The central conceptual shift, according to the authors, is the explicit separation between the diagnostic and the quantitative uses of lung ultrasound. Both roles existed in 2012, but the new documents make clear that they demand different approaches and different techniques. Diagnosis answers a categorical question through a focused, pattern-based scan. Quantification translates findings into a reproducible measure through a standardized, multi-zone acquisition of the chest. The distinction matters because the two applications are easily conflated at the bedside, and the techniques that serve one do not automatically serve the other.

Quantification, the commentary stresses, does not concern aeration alone. Semi-quantitative techniques now extend to pulmonary congestion, pneumothorax volume and pleural effusion. In purely diagnostic applications, dividing the chest into zones and counting artifacts is usually unnecessary, with one exception: the differential diagnosis between diffuse and focal interstitial syndrome, assessed over the antero-lateral chest, where a criterion based on the number of positive zones provides a first orientation in the acutely ill patient. The quantitative techniques for aeration and congestion, by contrast, rest on zone division and apply semi-quantification to each scanned region.

The most substantial advance over 2012, the authors argue, is prognostic. A within-patient trajectory of repeated measurements enables monitoring, but it also underpins stratification that positions an individual patient’s risk relative to others. The commentary draws a careful line between two roles that are often blurred: as a prognostic tool, lung ultrasound predicts outcome, associating a measurement with a patient’s risk at a given moment; as a tool for treatment guidance, it informs a therapeutic decision and, through that decision, may alter the clinical course. The evidence base for the prognostic role, the authors note, is currently more robust than the evidence for treatment guidance.

A second structural novelty is the articulation of three diagnostic levels. The first concerns fundamental signs, for which the update refines nomenclature and definitions. The second, described as the genuine novelty, uses additional features to sharpen differential diagnosis between conditions that share the same elementary pattern: the distribution and quantity of B-lines, the characteristics of the pleural line, the correlation with clinical severity, the shape, margins, air bronchograms and size of consolidations, and the complexity of pleural effusion in relation to any associated consolidation. The third level is the final clinical interpretation, in which sonographic findings are integrated into a wider diagnostic puzzle where lung ultrasound may be decisive but should never be sufficient on its own.

The authors are explicit about why these levels need spelling out: they are too often collapsed in practice. A sign identified at the first level does not by itself deliver a diagnosis, and errors attributed to lung ultrasound frequently arise not in image acquisition but in the interpretation that follows. Competence in the technique and competence in clinical interpretation are distinct skills. The sonographic finding is a premise for reasoning, not a conclusion, and treating it as one is a recurring source of diagnostic misattribution.

Perhaps the most striking reframing concerns the power of exclusion. While confirming specific diagnoses, particularly in dedicated settings such as acute respiratory distress syndrome, is identified as one of the most promising future applications in intensive care, the commentary emphasizes the role of lung ultrasound in ruling conditions out and thereby narrowing the differential. This rule-out power rests on the high sensitivity, and therefore high negative predictive value, of the elementary signs: their absence renders the corresponding condition improbable even where a positive finding alone would not confirm it. Exclusion is typically achieved at the elementary-sign level, whereas confirmation requires the interpretive steps that follow.

The clinical applications where the new evidence bites hardest include heart failure and dialysis. In decompensated heart failure, the number of B-lines correlates with severity and predicts rehospitalization and death both at admission and at discharge, capturing residual, often subclinical congestion that clinical evaluation alone can miss. Because this prognostic value is independent of ejection fraction, it underscores the separation between cardiac function and cardiogenic congestion, and it extends to moderate-to-severe aortic stenosis, ST-elevation myocardial infarction and cardiogenic shock. In chronic heart failure, B-lines can also support therapeutic guidance: in one randomized trial, therapy titrated on sonographic monitoring reduced hospitalization for decompensation. In dialysis patients, sonographic congestion is often uncoupled from signs and symptoms, meaning an asymptomatic patient is not always euvolemic, and this occult overload predicts death and cardiac events. The authors caution, however, that there is not yet robust evidence that titrating fluid removal according to sonographic congestion during dialysis improves survival.

In mechanically ventilated patients, the lung aeration score monitors the effects of positive end-expiratory pressure, recruitment maneuvers and drugs, and carries prognostic weight, but it must be read alongside respiratory mechanics and gas exchange and is meaningful only when repeated acquisitions are rigorously standardized. The score provides no more than a coarse semi-quantification: it correlates with gold standards and with severity of lung involvement, yet it remains highly approximate because of the artifactual nature of the sonographic signs and the absence of a linear progression between score classes. A specific blind spot is overdistension, since overinflation of already aerated regions produces no change in the score; assessing reduced pleural sliding may eventually help close this gap. Lung ultrasound may also help select the ventilatory strategy itself: by distinguishing focal from non-focal lung morphology, it can guide a tailored approach assigning higher PEEP and recruitment maneuvers to non-focal ARDS and lower PEEP with prone positioning to focal ARDS, a strategy currently being compared with conventional ventilation in the ongoing PEGASUS trial.

The new guidelines also introduce, for the first time, an explicit distinction between aeration and congestion scores as separate techniques, validated in different populations against different gold standards and assessed over different chest regions, along with the concept of fluid tolerance: the absence of congestion signs defines a margin of safety for fluid administration. Fluid tolerance and fluid responsiveness are distinct questions, and lung ultrasound speaks to the former, functioning as a safety limit rather than a decision to give fluid, and is best complemented by focused cardiac ultrasound. Semi-quantification of pneumothorax, graded from the laterality of the lung point, and approximation of pleural effusion volume extend the same move from binary reading to graded measure, aiding decisions about drainage and enabling bedside monitoring without repeated radiographic exposure. The authors close with a caveat that doubles as a research agenda: prognostic weight is established, but validated numerical cut-offs for severity, characterization and prognosis in cardiogenic pulmonary edema, interstitial lung disease and ARDS are still missing, and that gap weighs most heavily on treatment guidance rather than prognosis.

