradually reduce distress through narrative discussion and empathetic listening. The proposed study protocol plans to address the characteristics of narrative interventions including duration, content, and structure. It determines targeted patient populations by assessing patients' clinical diagnosis, psychosocial distress states, and cultural factors. The protocol also aims to determine how to measure and define targeted outcomes of narrative interventions. The protocol authors plan to use the U-M GPT (GPT-4.1, University of Michigan) large language model to extract narrative data, which will be followed by an assessment of the extracted data by a human reviewer.

Moderate Predictive Ability of Machine Learning for Achievement of Minimal Clinically Important Difference for the Pain and Healthy Utility Scores after Hip Arthroscopy: Analysis From the Femoroacetabular Impingement RandomiSed Controlled Trial (FIRST) and Embedded Prospective Cohort

Mon, 18 May 2026 12:00:00 GMT

Femoracetabular impingements (FAI) are a common cause of hip pain in young adults. This can lead to osteoarthritis at a young age, and the treatment often includes arthroscopic surgery to address the labrum. A trial termed FIRST, Femoracetabular Impingement RandomiSed controlled Trial, was a randomized controlled trial comparing arthroscopic osteochondroplasty to arthroscopic lavage and looked at the outcomes between these groups. This trial used pain scores as an outcome, VAS and EQ-5D, but more recently, there has bene a shift towards using MCID, the minimally clinical important difference, as a metric rather than a pain score. This study aims to train a predictive model to interpret and predict MCID scores following the treatment of FAI using data gathered from the FIRST trial. The FIRST trial included patients aged 18-50 with a total of 220 patients, and the FIRST trial used both the VAS and EQ-5D pain scale, which both range from 0-100, 100 being the best, 0 being immense pain. 110 more patients were included that were not a part of the FIRST trial, but the same markers were identified and tested. In this study, four models were trained to predict the MCID at 6 months and 12 months post-operative. 70% of the data was used for training the models, while 30% of the data was used to test the accuracy of the models. At 6 months using the VAS pain scale, the models XGBoost and logistic regression scored the best, with an AUC of .653 and .623. At 12 months, logistic regression and LASSO performed the best with an AUC of .777 and .762. For the EQ-5D pain scale, logistic regression scored the best AUC for 6 months and LASSO at 12 months. Interestingly, different models had different important data points that lead them to their prediction, where some included Outerbridge classification and age while others included traction time and BMI. In this study, logistic regression performed similarly to the learning models tested and had a similar AUC (accuracy). From this, it can be concluded that machine learning still has progress to make, and the authors note that with more data available, machine learning tends to fair better long term. However, several important key data points resulted from this study surrounding important markers in reaching MCID, such as male sex, younger age, and increased traction time, all contributing to better overall recovery.

DNA methylation biomarkers associated with early gastric cardia carcinogenesis

Fri, 22 May 2026 12:00:00 GMT

Gastric cardia cancer is a highly aggressive malignancy with consistently poor prognoses, underscoring the need for better characterization of its early molecular alterations. To address this gap, the researchers developed a machine learning model trained on 69 samples to validate tissue-based DNA methylation markers associated with early precancerous and neoplastic lesions of the gastric cardia. Their results revealed that epigenetic dysregulation increased as the cancer progressed, with the most pronounced changes occurring in precancerous lesions. Notably, the model accurately distinguished both precancerous gastric cardia lesions and GCC from normal tissue within the integrated dataset. Building on these findings, Liao et al. identified EDNRB and SALL1 promoter hypermethylation as promising tissue-based biomarkers of early neoplastic transformation, providing a stronger foundation for understanding gastric cardia precancerous lesions and cancer.