Subject of Research: Updated evidence-based recommendations for diagnostic and quantitative point-of-care lung ultrasound in critical care

Article Title: What’s new in point-of-care lung ultrasound: reading the new recommendations

Article References: Volpicelli, G., Gargani, L., & Zieleskiewicz, L. (2026). What’s new in point-of-care lung ultrasound: reading the new recommendations. Intensive Care Medicine. https://doi.org/10.1007/s00134-026-08649-2

Image Credits: AI Generated

DOI: 10.1007/s00134-026-08649-2

Keywords: lung ultrasound, point-of-care ultrasound, intensive care, B-lines, ARDS, heart failure, aeration score, congestion, fluid tolerance, pneumothorax, pleural effusion, prognostication

The correspondence responds to a scoping review by Munir and colleagues, published in the same journal in May 2026, which examined the clinical relevance of sleep disturbance, or insomnia, in patients with advanced cancer receiving palliative care. Scoping reviews are designed to map the breadth of existing evidence rather than pool it statistically, and the original review highlighted how frequently sleep problems are reported by this patient population while also exposing the fragmented ways in which researchers have defined and measured them. Yan and Zhang use their comment to push the conversation further, contending that the clinical relevance of insomnia is not merely a matter of prevalence but of consequence, and that the field needs sharper tools to capture both.

The technical heart of the debate lies in measurement. Sleep disturbance in palliative care has traditionally been assessed through patient-reported instruments, and one of the most widely used is the sleep item of the Edmonton Symptom Assessment System, a brief screening tool that asks patients to rate the severity of their symptoms on simple numeric scales. Research published in BMJ Supportive & Palliative Care by Yennurajalingam and colleagues in 2015 showed that patient-reported sleep disturbance in advanced cancer is common, that it can be predicted from other symptom profiles, and that the ESAS sleep item performs reasonably well as a screening instrument. But a single numeric rating compresses an enormously complex phenomenon, one that spans difficulty falling asleep, frequent awakenings, early morning waking, and non-restorative sleep, into a single point on a scale. Critics argue that this compression obscures the very features of insomnia that carry clinical meaning.

Objectively measured sleep tells a different story again. Actigraphy, a technique in which a small wrist-worn sensor records movement over days or weeks, allows researchers to derive rest-activity rhythms and sleep parameters without relying on subjective report. In an exploratory machine learning study published in Cancers in 2023, Patel, Davies, and colleagues, working with Dijk, combined actigraphy-derived rest-activity and sleep parameters with routine clinical data to explore prognostication in advanced cancer. The study suggested that objective sleep and circadian features, when fused with clinical variables through machine learning models, could contribute to predicting outcomes in this population. For Yan and Zhang, this line of evidence points toward a future in which sleep is not just a symptom to be rated but a physiological signal to be integrated into routine clinical assessment and even prognostic modeling.

The prognostic angle is what gives the correspondence its urgency. In advanced cancer, clinicians and families alike want to know how much time remains, and the search for reliable prognostic markers is one of palliative care’s most active research frontiers. Sleep disturbance may seem an unlikely candidate, but disrupted rest-activity rhythms are known to reflect underlying physiological deterioration, and studies have linked poor sleep quality in advanced lung cancer to anxiety and depression in both patients and caregivers, as He and colleagues reported in BMJ Supportive & Palliative Care in 2020. Mercadante and colleagues, writing in the International Journal of Clinical Practice in 2021, documented high rates of insomnia among patients with advanced lung cancer admitted to palliative care services, underscoring that the problem is not marginal but affects a substantial share of the population at the very end of life.

What makes the Yan and Zhang comment technically significant is its insistence on distinguishing between insomnia as a subjective complaint and sleep disturbance as an objective physiological phenomenon. The two do not always align. A patient may report severe insomnia while actigraphy shows relatively preserved sleep duration, or conversely may underreport disturbance that objective monitoring reveals to be profound. This dissociation matters clinically because the mechanisms driving the two may differ, ranging from pain and medication effects to circadian rhythm disruption and neuroendocrine changes associated with advanced disease. Treating them as a single entity, the correspondence implies, risks misdirecting both research and therapy, and the scoping review’s finding of heterogeneous definitions across studies only reinforces the point.

The therapeutic implications are equally consequential. Insomnia in palliative care is frequently managed with hypnotic medications, yet evidence for their effectiveness in this population is limited, and concerns about sedation, falls, and drug interactions weigh heavily in patients whose systems are already burdened by disease and polypharmacy. Non-pharmacological approaches, including cognitive behavioral therapy adapted for cancer populations, light exposure interventions, and structured sleep hygiene programs, offer alternatives, but their availability in palliative care settings remains uneven. Yan and Zhang’s argument that sleep deserves systematic clinical attention implies that screening should be routine, that assessment should distinguish subjective from objective disturbance, and that treatment pathways should be tailored to the mechanisms identified in each patient rather than applied generically.

There is also a broader scientific payoff to taking sleep seriously in this population. Sleep and circadian biology are increasingly recognized as windows into systemic physiology, and disrupted rest-activity rhythms have been associated with inflammation, metabolic dysregulation, and neuroendocrine stress across a range of diseases. In advanced cancer, where multiple organ systems are failing simultaneously, the pattern of sleep disruption may carry diagnostic information that conventional symptom scales cannot capture. The machine learning approach of Patel and colleagues represents an early attempt to exploit this information, and if future studies validate the prognostic value of actigraphy-derived sleep parameters, sleep monitoring could become as standard in palliative care as pain assessment is today.

For patients and families, the message from this correspondence is ultimately a human one. Sleepless nights in advanced cancer are not a trivial side effect to be endured in silence, but a symptom with measurable physiological correlates, prognostic significance, and treatable mechanisms. The comment by Yan and Zhang, published as a correspondence in Supportive Care in Cancer in October 2026, adds its voice to a growing chorus calling for sleep disturbance to be recognized as clinically relevant in its own right, not merely as a companion symptom to pain or depression. As the field moves from scoping reviews that map the evidence toward studies that test interventions and validate objective monitoring, the hope is that the nights of patients in palliative care will become a little less fragmented, and that clinicians will finally give sleep the attention its clinical relevance demands.