Prediction model for additional procedure requirement in flexible ureterorenoscopy using explainable artificial intelligence

Fri, 15 May 2026 12:00:00 GMT

This study developed machine learning models to predict which patients will need an additional procedure after flexible ureterorenoscopy (f-URS). The analysis included 656 patients treated between 2015 and 2025, with 180 patients (27.4%) requiring further treatment such as repeat f-URS, ureteroscopy, shock wave lithotripsy, or percutaneous nephrolithotomy. The authors tested 14 machine learning models and used feature selection methods including Boruta, LASSO, and ElasticNet. They also used explainability tools such as SHAP and LIME, to show which variables mattered most. The strongest predictor was the ureteropelvic junction–pelvis angle (UPJ-PA). Patients below 110° had a much higher intervention rate than those above it (84.3% vs 2.8%). Explainability analyses consistently ranked UPJ-PA as the most important feature, followed by FANS-UAS use and access sheath size. Robustness testing also showed that the model stayed stable even when noise was added to UPJ-PA measurements.

Estimate renal cell carcinoma recurrence rates using electronic health records

Fri, 15 May 2026 12:00:00 GMT

Hou et al present a study evaluating whether electronic health record (EHR) data could be used to accurately estimate recurrence rates in patients with renal cell carcinoma (RCC) after treatment. The researchers developed and tested an automated method that used clinical records, imaging information, pathology reports, and follow-up data to identify cancer recurrence events. They found that the EHR-based approach was effective in detecting RCC recurrence and produced estimates comparable to traditional manual chart review methods while requiring less time and effort. Overall, the study demonstrates that automated EHR analysis may provide a reliable and scalable tool for monitoring RCC outcomes in large patient populations.

Non-invasive profiling of the tumour microenvironment with spatial ecotypes

Sun, 17 May 2026 12:00:00 GMT

This study used a machine-learning framework to identify spatial ecotypes, which are recurring multicellular ecosystems within a tumor’s microenvironment defined by spatial organization of cells and their abundances. The researchers analyzed millions of single-cell and spatial gene expression profiles across different cancers and identified nine ecotypes with several conserved, predictive features. The authors also developed tools that can detect these ecotypes not only in tumor tissue, but also from cell-free DNA extracted from plasma. Some ecotypes were linked to stronger immune responses, while others were connected to tumor growth and worse clinical outcomes.

Multiple Targets and Pathways Non-Monotonically Regulate Lung Squamous Cell Carcinoma Migration in Response to Phthalates

Tue, 12 May 2026 12:00:00 GMT

This study investigated how phthalates, widely used plasticizer chemicals, may influence the progression of Lung Squamous Cell Carcinoma. Using network toxicology, machine learning, molecular docking, and laboratory experiments, the researchers identified four key genes linked to phthalate-associated LUSC: ANXA5, CDK4, MMP1, and SRD5A1. These genes are involved in cancer-related processes such as cell growth, invasion, hormone metabolism, and immune regulation. A diagnostic model based on these four genes showed strong predictive accuracy with an AUC of 0.96, suggesting that the gene signature may serve as a reliable biomarker for LUSC. Molecular docking simulations also suggested that benzyl butyl phthalate (BBP), a common phthalate, could bind directly to proteins encoded by ANXA5 and SRD5A1, supporting the possibility that phthalates interfere with cellular signaling pathways involved in cancer progression. The key experimental finding was that low, non-toxic concentrations of BBP promoted the migration of lung squamous carcinoma cells in both H226 and NCI-H1703 cell lines, which may indicate an increased metastatic potential. At 0.1 μM, BBP significantly increased the expression of ANXA5 and SRD5A1 and enhanced cell migration, while higher doses either reduced or reversed some of these effects. This “low-dose activation and high-dose inhibition” pattern is characteristic of endocrine-disrupting chemicals. The study also found that phthalate-associated gene changes were linked to altered immune cell infiltration, particularly increased neutrophils and M2 macrophages, which are associated with tumor progression. Overall, the researchers concluded that BBP may promote LUSC progression through multiple pathways involving cell migration, steroid metabolism, membrane signaling, and immune microenvironment changes, although further studies are needed to confirm direct causal effects in humans.