Subject of Research: Palliative Care

Article Title: Comment on: “The clinical relevance of sleep disturbance (‘insomnia’) in patients with advanced cancer receiving palliative care: a scoping review”

Article References: Yan, H., & Zhang, Y. (2026). Comment on: “The clinical relevance of sleep disturbance (‘insomnia’) in patients with advanced cancer receiving palliative care: a scoping review”. Supportive Care in Cancer, 34(11), Article 1115. https://doi.org/10.1007/s00520-026-11359-6

Image Credits: AI Generated

DOI: 10.1007/s00520-026-11359-6

Keywords: insomnia, palliative care, advanced cancer, sleep disturbance, Supportive Care in Cancer, scoping review, symptom assessment, actigraphy, prognostication, patient-reported outcomes, Comment, clinical

The editorial, led by Hipólito Nzwalo of the University of Algarve together with Miguel Barboza of Hospital Rafael Angel Calderon Guardia in Costa Rica and Nicola Logallo of Haukeland University Hospital in Norway, frames a persistent structural problem: most stroke trials and risk models have been built on younger, healthier cohorts. When clinicians apply those tools to patients with reduced physiological reserve, heavier comorbidity burdens, and different goals of care, accuracy degrades. In very old patients with intracerebral hemorrhage, for example, the main mortality scores retain acceptable overall discrimination but lose the ability to pinpoint which specific subgroups face the highest risk of death.

The consequences of this extrapolation are visible in everyday care. Evidence shows that older stroke survivors receive less guideline-compliant treatment than younger patients, regardless of how functionally independent they were on admission or at discharge. The editorial’s authors argue that treatment decisions should rest on biological rather than chronological age, a principle that several studies in the collection put to the test.

On the prevention side, several papers focused on risk stratification using readily available clinical data. One study examined the Weight-Adjusted Waist Index, an innovative obesity metric, as a predictor of first stroke in the elderly. Another found that combined levels of C-reactive protein and LDL cholesterol carried prognostic information, while a third investigated the glucose disposal rate, a surrogate marker of insulin resistance, in relation to stroke incidence. The appeal of these measures lies in their simplicity: they can be calculated from routine blood tests and anthropometric measurements without specialized equipment.

Two studies probed intermediate risk factors that sit upstream of stroke itself. Obesity, hypertension, triglycerides, glycosylated hemoglobin, sex, and fasting blood glucose were identified as predictors of arteriosclerosis, the arterial stiffening that underlies many cerebrovascular events. Separately, higher plasma levels of trimethylamine N-oxide, a metabolite produced by gut microbiota, were associated with a greater burden of cerebral white matter hyperintensities, the bright lesions on MRI that signal accumulating small-vessel damage. Importantly, the presence of these white matter changes was linked to early progression of stroke, particularly in older patients with anterior choroidal artery infarction, suggesting that brain imaging findings could sharpen acute risk assessment in the elderly.

Psychological factors emerged as another modifiable dimension of risk. Researchers found that an ascending trajectory of cognitive-affective symptoms, alongside unstable and persistently elevated trajectories of somatic depressive symptoms, were associated with increased stroke risk in older adults. The finding adds to growing recognition that depression and anxiety are not merely consequences of cerebrovascular disease but may actively contribute to its onset.

For prognostication after stroke has occurred, a series of studies highlighted the value of inflammatory and metabolic biomarkers calculated from standard blood counts. The neutrophil-percentage-to-albumin ratio, the triglyceride-glucose-body mass index, and the neutrophil-to-lymphocyte ratio all emerged as promising predictors of mortality in older stroke patients. In acute settings, the picture was similarly detailed: a risk model integrating cerebral circulation time and collateral status improved prediction of futile recanalization following mechanical thrombectomy, while distinct trajectories of the neutrophil-to-albumin ratio were associated with functional outcomes after endovascular treatment. The triglyceride-glucose-body mass index also predicted long-term all-cause mortality in elderly patients treated with intravenous thrombolysis, and the platelet-to-lymphocyte ratio proved an independent predictor of death in patients with severe acute ischemic stroke requiring intensive care.

Beyond biomarkers, the collection included a comprehensive review of deubiquitinases, enzymes that regulate the ubiquitin-proteasome system, a cellular machinery deeply implicated in cerebral ischemia-reperfusion injury, the cascade of damage that unfolds when blood flow is restored to oxygen-starved brain tissue. The review discusses the potential of deubiquitinase-targeted interventions as novel neuroprotective strategies, pointing toward therapies that could one day limit brain injury during reperfusion, a critical window in acute stroke care.

Several contributions directly challenged ageist assumptions by demonstrating that evidence-based interventions remain beneficial in the elderly. Extracranial vertebral artery stenting was shown to be both safe and effective in older individuals, as was intravenous thrombolysis for minor ischemic stroke. These findings carry practical weight: they support offering older patients the same reperfusion and endovascular options considered standard in younger populations, provided decisions are grounded in each patient’s biological status rather than an arbitrary age threshold.

Long-term cognitive outcomes received dedicated attention as well. One study identified three distinct post-stroke cognitive trajectories among elderly patients, characterized as high cognitive function with slow decline, low cognitive function with intermediate decline, and low cognitive function with rapid decline. Another found that the hemoglobin-to-red cell distribution width ratio, a simple calculation from a complete blood count, was associated with the risk of post-stroke cognitive impairment, offering clinicians an accessible biomarker for flagging patients at greater risk of cognitive deterioration. Taken together, the editors conclude that evidence for very old stroke patients remains weak, fragmented, and incomplete, and that tools derived from younger, less comorbid populations cannot simply be extrapolated upward. By spanning risk prediction, prognostication, acute management, and long-term outcomes, the research topic represents a concerted effort to close one of stroke medicine’s most persistent and consequential gaps.

Subject of Research: Stroke risk prediction, prognostication, and treatment in elderly and very old adults

Article Title: Editorial: Stroke research in the elderly: addressing ageism and prognostication

Article References: Nzwalo, H., Barboza, M., & Logallo, N. (2026). Editorial: Stroke research in the elderly: addressing ageism and prognostication. Frontiers in Stroke, 5, Article 1978900. https://doi.org/10.3389/fstro.2026.1978900

Image Credits: AI Generated

DOI: 10.3389/fstro.2026.1978900

Keywords: stroke, elderly, ageism, prognostication, biomarkers, intracerebral hemorrhage, thrombolysis, thrombectomy, white matter hyperintensities, cognitive decline, inflammation, insulin resistance

The outcome measure at the heart of the study is the Cerebral Performance Category, or CPC, a five-level scale that neurointensivists use to summarize how well a patient’s brain is functioning. Categories 1 through 3 denote good outcomes, ranging from a return to normal life or moderate disability, while categories 4 and 5 denote poor outcomes, spanning severe disability, a vegetative state, or death. Assigning a CPC score requires a trained human to synthesize days or weeks of clinical information, which makes it labor-intensive and difficult to standardize across hospitals. That bottleneck matters because large-scale research on cardiac arrest prognostication, including multi-center trials and quality-improvement programs, depends on consistently labeled outcomes for thousands of patients. If an algorithm could assign these categories automatically, researchers could analyze far larger cohorts and clinicians could track outcomes in near real time.

To test that idea, the team conducted a retrospective cohort study of adult patients who were comatose after either in-hospital or out-of-hospital cardiac arrest at three academic hospitals in the United States. The dataset comprised 357 patients in total, split into a training set of 249 patients on which the model learned, and a holdout set of 108 patients on which its performance was evaluated. The raw material was the full corpus of clinical notes in each patient’s electronic health record, from admission through discharge. Before any modeling, the text was de-identified to strip out protected health information, a critical step given the privacy stakes of mining free-text records. The pipeline then transformed the notes into numerical features and fed them into a logistic regression classifier, a relatively simple and transparent model chosen deliberately over more opaque deep learning architectures so that the researchers could inspect what the model was actually learning.

The performance figures are the study’s headline. Using all clinical notes across the entire hospitalization, the model achieved an overall accuracy of 81 percent in distinguishing good from poor neurological outcomes, with an area under the receiver operating characteristic curve of 0.90 and an area under the precision-recall curve of 0.89. In practical terms, an AUROC of 0.90 means that if you picked one patient with a good outcome and one with a poor outcome at random, the model would assign a higher risk score to the poorer-outcome patient nine times out of ten. The AUPRC figure is particularly meaningful in this setting because good and poor outcomes are not evenly balanced in the cohort; precision-recall performance is less inflated by class imbalance and therefore a stricter test of real-world utility. For a model built from nothing but free text, these numbers place it in the same performance territory as many prognostic tools built on carefully curated physiological data.

Perhaps the most provocative finding emerged when the researchers restricted the model’s input to notes documented only within the first 24 hours after hospitalization. Accuracy dipped to 74 percent, with the AUROC and AUPRC both falling to 0.83, but the model still performed well above chance. That means substantial prognostic information about a patient’s eventual neurological outcome is embedded in the very first day of clinical documentation, long before the outcome is known. The authors interpret this signal as a mixture of three ingredients: the patient’s underlying biology, which manifests in early exam findings and physiological derangements; the clinician’s structured assessment of those findings; and potentially a third, more troubling component, early prognostic framing, in which clinicians’ initial expectations about recovery color the language they use and the care they document.

That third component connects to one of the most debated issues in neurocritical care: the role of clinician bias and self-fulfilling prophecy in cardiac arrest prognostication. Prior research has shown that early withdrawal of life-sustaining therapy is common after cardiac arrest and may result in deaths that would not otherwise occur, and that providers can be overconfident in their early outcome predictions. The new study found that the model demonstrated higher precision in predicting poor neurological outcomes particularly among patients who underwent withdrawal of life-sustaining therapy. In other words, the model was especially good at detecting poor outcomes in precisely the group where the outcome may have been shaped, at least in part, by the decision to withdraw care. This raises a subtle question about what the model is truly measuring: the patient’s intrinsic recovery potential, or the trajectory that clinical decision-making set in motion. The authors are careful on this point, and their framing of early prognostic signals as a blend of biology and clinician assessment acknowledges the entanglement rather than claiming the model has solved it.

Technically, the choice of logistic regression over a large language model is worth unpacking. Modern clinical NLP increasingly relies on transformer-based models that can capture context and nuance in text, but they are harder to interpret and validate in high-stakes medical settings. By using a simpler model on engineered text features, the team could examine which words and phrases drove predictions, an essential property when the goal is to understand what clinical documentation reveals about prognosis. The researchers have also made the entire NLP pipeline, including preprocessing, modeling, and evaluation code, publicly available on GitHub, which lowers the barrier for other groups to replicate the approach on their own patient populations and to scrutinize the method for hidden artifacts. Reproducibility of this kind is rare and valuable in clinical machine learning, where models often fail to generalize when moved between hospitals with different documentation practices.

The study is candid about its limitations and about the work that remains. The cohort of 357 patients from three academic centers is modest by machine learning standards, and the model’s performance will need validation in external, more diverse populations before any clinical deployment. The authors note that future research should incorporate multimodal data, combining text with EEG, imaging, and laboratory values, and should employ interpretable advanced language models to push accuracy higher toward fully automatable CPC derivation. There is also the question of what such a tool should be used for. The authors position it primarily as a research instrument, a way to generate consistently labeled outcome data at scale for prognostication studies and quality improvement, rather than as a bedside oracle that would tell families whether a loved one will wake up. Given the documented dangers of premature prognostication and early withdrawal of support, that restraint seems well judged.