Using tree-based ensemble methods to produce a population-based mortality risk score in Ontario, Canada

Sat, 16 May 2026 12:00:00 GMT

This retrospective, population-based single-center study asked whether tree-based machine learning models could improve prediction of 1-year all-cause mortality in the general adult population compared with traditional statistical methods, using ensemble tree-based machine learning models to perform mortality risk prediction. Researchers analyzed 12,080,801 adult Ontarians alive as of January 1, 2022, using administrative healthcare data, including hospital discharge records, ambulatory care encounters, physician billing claims, laboratory data, cancer registry data, and long-term care assessments collected from provincial databases in Ontario, Canada. Predictors were demographic characteristics, Charlson-style comorbidities, healthcare utilization measures, physician billing indicators, and Activities of Daily Living (ADL) scores. Preprocessing included retaining continuous variables without scaling, assigning missing ADL scores to a separate category, and transforming categorical variables using one-hot encoding. The models tested included logistic regression, random forest, ExtraTrees, AdaBoost, gradient boosting, Newton boosting/XGBoost-style models, CatBoost, and explainable boosting machines, compared primarily against standard logistic regression. The best-performing model was CatBoost, which achieved an AUROC of 0.933, PR-AUC of 0.281, Brier score of 0.0083, and Integrated Calibration Index equivalent (ICIeq) of 0.0003, outperforming logistic regression, which achieved an AUROC of 0.926 and PR-AUC of 0.256. External temporal validation in a 2024 cohort maintained an AUROC of 0.933 with a slightly lower PR-AUC of 0.254 and modest calibration drift. The analysis showed that 1.0% of the cohort died within one year, while 51.3% were female and the mean age was 49.0 years. Diabetes without complications was the most prevalent comorbidity at 4.3%, followed by primary cancer at 2.1% and diabetes with complications at 2.0%. Mortality rates were highest among patients receiving palliative care (36%), residing in long-term care (27%), experiencing dementia (26%), pressure injury (25%), delirium (21%), metastatic cancer (20%), and stage 4–5 chronic kidney disease (16–17%). Secondary analyses included calibration assessment, Kaplan-Meier survival stratification, explainability modeling using feature importance and permutation feature importance, marginal effect estimation, sensitivity analyses evaluating alternate cancer and kidney disease definitions, and comparison of 70/30 train-test splitting with 10-fold cross-validation. Additional results demonstrated that age, outpatient hospital visits, sex, healthcare utilization frequency, ADL score, palliative care status, and hospitalization burden were among the strongest predictors of mortality. Marginal effect analyses showed that palliative care was associated with an absolute mortality risk increase of 4.03% and a relative marginal effect of 438%, while moderate-to-severe liver disease and metastatic cancer increased mortality risk by 2.60% and 1.53%, respectively. Explainable boosting machine analyses further demonstrated that age interacted strongly with multiple clinical variables in predicting mortality risk. Limitations include reliance on administrative coding data, incomplete disease granularity, such as cancer staging and subtype information, limited computational resources preventing exhaustive hyperparameter tuning, and potential algorithmic bias introduced by healthcare utilization variables that may underrepresent underserved populations. Although external validation was performed using a later Ontario cohort, subgroup fairness analyses were limited, and advanced bias mitigation techniques were not implemented. Findings reflect predictive discrimination and calibration within large-scale administrative datasets rather than direct improvements in patient-centered outcomes and therefore do not establish clinical efficacy.

AI-Based Medical Decision Support: Exploring the Data Gap

Fri, 15 May 2026 12:00:00 GMT

AI is increasingly being used to aid physicians in clinical decision making, however its real world impact is still limited. The biggest issue lies in AI's holistic efficacy stemming from input data used in training. This data is often fragmented and incomplete. Most AI systems rely on electronic medical records(EMRs), which are designed for billing rather than detailed medical understanding. Due to this, important patient information is often missing, scattered, or unrecorded in a consistent manner. Even advanced tools like natural language processing and large language models can only work as well as the input data. Medicine also lacks continuous, real time physiological data, especially for muscle and joint health where measurements are subjective. Future progress in medical AI will depend on better data collection, especially through wearable technologies such as sensors and continuous monitoring technologies.