Still, the broader implications are considerable. Cardiac arrest affects hundreds of thousands of people each year in the United States alone, and neurological outcome remains the dominant determinant of long-term quality of life among survivors. The American Heart Association has issued formal standards for how prognostication studies should be conducted, precisely because the field is littered with overconfident early predictions. An automated system that can derive standardized outcome categories from routine documentation could accelerate the search for better prognostic models, enable continuous auditing of how outcomes vary across hospitals and patient groups, and eventually support clinicians with a second, data-driven opinion grounded in the full record rather than a single exam. The finding that the first 24 hours of notes already carry most of the signal is both an opportunity, for earlier and better-informed decision-making, and a warning, that the language clinicians write on day one may be quietly shaping the outcomes they later record. Either way, the study makes a compelling case that the medical record’s unstructured text is not just documentation of care, but a rich, largely untapped data source for understanding and predicting recovery of the injured brain.

Subject of Research: Automated prediction of neurological outcomes after cardiac arrest using natural language processing of electronic health record notes

Article Title: Automated Derivation of Cerebral Performance Category at Hospital Discharge After Cardiac Arrest Using Natural Language Processing and Machine Learning

Article References: Furlow, S., Houston, P., Bao, K., Zhou, B., Velasquez, G., Peterson, K., Hayman, A., Bains, A., Lieberman, O., Hendrickson, C. M., Zier, L., Singhal, N. S., Hemphill, J. C., Zafar, S., Westover, M. B., Fernandes, M., & Amorim, E. (2026). Automated Derivation of Cerebral Performance Category at Hospital Discharge After Cardiac Arrest Using Natural Language Processing and Machine Learning. Neurocritical Care. https://doi.org/10.1007/s12028-026-02650-9

Image Credits: AI Generated

DOI: 10.1007/s12028-026-02650-9

Keywords: cardiac arrest, natural language processing, machine learning, cerebral performance category, electronic health records, neurocritical care, prognostication, clinical notes, logistic regression, coma, neurological outcome, predictive medicine

The research, led by a team at Boston Medical Center and Boston University’s Chobanian and Avedisian School of Medicine, set out to do something deceptively simple: train artificial intelligence models to recognize the electrical fingerprint of cerebral edema in routine electroencephalography recordings. EEG is already standard practice in post-cardiac arrest care. Comatose survivors are typically monitored continuously for days, both to detect seizures and to help clinicians judge the extent of brain injury. Yet the vast majority of that data is reviewed visually by human experts, who can extract only a fraction of the information embedded in the multichannel voltage traces. The new work demonstrates that deep learning models can mine those same recordings for evidence of a complication that has, until now, required imaging to confirm.

The study was retrospective and single-center, drawing on the records of adult patients resuscitated from cardiac arrest between 2016 and 2024 who had undergone both neuroimaging and EEG monitoring as part of their ordinary clinical care. No additional tests were performed; the researchers simply reanalyzed data that already existed. From this cohort, 124 patients formed the detection dataset. Their median age was 53 years, roughly 60 percent were male, and just over half—65 patients, or 52.4 percent—went on to develop diffuse cerebral edema visible on imaging. That high prevalence reflects the sobering reality of post-arrest neurology: brain swelling is not a rare complication but a central feature of hypoxic-ischemic injury, the cascade of damage that follows oxygen deprivation.

Technically, the team compared two families of neural network architectures that represent different philosophies of time-series analysis. The first, long short-term memory networks, or LSTMs, are recurrent models that process sequences step by step, carrying information forward through gated memory cells. They have been a mainstay of sequence modeling since their introduction in 1997, but they struggle with very long-range dependencies and process data sequentially rather than in parallel. The second architecture, the transformer, relies on self-attention mechanisms that allow the model to weigh relationships between any points in the input sequence simultaneously, regardless of distance. Transformers underpin modern large language models, and their application to EEG represents a growing trend in neurotechnology. In this study, the transformer won decisively, outperforming the LSTM both for classifying established edema and for predicting it before radiographic recognition.

The detection model’s inputs were four-hour segments of EEG recorded more than 24 hours after the arrest, a window chosen to avoid the earliest, most volatile period of post-arrest physiology. The best-performing configuration—a transformer operating on 4-hour EEG segments—achieved a median area under the receiver operating characteristic curve of 83.5 percent, with 75.0 percent accuracy, 80.0 percent sensitivity, and 70.0 percent specificity. In clinical terms, the model correctly identified four out of five patients with edema while flagging a manageable rate of false alarms. Those numbers fall short of diagnostic certainty, but the researchers emphasize that the intended use is not autonomous diagnosis. Rather, the model would serve as a screening layer over continuous EEG, flagging patterns for clinician review in real time—much as automated systems already flag suspicious findings in cardiac monitoring.

The more provocative result came from the study’s secondary analysis. If edema leaves an electrical signature, could that signature appear before a CT scan would reveal the swelling? To test this, the team identified 19 patients who ultimately developed diffuse cerebral edema and matched them with 19 referents who did not, controlling for age, sex, whether the arrest was witnessed, and the timing of the EEG recordings. They then fed the models EEG segments that preceded the radiographic recognition of edema in the affected patients. The top-performing prediction model used 8-hour EEG segments and achieved a median AUC of 80.0 percent, with 70.0 percent accuracy, 80.0 percent sensitivity, and 60.0 percent specificity. In other words, the brain’s electrical activity betrayed the onset of swelling before conventional imaging could confirm it.

The biological rationale for this is grounded in cellular physiology. Cerebral edema after cardiac arrest is predominantly cytotoxic: neurons and glial cells, starved of energy, fail to maintain the ion gradients that keep water inside blood vessels and outside cells. As adenosine triphosphate supplies dwindle, sodium and calcium pour into cells, water follows osmotically, and the cells swell. This swelling changes the electrical properties of neural tissue—altering conduction velocities, dampening synaptic transmission, and reshaping the synchronized rhythms that EEG records. Biophysical models of cytotoxic cell swelling developed in recent years have predicted exactly these kinds of changes in signal amplitude and frequency content. The machine learning model, in effect, learned to recognize the macroscopic electrical consequences of microscopic cellular disaster.