From prediction to practice: closing the translation gap in artificial intelligence for anesthesia

Fri, 15 May 2026 12:00:00 GMT

As advancements in artificial intelligence (AI) continue to emerge in medicine, there still exists a gap between the promising potential of AI in the field of anesthesiology and its actual application in the clinical setting. This review addresses the causes of this translational gap in order to provide practical solutions to support the integration of AI in anesthesiology. It is often the case that contemporary AI models are trained on homogenous datasets that overlook rare patient presentations, but these are the clinical scenarios in which accuracy becomes crucial. Despite growing findings in the applicability of AI in anesthesia, there still exists a lack of robust, large-scale studies to demonstrate clinical relevance. This paired with potential ambiguity when it comes to understanding how AI models operate could reduce anesthesia providers' confidence in using these tools to aid in clinical decision making. Strategies to bridge this translational gap include conducting rigorous prospective studies to validate the practicality of AI tools in anesthesia delivery. Collaboration across disciplines should be encouraged to seamlessly integrate AI into existing clinical workflow. The future of AI in anesthesia is projected to involve proactively incorporating AI tools into anesthesia platforms and using AI to combine sources of information from various data collecting methods to provide patient-centered anesthesia.

Prognostic value of stress hyperglycemia ratio, hemoglobin glycation index, and glycemic variability for postoperative atrial fibrillation: a machine learning-based prediction model

Fri, 15 May 2026 12:00:00 GMT

This article presents a good case for machine learning to help reduce postoperative atrial fibrillation (POAF) in cardiac surgery patients. After analyzing 2,177 patients, they found that variables such and stress hyperglycemia ratio (SHR), glycemic variability (GV), and hemoglobin glycation index (HGI) were linked to higher rates of POAF. They indicated that SHR and HGI were independent predictors of POAF and that SHR and GV had non-linear relationships with POAF, suggesting that once these values got to a certain level, risk significantly increased. Importantly higher values of HGI displayed a protective effect.

Clinic-first sepsis recognition in the ICU: a proteomics-guided, parsimonious model with independent validation

Tue, 12 May 2026 12:00:00 GMT

This article aims to define key variables that contribute to potential sepsis infections and integrate this data into a learning model in hopes to alert hospital personnel of these patients. To do this, the authors looked at 55 patients (38 with sepsis and 17 critically ill but non-septic) within 48 hours of critical illness onset and looked at specific plasma proteins through proteomics. The data used includes lab measurements, infection characteristics, and other clinical data and scores. The blood was obtained within 4 hours of enrollment, and the first batch was used to analyze inflammatory markers while the second batch was used to interpret specific proteins. The protein expression was mapped using a heat map and the 12 key septic inflammatory markers can clearly be seen and identified (such as LILRA3, LCN2, and B2M). The final results indicated that sepsis can be predicted using proteins along with other key data such as creatinine. The area under the curve was found to be .73 from the machine learning model, where .5 is random chance, which indicates this learning model was clinically useful in determining patients who are at a high risk for sepsis. There are many markers that have been known to be related to sepsis, however, new inflammatory markers, such as IGFBP6, have yet to be fully explored and their mechanism of action in precipitating sepsis. Overall, using inflammatory markers and other clinical data can yield scores high enough to predict patients who are at a high risk of going into septic shock which can aid in giving physicians and healthcare teams a forewarning about potential life-threatening complications.

Enhancing the functionality of soft continuum robots for minimally invasive and endoluminal interventions: a review

Fri, 27 Mar 2026 12:00:00 GMT

The paper by Bacchetti et al. (2026) provides a comprehensive review of the development and clinical potential of soft continuum robots (SCRs) for minimally invasive and endoluminal medical procedures. The authors argue that while soft robotic systems offer superior flexibility and compliance compared to traditional rigid instruments, their widespread clinical adoption is currently limited by challenges in functional integration, miniaturization, and reliability. Soft continuum robots are inspired by biological structures such as octopus arms and elephant trunks, enabling them to navigate complex and tortuous anatomical pathways with continuous curvature rather than discrete joints. The review evaluates key technical components required to make these robots clinically viable, including actuation mechanisms (tendon-driven, fluidic, smart-material, and magnetic systems), embedded sensing technologies for shape and force feedback, and advanced modeling frameworks, such as Cosserat rod theory, for real-time control. The authors emphasize that integrating these systems into a single compact device—capable of navigation, sensing, and therapeutic intervention—remains the major engineering bottleneck. Additionally, issues such as sterilization, durability of soft materials, and regulatory approval processes continue to slow the translation of laboratory prototypes into clinical practice.