Why does earlier detection matter? Cerebral edema after cardiac arrest has long been viewed with therapeutic pessimism, often interpreted as a marker of catastrophic, irreversible injury. But a growing body of research argues that at least some components of post-arrest edema are treatable complications rather than fixed verdicts. Neurocritical care guidelines already outline acute treatments for cerebral edema, including osmotic therapy and blood pressure management, and there is active debate about whether aggressive intervention could tip some patients away from devastating outcomes. An EEG-based early warning system would give clinicians a head start—hours, potentially—during which interventions could be escalated, monitoring intensified, and families engaged in more informed conversations about what is happening inside the skull. The authors note that such a tool could inform discussions about both reversible and irreversible brain injury, a distinction that carries enormous weight in intensive care decision-making.

The study’s limitations are those typical of first-in-field work. It was retrospective and conducted at a single center, meaning the models were trained and tested on data from one institution’s EEG equipment, recording protocols, and patient population. External validation on independent cohorts from other hospitals is essential before any clinical deployment, and the authors are explicit about this. The prediction analysis, while conceptually striking, involved only 38 patients—a sample small enough that the confidence intervals around its performance estimates are wide. EEG is also vulnerable to artifacts in the intensive care environment: muscle activity, eye movements, electrode impedance shifts, and electrical interference can all contaminate recordings, and the team employed automated artifact rejection methods to mitigate this. How well the models generalize to noisier, real-world data streams remains to be demonstrated.

Even so, the study lands at a moment of genuine momentum. Deep learning applied to EEG after cardiac arrest has already shown promise for predicting neurological outcome, and recent American Heart Association and European Resuscitation Council guidelines have highlighted the growing role of multimodal, data-driven prognostication. What distinguishes this work is its target: not outcome prediction, which informs decisions about withdrawing care, but detection of a dynamic, potentially modifiable complication. If future studies confirm that EEG changes reliably precede radiographic edema, continuous EEG monitoring could evolve from a passive observation tool into an active surveillance system—alerting clinicians to brain swelling while there is still time to act. For the hundreds of thousands of people who suffer cardiac arrest each year worldwide, that shift could mean the difference between a brain that recovers and one that does not.

Subject of Research: Machine learning detection of cerebral edema using EEG in post-cardiac arrest patients

Article Title: Automated Cerebral Edema Detection using Electroencephalography in Post-cardiac Arrest Patients

Article References: Stafford, R. A., Srivastava, V., Altaf, A. Z., Tang, L., Sachdeva, A. S., Elman, D., Mallinger, L. A., Reinert, A. L., Madipally, S. S. K., Salvati, L. R., Wahlster, S., Chen, D., Gilmore, E. J., Greer, D. M., Cheng, H., & Ong, C. J. (2026). Automated Cerebral Edema Detection using Electroencephalography in Post-cardiac Arrest Patients. Neurocritical Care. https://doi.org/10.1007/s12028-026-02668-z

Image Credits: AI Generated

DOI: 10.1007/s12028-026-02668-z

Keywords: cerebral edema, electroencephalography, cardiac arrest, machine learning, transformer, deep learning, LSTM, neurocritical care, hypoxic-ischemic brain injury, computed tomography, brain swelling, prognostication

The tool in question is the FUNC score, short for the Functional Outcome in Patients with Primary Intracerebral Hemorrhage score, developed by neurologists at Massachusetts General Hospital and first published in the journal Stroke in 2008. The score distills a handful of variables available at hospital admission — the volume and location of the bleed on initial imaging, the patient’s age, the Glasgow Coma Scale score on arrival, and whether the patient was already taking an anticoagulant medication such as warfarin — into a single number ranging from zero to eleven. In its original validation, the score predicted whether patients would achieve functional independence, defined as a Glasgow Outcome Scale score of four or higher, ninety days after the hemorrhage. That three-month window, however, has increasingly been recognized as an arbitrary snapshot of what is often a much longer recovery.

Recovery after brain injury is not a fixed event but a moving target. Emerging evidence, including a growing body of work on recovery trajectories after intracerebral hemorrhage, shows that many patients continue to improve between three and twelve months after the bleed. In some cohorts, a meaningful proportion of survivors who were dependent at three months walk, feed themselves, and manage their own affairs a year later. If clinicians base early prognostic conversations only on tools validated at three months, they risk anchoring families to an overly pessimistic view — or, conversely, they may lack the statistical justification to make confident statements about long-term independence at all. Recognizing this gap, a team led by Joel Neves Briard, a postdoctoral researcher in Columbia’s Department of Neurology, and senior author Jan Claassen, professor of neurology and critical care specialist, set out to ask whether the FUNC score could be stretched, without distortion, to a twelve-month horizon.

The team conducted a single-center prospective cohort study, enrolling adult patients admitted with primary intracerebral hemorrhage between February 2009 and January 2018. Prospective enrollment matters in this kind of research: rather than reconstructing cases from medical records after the fact, investigators collected structured data on patients as they arrived and followed them forward, reducing the bias and missingness that plague retrospective analyses. All patients included in the analysis had a primary hemorrhage — meaning the bleed was not secondary to trauma, aneurysm rupture, or vascular malformation — and each patient’s FUNC score was calculated at admission using the same variables that have anchored the score since its debut. Twelve months later, surviving patients’ functional status was assessed using the Glasgow Outcome Scale, a five-point measure that ranges from death to good recovery and that classifies patients scoring four or five as broadly independent.

The cohort ultimately comprised 535 patients, a substantial sample for this disease. The patients were, on average, seriously ill: their median age was 68 years, 44 percent were women, and the median National Institutes of Health Stroke Scale score at presentation was 16, a level indicating moderate to severe neurological impairment. Median FUNC score in the cohort was 8 on the eleven-point scale. Among the 445 patients with a known Glasgow Outcome Scale assessment at twelve months, 99 — roughly 22 percent — had achieved functional independence. That figure alone carries a message for families: even in a cohort skewing severely ill, more than one in five patients was living independently a year after a brain bleed, a reminder that early appearances can mislead.