Restoring rapid natural bimanual typing with a neuroprosthesis after paralysis

Fri, 20 Mar 2026 12:00:00 GMT

An interruption in communication between the brain and muscles can result from spinal cord injuries, which oftentimes end in paralysis. Due to paralysis, individuals may lose the ability to type or communicate quickly or efficiently. Brain-computer interfaces can decode neural activity and transmit information to the computer. There are previous limitations to brain-computer interfaces, including the need for simplified tasks or slower communication. The study aims to decode intended finger movements directly from brain signals. The participants included individuals with chronic paralysis due to spinal cord injury, and intracortical microelectrode arrays were implanted in the motor cortex using algorithms. The participant attempted to type with both hands, and the data measured were typing speed and accuracy. The results showed successful neural encoding of distinct finger movements in both hands. The participants achieved rapid typing speeds, and performance proved more beneficial than that of previous brain-computer interface systems. The study shows how complex motor intentions are preserved in the brain after paralysis and how brain-computer interfaces can decode finger movements for natural tasks such as typing. However, the study included only one participant, so its generalizability is limited, and the system requires surgically implanted electrodes, which could pose a challenge for future participants.

Distance-based temporal similarity metrics for adaptive channel selection in multi-modal EEG-fNIRS BCI frameworks

Fri, 20 Mar 2026 12:00:00 GMT

The study investigates how to improve the computational efficiency of hybrid BCI systems that combine EEG and functional near-infrared spectroscopy (fNIRS) without reducing performance. The researcher come up with a channel selection algorithm that is based on distance metrics and uses electrode pairing and comparison of mean versus median threshold. It then evaluates the values on datasets with motor imagery, metal arithmetic and P300 tasks with a multitude of classifiers. They found that this method reduces the channel number by over 50% while keeping or improving the accuracy with median based thresholding performing better on noise sensitive tasks with a peak of 94.36%, P300 with LDA, and 72.93% mental arithmetic with SVM. Further analysis confirmed that the channels that were selected corresponded to task relevant cortical regions and the reduced channel lowered the decision latency to 0.11-0.20 seconds. The author concluded that the distance based channel section is a simpler and effective alternative to complex optimization for handling high dimensional multimodal BCI data.

Effects of repetitive transcranial magnetic stimulation on electroencephalographical measures of poststroke upper limb dysfunction: study protocol for a randomized controlled trial

Tue, 17 Mar 2026 12:00:00 GMT

This study discusses patients who have experienced strokes and currently struggle with upper limb malfunction, and investigates how repetitive transcranial magnetic stimulation could be used in aiding motor recovery for these patients. The study randomizes 42 stroke patients with upper-limb dysfunction into a high-frequency rTMS group and a low-frequency rTMS group, with a 1:1 ratio. The high-frequency group used a 90% resting motor threshold and 10 Hertz, acting on the ipsilesional M1 for a total of 1500 pulses over 2 weeks; the low-frequency group used the same experimental parameters, except that 1 Hertz was used, acting on the contralesional M1. The National Institutes of Health Stroke Scale (NIHSS), motor deficit (Fugl-Meyer Assessment Upper Extremity (FMA-UE), Modified Barthel Index (MBI), and resting-state electroencephalogram (EEG) were all recorded at baseline and within 1 week after rTMS. The study contributes to the medical literature of the brain’s neurophysiological reactions to high-frequency and low-frequency rTMS changes in people who have suffered a stroke.

Brain-Adhesive Bioelectronics With Shape-Morphable and Biodegradable Properties for Stable Brain Signal Monitoring

Mon, 09 Mar 2026 12:00:00 GMT

This study follows the development of a biodegradable and brain-adhesive electrocorticogram (ECoG) created for temporary monitoring of brain activity. They designed a morphable device built on a polyurethane elastomer integrated with polycarbonate to allow for flexibility and biodegradability while a tissue adhesive hydrogel made sure there was conformal cortical attachment and ultrathin molybdenum electrodes attached in an open mesh structure to allow for stable signal recording with minimal background noise. Testing showed that the sensor remained stable and demonstrated strong compatibility with cell viability and was able to naturally degrade. In vivo experiments demonstrated that it could, with high precision, record baseline neural activity, somatosensory evoked potential, and epileptiform discharges caused by 4-aminopyridine. The research highlights that limitations of nondegradable ECoG devices are fixed through the combination of biodegradable and mechanical compliance.