The statistical heart of the study lay in testing whether the FUNC score’s predictions still held at one year. The researchers used logistic regression to calculate the area under the receiver operating characteristic curve, or AUC, the standard gauge of a prediction model’s ability to separate those who will experience an outcome from those who will not. An AUC of 0.5 indicates performance no better than a coin flip, while 1.0 represents perfect discrimination. For twelve-month functional independence, the FUNC score achieved an AUC of 0.79, with a 95 percent confidence interval of 0.75 to 0.84 — a level of discrimination broadly comparable to the score’s originally validated three-month performance. In practical terms, the score’s simple admission-time arithmetic remained nearly as informative about the one-year future as about the three-month one.

Discrimination, however, is only half the story. A model can rank patients correctly yet still misjudge how likely the outcome actually is, which is why the researchers also evaluated calibration — the agreement between predicted probabilities and observed frequencies — using a calibration plot, alongside the Brier score, a composite measure of both discrimination and calibration that penalizes confident wrong answers. The calibration plot showed reasonable agreement between predicted and observed outcomes across the probability range, with modest evidence that the score slightly overestimated the likelihood of independence at the low end of predicted probabilities. The Brier score came in at 0.15, a respectable figure for a binary clinical outcome. To address the patients lost to follow-up, the team handled missing twelve-month outcomes using multiple imputation by chained equations, a modern statistical technique that fills gaps by modeling each incomplete variable conditional on all others, preserving uncertainty rather than pretending missing data do not exist.

Finally, the investigators asked a question that has become central to the modern evaluation of clinical prediction tools: does the score actually help make decisions? Using decision curve analysis, a method developed by Vickers and Elkin that quantifies the net clinical benefit of acting on a model’s predictions across a range of risk thresholds, the team found that the FUNC score delivered measurable net benefit across threshold probabilities from 5 to 50 percent. In other words, within the range of probabilities most relevant to real clinical conversations, basing decisions on the score beat both the strategy of assuming every patient will be independent and the strategy of assuming none will be — the two extremes any prognostic tool must outperform to be worth the bedside clinician’s attention.

The implications reach beyond the family meeting. The authors point to two principal applications. The first is counseling: armed with a validated twelve-month prediction, clinicians can offer families a more honest account of long-term prospects, tempering the nihilism that has historically shadowed intracerebral hemorrhage care. This matters because premature withdrawal of life-sustaining therapy, driven in part by excessively pessimistic early prognostication, remains a documented concern in hemorrhagic stroke, and formal guidelines for neuroprognostication in critically ill adults with intracerebral hemorrhage explicitly caution against self-fulfilling prophecies. The second application is methodological. The FUNC score can support sliding dichotomy, a trial-design strategy in which the outcome threshold for success is adjusted to each patient’s predicted baseline risk — an approach pioneered in traumatic brain injury research that increases statistical power by recognizing that a modest recovery for the sickest patients may be as meaningful as a full recovery for the mildest ones.

The researchers are careful about limits. This was a single-center study at a major academic medical center, and prediction models notoriously lose calibration when transplanted to different populations and care environments; external validation across centers is needed before the twelve-month FUNC score becomes standard practice. The field’s own reporting standards, codified in the TRIPOD+AI statement for clinical prediction models, demand exactly this kind of transparency about performance, calibration, and intended use. Still, the study’s message is striking in its simplicity. A tool invented nearly two decades ago, computed from five variables a clinician can gather before the patient has even reached the intensive care unit, appears to hold its predictive power across the full year in which the brain does its slow, stubborn work of healing. For the families sitting in hospital corridors, that is not just a statistical achievement — it is a reason to keep asking the question, and to keep hoping the answer changes.

Subject of Research: Prediction of 12-month functional independence after primary intracerebral hemorrhage using the FUNC score

Article Title: Extension of the FUNC Score for Prediction of 12-Month Functional Independence after Primary Intracerebral Hemorrhage

Article References: Neves Briard, J., Kansara, V., Shen, Q., Song, Y. L., Cami, A. B., Velazquez, A. G., Esposito, J. M., Klein, A. J., Ghoshal, S., Agarwal, S., Park, S., Connolly, E. S., Roh, D., & Claassen, J. (2026). Extension of the FUNC Score for Prediction of 12-Month Functional Independence after Primary Intracerebral Hemorrhage. Neurocritical Care. https://doi.org/10.1007/s12028-026-02661-6

Image Credits: AI Generated

DOI: 10.1007/s12028-026-02661-6

Keywords: intracerebral hemorrhage, stroke, FUNC score, prognostication, functional outcome, Glasgow Outcome Scale, neurocritical care, prediction model, recovery, brain bleed, Extension, FUNC

Researchers at Tampere University Hospital set out to answer a question that had not been addressed before: whether temporal muscle thickness, or TMT, measured on ordinary computed tomography scans, could predict survival in patients aged 70 or older who received brain radiotherapy for metastatic cancer. The question matters because prognostication in this population is exceptionally difficult and exceptionally important. Brain metastases affect an estimated 10 to 40 percent of patients with solid tumors, and population studies place the incidence between 8.3 and 14.3 per 100,000 people, a figure that almost certainly underestimates the true burden because not all cases are registered. Median survival across the field ranges from as little as two months to 21 months depending on the primary cancer, and for patients over 65, Surveillance, Epidemiology, and End Results data indicate a median survival of four months or less for nearly all common malignancies, including breast, colorectal, esophageal, renal, and lung cancer and melanoma.

Accurate prognostic tools would allow clinicians to tailor treatment intensity intelligently. Patients with a favorable outlook might benefit from more aggressive interventions, while those expected to survive only weeks could be spared the burdens of hospitalization, repeated radiotherapy sessions, or chemotherapy in the final weeks of life, interventions that may degrade quality of life without meaningfully extending it. Honest prognostic information also underpins shared decision-making, giving patients and families a realistic basis on which to weigh intensive treatment against palliative care. The Tampere team reasoned that if a simple measurement already embedded in the treatment-planning scan could sharpen these predictions, it would be a genuinely useful clinical instrument.

The study, published in Cancer Reports, took the form of a retrospective cohort analysis of consecutive patients aged 70 years or older who underwent brain radiotherapy, with or without prior surgical resection, for brain metastases between 2014 and 2022. From the hospital’s electronic radiotherapy database, 249 patients met the inclusion criteria: 136 men with a mean age of 76.0 years and 113 women with a mean age of 76.4 years. Lung cancer dominated the cohort, accounting for 113 patients or 45.4 percent, followed by breast cancer with 29 patients and melanoma with 27. Most participants, 226 in total, received radiotherapy alone, split between whole-brain radiotherapy, given to 151 patients, and stereotactic radiotherapy, delivered to 98, while 23 patients underwent surgical resection followed by postoperative radiation.

Measuring the temporalis muscle requires surprising methodological care. A single observer, blinded to patient outcomes, assessed TMT on each patient’s first radiotherapy planning head CT, or on a diagnostic CT performed within the preceding month. The measurement plane had to be oriented parallel to the falx cerebri, the membrane dividing the brain’s hemispheres, and tangential to the floor of the middle cranial fossa, with the image windowed to a narrow setting of 80 Hounsfield units width and 40 length to sharpen the muscle’s boundaries. On the first axial slice above the bony orbit, a tangent line was drawn from the muscle’s anterior attachment to the skull, and thickness was recorded at the muscle’s thickest point perpendicular to that line, deliberately excluding fat, fascia, and blood vessels. The average of left and right measurements was used in the analyses. To confirm the technique’s reliability, a second observer independently measured a random sub-cohort of 19 patients, and the agreement between readers was excellent, with an intraclass correlation coefficient of 0.942 for single measurements and 0.970 for averages—figures indicating that the measurement itself is robust and reproducible.

Because men naturally have thicker temporal muscles than women, the researchers derived sex-specific thresholds for low TMT by maximizing Youden’s index, a statistical measure balancing sensitivity and specificity, for three-month survival. The resulting cut-offs were 4.3 millimeters for men, with 46.7 percent sensitivity and 74.6 percent specificity, and 3.975 millimeters for women, with 64.6 percent sensitivity and 49.2 percent specificity. These thresholds are notably lower than those reported in earlier studies of younger patients, a difference the authors attribute to the advanced age of their cohort, in which temporal muscle wasting is already widespread. By these criteria, 35.3 percent of the men and 56.6 percent of the women had low temporal muscle thickness.

The survival picture in the cohort was stark. Median overall survival was just 90 days, or roughly three months, and by the one-year mark 86.7 percent of the patients had died. Patients with normal TMT did live somewhat longer on average than those with low TMT, with a median of 103 days versus 75 days, and in univariate analyses low TMT was significantly associated with worse survival at three months, with a hazard ratio of 1.54, and at six months, with a hazard ratio of 1.41. At one month and twelve months, no association reached statistical significance. But the crucial test came when the researchers adjusted their models for other clinical variables: age, body mass index, surgical treatment, and type of radiotherapy. Once these factors were accounted for, the relationship between TMT and survival evaporated entirely at every time horizon, with the adjusted three-month hazard ratio falling to 1.40 and losing significance.

What did predict survival, independently, were other clinical variables. Receiving whole-brain radiotherapy rather than stereotactic radiotherapy carried a substantially higher risk of death at every time point, with an adjusted hazard ratio of 5.98 for one-month survival and approximately 2.3 at the longer horizons. This likely reflects treatment selection rather than harm: whole-brain radiotherapy is generally reserved for patients with multiple metastases and poorer prognosis, while stereotactic techniques are favored for those with limited, more favorable disease. Similarly, patients who did not undergo surgical resection had markedly higher mortality across all time frames, since surgery is typically offered only to younger patients with solitary lesions and better expected outcomes. Higher body mass index was protective for one-month survival, with a hazard ratio of 0.31, and advancing age independently worsened six- and twelve-month survival. The authors are candid that these patterns largely mirror the realities of clinical decision-making.

The contrast with previous research is instructive. Of six earlier studies evaluating TMT in patients with brain metastases, three found that low thickness predicted shorter survival, while others found no association once confounders were considered. Crucially, those studies examined patients on average more than a decade younger, with median survival extending from three to eleven months, and several included only surgically treated patients. The Tampere cohort, by contrast, was older, sicker, and predominantly managed with radiotherapy alone. The researchers also tested whether TMT predicted survival within their largest tumor subgroup, the 113 lung cancer patients, and found no significant association at any time point, which discouraged further subgroup analyses. Additional limitations deserve mention: performance status, number of brain lesions, intracranial tumor volume, extracranial disease status, and systemic therapy were not incorporated into the models, and the derived sex-specific cut-offs may risk overfitting, making external replication essential.

The study’s conclusion is refreshingly direct: temporal muscle thickness should not be regarded as a prognostic factor in patients aged 70 or older with brain metastases treated with radiotherapy. The finding is a valuable corrective to a growing enthusiasm for imaging-derived body-composition markers, reminding clinicians and researchers that a measurement can be technically reliable, easily obtained, and biologically plausible yet still fail to add independent predictive value in the population that matters. In very old, profoundly ill patients whose median survival is measured in weeks, disease aggressiveness and treatment selection appear to overwhelm any signal carried by skeletal muscle reserves. For this vulnerable group, prognostic conversations must continue to rest on established clinical variables, while the search for better tools goes on.

Subject of Research: Temporal muscle thickness as a prognostic marker for survival in older patients with brain metastases treated with radiotherapy

Article Title: Association Between Temporal Muscle Thickness and Survival in Older Patients With Brain Metastases Undergoing Radiation Therapy

Article References: Pikkarainen, L., Korhonen, T. K., Tolonen, A., Pesonen, E. K., Hernandez, N., Skyttä, T., & Arponen, O. (2026). Association Between Temporal Muscle Thickness and Survival in Older Patients With Brain Metastases Undergoing Radiation Therapy. Cancer Reports, 9(9), Article e70671. https://doi.org/10.1002/cnr2.70671

Image Credits: AI Generated

DOI: 10.1002/cnr2.70671

Keywords: temporal muscle thickness, brain metastases, radiotherapy, survival, sarcopenia, cachexia, prognostication, computed tomography, whole-brain radiotherapy, stereotactic radiotherapy, older patients, Cancer Reports