Motor imagery EEG signal classification using minimally random convolutional kernel transform and hybrid deep learning

Fri, 06 Mar 2026 12:00:00 GMT

This study investigates improved methods for classifying motor-imagery (MI) electroencephalography (EEG) signals, which are widely used in brain–computer interface (BCI) systems. Motor imagery refers to the mental simulation of movements (e.g., imagining moving the left or right hand), which produces measurable neural activity patterns. Accurate classification of these signals is essential for translating brain activity into commands for assistive technologies. The researchers evaluated a computational framework combining MiniRocket feature extraction with deep learning architectures, including convolutional neural networks (CNN) and long short-term memory (LSTM) networks. MiniRocket is a time-series feature-extraction algorithm that applies thousands of random convolutional kernels to EEG signals and computes summary features (such as the proportion of positive values) to efficiently represent temporal patterns. Compared with earlier algorithms like ROCKET, MiniRocket maintains a high-dimensional feature representation while reducing computational cost. The study used publicly available EEG motor-imagery datasets, particularly the PhysioNet MI-EEG dataset, with preprocessing steps including band selection (μ and β rhythms), signal segmentation, and independent component analysis to isolate relevant neural signals. Data were split into training, validation, and testing sets while preventing subject-level data leakage. Results showed that combining MiniRocket with hybrid deep-learning models (CNN–LSTM) achieved strong classification performance for motor-imagery EEG signals while maintaining computational efficiency. The model captured both spatial features (via CNN layers) and temporal dependencies (via LSTM or dilated convolutions). The authors also observed substantial inter-subject variability, meaning classification accuracy differed across individuals, highlighting a major challenge in BCI systems. Overall, the study demonstrates that efficient time-series feature extraction combined with deep neural networks can significantly improve motor-imagery EEG decoding, potentially enabling faster and more scalable brain-computer interface systems.

Anesthetic Management of Patients for Vagal Nerve Stimulator Placement: A Narrative Review

Thu, 05 Mar 2026 12:00:00 GMT

This paper aims to provide an overview of anesthetic management considerations for vagal nerve stimulator (VNS) implantation. Ultimately, the goal is to help anesthesiologists improve patient safety and operative outcomes during VNS implantation. VNS is a neuromodulation therapy that has many purposes, including drug-resistant epilepsy. This therapy works by sending electrical impulses through the vagus nerve to the brain and can influence neural pathways related to seizures and mood. The implantation involves an incision to the left vagus nerve and placement of electrodes around this nerve, while the pulse generator is implanted into the upper chest. The authors conducted a literature review surrounding clinical trials, cohort studies, review articles, and device safety analysis. They evaluated patient histories, as well as potential interactions and risks associated with VNS and surgical equipment. The findings show that general anesthesia with endotracheal intubation is the most common approach. The Vagus nerve manipulation can result in bradycardia, hypotension, and, rarely asystole. The postoperative findings show that patients are first monitored for factors such as airway obstruction and seizure occurrence, and that their pain control should balance the need to avoid respiratory depression with the need to provide adequate analgesia. The VNS implantation introduces unique anesthetic challenges due to device characteristics and surgical locations near the airway. The authors emphasize that there is a need for standardized management strategies for VNS therapy.

Neurotechnological cognitive enhancement and human rights: a complex dynamic between empowerment and constraint

Fri, 27 Feb 2026 12:00:00 GMT

The article focuses on the ethical and human rights issues in neurotechnological cognitive enhancers (NCEs) like neurofeedback systems. It focuses on the implications of where people have protected freedoms to use the technology for cognitive enhancement. The writer analyzes the paper, beginning by defining and explaining the trajectory of NCEs and then evaluating their implications. It identified a problem between empowerment and constraint, human rights supporting autonomy, but limitations being based on human equality and dignity. The author concludes that NCE is both enabled and constrained by human rights laws, requiring a new framework to protect individuals while they are regulated.